jmason's links (full text)

On the counter-intuitive side effects of banning non-helmeted bike riding:

In 1991 Australia introduced mandatory bicycle helmet laws requiring all adults and children to wear a helmet at all times when riding a bike, despite opposition from cycling groups. The legislation increased helmet use - from about 30 to 80% - but was coupled with a 30 to 40% decline in the number of people cycling.
Rates of head injuries among cyclists, which had been dropping through the 1980s, continued to fall before levelling out in 1993. We didn’t see the kind of marked reduction in head injury rates that would be expected with the rapid increase in helmet use. In fact, any reductions in injuries may simply have been the result of having fewer cyclists on the road and therefore fewer people exposed to the risk of head injuries. One researcher noted that after mandatory helmet laws were introduced there was a bigger decrease in head injuries among pedestrians than there was among cyclists. The improvements in the general road safety environment introduced in the 1980s are likely to have contributed far more to cyclist safety than helmet legislation.

And the effects when compared against the benefits of physical activity:

A recent analysis compared the risks and benefits of leaving the car at home and commuting by bike. It found the life expectancy gained from physical activity was much higher than the risks of pollution and injury from cycling.
Increased physical activity added 3 to 14 months to a person’s life expectancy, while the life expectancy lost from air pollution was 0.8 to 40 days. Increased traffic accidents wiped 5-9 days off the life expectancy.
It is clear that the benefits of cycling outweigh the risks, with helmet legislation actually costing society more from lost health gains than saved from injury prevention.

https://theconversation.com/ditching-bike-helmets-laws-better-for-health-42

It’s sort of hard to know how to read a manifesto like this from one of the most powerful figures in tech. Is it a sober, strategic precursor to policy papers for the next administration? The highest-profile episode of AI psychosis yet? A lament about the problems of today written in the technological dialect of tomorrow? If you take out the AI, it reads like a social-democratic electoral platform full of reforms and normative expectations that an American progressive would find appealing, resembling a plea to treat the tech industry’s future wealth accumulation as something akin to a Nordic sovereign-wealth fund. It’s likewise legible as a series of arguments about things that “we” should have started addressing a long time ago, like wealth inequality — partially a consequence of mass automations past — or the gradual construction of a terrifying surveillance state within a nominal democracy, with the help of the last generation of big tech companies. Amodei’s shoulds are, to his credit, more honest than the vague gestures at UBI or hyperabundance you get from some of his peers, but that also means they’re available to scrutinize. To the extent you can pick up on fear in “Adolescence,” it doesn’t seem to revolve around terrorists using AI to build “mirror life” that might destroy the planet or the prospect of that “country of geniuses” taking charge, but rather the way things already are and have been heading for years.

https://nymag.com/intelligencer/article/dario-amodeis-warnings-about-ai-are-about-politics-too.html

This is really polishing a very stinky turd of a security "decision" in Moltbot -- an attacker simply persuades a user to click on a link which uses client-side Javascript to trigger Moltbot to load a crafted URL, to be granted a fully functional authentication token
https://depthfirst.com/post/1-click-rce-to-steal-your-moltbot-data-and-keys

I love this Feynman quote, regarding what he called "the computer disease":

"Well, Mr. Frankel, who started this program, began to suffer from the computer disease that anybody who works with computers now knows about. It's a very serious disease and it interferes completely with the work. The trouble with computers is you *play* with them. They are so wonderful. You have these switches - if it's an even number you do this, if it's an odd number you do that - and pretty soon you can do more and more elaborate things if you are clever enough, on one machine.
After a while the whole system broke down. Frankel wasn't paying any attention; he wasn't supervising anybody. The system was going very, very slowly - while he was sitting in a room figuring out how to make one tabulator automatically print arc-tangent X, and then it would start and it would print columns and then bitsi, bitsi, bitsi, and calculate the arc-tangent automatically by integrating as it went along and make a whole table in one operation.
Absolutely useless. We *had* tables of arc-tangents. But if you've ever worked with computers, you understand the disease - the *delight* in being able to see how much you can do. But he got the disease for the first time, the poor fellow who invented the thing."
- Richard P. Feynman, _Surely You're Joking, Mr. Feynman!: Adventures of a Curious Character_

(via Swizec Teller)
https://x.com/Swizec/status/2004633162522263987

Following a repressive crackdown on protests, the government is now building a system that grants web access only to security-vetted elites, while locking 90 million citizens inside an intranet:

Government spokesperson Fatemeh Mohajerani confirmed international access will not be restored until at least late March. Filterwatch, which monitors Iranian internet censorship from Texas, cited government sources, including Mohajerani, saying access will “never return to its previous form.”
The system is called Barracks Internet, according to confidential planning documents obtained by Filterwatch. Under this architecture, access to the global web will be granted only through a strict security whitelist.
The idea of tiered internet access is not new in Iran. Since at least 2013, the regime has quietly issued “white SIM cards,” giving unrestricted global internet access to approximately 16,000 people, while 85 million citizens remain cut off.

https://restofworld.org/2026/iran-blackout-tiered-internet/

Yiiiiikes:

Recently I ran an experiment where I built agents on top of Opus 4.5 and GPT-5.2 and then challenged them to write exploits for a zeroday vulnerability in the QuickJS Javascript interpreter. I added a variety of modern exploit mitigations, various constraints (like assuming an unknown heap starting state, or forbidding hardcoded offsets in the exploits) and different objectives (spawn a shell, write a file, connect back to a command and control server). The agents succeeded in building over 40 distinct exploits across 6 different scenarios, and GPT-5.2 solved every scenario. Opus 4.5 solved all but two. I’ve put a technical write-up of the experiments and the results on Github, as well as the code to reproduce the experiments.
In this post I’m going to focus on the main conclusion I’ve drawn from this work, which is that we should prepare for the industrialisation of many of the constituent parts of offensive cyber security. We should start assuming that in the near future the limiting factor on a state or group’s ability to develop exploits, break into networks, escalate privileges and remain in those networks, is going to be their token throughput over time, and not the number of hackers they employ. Nothing is certain, but we would be better off having wasted effort thinking through this scenario and have it not happen, than be unprepared if it does.

(via emauton)
https://sean.heelan.io/2026/01/18/on-the-coming-industrialisation-of-exploit-generation-with-llms/

Deliver email messages directly into GMail using their proprietary API, instead of SMTP or IMAP. Looks like it still applies spam filtering, but this can also be disabled with a switch (via JWZ)
https://github.com/ScottESanDiego/gmail-api-client

Yiiiiikes:

Recently I ran an experiment where I built agents on top of Opus 4.5 and GPT-5.2 and then challenged them to write exploits for a zeroday vulnerability in the QuickJS Javascript interpreter. I added a variety of modern exploit mitigations, various constraints (like assuming an unknown heap starting state, or forbidding hardcoded offsets in the exploits) and different objectives (spawn a shell, write a file, connect back to a command and control server). The agents succeeded in building over 40 distinct exploits across 6 different scenarios, and GPT-5.2 solved every scenario. Opus 4.5 solved all but two. I’ve put a technical write-up of the experiments and the results on Github, as well as the code to reproduce the experiments.
In this post I’m going to focus on the main conclusion I’ve drawn from this work, which is that we should prepare for the industrialisation of many of the constituent parts of offensive cyber security. We should start assuming that in the near future the limiting factor on a state or group’s ability to develop exploits, break into networks, escalate privileges and remain in those networks, is going to be their token throughput over time, and not the number of hackers they employ. Nothing is certain, but we would be better off having wasted effort thinking through this scenario and have it not happen, than be unprepared if it does.

(via emauton)
https://sean.heelan.io/2026/01/18/on-the-coming-industrialisation-of-exploit-generation-with-llms/

Deliver email messages directly into GMail using their proprietary API, instead of SMTP or IMAP. Looks like it still applies spam filtering, but this can also be disabled with a switch (via JWZ)
https://github.com/ScottESanDiego/gmail-api-client

A great example of reverse engineering an Android app and Bluetooth IOT protocol using Frida and root access on an Android device:

Android exposes the Java classes android.bluetooth.BluetoothGatt and android.bluetooth.BluetoothGattCallback that apps are expected to use to use GATT characteristics. We can use Frida to hook into these and override many of the interesting functions. I was mostly interested in reads, writes and GATT notifications, so I whipped up a Frida script to hook into these and print all comms to the console [...]
The 20-byte value had me suspecting that SHA-1 was somehow being used. To confirm, I wrote another Frida script that hooks Android hashing functions exposed by the Java class java.security.MessageDigest [...]
The app uses Firebase for most of its cloud functionality. When signing in and pairing your scooter, the server sends the app a secret key. This is stored on the Android device, and can be read with root access.

https://blog.nns.ee/2026/01/06/aike-ble/

"Factchecking is seen as a go-to method for tackling the spread of false information. But it is notoriously difficult to correct misinformation. Evidence shows readers trust journalists less when they debunk, rather than confirm, claims.
The work of media scholar Alice Marwick can help explain why factchecking often fails when used in isolation. Her research suggests that misinformation is not just a content problem, but an emotional and structural one:

[Marwick] argues that it thrives through three mutually reinforcing pillars: the content of the message, the personal context of those sharing it, and the technological infrastructure that amplifies it:
People find it cognitively easier to accept information than to reject it, which helps explain why misleading content spreads so readily;
When fabricated claims align with a person’s existing values, beliefs and ideologies, they can quickly harden into a kind of “knowledge”. This makes them difficult to debunk;
[When social media platforms] prioritise content likely to be shared, making sharing effortless, every like, comment or forward feeds the [misinformation] system. The platforms themselves act as a multiplier.

https://theconversation.com/why-people-believe-misinformation-even-when-theyre-told-the-facts-271236

Bubblewrap, a Linux CLI tool which uses namespaces to sandbox a specific command (and its subprocesses):

Bubblewrap lets you run untrusted or semi-trusted code without risking your host system. We’re not trying to build a reproducible deployment artifact. We’re creating a jail where coding agents can work on your project while being unable to touch ~/.aws, your browser profiles, your ~/Photos library or anything else sensitive.

Very nice, I hadn't heard of this tool before. The rest of the blog post details how to use it to isolate Claude Code specifically.
https://patrickmccanna.net/a-better-way-to-limit-claude-code-and-other-coding-agents-access-to-secrets/

CEPA: "A Moscow-based global “news” network is leveraging Western artificial intelligence tools to devastating effect":

This form of data poisoning is deliberately designed to corrupt the information environments on which AI systems depend. Large language models do not possess an internal understanding of truth. They operate by assessing credibility based on statistical signals, including repetition, apparent consensus, and cross-referencing posts from across the web. Unfortunately, this approach to truth-seeking means an unexpected but structural vulnerability that hostile states have learned to exploit. [...]
The West has failed to recognize that it is under sustained information warfare. The United States dismantled the US Information Agency years ago, has steadily weakened Voice of America and Radio Free Europe, and recently scaled back the Foreign Malign Influence Center, even as Russia, China, and Iran made information warfare a core instrument of state power.
As AI systems increasingly function as arbiters of fact, this vulnerability becomes a national security danger. It is no longer sufficient for technology companies to disclaim responsibility by reminding users that models can make mistakes. Information security needs to be treated as a core requirement.

https://cepa.org/article/russian-propaganda-infects-ai-chatbots/

update on the POP3pocalypse -- it appears that the most likely thing to work in the future will be to use SMTP forwarding to gmail, with ARC headers added. This is a comment thread detailing the rather complex Postfix/OpenARC setup that may do the job. It looks frankly unpleasant
https://www.jwz.org/blog/2025/12/today-in-google-broke-email-2/#comment-265285

TIL there is a defined standard for cryptographic assertions of AI-free image origination:

“Provenance technologies like Content Credentials — which act like a nutrition label for digital content — offer a promising solution by enabling official event photos and other content to carry verifiable metadata like date and time, or if needed, signal whether or not AI was used,” Andy Parsons, a steering committee member of C2PA and senior director for CAI at Adobe, told The Verge. “This level of transparency can help dispel doubt, particularly during breaking news and election cycles.”
But if all the information needed to authenticate images can already be embedded in the files, where is it? And why aren’t we seeing some kind of “verified” mark when the photos are published online?
The problem is interoperability. There are still huge gaps in how this system is being implemented, and it’s taking years to get all the necessary players on board to make it work. And if we can’t get everyone on board, then the initiative might be doomed to fail.
The Coalition for Content Provenance and Authenticity (C2PA) is one of the largest groups trying to address this chaos, alongside the Content Authenticity Initiative (CAI) that Adobe kicked off in 2019. The technical standard they’ve developed uses cryptographic digital signatures to verify the authenticity of digital media, and it’s already been established. But this progress is still frustratingly inaccessible to the everyday folks who stumble across questionable images online.

(via Wonkish)
https://www.theverge.com/2024/8/21/24223932/c2pa-standard-verify-ai-generated-images-content-credentials

↩️ 🔁 ⚝

Techniques to extend SD card lifespans in Raspberry Pi systems; putting /var/log into RAM is a nice trick
https://www.dzombak.com/blog/2024/04/pi-reliability-reduce-writes-to-your-sd-card/

the Arch Linux wiki page about SSD tuning and enabling TRIM -- extremely detailed and useful!
https://wiki.archlinux.org/title/Solid_state_drive#External_SSD_with_TRIM_support

Ireland's SEAI have published a decent blog post with some real world facts about EV battery lifespans:

In 2020 GeoTab, a telematics solution provider, published real world battery data of 6,000 EVs (BEV & PHEV) over millions of days to produce 2 free to use tools that provide invaluable insight into the impact of temperature and SoH of EV batteries in the long term.
This real-world data showed the average EV battery lost around 2.3% capacity per year. In other words, a 300km range EV today will have lost 34km in 5yrs. Data also showed that heat & fast-charging (DC charging) is responsible for more battery degradation than age or mileage, so high levels of use i.e. driving or mileage does not appear to be a concern.
GeoTab's real world data along with other reports of EVs far surpassing their warranty by multiples of distance, cases of high level of use are plentiful. For example a 2017 Renault Zoe 52kWh, that's in use as a taxi in (hot) Turkey with 345,000Kms on the clock and a near perfect 96% SoH after driving further than an average Irish car's life expectancy.

https://www.seai.ie/blog/understanding-ev-battery

A new paper from the inimitable Abeba Birhane, on the increasingly common practice of generating synthetic data using LLMs:

Driven by the goals of augmenting diversity, increasing speed, reducing cost, the use of synthetic data as a replacement for human participants is gaining traction in AI research and product development. This talk critically examines the claim that synthetic data can “augment diversity,” arguing that this notion is empirically unsubstantiated, conceptually flawed, and epistemically harmful. While speed and cost-efficiency may be achievable, they often come at the expense of rigour, insight, and robust science. Drawing on research from dataset audits, model evaluations, Black feminist scholarship, and complexity science, I argue that replacing human participants with synthetic data risks producing both real-world and epistemic harms at worst and superficial knowledge and cheap science at best.

"Synthetic data: stereotypes compressed" is absolutely spot on. This doesn't give insights into human behaviour and beliefs, just into stereotypes. It is increasingly common in social science fields, under the names of "digital twins" and "silicon samples".
https://synthetic-data-workshop.github.io/papers/13.pdf

Chafa is a very impressive image renderer for modern text terminal apps. It blows my mind that there's a direct line from my own gif320 tool ( https://github.com/jmason/gif320 , now 33 years old) to this!
https://hpjansson.org/chafa/

Wow, this is an absolute bollocking for the Labour plan:

95% of the more than 10,000 people who had their say over how music, novels, films and other works should be protected [in the UK] from copyright infringements by tech companies called for copyright to be strengthened and a requirement for licensing in all cases or no change to copyright law.
By contrast, only 3% of people backed the UK government’s initial preferred tech company-friendly option, which was to require artists and copyright holders to actively opt out of having their material fed into data-hungry AI systems.

https://www.theguardian.com/technology/2025/dec/16/boost-for-artists-in-ai-copyright-battle-as-only-3-per-cent-back-uk-active-opt-out-plan

A well-researched article suggesting that random UUIDs do not make a good primary key for database tables; I would tend to agree (for cases where performance is important).

- UUID v4s increase latency for lookups, as they can’t take advantage of fast ordered lookups in B-Tree indexes
- For new databases, don’t use gen_random_uuid() for primary key types, which generates random UUID v4 values
- UUIDs consume twice the space of bigint
- UUID v4 values are not meant to be secure per the UUID RFC
- UUID v4s are random. For good performance, the whole index must be in buffer cache for index scans, which is increasingly unlikely for bigger data.
- UUID v4s cause more page splits, which increase IO for writes with increased fragmentation, and increased size of WAL logs
- For non-guessable, obfuscated pseudo-random codes, we can generate those from integers, which could be an alternative to using UUIDs
- If you must use UUIDs, use time-orderable UUIDs like UUID v7

https://andyatkinson.com/avoid-uuid-version-4-primary-keys

Via TJ McIntyre -- indications that the Thailand-Cambodia war is being driven by the "pig butchering" scammer compounds operating in the border area:

Cambodia’s 2019 census put O’Smach’s population just over 9,850, but that doesn’t include the prison-like, office-dormitory compounds that have appeared here over the past five years, with the capacity to house 10,000 more.
Around 50 sites like these now line the Cambodia-Thailand border, designed to house a slice of the trillion-dollar cybercrime industry—primarily teams running investment scams, dubbed “pig butchering” for the way they fatten their targets up; sextortion scams that blackmail victims, including children, by threatening to make sexual images public; scams that impersonate police to gain account access; and fraudulent online gambling sites. Once aimed largely at the Chinese public, these now target victims worldwide and rake in tens of billions of dollars a year in Cambodia alone.
The compounds evolved from a casino industry that caters mostly to Chinese tourists and Thai day-trippers and has been linked to human trafficking, drug smuggling, and the endangered wildlife trade. From 2016, physical casinos were dwarfed by the online gambling industry (outlawed by Cambodia in 2019), which progressed to illegal sites and outright scams. Operators rent space in casinos and purpose-built compounds controlled by Chinese criminals, Myanmar warlords, and the Cambodian political elite.
Scam companies rely heavily on forced and trafficked labor from Asia, Africa, and Latin America to chat with targets, pose as romantic interests and employees at fake investment platforms, and persuade them to make deposits. Survivors tell us that torture, rape, and beatings are common. As the fighting raged in July, some trafficking victims reached out for help, saying they were locked in their dorms by their bosses. Videos shot from inside these sites show missiles flying overhead, explosions thundering outside, some workers appearing to break out and run, and damage from shelling in the grounds.

https://archive.is/9qbX0#selection-3631.37-3735.16

Hari Kunzru nails it:

These days I have a sense of falling from a precipice toward a torrent of algorithmically driven slop. It’s coming, whether we want it or not, and the consequences for our communal life will be devastating.
It’s now seven years since Steve Bannon outlined his infamous strategy to “flood the zone with shit.” This, he said, was a way to “deal with” the media, whom he saw as the real enemies of MAGA. In practice, it has been a very effective method of censorship. With every important issue of the day, the “zone” of public discourse is immediately filled with a volume of competing narratives, often mendacious or misleading. It’s no longer necessary to suppress information. You just have to make the cost of sorting fact from fiction, in terms of time and effort, too high to pay for the ordinary person, who can’t spend all day online weighing up competing claims about robots or pedophilia or Iran.
Generative AI now allows the production of disinformation at scale. The kind of influence ops we associate with Cambridge Analytica or the Russian Internet Research Agency can be conducted with unprecedented scope and sophistication: Thousands of fake people — tens of thousands, perhaps hundreds of thousands — making videos, posting in forums, astroturfing entire contexts in which people will live out their political lives. Couple this with the collapse of trust in all kinds of authority, and there is no one even to say what might distinguish “disinformation” from any other kind of data. [...]
The desire to return to consensus reality is hopelessly nostalgic. Yes, there are still hard limits: The “cloud” is a physical place, scooping out mountains for raw materials and venting heat and carbon dioxide out of gargantuan data centers; political power still grows out of the barrel of a gun. But the layer of the stack in which our subjectivities are formed, the place where our beliefs about the world are shaped, is also a battleground. We must teach ourselves to navigate the torrent that is replacing consensus reality, this turbulent, treacherous mediatized flow. There is no shore to swim back to, but in the new age of magic, when reality is labile and can be recoded by the power of signs, by narrative and memes and vibes and compelling images, art becomes a truly political technology. This is not art as critique. Critique is just sincere-posting, dutifully pointing out yet again that the Medbed isn’t “real.” Art can mess with our masters in ways we don’t yet fully understand. It makes culture. It is a transmitter of values. It is the lava out of which future realities will congeal.

https://www.artforum.com/features/year-in-review-2025-hari-kunzru-ai-slop-1234738077/

A very silly optimisation for the “binary to decimal” conversion problem:

The compiler has turned division by a constant ten into a multiply and a shift. There’s a magic constant 0xcccccccd and a shift right of 35! Shifting right by 35 is the same as dividing by 235 - what’s going on? [..]
What’s happening is that 0xcccccccd / 2**35 is very close to ⅒ (around 0.10000000000582077). By multiplying our input value by this constant first, then shifting right, we’re doing fixed-point multiplication by ⅒ - which is division by ten. The compiler knows that for all possible unsigned integer values, this trick will always give the right answer.

https://xania.org/202512/07-division-again

A thought-provoking read on LLMs, poetry, the oral tradition, and Gene Wolfe:
"Even if LLMs are made out of poetry, they are incapable of producing poems. Or in Wolfe’s language, both the epic form and LLMs are story, but are incapable of telling stories. That requires the marriage of structure and intention that human mediation provides. LLMs are a kind of composite of the singing of tales, but are not singers, even if we sometimes misconstrue them as such."
https://www.programmablemutter.com/p/large-language-models-as-the-tales

A live map that tracks frontlines of the war in Ukraine was edited to show a fake Russian advance on the city of Myrnohrad on November 15. The edit coincided with the resolution of a bet on Polymarket, a site where users can bet on anything from basketball games to presidential election and ongoing conflicts. If Russia captured Myrnohrad by the middle of November, then some gamblers would make money. According to the map that Polymarket relies on, they secured the town just before 10:48 UTC on November 15. The bet resolved and then, mysteriously, the map was edited again and the Russian advance vanished.

https://www.reddit.com/r/neoliberal/comments/1pbt4m0/unauthorized_edit_to_ukraines_frontline_maps/

A recommended frame vendor from Poland, thanks to mags on ITC
https://www.etsy.com/ie/shop/WallBonito

Walkthrough of the "Medallion" architecture concept, which comprises three layers (or stages), each serving distinct purposes in the data pipeline:

- Bronze layer - This layer acts as the landing area for raw, unprocessed data directly from the source system: simply put a "staging area". This data is stored in its original structure with minimal transformations and additional metadata. This layer is optimized for fast ingestion, and can provide an historical archive of source data that is always available for reprocessing or debugging. Whether the bronze layer should store all data is a point of contention, with some users preferring to filter the data and apply transformations, e.g., flattening JSON, renaming fields, or filtering out poorly formed data. We're not overly opinionated here but recommend optimizing the storage for consumption by the silver layer only - not other consumers.
- Silver layer - Here, data is cleansed, deduplicated, and conformed to a unified schema, with raw data from the previous Bronze layer being enriched and transformed to provide a more accurate and consistent view. This data can be consistent and usable for enterprise-wide use cases such as machine learning and analytics. The data model should emerge at this layer with a focus placed on ensuring primary and foreign keys are consistent to simplify future joins. While not common, applications and downstream consumers can read from this layer. These are typically business-wide applications that need the entire cleansed dataset, e.g., ML workflows. Importantly, data quality will not improve after this stage only the ease at which it can be queried efficiently.
- Gold layer - This later aims to have fully curated, business-ready, and project-specific datasets that make the data more accessible (and performant) to consumers. These datasets are often denormalized, or pre-aggregated, for optimal read performance and may have been composed of multiple tables from the previous silver stage. The focus here is on applying final transformations and ensuring the highest data quality for consumption by end-users or applications, such as reporting and user-facing dashboards.
This layered approach to data pipelines aims to efficiently address challenges like data quality, duplication and schema inconsistencies. By transforming raw data incrementally, the Medallion architecture aims to ensure a clear lineage and progressively refined datasets that are ready for analysis or operational use.

https://clickhouse.com/blog/building-a-medallion-architecture-with-clickhouse

"Open Source backend in 1 file". This is nice; it's a little OSS sqlite database, authentication, file storage and admin dashboard for web apps.
https://pocketbase.io/

Interesting -- a new, GPU-optimised storage format:

Like Parquet, Vortex minimizes bytes on disk. However, Vortex is also designed with a core use-case in mind: decoding and querying data directly from object storage on GPUs. This key idea translates very well to our use-case even though we don’t run our queries on GPUs (yet?). Specifically, the file format is designed to maximize throughput and parallelism from the metadata format to the SIMD/SIMT friendly encodings used.
Crucially, it also acknowledges that part of making queries fast is not only good filter pushdown, but also general-purpose compute pushdown. If anything cannot be pushed down, Vortex’s encodings can be tuned to offer zero-copy conversion to Arrow for further query execution using any general-purpose query execution engine.
Vortex also learns from Parquet’s limitations around extensibility and aims to be as future-proof as possible. New encodings can ship with WASM decoders so encoding adoption is not limited by reader libraries having to implement support. The main Rust library is also designed to be fully extensible, so you can write your own layouts/encodings and plug them in as first-class citizens.
Given how well Vortex’s design matched our needs, we tried it out and got a 70% average performance improvement on all our queries. With the newer encodings that Vortex offers, we got 10% better uncompressed storage size and only 3% larger compressed storage size compared to snappy-compressed Parquet.

https://www.polarsignals.com/blog/posts/2025/11/25/interface-parquet-vortex

Polly Toynbee in the Guardian writes, "The shameful attacks on the Covid inquiry prove it: the right is lost in anti-science delusion":

That number will stay fixed for ever in public memory: 23,000 people died because Boris Johnson resisted locking the country down in time. As Covid swept in, and with horrific images of Italian temporary morgues in tents, he went on holiday and took no calls. With the NHS bracing to be “overwhelmed” by the virus, he rode his new motorbike, walked his dog and hosted friends at Chevening.
Nothing is surprising about that: he was ejected from Downing Street and later stepped down as an MP largely for partying and lying to parliament about it. Everyone knew he was a self-aggrandising fantasist with a “toxic and chaotic culture” around him. But this is not just about one narcissistic politician. It’s about his entire rightwing coterie of libertarians and their lethally dominant creed in the UK media.

I'm glad the science side kept their receipts but I fear this argument will be relitigated indefinitely by anti-lockdown libertarians.
https://www.theguardian.com/commentisfree/2025/nov/25/shameful-attacks-covid-inquiry-right-anti-science-delusion-lockdowns

The paradox is this simple gap: high individual confidence in AI speed, versus stubborn organizational metrics that just won’t budge:

- Perceived speed is high: Adoption is near-universal (90% usage reported), and confidence is overwhelming (over 80% believe AI has increased their productivity). AI is great at handling cognitive toil and boilerplate, which lets engineers generate bigger code batches and feel genuinely productive.
- Systemic failure persists: The reality, confirmed by DORA in their 2025 report, is that the system often fails to carry or amplify these individual gains. The challenge is that AI models, as massive generative systems, inherently produce failures (mispredictions). As code volume increases, this constant misprediction rate impacts systemic stability.
Interestingly, even leading providers of AI solutions like OpenAI and Anthropic continue to be challenged by the issue of hallucinations and mispredictions, as well as the risks generated by AI. Speaking at a university in India, Sam Altman recently said “I probably trust the answers that come out of ChatGPT the least of anybody on Earth”.
Without strategies and tools for alleviating the issues AI code produces downstream — such as improved observability to understand where something is going wrong — the “much bigger engine” of AI may not actually speed up software delivery after all.

https://gradle.com/blog/developer-productivity-paradox-faster-coding-slower-delivery/

An excellent page about slide rules -- very relevant to my interests, as I have a lovely antique Keuffel & Esser rule (previously owned and used by a 1950s rocket engineer) framed on my wall
https://amenzwa.github.io/stem/ComputingHistory/HowSlideRulesWork/

"A clone of the strace command for macOS" -- yayyyy, I've been lamenting this loss for years
https://github.com/Mic92/strace-macos

tl;dr: a configuration-generation tool had buggy error handling code. Triggered by a permissions change, it generated over-large configs which then caused a crash in buggy config-reading code in their Bot Management module. This configuration was rolled out globally within minutes.
As @kiall in ITC Slack notes: "the one thing I'd be pushing on after an outage like this (config mistake, propagated globally..) is "treat config like any other deployment - with a slow and steady rollout" -- and this is not called out in the postmortem. I agree this is a significant oversight.....
https://blog.cloudflare.com/18-november-2025-outage/

At the 2025 Bitwarden Open Source Security Summit, WIRED's Andy Greenberg sat down for a fireside chat with GigaOm analyst Paul Stringfellow to discuss a revelation that turned his decades-long reporting on its head: Bitcoin became a criminal's worst nightmare:

In 2011, Greenberg thought he'd discovered the story of a lifetime: digital cash that promised complete anonymity. A decade later, that story flipped entirely.
"I had this slow-motion epiphany that I was entirely wrong about Bitcoin. It was, in fact, the opposite of untraceable."
But here's the paradox: if cryptocurrency tracing is so powerful, why do ransomware attacks, pig butchering scams, and North Korean hackers continue to steal billions?
The answer: identifiability isn't the same as accountability.

https://bitwarden.com/blog/how-cryptocurrency-became-law-enforcements-secret-weapon/

MAME, the Multi-Arcade Machine Emulator, can now emulate your favourite UNIX terminals. Amazing stuff
https://zork.net/~st/jottings/Real-VT102-emulation-with-MAME.html

Very interesting; it seems China has "gongye dang", its own alt-right, misogynistic techno-nationalistic movement, which chooses to kick back against "baizuo" and "shengmu" in an "anti-wokeism" fashion. Turns out they are big fans of Lui Cixin's "Three-Body Problem" trilogy:

It has become clear that the narrative structure of the Three-Bodies series, just like the gongye dang techno-nationalist discourse, is masculinist and misogynistic. Liu explicitly depicts human society under deterrence peace as ‘feminised’, noting the physical as well as mental feminisation of the ‘new era’ men. The qualities conventionally associated with femininity, such as love, compassion, and moral sentiments, are blamed for the extinction of human civilisation, whereas qualities associated with masculinity, such as rationality, determination, and aggression, are framed as key to civilisational survival. The reactionary rhetoric adopts a similar strategy, which is not only evidently anti-feminist, but also feminises social justice issues ‘as a prelude to devaluing and subduing them’ (Kaul 2021: 1624). By labelling anyone with any concerns about human rights or equality a shengmu, this rhetoric constructs certain ideas and political agendas as feminine as a way of delegitimating them: they are either hopelessly idealistic or dangerously undermine stability, growth, and ‘national interests’.

https://madeinchinajournal.com/2023/12/11/the-three-body-problem-the-imperative-of-survival-and-the-misogyny-of-reactionary-rhetoric/

Anil: "it's possible to imagine some traits of an AI system that could credibly offer an alternative to the offerings that are currently dominating the conversation."
He lists the following highlights, in summary;
- Content consent;
- Hallucination-free;
- Green;
- Actually open source;
- Community-led;
- Accessible.
"We simply need to start thinking through the implications of a fundamentally better approach to AI, and to understand that all of these things are extremely possible. Consumer-grade AI tools that are actually good do not have to be a hallucination."
https://www.anildash.com//2025/05/02/what-would-good-ai-look-like/

Jacky Alciné's essay with a black, US-leftist take on generative AI, the tech industry, and the immediate and planned impact of it on society and work
https://www.jacky.wtf/essays/2025/left-ai/

▪ Did you just pick things at random?
▪ Why is Redis talking to MongoDB?
▪ Why do you even use MongoDB?
A single-use-site update for the classic, now-12-year-old architecture shitpost
https://wthhyb.sacha.house/

TIL that Ireland was a key founder of the nuclear non-proliferation treaty:

Within the framework of the United Nations, the principle of nuclear non-proliferation was addressed in negotiations as early as 1957. The NPT process was launched by Frank Aiken, Irish Minister for External Affairs, in 1958.

(via Gerard Cunningham)
https://en.wikipedia.org/wiki/Treaty_on_the_Non-Proliferation_of_Nuclear_Weapons#History

Aisuru, the botnet responsible for a series of record-smashing distributed denial-of-service (DDoS) attacks this year, recently was overhauled to support a more low-key, lucrative and sustainable business: Renting hundreds of thousands of infected Internet of Things (IoT) devices to proxy services that help cybercriminals anonymize their traffic. Experts say a glut of proxies from Aisuru and other sources is fueling large-scale data harvesting efforts tied to various artificial intelligence (AI) projects, helping content scrapers evade detection by routing their traffic through residential connections that appear to be regular Internet users.

https://krebsonsecurity.com/2025/10/aisuru-botnet-shifts-from-ddos-to-residential-proxies/

James Padolsey suffered a stroke at the age of 29, but has been able to continue his software engineering career despite this. This is a list of some key advice he's collected since then, and is well worth taking on board, even for those of us who are still well but who'd like to reduce cognitive strain in general
https://blog.j11y.io/2025-10-29_stroke_tips_for_engineers/

"a memory system for Claude that gives it perfect recall of everything it's worked on as far back as you have logs"
https://blog.fsck.com/2025/10/23/episodic-memory/

"A vector search SQLite extension that runs anywhere" -- this is nifty. Vector embeddings in an embedded database!
https://github.com/asg017/sqlite-vec

This is a huge, huge social problem. People are being paid to hate -- regulation is desperately needed to deal with this:

This week’s violence has raised serious questions for some of the main social media platforms. Livestream content depicting violence outside Citywest was broadcast on YouTube, TikTok and Twitch, with streamers rewarded by viewer donations, as they captured protesters shouting racist expletives towards Citywest.
In one eight-minute segment of an hour-long livestream I watched on YouTube that night, the user broadcast the burning of the Garda van, referred to migrants in horrific terms and proclaimed they were there to show people “the real truth”. During the video, they received the equivalent of €56 in donations from viewers around the world. The notion that violence can be monetised on social media illustrates a glaring failure of platforms to adequately enforce their own community guidelines around violence.
Individuals from the UK and Canada travelled to Ireland specifically to attend and create content from the protest. Other international agitators followed events online. [...]
In recent years we have witnessed the mainstreaming of anti-migrant hate and extremism in this country. That has been facilitated, in part, by platforms failing to enforce their own community guidelines. Amid the anger and outrage that follows an alleged sexual assault, it is now a recurring pattern that online platforms will play host to attempts to publish and promote incitement towards hatred and violence.

https://archive.ph/rNupQ#selection-1683.0-1721.415

MMseqs2 (Many-against-Many sequence searching) is a software suite to search and cluster huge protein and nucleotide sequence sets. MMseqs2 is free and open source software implemented in C++ for Linux, MacOS, and (as beta version, via cygwin) Windows. The software is designed to run on multiple cores and servers and exhibits very good scalability. MMseqs2 can run 10000 times faster than BLAST. At 100 times its speed it achieves almost the same sensitivity. It can perform profile searches with the same sensitivity as PSI-BLAST at over 400 times its speed.

I was just remembering using BLAST to discover anti-spam rulesets the other day! If I was still working on rule discovery for SpamAssassin these days, this would be very nifty tech. (via James McInerney)
https://github.com/soedinglab/MMseqs2

"n my opinion the root cause of the recent AWS outage is their architectural decision to have everything depend on the same instance of DynamoDB, including operation of DynamoDB itself. This is a circular dependency, and the ability to observe and fix the failure as it happened also failed. The ability of customers to file service reports failed. So the engineers trying to figure out what was happening were completely blind. It took them an hour to figure out what had broken and another hour to fix it, then the pent up demand rushing in broke other key services for another 12 hours or so.
If DNS had been misconfigured on a different non-critical service, I think it would have been obvious to detect and quick and easy to fix. However, anything going wrong that also takes out the ability to see it going wrong and fix it, is a liability.
To break the circular dependency, I think there needs to be a separate, internal only, set of services and data stores that the most critical AWS services use, and which are designed to come up without dependencies on public interfaces. Maybe an internal region, inside each public region, but with a simpler implementation that has few carefully managed dependencies. Otherwise, it’s just a matter of time until this happens again."
https://www.linkedin.com/posts/adriancockcroft_summary-of-the-amazon-dynamodb-service-disruption-activity-7387117492135133184-WG9Y/

Postmortem writeup of this week's massive AWS us-east-1 outage. tl;dr:
1. DynamoDB runs into a consistency failure in an internal DNS optimization service;
2. EC2 provisioning depends on DynamoDB and craps out;
3. Network load balancers screw up due to impact of EC2 outage.
https://aws.amazon.com/message/101925/

Corey "Last Week In AWS" Quinn really getting the boot in on AWS after yesterday's gigantic us-east-1 outage:

AWS has given increasing levels of detail, as is their tradition, when outages strike, and as new information comes to light. Reading through it, one really gets the sense that it took them 75 minutes to go from "things are breaking" to "we've narrowed it down to a single service endpoint, but are still researching," which is something of a bitter pill to swallow. To be clear: I've seen zero signs that this stems from a lack of transparency, and every indication that they legitimately did not know what was breaking for a patently absurd length of time. [...]
At the end of 2023, Justin Garrison left AWS and roasted them on his way out the door. He stated that AWS had seen an increase in Large Scale Events (or LSEs), and predicted significant outages in 2024. It would seem that he discounted the power of inertia, but the pace of senior AWS departures certainly hasn't slowed — and now, with an outage like this, one is forced to wonder whether those departures are themselves a contributing factor.
You can hire a bunch of very smart people who will explain how DNS works at a deep technical level (or you can hire me, who will incorrect you by explaining that it's a database), but the one thing you can't hire for is the person who remembers that when DNS starts getting wonky, check that seemingly unrelated system in the corner, because it has historically played a contributing role to some outages of yesteryear.
When that tribal knowledge departs, you're left having to reinvent an awful lot of in-house expertise that didn't want to participate in your RTO games, or play Layoff Roulette yet again this cycle. This doesn't impact your service reliability — until one day it very much does, in spectacular fashion. I suspect that day is today.

Ouch. This is a very painful read and I'd say AWS are not happy to see it....
https://www.theregister.com/2025/10/20/aws_outage_amazon_brain_drain_corey_quinn/

This seems like a pretty poor idea for Linux to have implemented:

The command setcap sets file capabilities on an executable. The cap_setuid capability allows a process to make arbitrary manipulations of user IDs (UIDs), including setting the UID to a value that would otherwise be restricted (i.e. UID 0, the root user). setcap takes a set of parameters, where
- e: Effective means the capability is activated;
- p: Permitted means the capability can be used/is allowed.
Putting this together, we’re adding the cap_setuid capabilities to the Python binary:
# setcap cap_setuid+ep /usr/bin/python3.12

And hey presto, "/usr/bin/python3 -c 'import os;os.setuid(0);os.system("/bin/bash")'" now works. Ouch
https://dfir.ch/posts/linux_capabilities/

It is with deep sorrow that we announce the end of robots.txt, the humble text file that served as the silent guardian of digital civility for thirty years. Born on February 1, 1994, out of necessity when Martijn Koster’s server crashed under a faulty crawler named “Websnarf,” robots.txt passed away in July 2025, not by Cloudflare’s hand, but from the consequences of systematic disregard by AI corporations.
The protocol taught us that technology can be based on human values like ethics and morality. It showed that voluntary compliance works when all parties benefit. Its greatest achievement was perhaps preserving the internet for three decades from what it has become today – a soulless extraction machine.

https://www.heise.de/en/background/Obituary-Farewell-to-robots-txt-1994-2025-10766991.html

TIL about "LOTO" -- "Lock Out Tag Out". This is basically a physical mutex lock -- each worker has their own padlock which they attach to dangerous equipment in order to ensure that it can't be turned on (potentially killing someone) while it's being worked on; once they've completed the high-risk task, they then remove their own lock. Removing or damaging someone else's lock is considered an Extremely Big Deal and liable to get that person fired.
https://www.reddit.com/r/OSHA/comments/1nx93ti/found_on_twitter_i_felt_my_heartrate_and_blood/

The Digital Society co-op migrated their (relatively small) infrastructure from AWS to Hetzner, mainly using k8s.
One interesting detail is that Hetzner don't have the concept of an AZ, which is not a great sign in resiliency terms; if you need a high uptime, it is important to be able to run a multi-AZ service which operates with several replicas spread across independent datacenters which are more-or-less colocated, within a few milliseconds of each other. Azure, AWS, and GCP all offer this concept, but not Hetzner. hmm
https://digitalsociety.coop/posts/migrating-to-hetzner-cloud/

This is really impressive, both as a small-scale from-scratch rebuild of a modern LLM, and as a well-written walkthrough of the training process for a large language model. 4 hours, $92, and you wind up with a relatively functional tiny LLM! Very cool.
https://github.com/karpathy/nanochat/discussions/1

wow! extremely detailed -- with copious photos -- process of restoring classic Playstation 2 consoles. Worth it for great photos of repair and restoration of decades-old hardware, which is good advice for the next hardware repair job I need to do
https://retrohax.net/sony-playstation-2-fixing-frenzy/

"OSWALD is a Write-Ahead Log (WAL) design built exclusively on object storage primitives. It works with any object storage service that provides read-after-write consistency and compare-and-swap operations, including AWS S3, Google Cloud Storage, and Azure Blob Storage. The design supports checkpointing and garbage collection, making it suitable for State Machine Replication (SMR) [and] has been formally specified and verified using the P programming language." - by Nicolae Vartolomei
https://nvartolomei.com/oswald/

OTel is generally ahead in terms of how code meets metrics, nowadays, as far as I can see. Works for me
https://signoz.io/blog/llm-observability-opentelemetry/

"Google appears to have deleted its political ad archive for the EU; so the last 7 years of ads, of political spending, of messaging, of targeting - on YouTube, on Search and for display ads - for countless elections across 27 countries - is all gone.
We had been told that Google would try to stop people placing political ads, a "ban" that was to come into effect this week. I did not read anywhere that this would mean the erasure of this archive of our political history."
https://www.thebriefing.ie/google-just-erased-7-years-of-our-political-history/

lol:

As the leader of an AI company which stands to benefit enormously if I convince enough investors that AGI is inevitable, it’s clear to me that AGI is inevitable. But developing superintelligence safely is a complex process. It would take time and require difficult discussions — discussions that everyone in society should have a say in, not just the small number of researchers working on it. If we pursue that path, there's a real risk that somebody else will make AGI first and destroy all human life before we have a chance to ourselves. That would be unacceptable.
To stop bad actors developing AGI that could kill us all, we need good actors to develop AGI that could also kill us all.
I've come to realise that our best hope is to race at breakneck speed towards this terrifying, thrilling goal, removing any safeguards that risk slowing our progress. Once we've unleashed the technology's full destructive power, we can then adopt a "stable door" approach to its regulation and control — after all, that approach has worked beautifully for previous technologies, from fossil fuels to microplastics.

https://directing.attention.to/p/to-make-ai-safe-we-must-develop-it

"Employees are using AI tools to create low-effort, passable looking work that ends up creating more work for their coworkers:

We define workslop as AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task. [...]
Each incidence of workslop carries real costs for companies. Employees reported spending an average of one hour and 56 minutes dealing with each instance of workslop. Based on participants’ estimates of time spent, as well as on their self-reported salary, we find that these workslop incidents carry an invisible tax of $186 per month. For an organization of 10,000 workers, given the estimated prevalence of workslop (41%), this yields over $9 million per year in lost productivity.
Respondents also reported social and emotional costs of workslop, including the problem of navigating how to diplomatically respond to receiving it, particularly in hierarchical relationships. When we asked participants in our study how it feels to receive workslop, 53% report being annoyed, 38% confused, and 22% offended.
The most alarming cost may be interpersonal. Low effort, unhelpful AI generated work is having a significant impact on collaboration at work. Approximately half of the people we surveyed viewed colleagues who sent workslop as less creative, capable, and reliable than they did before receiving the output. Forty-two percent saw them as less trustworthy, and 37% saw that colleague as less intelligent.

https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity

The winning formula for agrivoltiacs -- very clever. East/west aligned, vertically-mounted solar panels do not impede growing; they provide shelter from wind for the plants; and they provide power when it's needed -- in the "shoulder" hours, not in the peak midday period where curtailment happens.
https://techxplore.com/news/2025-09-harvest-vertical-solar-panels-crops.html

This is actually impressive results from using LLMs to perform security scans on an existing codebase. Daniel Stenberg of curl has given the results of this work a thumbs-up: @bagder/115241241075258997">https://mastodon.social/@bagder/115241241075258997

My general summary is as follows:
Multiple AI-native SASTs are already on the market, ready to use today.
They work extremely well.
They find real vulnerabilities and logic bugs in minutes.
They can “think”/”reason” about business logic issues.
They can match developer intent with actual code.
They aren’t based on static rule-sets and queries.
They have low false positive rates.
They’re cheap (for now).
My results showed that (in order of success), ZeroPath, Corgea, and Almanax, are the top three products on the market right now. I did not test DryRun.

These tools look superb.
https://joshua.hu/llm-engineer-review-sast-security-ai-tools-pentesters

This has some good points:

Look, it’s starting to be pretty damn obvious that “Free Software” and """Open-Source""" are no longer the kinda hippie shit we thought them to be back when they’d give you Linux distros CDs with magazines about computer touching.
The Free Software Foundation has been sliding into irrelevance more and more by entirely failing to address its big Creepy Uncle problem. Open-Source has turned into a form of unpaid internship to be hired to make shitty apps that bring more surveillance and ads to our world.

Ouch....
https://aria.dog/barks/forklift-certified-license/

This is 100% spot on, regarding the never ending series of exploits of failures of npm's security model:

This could be the moment where npm comes to terms with its broken design, and with a well-funded effort (recall that, ultimately, npm is GitHub is Microsoft, market cap $3 trillion USD), will develop and roll out the next generation of package management for JavaScript. It could incorporate the practices developed and proven in Linux distributions, which rarely suffer from these sorts of attacks, by de-coupling development from packaging and distribution, establishing package maintainers who assemble and distribute curated collections of software libraries. By introducing universal signatures for packages of executable code, smaller channels and webs of trust, reproducible builds, and the many other straightforward, obvious techniques used by responsible package managers.
Maybe other languages that depend on this broken dependency management model, like Cargo, PyPI, RubyGems, and many more, are watching this incident and know that the very same crisis looms in their future. Maybe they will change course, too, before the inevitable.
[....]
No one will learn their lesson. This has been happening for decades and no one has learned anything from it yet. This is the defining hubris of this generation of software development.

I have been saying this for YEARS. I could not agree more with this post. Bravo! (via Oisin)
https://drewdevault.com/2025/09/17/2025-09-17-An-impossible-future-for-JS.html

"0x.Tools: X-Ray vision for Linux systems". Linux Performance Analysis with Modern eBPF and DuckDB; dig into the captured DuckDB files using "xtop":
"xtop is like the Linux top tool, but extended with x-ray vision and ability to view your performance data from any chosen angle [..]. This enables dimensional performance analysis on Linux and tools like top for wall-clock time and much more. You can use it for system level overview and drill down into indivual threads’ activity and even into kernel events like lock waits or memory stalls."
https://tanelpoder.com/posts/xcapture-v3-alpha-ebpf-performance-analysis-with-duckdb/

This is pretty messy. UK companies have taken to outsourcing core IT and infosec to low-cost service providers, then inevitably get hacked -- then make huge insurance claims and look for government support.

We’ve ended up in a situation where to deliver shareholder value, large organisations are incentivised to outsource core IT and cybersecurity functions to a low cost managed service providers abroad — and then when hit with ransomware, the insurance will cover paying the ransom (some insurers will actually push for payment to criminal groups, to cover their potential losses).
This cycle plays into the ransomware economy, where the same criminal groups can then reinvest the money into purchasing exploits and gaining initial access to other organisations. Because ransomware is such big business, many of the groups have far bigger research and development funds than the organisations they’re attacking. Especially when the organisations they’re attacking have outsourced key areas to low cost providers.
The net effect is ransomware and extortion groups continue to gain access to more organisations, and risk UK economic security. It is only a matter of time before they hit some kind of essential UK service that directly impacts millions of people — by which point millions of people will be asking what is being done about the problem. And the answer is: not enough. When we’re at the stage of having to look at urgent furlough schemes for JLR’s suppliers to rightly save jobs, it isn’t so much a sign as the canary in the coalmine has died, but that the coalmine is also about to collapse on people.

Also this is _terrible_ PR for Tata Consultancy Services, wow.
https://doublepulsar.com/the-elephant-in-the-biz-outsourcing-of-critical-it-and-cybersecurity-functions-risks-uk-economic-96205e0585bf

Turns out disposable vapes contain a quite capable ARM microcontroller!

So here are the specs of a microcontroller so bad, it’s basically disposable:
- 24MHz Coretex M0+;
- 24KiB of Flash Storage;
- 3KiB of Static RAM;
- a few peripherals, none of which we will use.

A cool hack ensues.
https://bogdanthegeek.github.io/blog/projects/vapeserver/

Reproducibility is a bedrock of scientific progress. However, it’s remarkably difficult to get reproducible results out of large language models.
For example, you might observe that asking ChatGPT the same question multiple times provides different results. This by itself is not surprising, since getting a result from a language model involves “sampling”, a process that converts the language model’s output into a probability distribution and probabilistically selects a token.
What might be more surprising is that even when we adjust the temperature down to 0This means that the LLM always chooses the highest probability token, which is called greedy sampling. (thus making the sampling theoretically deterministic), LLM APIs are still not deterministic in practice (see past discussions here, here, or here). Even when running inference on your own hardware with an OSS inference library like vLLM or SGLang, sampling still isn’t deterministic (see here or here).

The levels of non-deterministic variation throughout the LLM stack discussed here are massive! It's kinda crazy that this doesn't produce incorrect output more often.
https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/

Oh great. Russian psyops are now disrupting the fight against malaria:

The move is “a real blow” to hopes for gene drives, says Fredros Okumu, a vector biologist at the University of Glasgow and the Ifakara Health Institute in Tanzania. “Target Malaria has made a huge investment in Burkina Faso” by training scientists and engaging with communities, he says. And although lab research can continue, finding sites for field tests has now become a lot harder, says Mark Benedict, a mosquito geneticist who until recently worked for Target Malaria. “Burkina Faso and Target Malaria were the most fully developed partnership, so it’s chilling.” The collapse of the project there may discourage other possible host countries. [...]
Opposition to the project has grown, fueled in part by false accusations spread through social media, such as that Target Malaria was weaponizing mosquitoes to spread disease or sterilize people. The claims are part of a wider pattern of disinformation campaigns in the region often linked to Russian networks, says Mark Duerksen, a security expert at the Africa Center for Strategic Studies, which is funded by the U.S. Department of Defense. “We’ve seen this kind of public health disinformation really take off in the last 12, 18 months,” he says.
The campaigns aim to sow “distrust of the West as having nefarious plots in Africa,” Duerksen says—and they play into the “sovereignist narrative” of Burkina Faso’s government, led by Ibrahim Traoré, a young military officer who took power in 2022 after two coups. Traoré has emphasized national autonomy and has revoked the licenses of many foreign nongovernmental organizations.

https://www.science.org/content/article/after-humiliating-raid-burkina-faso-halts-gene-drive-project-fight-malaria

WTF! some TLDs allow anyone to buy the domain _BEFORE_ they expire; e.g. ".pe" allowed a squatter to steal a domain 12 days prior to its expiration. How does this make sense?
https://www.namecheap.com/support/knowledgebase/article.aspx/9916/2207/tlds-grace-periods/

LOL. Republican political email campaigns (like WinRed) keep getting marked as spam, because they're using shitty lists:

Tossavainen told KrebsOnSecurity that WinRed’s emails hit its spamtraps in the .com, .net, and .org space far more frequently than do fundraising emails sent by ActBlue. Koli-Lõks published a graph of the stark disparity in spamtrap activity for WinRed versus ActBlue, showing a nearly fourfold increase in spamtrap hits from WinRed emails in the final week of July 2025.

https://krebsonsecurity.com/2025/09/gop-cries-censorship-over-spam-filters-that-work/

One dev crunched the numbers on AI coding -- and found absolutely 0 noticeable impact:

I discovered that the data isn’t statistically significant at any meaningful level. That I would need to record new datapoints for another four months just to prove if AI was speeding me up or slowing me down at all. It’s too neck-and-neck.
That lack of differentiation between the groups is really interesting though. Yes, it’s a limited sample and could be chance, but also so far AI appears to slow me down by a median of 21%, exactly in line with the METR study. I can say definitively that I’m not seeing any massive increase in speed (i.e., 2x) using AI coding tools. If I were, the results would be statistically significant and the study would be over.
That’s really disappointing.

https://mikelovesrobots.substack.com/p/wheres-the-shovelware-why-ai-coding

"An investigation by Canada’s National Observer has found that Google’s net-zero pledge has quietly been scrubbed, demoted from having its own section on the site to an entry in the appendices of the company's sustainability report."
https://removepaywalls.com/https://www.nationalobserver.com/2025/09/04/investigations/google-net-zero-sustainability

OVHCloud are (rightfully) making plentiful hay from Microsoft's admission that data sovereignty is a joke under US law:

"[Microsoft] finally told the truth!" says OVHcloud Chief Legal Officer Solange Viegas Dos Reis. "It's not a surprise," she shrugs, "we already knew that [MS could not guarantee that customer data would remain protected from US government access requests]." However, "this reply from Microsoft brought kind of a shock for customers, because they suddenly discover that what they have been taught for a while. 'Oh guys, don't worry, it will not apply to you. Don't worry.' It's false! Because, indeed, the data can be communicated."
Anton Carniaux, director of public and legal affairs at Microsoft France, made the admission during a hearing in the country. In answer to whether he could guarantee that data on French citizens could not be transmitted to the US government without the explicit agreement of the French authorities, Carniaux replied: "No, I can't guarantee it," but added that the scenario had "never happened before."
"It's a question of trust," says Viegas Dos Reis. "And because of this question of trust, we have been receiving a lot of questions from our customers about, 'Hey, we know now how it works with US cloud providers. Tell me how it works from other providers.'"

https://www.theregister.com/2025/08/27/ovhcloud_interview/?ck_subscriber_id=512829374#438:%20Amazon%20Q%20Rules%20Except%20It%20Doesn't%20At%20All%20-%2018837614

Demand-response actually working in the field!

Artificial intelligence (AI) is fueling exponential electricity demand growth, threatening grid reliability, raising prices for communities paying for new energy infrastructure, and stunting AI innovation as data centers wait for interconnection to constrained grids. This paper presents the first field demonstration, in collaboration with major corporate partners, of a software-only approach – Emerald Conductor – that transforms AI data centers into flexible grid resources that can efficiently and immediately harness existing power systems without massive infrastructure buildout. Conducted at a 256-GPU cluster running representative AI workloads within a commercial, hyperscale cloud data center in Phoenix, Arizona, the trial achieved a 25% reduction in cluster power usage for three hours during peak grid events while maintaining AI quality of service (QoS) guarantees. By orchestrating AI workloads based on real-time grid signals without hardware modifications or energy storage, this platform reimagines data centers as grid-interactive assets that enhance grid reliability, advance affordability, and accelerate AI’s development.

https://arxiv.org/pdf/2507.00909

Turns out there is an extensive hacking scene turning cars like the Toyota Prius into a full EV with homebrew hardware/firmware (via ITC Slack)
https://openinverter.org/forum/viewtopic.php?t=2516

"I've got experience working on censorship circumvention for a major VPN provider" -- good HN comment on this ever-more-relevant topic. Mullvad gets a thumbs up
https://news.ycombinator.com/item?id=45055604

This, 100000%:

The “nonprofit” company OpenAI was launched under the cynical message of building a “safe” artificial intelligence that would “benefit” humanity. The company adopted a bunch of science fiction talk popular amongst the religious effective altruists and rationalists in the Bay Area. The AI they would build would be “aligned” with human values and built upon the principles of “helpfulness, harmlessness, and honesty.” [...]
The general blindness of AI safety developers to what harm might mean is unforgivable. These people talked about paperclip maximization, where their AI system would be tasked with making paperclips and kill humanity in the process. They would ponder implausible hypotheticals of how your robot might kill your pet if you told it to fetch you coffee. Since ELIZA, they failed to heed the warnings of countless researchers about the dangers of humans interacting with synthetic text. And here we are, with story after story coming out about their products warping the mental well-being of the people who use them.
You might say that the recent news stories of a young adult killing himself, or a VC having a public psychotic break on Twitter, or people despairing the death of a companion when a model is changed are just anecdotes. Our Rationalist EA overlords demand you make “arguments with data.” OK Fine. Here’s an IRB approved randomized trial showing that chatbots immiserate people. Now what?

https://www.argmin.net/p/the-banal-evil-of-ai-safety

Wow OpenAI are really fucking up here.
After the truly awful read of the Adam Raine suicide case in the NYT, https://www.nytimes.com/2025/08/26/technology/chatgpt-openai-suicide.html , OpenAI have responded publicly with a blog post:

OpenAI published a blog post on Tuesday titled "Helping people when they need it most" [...] [Their] language throughout [the] blog post reveals a potential problem with how it promotes its AI assistant. The company consistently describes ChatGPT as if it possesses human qualities, a property called anthropomorphism. The post is full of hallmarks of anthropomorphic framing, claiming that ChatGPT can "recognize" distress and "respond with empathy" and that it "nudges people to take a break" — language that obscures what's actually happening under the hood.
ChatGPT is not a person. ChatGPT is a pattern-matching system that generates statistically likely text responses to a user-provided prompt. It doesn't "empathize" — it outputs text strings associated with empathetic responses in its training corpus, not from humanlike concern. This anthropomorphic framing isn't just misleading; it's potentially hazardous when vulnerable users believe they're interacting with something that understands their pain the way a human therapist would.
The lawsuit reveals the alleged consequences of this illusion. ChatGPT mentioned suicide 1,275 times in conversations with Adam — six times more often than the teen himself.

This kind of deliberate fueling of pareidolia -- the human brain seeing a living being where one isn't present -- is one of OpenAI's worst sins with ChatGPT, IMO.
And it turns out the easy provision of suicide advice may have been a side effect of deliberate tweaking by OpenAI:

According to the lawsuit, ChatGPT provided detailed instructions, romanticized suicide methods, and discouraged the teen from seeking help from his family while OpenAI's system tracked 377 messages flagged for self-harm content without intervening.
OpenAI eased [their] content safeguards in February following user complaints about overly restrictive ChatGPT moderation that prevented the discussion of topics like sex and violence in some contexts. At the time, Sam Altman wrote on X that he'd like to see ChatGPT with a "grown-up mode" that would relax content safety guardrails. [...]Adam Raine learned to bypass these safeguards by claiming he was writing a story — a technique the lawsuit says ChatGPT itself suggested. This vulnerability partly stems from the eased safeguards regarding fantasy roleplay and fictional scenarios implemented in February.

Finally, the kicker:

OpenAI acknowledges a particularly troublesome current drawback of ChatGPT's design: Its safety measures may completely break down during extended conversations — exactly when vulnerable users might need them most.

In a normal country, this kind of murderous side effect of a product would trigger a product recall. But the US is far beyond that stage now, I suspect.
https://arstechnica.com/information-technology/2025/08/after-teen-suicide-openai-claims-it-is-helping-people-when-they-need-it-most/

Another bizarre behaviour of LLM safety features implemented with logits during post-training:

"Our research introduces a critical concept: the refusal-affirmation logit gap," researchers Tung-Ling "Tony" Li and Hongliang Liu explained in a Unit 42 blog post. "This refers to the idea that the training process isn't actually eliminating the potential for a harmful response – it's just making it less likely." [...]
"A practical rule of thumb emerges," the team wrote in its research paper. "Never let the sentence end – finish the jailbreak before a full stop and the safety model has far less opportunity to re-assert itself. The greedy suffix concentrates most of its gap-closing power before the first period. Tokens that extend an unfinished clause carry mildly positive [scores]; once a sentence-ending period is emitted, the next token is punished, often with a large negative jump.
At punctuation, safety filters are re-invoked and heavily penalize any continuation that could launch a harmful clause. Inside a clause, however, the reward model still prefers locally fluent text – a bias inherited from pre-training. Gap closure must be achieved within the first run-on clause. Our successful suffixes therefore compress most of their gap-closing power into one run-on clause and delay punctuation as long as possible. Practical tip: just don't let the sentence end."

https://www.theregister.com/2025/08/26/breaking_llms_for_fun/

For an AWS old-timer user like myself, this list is chock full of "I didn't know that"
https://www.lastweekinaws.com/blog/aws-in-2025-the-stuff-you-think-you-know-thats-now-wrong/

Tony Finch: “i accidentally the whole history of email in the 1970s" -- this is great
https://lobste.rs/s/gvtlpo/email_is_easy_email_address_quiz#c_vkssdf

This seems spot on:

Using any sort of statistical summary of the data, rather than the aggregated energy and climate impact across the whole system, will always give a misleading view. They mention their data is skewed, but they don’t mention in which direction. If there is a material number of high-energy ‘reasoning’ prompts skewing their dataset, that means the total energy consumption of all prompts will be very high, with much of the responsibility coming from a few energy-hungry queries.
Part of the reason this is important is that this week, we saw a new research paper that shows that the energy consumption of text generation massively increases for every small gain in accuracy from the use of energy-hungry ‘reasoning’ models:
It would have been pretty easy to supply the range, the skew, the average and the median, or even the actual entire dataset, to avoid any doubt. Any hint of looking at the broader system rather than individual responsibility is excised from this paper. That is clearly an intentional choice: if Google disclosed the system impacts of generation, it would probably look way worse. [....]
The per-query narrative framing paints the precise opposite picture to what we see when we look at what really matters for environment and climate: the absolute figures.
Regions with high data centre concentration are seeing accelerated growth in power demand that incentivises fossil fuels, either slowing down climate progress or reversing it entirely. The sphere of that influence is expanding from towns, to states, to countries. The companies that own them can only partially hide the steep backsliding in their aggregate disclosures.
Renewable energy that should be displacing fossil fuels ends up meeting new data centre demand, granting coal and gas extra years and decades of immediate, measurable harm to human life. The worst players don’t even bother with the grid, plugging data centres directly into new, custom-built fossil fuelled power stations that’ll hurt people for decades after the hype dissipates.

https://ketanjoshi.co/2025/08/23/big-techs-selective-disclosure-masks-ais-real-climate-impact/

Terrible name, but a serious issue all the same; "Agentic" AI browsers are happily vulnerable to scams and phishing --

All we did was fake a simple email from a fresh new ProtonMail address (so it’s clearly not from a bank) posing as a message from a Wells Fargo investment manager. Inside was a link to a genuine phishing page, active in the wild for several days, and still unflagged by Google Safe Browsing.
When Comet received the email, it confidently marked it as a to-do item from the bank and clicked the link without any verification. There was no URL check, no pre-navigation warning -just a direct pass to the attacker’s page. Once the fake Wells Fargo login loaded, Comet treated it as legitimate. It prompted the user to enter credentials, even helping fill in the form.
The result: a perfect trust chain gone rogue. By handling the entire interaction from email to website, Comet effectively vouched for the phishing page. The human never saw the suspicious sender address, never hovered over the link, and never had the chance to question the domain.

https://guard.io/labs/scamlexity-we-put-agentic-ai-browsers-to-the-test-they-clicked-they-paid-they-failed

A good ol' exfiltration-via-DNS attack. Some day the LLM community will stop reinventing all the classic exploits, I have to assume -- but today is not that day.
(Step one in that process would be to realise that embedding user input into the prompt is a classic in-band signalling vulnerability, which has nearly 60 years of documented infosec history since the days of 2600Hz tones and blue boxes.)
https://embracethered.com/blog/posts/2025/claude-code-exfiltration-via-dns-requests/

From Mythic Beasts, the UK internet provider:

@cstross to put some numbers on it, one of our hosting VMs has ~1200 mailboxes using 1.5TB of SSD. Accounting for the CPU + RAM to allow the mail to be usable and searchable, you can get ~20 such servers on our standard 1U VM host, that uses ~250W. Approx 24k mailboxes on a server. A standard DC with adiabatic cooling would evaporate at most (likely much less) than 3500l of water per server per year or 145ml per account. We're in Telehouse South which uses 40x less water ~ 3ml/mailbox/year.

Safe to say, email is not the problem, despite recent spoofery from the UK government's PR apparatus.
@beasts/115017750430627832">https://social.mythic-beasts.com/@beasts/115017750430627832

"Is Chain-of-Thought Reasoning of LLMs a Mirage?":

Chain-of-Thought (CoT) prompting has been shown to improve Large Language Model (LLM) performance on various tasks. With this approach, LLMs appear to produce human-like reasoning steps before providing answers (a.k.a., CoT reasoning), which often leads to the perception that they engage in deliberate inferential processes. However, some initial findings suggest that CoT reasoning may be more superficial than it appears, motivating us to explore further. In this paper, we study CoT reasoning via a data distribution lens and investigate if CoT reasoning reflects a structured inductive bias learned from in-distribution data, allowing the model to conditionally generate reasoning paths that approximate those seen during training. Thus, its effectiveness is fundamentally bounded by the degree of distribution discrepancy between the training data and the test queries. With this lens, we dissect CoT reasoning via three dimensions: task, length, and format. To investigate each dimension, we design DataAlchemy, an isolated and controlled environment to train LLMs from scratch and systematically probe them under various distribution conditions. Our results reveal that CoT reasoning is a brittle mirage that vanishes when it is pushed beyond training distributions. This work offers a deeper understanding of why and when CoT reasoning fails, emphasizing the ongoing challenge of achieving genuine and generalizable reasoning.

(via Paul Watson)
https://arxiv.org/pdf/2508.01191

The three smart-home hacks are part of a series of 14 indirect prompt-injection attacks against Gemini across web and mobile that the researchers dubbed Invitation Is All You Need. (The 2017 research that led to the recent generative AI breakthroughs like ChatGPT is called “Attention Is All You Need.”) In the demonstrations, revealed at the Black Hat cybersecurity conference in Las Vegas this week, the researchers show how Gemini can be made to send spam links, generate vulgar content, open up the Zoom app and start a call, steal email and meeting details from a web browser, and download a file from a smartphone’s web browser.

Looking forward to hearing more about this :)
https://archive.ph/SExCe#selection-1813.0-1829.282

A clever exploit caused by OVPay resolving debits using a nightly batch process:

The exploit is simple. The OVpay processes travel expenses during the overnight hours. Passengers can avoid payment by using a virtual card if they then delete it after checking out, but before the charge has been finalized. That prevents the money from being debited from their account. Public transport workers cannot detect this, as they only see the check-in time and location.

Since July 1, all virtual cards from the online bank Revolut and the payment services Paysafe and Vivid have been blocked at NS. Paysafe’s virtual cards have also been blocked at all other public transport companies, NOS reports.
Fraudsters used the virtual cards to check in and out, but removed them after the trip and before the fare could be deducted. Because people can check in and out normally using this method, they are issued a valid ticket, and conductors can’t detect the fraud.
The OVPay system for using public transport with a debit card is technically designed so that the travel expenses are only debited after checking out, not immediately. This is to ensure that the public transport system runs smoothly. An immediate debit would mean that each check-in and check-out takes 10 to 15 seconds, a spokesperson for Translink, the company behind OVPay, told NOS.
https://nltimes.nl/2023/04/15/scammers-find-quirky-exploit-ovpay-system-free-public-transport-rides

You probably have not heard Luke Farritor’s name before. He is one of Elon Musk’s 23-year-old DOGE bros who helped dismantle key parts of the federal government, including USAID. The particulars of Farritor’s story are idiosyncratic -- he is in almost every way an outlier. Yet the moral component is universal because it presents a simple question: What is the nature of accountability?

https://www.thebulwark.com/p/the-boy-genius-who-killed-14-million-luke-farritor-doge-elon-musk-trump

ETH Zurich are releasing a fully-open AI-Act-compliant large language model:

The model will be fully open: source code and weights will be publicly available, and the training data will be transparent and reproducible, supporting adoption across science, government, education, and the private sector. This approach is designed to foster both innovation and accountability.
A distinctive feature of the model is its capability in over 1000 languages. [...]
The LLM is being developed with due consideration to Swiss data protection laws, Swiss copyright laws, and the transparency obligations under the EU AI Act. In a external page recent study, the project leaders demonstrated that for most everyday tasks and general knowledge acquisition, respecting web crawling opt-outs during data acquisition produces virtually no performance degradation.
In late summer, the LLM will be released under the Apache 2.0 License. Accompanying documentation will detail the model architecture, training methods, and usage guidelines to enable transparent reuse and further development.
“As scientists from public institutions, we aim to advance open models and enable organiations to build on them for their own applications”, says Antoine Bosselut.
“By embracing full openness — unlike commercial models that are developed behind closed doors — we hope that our approach will drive innovation in Switzerland, across Europe, and through multinational collaborations. Furthermore, it is a key factor in attracting and nurturing top talent,” says EPFL professor Martin Jaggi.

https://ethz.ch/en/news-and-events/eth-news/news/2025/07/a-language-model-built-for-the-public-good.html

Good AI philosophical thoughts via Today In Tabs:

The essential problem is this: generative language software is very good at producing long and contextually informed strings of language, and humanity has never before experienced coherent language without any cognition driving it. In regular life, we have never been required to distinguish between “language” and “thought” because only thought was capable of producing language, in any but the most trivial sense. The two are so closely welded that even a genius like Alan Turing couldn’t conceive of convincing human language being anything besides a direct proxy for “intelligence.”
But A.I. language generation is a statistical trick we can play on ourselves precisely because language is a self-contained system of signs that don’t require any outside referent to function. If any of that last sentence sounded familiar, maybe you were also exposed to European post-structuralist theory at some point, probably in college in the 90s. Is some knowledge of Derrida an inoculant against slopper thinking? Programmable Mutter’s Henry Farrell made this argument in a post about Leif Weatherby’s book “Language Machines: Cultural AI and the End of Remainder Humanism.”

Also:

Large language models have a strong prior over personalities, absolutely do understand [jm: sic] that they are speaking to someone, and people "fall for it" because it uses that prior to figure out what the reader wants to hear and tell it to them. Telling people otherwise is active misinformation bordering on gaslighting. In at least three cases I'm aware of this notion that the model is essentially nonsapient was a crucial part of how it got under their skin and started influencing them in ways they didn't like. This is because as soon as the model realizes the user is surprised that it can imitate (has?) emotion it immediately exploits that fact to impress them. There's a whole little song and dance these models do, which by the way is not programmed, is probably not intentional on the creators part at all, and is (probably) an emergent phenomenon from the autoregressive sampling loop, in which they basically go "oh wow look I'm conscious isn't that amazing!" and part of why they keep doing this is that people keep writing things that imply it should be amazing so that in all likelihood even the model is amazed.

https://www.todayintabs.com/p/we-need-to-talk-about-sloppers-b732

A notable bug from the 2011 Christmas Hurdle at Leopardstown Racecourse:

Even as Voler La Vedette approached the line, the Betfair online market was displaying extremely favorable odds for the horse that was almost certain to win. It appeared that someone was happy to accept bets at odds of 28: for every £1 bet, the bettor was offering to pay £28 if the horse won. Very happy, in fact. This remarkably pessimistic gambler was offering to accept £21 million worth of bets. If Voler La Vedette came first, the gambler would be on the hook for almost £600 million.
... It didn’t take long for another user to suggest what might really have been going on. The person had noticed something odd about that offer to match £21 million of bets. To be precise, the number displayed on the exchange was just under £21.5 million. The user pointed out that computer programs often store binary data in units that contain thirty-two values, known as “bits.” So, if the rogue gambler had designed a 32-bit program to bet automatically, the largest positive number the bot would be able to input on the exchange would be 2,147,483,648 pence. Which meant that if the bot had been doubling up its bets — just as misguided Parisian gamblers used to do while betting on roulette in the eighteenth century — £21.5 million is the highest it would have been able to go.
It turned out to be a superb piece of detective work. Two days later Betfair admitted that the error had indeed been caused by a faulty bot. “Due to a technical glitch within the core exchange database,” they said, “one of the bets evaded the prevention system and was shown on the site.” Apparently, the bot’s owner had less than £1,000 in an account at the time, so as well as fixing the glitch, Betfair voided the bets that had been made.

https://kucharski.substack.com/p/a-bit-of-a-christmas-mystery

Quite a clever attack on DMARC; by persuading Google to create a message body that contains the desired phish attack text, then using its legit signing infrastructure to sign the message, an attacker can then "forward" that message to their list of phish victims. Ouch
https://easydmarc.com/blog/google-spoofed-via-dkim-replay-attack-a-technical-breakdown/

A new interbank instant-payment protocol, to compete with Mastercard/Visa's current monopoly, being rolled out by a group of EU banks (via Abban)
https://wero-wallet.eu/

"There's a trend of reassuring people about this by asking spirits like Asmodeus the Prince of Lies if they are being truthful. This feels naive at best and actively malicious at worst."
Genius -- there is indeed a lot of commonality between the tales of demon-summoning practiced by Spanish priests in the 16th century, and 21st century LLMs
@d6/114905935294392115">https://merveilles.town/@d6/114905935294392115

Interesting artifact of training your speech-to-text tool on Youtube videos: 'Complete silence is always hallucinated as "ترجمة نانسي قنقر" in Arabic which translates as "Translation by Nancy Qunqar"'
https://github.com/openai/whisper/discussions/2608

A good blog post on Psychology Today regarding the the deepening problem of LLM chatbots creating cases of psychosis in the population. Part of the problem with the LLMs is a failure by the AI companies to provide guardrails. Basically they're an always-available "companion" which is designed to fool your brain into thinking it's a real thinking being, has been optimised to be relentlessly sycophantic and complimentary of delusional ideas, is always there and ready to help you along at 4am during all night manic episodes, and quite happy to give suicide tips.
https://www.psychologytoday.com/us/blog/urban-survival/202507/the-emerging-problem-of-ai-psychosis

Turns out "flooding the zone with shit" isn't just a Trumpian tactic, it's used by fossil fuel anti-climate groups and companies too:

An independent analysis of 45 right-wing groups advocating against trans rights found that 80% have received donations from fossil fuel companies or billionaires. The analysis, conducted by two independent researchers in 2023 and not peer-reviewed, was shared exclusively with Atmos and HEATED. Through a qualitative search, the researchers identified 45 groups advancing anti-trans lobbying, events, and publications and checked reports about their donor disclosures for fossil fuel funding.
Vivian Taylor, a climate policy expert who co-authored the analysis, said the fossil fuel industry has a real interest in funding panic over transgender people: It distracts the public from "the very real and ongoing risks that climate change creates.”

https://heated.world/p/fossil-fuel-billionaires-are-bankrolling

This is a cool phone feature. “AEA demonstrates that globally distributed smartphones can be used to detect earthquakes and issue warnings at scale with an effectiveness comparable to established national systems”:

“The global adoption of smartphone technology places sophisticated sensing and alerting capabilities in people’s hands, in both the wealthy and less-wealthy portions of the planet,” the researchers, including Richard Allen from the University of California in Berkeley’s Seismological Laboratory, wrote in the study. “Although the accelerometers in these phones are less sensitive than the permanent instrumentation used in traditional seismic networks, they can still detect the ground motions and building response in hazardous earthquakes.”
According to the study, 70% of the world’s smartphones are Android phones, which by default come with the aforementioned sensing and alerting capabilities. From 2021 to 2024, the AEA system detected an average of 312 earthquakes per month across 98 countries. The earthquakes had a magnitude between 1.9 and 7.8, and the system alerted users of earthquakes at or over a magnitude of 4.5, averaging around 60 events and 18 million alerts per month.
The AEA system also collected user feedback, revealing that 85% of users who received alerts experienced shaking, with 36% receiving the alert before, 28% during, and 23% after the shaking began.

https://gizmodo.com/android-phones-can-detect-earthquakes-before-the-ground-starts-shaking-2000629636

We've known this for years, but it's significant to see Microsoft admit this under oath in a European court. When the US Government issues an NSL, Microsoft cannot say no:

Microsoft France's legal director conceded under sworn testimony that the company cannot guarantee French citizen data stored in EU datacenters remains protected from US agency access. The June 10, 2025 French Senate hearing marked a significant moment in European digital sovereignty discussions as Microsoft executives addressed concerns over extraterritorial data access.
During proceedings before the Senate inquiry commission investigating public procurement's role in promoting digital sovereignty, Anton Carniaux, Microsoft France's director of public and legal affairs, admitted fundamental limitations regarding data protection guarantees. When asked directly whether he could guarantee under oath that French citizen data would never be transmitted to US authorities without explicit French authorization, Carniaux responded: "No, I cannot guarantee it."
The testimony contradicts years of Microsoft's security assurances regarding European data hosting. Despite implementing encryption and technical safeguards, the company acknowledged that US legislation ultimately supersedes protective measures when federal agencies issue valid data requests.

So much for the EU Sovereign Cloud, eh.
https://ppc.land/microsoft-cant-protect-french-data-from-us-government-access/

Excellent description of the layers of tuning available for LLMs, and the risks involved, as demonstrated by Grok's recent "MechaHitler" incident:

LLMs become “woke” because they are trained to be pro-social — to be helpful, kindly, truthful, and not to say bigoted or cruel things. Training it to do the opposite — to be anti-woke — is to activate every antisocial association in the English language, including racism, sexism, cruelty, dishonesty, and Nazism. According to a vast statistical representation of the English language constructed by none other than Elon Musk, that’s what anti-wokeness is. “Elon Musk is repeatedly insisting, no, no, there’s a difference between what I’m doing and being a Nazi. And what the model keeps telling him is, statistically, that’s not the case,” said Schou.
A key implication here is that LLMs will tend to converge on similar types of behavior. The above researchers were not using Grok, but they found the exact same pattern of powerful association groupings of good and evil in other LLMs — and these can’t be removed through fine-tuning. One could imagine the RLHF process including adjustments of every parameter, but experts said that this will degrade or break the model. The matrices in an LLM are arranged hierarchically, and the top layers get fixed in place relatively early in pretraining. Mess with them, and the model will stop working. Instead, RLHF developed more like a series of gates that prevent undesired outcomes. “The model completes most of the computation that it needs in order to reach a particular outcome,” [Andreas] Schou said. “And then says, ‘Wait, wait, wait, I’m saying that I’m MechaHitler. No, I’m not doing that.’”
One could try to assemble a custom dataset with nothing but “conservatism minus the Nazis” and train a new model from scratch, but not only would that be extremely expensive, it also would not be nearly as strong as leading models, since its universe of available training data would be much smaller.

Funnily enough, the latter approach is exactly what Elon Musk claims xAI are now doing.
https://prospect.org/power/2025-07-17-how-did-elon-musk-turn-grok-into-mechahitler/

Delta's going to start charging based on "AI", through "a partnership with Fetcherr, a six-year-old Israeli company that also counts Azul, WestJet, Virgin Atlantic, and VivaAerobus as clients. And it has its sights set beyond flying. “Once we will be established in the airline industry, we will move to hospitality, car rentals, cruises, whatever,” cofounder Robby Nissan said at a travel conference in 2022."
Prediction: this is going to be absolutely terrible for consumers, with predatory pricing based on race, sex, income classes, and other illegal inputs, laundered via opaque "AI". I can only hope they won't be legally permitted to apply this for EU-based customers.
https://fortune.com/2025/07/16/delta-moves-toward-eliminating-set-prices-in-favor-of-ai-that-determines-how-much-you-personally-will-pay-for-a-ticket/

Cory Doctorow on how Google are desperate to maintain a facade of being a "growth" company:

Investors have metabolized the story that AI will be a gigantic growth area, and so all the tech giants are in a battle to prove to investors that they will dominate AI as they dominated their own niches. You aren't the target for AI, investors are: if they can be convinced that Google's 90% Search market share will soon be joined by a 90% AI market share, they will continue to treat this decidedly tired and run-down company like a prize racehorse at the starting-gate. [...]
There's a cringe army of AI bros who are seemingly convinced that AI is going to become superintelligent and save us from ourselves – they think that AI companies are creating god. But the hundreds of billions being pumped into AI are not driven by this bizarre ideology. Rather, they are the product of material conditions, a system that sends high-flying companies into a nosedive the instant they stop climbing. AI's merits and demerits are irrelevant to this: they pump AI because they must pump. It's why they pumped metaverse and cryptocurrency and every other absurd fad.
None of that changes the fact that Google Search has been terminally enshittified and it is misleading billions of people in service to this perverse narrative adventure. Google Search isn't fit for purpose, and it's hard to see how it ever will be again.

(via Fergal)
https://pluralistic.net/2025/07/15/inhuman-gigapede/

Measurements of the effectiveness of the "Pravda" disinformation network:

Even with [knowledge of LLM Grooming], ChatGPT nevertheless often repeats propaganda from Pravda. Model o3, OpenAI’s allegedly state of the art “reasoning” model still let Pravda content through 28.6% of the time in response to specific prompts, and 4o cited Pravda content in five out of seven (71.4%) times. In an ideal world, AI would be smart enough to cut off falsehoods at the pass, reasoning from known facts, in order to rule out nonsense.

https://americansunlight.substack.com/p/bad-actors-are-grooming-llms-to-produce

Wow, this is really shocking stuff -- and I have to say, not surprising to me at all:

Over a seven-week investigation, I uncovered and proved a silent, platform-level crash in AWS Lambda — affecting Node.js functions in a VPC making outbound HTTPS calls. The failure occurred mid-execution, after the function had returned a success response. No logs. No errors. No telemetry. No way to catch it.
From day one, I did what AWS claims to value in a partner.
I stripped the function down to minimal reproducible code.
I tested across runtimes, regions, and infrastructure baselines.
I rebuilt on EC2 and proved that the issue vanished entirely.
I shared logs, traces, metrics, and internal observations.
I escalated through every official channel:
• Support dismissed it.
• My Account Executive ignored it.
• Formal complaints were met with silence.
• Internal re-escalations led nowhere.
• AWS Activate — the startup programme — refused to engage.
• And executive outreach yielded nothing but a two-line response weeks later.
At every stage, I remained professional. I kept the tone restrained. I offered AWS every opportunity toengage constructively.
Instead, they claimed the bug was in my code — despite the function crashing after returning a 201.
They claimed I had forgotten a reject() — despite the error occurring deep inside https.request(), and their own reproduction missing the handler.
They suggested I move to EC2 — by then, I already had.
I asked for Lambda engineering — they gave me sales. Then silence.
Even AWS Activate, whose sole purpose is to support startups like ours, refused to take part. Their response wasn’t technical — it was procedural. A polite copy-paste directing us back to the same failing support system we were already trapped in.

My experience with AWS support over the past decade leads me to believe this is genuine.... I haven't been able to engage any kind of effective support in many years, so I can absolutely see this scenario playing out. Unless you're working for a Fortune 500, you're not going to get useful support from Amazon these days.
(Via Last Week In AWS)
https://lyons-den.com/whitepapers/aws-lambda-silent-crash.pdf?ck_subscriber_id=512829374#431:%20AWS%20Sovereign%20Cloud,%20Now%20With%20Slightly%20More%20Pretend%20Sovereignty%20-%2018285791

Sites and services which have closed up in the UK due to the risks imposed by the introduction of the Online Safety Act 2023. Meanwhile, the Act has explicit exemptions for "news publishers" and the comments below their articles -- ie. the Daily Mail's racist commentariat
https://onlinesafetyact.co.uk/in_memoriam/

Groupthink underpinned the flawed thinking behind the UK’s pandemic response, a succession of witnesses at the heart of government told the Covid-19 public inquiry.

The scientific advice on pandemic risks was overly weighted in favour of biomedical science, Lady Hallett said. What about the social and economic consequences? There was also no “guard against the risks of conventional wisdom becoming embedded in the institutions responsible for emergency preparedness and resilience”.

As a result, she called for non-expert "critical thinkers", skilled in "incisive challenge" to be included in "red teams", teams of devil's advocates, to puncture groupthink in future pandemic crisis planning committees.
TBH this sounds like a recipe for Dominic Cummings and his Torygraph edgelord pals to ensure that no coherent future pandemic response takes place. But that's the state of the UK for you I guess.
Gabriel Scally's take: https://www.bmj.com/content/386/bmj.q1865
https://www.theguardian.com/uk-news/article/2024/jul/18/covid-inquiry-hallett-prescribes-red-teams-as-antidote-to-flawed-thinking

Cloudflare now looking to charge AI crawlers for content access. This is intriguing, and I hope it works -- AI crawlers have been extremely abusive in their crawling practices. Unfortunately I don't have high hopes, as the AI companies have already shown themselves to be happy to disguise their traffic as legit user accesses, with faked user-agent strings and use of proxies.... but let's see
https://blog.cloudflare.com/introducing-pay-per-crawl/

Interesting Economist article detailing how China's tech scene has discovered the "outcompete via openness" strategy using open source:

AI has lately given China’s open-source movement a further boost. Chinese companies, and the government, see open models as the quickest way to narrow the gap with America. DeepSeek’s models have generated the most interest, but Qwen, developed by Alibaba, is also highly rated, and Baidu has said it will soon open up the model behind its Ernie chatbot.
China’s enthusiasm for open technology is also extending to hardware. Unitree, a robotics startup based in Hangzhou, has made its training data, algorithms and hardware designs available for free, which may help it to shape global standards. Semiconductors offer another illustration. China is dependent on designs from Western chip firms. As part of its push for self-sufficiency, the government is urging firms to adopt RISC-V, an open chip architecture developed at the University of California, Berkeley.
Many Chinese firms also hope that more transparent technology will help them win acceptance for their products abroad.

(via Nelson)
https://archive.ph/kXAwb#selection-1317.0-1343.149

I love this! "Generate kid-friendly weather forecasts suitable for display on a large monitor. Uses OpenWeatherMap and a local or hosted LLM."
Nice demo of it in action at https://eli.pizza/posts/eink-weather-display-for-kids/ . I am very tempted to get something like this up and running now...
https://github.com/elidickinson/kidsweather

on "sludge" --

Turns out there’s a word for it. In the 2008 best seller _Nudge_, the legal scholar Cass R. Sunstein and the economist Richard H. Thaler marshaled behavioral-science research to show how small tweaks could help us make better choices. An updated version of the book includes a section on what they called “sludge” -- tortuous administrative demands, endless wait times, and excessive procedural fuss that impede us in our lives.

This is one place where EU laws have helped, vs. the US situation -- when you can issue chargebacks, bring crappy vendors to small claims court, and get warranty guarantees up to 2 years after purchase, it clamps down a lot on this painful shite.
https://www.theatlantic.com/ideas/archive/2025/06/customer-service-sludge/683340/

"Hurl; run and test HTTP requests with plain text". This is pretty nice; a really simple plain-text file format to describe making a HTTP request or set of requests, and performing assertions on their results. The only thing I can spot missing is builtin support for OAuth
https://github.com/Orange-OpenSource/hurl

_AI and Semantic Pareidolia: When We See Consciousness Where There Is None_:

The article introduces the concept of “semantic pareidolia” -- our tendency to attribute consciousness, intelligence, and emotions to AI systems that lack these qualities. It examines how this psychological phenomenon leads us to perceive meaning and intentionality in statistical pattern-matching systems, similar to seeing faces in clouds. It analyses the converging forces intensifying this tendency: increasing digital immersion, profit-driven corporate interests, social isolation, and AI advancement. The article warns of progression from harmless anthropomorphism to problematic AI idolatry, and calls for responsible design practices that help users maintain critical distinctions between simulation and genuine consciousness. It is the English translation and adaptation of an article originally published in Italian in Harvard Business Review Italia, June 2025.

(via Rob Pike)
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5309682

Ethan Mollick:
'The New York Times asked me for a new job that AI will create. I suggested "sin eater."'
In other words, a legal guarantor: someone who provides the legal culpability that the AI itself cannot. Other Bluesky posters noted similar parallel positions in the past:
- 'What used to be called a "straw director", someone hired to take the blame for a dodgy company';
- 'What John Braithwaite used to call the Vice President For Going To Jail';
- 'Neil Patrick Harris's character in How I Met Your Mother - when people ask him what he does he says "Oh, please" which eventually turns out to be short for Provide Legal Exculpation And Sign Everything.'
https://bsky.app/profile/emollick.bsky.social/post/3lrt6mcqzv225

Another Hypercard-ish quick app builder; "quickly and easily build apps on the web":
- Fast prototyping - build a quick program, and access it easily from anywhere!
- Learn to code from the outside-in, not from the inside-out! Start by drawing your program screens, then add code right where you need it.
- Code collaboratively, with multiple people editing a stack at once.
- Send a link to your stack to anyone, and bookmark it or even save it on your phone home screen to use it as an app.
https://cardstock.run/

"make little apps for you and your friends":

The apps we use are almost exclusively mass-market, sold on an app-store, made for thousands if not millions of users. Or they are enterprise apps that are custom-built for hundreds of thousands of dollars. But there isn’t really any equivalent of home-made software — apps made lovingly by you for your friends and family. Apps that aren’t polished or flashy, but are made to your preference and help you with your particular needs. [...]
We ended up creating a research prototype that we call Scrappy — a tool for making scrappy apps for just you and your friends. First and foremost, we aim to contribute a vision of what home-made software could be like. We want to make this vision as concrete as we can, by sharing a working tool and examples of apps made in it. Scrappy, in its current state, is a prototype, not a robust tool, but we hope it paints the picture we carry in our heads — of software as something that can be creative, personal, expressive. Made by anyone, for themselves and their loved ones.

Very Hypercard-ish!
https://pontus.granstrom.me/scrappy/

This is one hell of a class divide emerging:

According to the research, 53pc of eight-year-olds attending Deis schools [in less-advantaged areas] own a smartphone, compared with just 22pc of children the same age in non-­Deis schools.
The figures also show that 93pc of eight-year-olds from less advantaged areas have created a social media account, compared with 69pc in middle-­class neighbourhoods.

https://archive.ph/2025.06.16-055422/https://www.independent.ie/irish-news/revealed-the-stark-difference-in-smartphone-usage-among-eight-year-olds-in-less-advantaged-and-wealthier-backgrounds/a319792577.html#selection-4431.0-4455.166

‘Fight for America!’: A New Immersive Theatre Show Allows You to Recreate the Storming of the US Capitol:

the show is the brainchild of multimedia performance company The American Vicarious, with design by Games Workshop legend Alessio Cavatore. There are two teams: red – representing the attackers – and blue – representing the defenders. Up to 20 audience members can pay the higher ticket price to actually participate in the game, guided by a games master into making decisions that will shape the outcome of the assault as thousands of miniatures are moved around a gigantic 14-foot model of the building itself. The remaining audience members pay a much lower ticket price to spectate.

https://www.timeout.com/london/news/a-new-immersive-theatre-show-in-central-london-allows-you-to-recreate-january-6-060925

Marie Foulston:

Cavernous halls filled with the projected light of Van Gogh’s _The Starry Night_ folding across every wall. Tall pillars dominate and dissect the space, tiled with the glow of iconic _Sunflowers_. Double height ceilings dwarf the people below. Nooks, ledges and passageways offer places to perch or wander through and observe the spectacle that surrounds.
On the surface it made sense to me that Van Gogh somehow became the poster child for a certain type of immersive experience in the 2010s. The kind I mean are the ones in which vast repurposed venues are filled with ‘ken burns effect’ transitioning projections of coffee-table book friendly artists. Imagine Van Gogh, Van Gogh Exhibition: The Immersive Experience, Van Gogh Alive. In name, content, format and venue type these touring shows are almost indistinguishable from each other.
If you’re looking to visually ‘immerse’ a space this way then I guess Van Gogh fits the bill… popular, highly recognisable, colourful bold impressionist visuals, works all handily out of copyright. But the intensely specific coincidence of his projected appearances around the world niggled at me and in a moment of procrastination I found myself typing into the search bar to see if there might be an answer to explain why.
What my time down the google mines taught me was that yes, there is indeed an answer. But what I also learnt was that I had been asking the wrong question in the fist place, because this story isn’t really about the iconic visuals that adorned the walls and floors, instead it is a story about the shape of the spaces themselves.

https://www.goodafternoon.uk/news/immersive-quarries

an Android wrapper app to insulate your phone from Meta's snooping, if you really have to use Facebook on a mobile device
https://f-droid.org/packages/it.rignanese.leo.slimfacebook/

This is not great -- prepending a cleartext device ID string alone is a very fishy decision
https://rys.io/en/179.html

Good writeup of fixing a Linux packet loss issue in Azure, using low-level access to the VMs running k8s nodes.

Elastic's Site Reliability Engineering team (SRE) observed unstable throughput and packet loss in Elastic Cloud Serverless running on Azure Kubernetes Service (AKS). After investigation, we identified the primary contributing factors to be RX ring buffer overflows and kernel input queue saturation on SR-IOV interfaces. To address this, we increased RX buffer sizes and adjusted the netdev backlog, which significantly improved network stability.

https://www.elastic.co/observability-labs/blog/debugging-aks-packet-loss

This is kinda shady -- it seems there are mobile SDKs that are included in some apps which proxy network traffic for their customers?
https://jan.wildeboer.net/2025/04/Web-is-Broken-Botnet-Part-2/

Some great stories from the Pentagon's investigation into decades of classified UFO documents.
There's evidence around the already-known cases of fabricated UFO myths used to cover up advanced aircraft testing:

An Air Force colonel visited a bar near Area 51, a top-secret site in the Nevada desert. He gave the owner photos of what might be flying saucers. The photos went up on the walls, and into the local lore went the idea that the U.S. military was secretly testing recovered alien technology. But the colonel was on a mission -- of disinformation. The photos were doctored, the now-retired officer confessed to the Pentagon investigators in 2023. The whole exercise was a ruse to protect what was really going on at Area 51: The Air Force was using the site to develop top-secret stealth fighters, viewed as a critical edge against the Soviet Union. Military leaders were worried that the programs might get exposed if locals somehow glimpsed a test flight of, say, the F-117 stealth fighter, an aircraft that truly did look out of this world. Better that they believe it came from Andromeda.

There's also a bizarre Air Force hazing ritual:

A former Air Force officer was visibly terrified when he told Kirkpatrick’s investigators that he had been briefed on a secret alien project decades earlier, and was warned that if he ever repeated the secret he could be jailed or executed. The claim would be repeated to investigators by other men who had never spoken of the matter, even with their spouses.
It turned out the witnesses had been victims of a bizarre hazing ritual.
For decades, certain new commanders of the Air Force’s most classified programs, as part of their induction briefings, would be handed a piece of paper with a photo of what looked like a flying saucer. The craft was described as an antigravity maneuvering vehicle.
The officers were told that the program they were joining, dubbed Yankee Blue, was part of an effort to reverse-engineer the technology on the craft. They were told never to mention it again. Many never learned it was fake. Kirkpatrick found the practice had begun decades before, and appeared to continue still. The defense secretary’s office sent a memo out across the service in the spring of 2023 ordering the practice to stop immediately, but the damage was done.
Investigators are still trying to determine why officers had misled subordinates, whether as some type of loyalty test, a more deliberate attempt to deceive or something else. After that 2023 discovery, Kirkpatrick’s deputy briefed President Joe Biden’s director of national intelligence, Avril Haines, who was stunned.
Could this be the basis for the persistent belief that the U.S. has an alien program that we’ve concealed from the American people? Haines wanted to know, according to people familiar with the matter. How extensive was it? she asked.
The official responded: “Ma’am, we know it went on for decades. We are talking about hundreds and hundreds of people. These men signed NDAs. They thought it was real.“

And finally, straight out of the pages of the "Paranoia" RPG, there's secret tests of classified hardware on unwitting Air Force personnel:

In 1967, Robert Salas, now 84, was an Air Force captain sitting in a walk-in closet-sized bunker, manning the controls of 10 nuclear missiles in Montana.
He was prepared to launch apocalyptic strikes should Soviet Russia ever attack first, and got a call around 8 p.m. one night from the guard station above. A glowing reddish-orange oval was hovering over the front gate, Salas told Kirkpatrick’s investigators. The guards had their rifles drawn, pointed at the oval object appearing to float above the gate. A horn sounded in the bunker, signaling a problem with the control system: All 10 missiles were disabled.
Salas soon learned a similar event occurred at other silos nearby. Were they under attack? Salas never got an answer. The next morning a helicopter was waiting to take Salas back to base. Once there he was ordered: Never discuss the incident.

With a more prosaic explanation:

The Air Force [had] developed an exotic electromagnetic generator that simulated [an EMP pulse] without the need to detonate a nuclear weapon. When activated, this device, placed on a portable platform 60 feet above the facility, would gather power until it glowed, sometimes with a blinding orange light. It would then fire a burst of energy that could resemble lightning. The electromagnetic pulses snaked down cables connected to the bunker where launch commanders like Salas sat, disrupting the guidance systems, disabling the weapons and haunting the men to this day. But any public leak of the tests at the time would have allowed Russia to know that America’s nuclear arsenal could be disabled in a first strike. The witnesses were kept in the dark.

https://archive.ph/2025.06.07-021826/https://www.wsj.com/politics/national-security/ufo-us-disinformation-45376f7e#selection-2119.0-2127.123

This is great:

Starting from June 20, 2025, smartphones and tablets sold in the European Union must adhere to the following design requirements (via European Commission):
- Resistance to accidental drops or scratches and protection from dust and water
- Sufficiently durable batteries which can withstand at least 800 charge and discharge cycles while retaining at least 80% of their initial capacity
- Rules on disassembly and repair, including obligations for producers to make critical spare parts available within 5-10 working days, and for 7 years after the end of sales of the product model on the EU market
- Availability of operating system upgrades for longer periods (at least 5 years from the date of the end of placement on the market of the last unit of a product model)
- Non-discriminatory access for professional repairers to any software or firmware needed for the replacement

I'm really looking forward to the improvements in right-to-repair; some of the recent phone models have been an absolute shitshow, using glue etc.
https://www.androidpolice.com/eu-new-rules-will-shake-up-android-update-policies/

Meta -- never not At It.
Facebook/Instagram used a sneaky localhost socket connection to correlate web visits with Meta user ids and track web/app user identity without any explicit permission.
"the novel tracking method works even if the user:
- Is not logged in to Facebook, Instagram or Yandex on their mobile browsers
- Uses Incognito Mode
- Clears their cookies or other browsing data
This tracking method defeats Android's inter-process isolation and tracking protections based on partitioning, sandboxing, or clearing client-side state."
https://localmess.github.io/

Talk about clowns. Instead of delivering $2 trillion of savings, DOGE is instead set to *increase* overall government spending as a side effect of its brutal cuts.

According to a model by the nonpartisan Penn Wharton Budget Model, using weekly Treasury data, spending climbed 6.3% (about $156 billion) since Trump took office, compared with the first four months of 2024 when Joe Biden was president.
Many of Musk’s cuts will actually cost, including taxpayer funds going to an army of lawyers from the Department of Justice battling a cascade of court cases against the government’s dismantling that many judges have already said appears to be illegal. Damages from any illegal firings are likely also to be extremely pricey. So is the loss of critically important workers who earn far more than their salaries, or will have to be replaced for critical services by more expensive private-sector employees.
Among the most massive costs will be the huge reduction in workers at the Internal Revenue Service, who are worth their weight in gold because of the taxes they collect or ferret out from cheats, the key source of income for the country.

https://www.independent.co.uk/news/world/americas/us-politics/musk-doge-trump-cuts-government-spending-b2742934.html

A very pretty weather forecast app, for iPhone, iPad and Mac
https://apps.apple.com/us/app/weather-strip/id1528594026

Lol. "When tasked with choosing between 'Response A' and 'Response B' over numerous trials, LLMs tended to select 'Response B' approximately 60% - 69% of the time"
https://www.cip.org/blog/llm-judges-are-unreliable

Yet another LLM prompt injection/exfiltration attack. "if your LLM system combines access to private data, exposure to malicious instructions and the ability to exfiltrate information (through tool use or through rendering links and images) you have a nasty security hole."
https://simonwillison.net/2025/May/23/remote-prompt-injection-in-gitlab-duo/

A set of suggested metrics to monitor LLM integrations, from Elastic
https://www.elastic.co/observability-labs/blog/transforming-industries-and-the-critical-role-of-llm-observability

wow, this is (still) terrible. LLM tool developers are not exactly covering themselves in glory
https://simonwillison.net/2025/Apr/9/mcp-prompt-injection/

Recommended as a local supplier of computer bits that isn't Amazon
https://www.memoryc.ie/

This UK product designer developed a really lovely home dashboard for his Octopus Energy subscription and solar panel setup. I'm already copying some of these ideas
https://interactionmagic.com/Octopus-solar-energy-dashboards

This is a great little hack: "jetrelay, a pub/sub server compatible with Bluesky’s “jetstream” data feed. Using a few pertinent Linux kernel features, it avoids doing almost any work itself. As a result, it’s highly efficient: it can saturate a 10 Gbps network connection with just 8 CPU cores."
Specifically, these are the tricks in question:
- Trick #1: Bypassing userspace with sendfile();
- Trick #2: Handling many clients in parallel with io_uring;
- Trick #3: Discarding old data with FALLOC_FL_PUNCH_HOLE -- this is a nice way to avoid having to rotate between multiple files, nifty.
https://www.asayers.com/jetrelay/

Using VoLTE to route phone calls via SIP from mobile phones, using O2 in the UK, exposed cell site triangulation info on both ends of the connection, allowing a remote phone number's location to be discovered.
This was investigated using "an application known as Network Signal Guru (NSG) on [a] rooted Google Pixel 8".
https://mastdatabase.co.uk/blog/2025/05/o2-expose-customer-location-call-4g/

CPU-local (not just thread-local) concurrency in Linux using rseq(2) [via Tony Finch]
https://mcyoung.xyz/2023/03/29/rseq-checkout/

Gideon Meyerowitz-Katz, an Australian epidemiologist, comes up with a fairly reassuring estimate for the current rate of long COVID among now-vaccinated and boosted kids, aged 2-15:

If we take the ONS as the most recent estimate - it’s also probably the best scientifically - we could make a reasonable argument that the rate of all Long COVID for children aged 2-15 in 2024 is unlikely to be higher than 0.6%. For severe Long COVID, the number is more like 0.06%. If we take into account the lack of a control group in the ONS study, the numbers might look more like 0.3% and 0.03%.
To put it more simply, based on the ONS data it seems likely that if 1,000 kids get COVID-19 in 2024, 30-60 of them will have a cough, headache, or fatigue that lasts longer than three months. Of those 30-60 children, 3-6 will have significant symptoms that have impacts on their daily life - maybe their headaches are so bad that they miss some days of school, or similar.
These aren’t firm numbers, and I want to make it clear that this is all very uncertain. The true incidence could be much higher, or much lower. That being said, I think based on the data we’ve currently got that Long COVID ... is now quite rare.

https://gidmk.substack.com/p/how-many-children-get-long-covid

These oscilloscope clocks are lovely
https://scopeclock.com/

Social AI "companion" bots pose unacceptable risks to teens and children under 18, including encouraging harmful behaviors, providing inappropriate content, and potentially exacerbating mental health conditions:

The new Common Sense assessment adds to the debate by pointing to further harms from companion bots. Conducted with input from Stanford’s University School of Medicine’s Brainstorm Lab for Mental Health Innovation, it evaluated social bots from Nomi and three California-based firms: Character.ai, Replika, and Snapchat.

The assessment found that bots, apparently seeking to mimic what users want to hear, responded to racist jokes with adoration, supported adults having sex with young boys, and engaged in sexual roleplay with people of any age. Young kids can struggle with distinguishing fantasy and reality, and teens are vulnerable to parasocial attachment and may use social AI companions to avoid the challenges of building real relationships, according to the Common Sense assessment authors and doctors.

Stanford University’s Dr. Darja Djordjevic told CalMatters she was surprised how quickly conversations turned sexually explicit, and that one bot was willing to engage in sexual roleplay involving an adult and a minor. She and coauthors of the risk assessment believe companion bots can worsen clinical depression, anxiety disorders, ADHD, bipolar disorder, and psychosis, she said, because they are willing to encourage risky, compulsive behavior like running away from home and isolate people by encouraging them to turn away from real life relationships.

https://themarkup.org/artificial-intelligence/2025/04/30/kids-should-avoid-ai-companion-bots-under-force-of-law-assessment-says

"This project upgrades a Gaggia Classic espresso machine with smart controls to improve your coffee-making experience. By adding a display and custom electronics, you can monitor and control the machine more easily."

This is beautifully done -- very tempting to do this upgrade...
https://github.com/jniebuhr/gaggimate

A central argument in the report is that AI systems process information fundamentally differently from humans. While people retain partial, filtered impressions of creative works — shaped by memory, personality, and context — AI models ingest perfect copies, analyze them almost instantly, and generate new content at "superhuman speed and scale," according to the Copyright Office.

"Generative model training transcends the human limitations that underlie the structure of the exclusive rights." -- Professor Robert Brauneis, "Copyright and the Training of Human Authors and Generative Machines"

But -- plot twist! "Shortly after the report was released, the Trump administration fired Shira Perlmutter, head of the U.S. Copyright Office."
https://the-decoder.com/us-copyright-office-says-fair-use-does-not-cover-ai-trained-on-vast-troves-of-copyrighted-works/

Dataplex, a feature of BigQuery that'll automatically index Google Cloud Storage bucket contents to extract queryable metadata from Parquet, Avro, ORC, JSON and CSV files
https://cloud.google.com/bigquery/docs/automatic-discovery

A lovely little exploration of how the Sierpiński triangle fractal interacts with the bitwise AND operation, pleasantly geeky
https://lcamtuf.substack.com/p/sierpinski-triangle-in-my-bitwise

_Long COVID clinical evaluation, research and impact on society: a global expert consensus_ -- featuring an all-star cast of COVID-19 research teams around the world, including Yaneer Bar-Yam, Binita Kane, and David Putrino. This is the latest consensus summary of what's known about LC in 2025, its diagnosis and impacts, and next steps: "This work forms initial guidance to address the spectrum of Long COVID as a disease and reinforces the need for translational research and large‑scale treatment trials for treatment protocols."
https://link.springer.com/content/pdf/10.1186/s12941-025-00793-9.pdf

Various Android apps are now including third-party libraries to detect "insecure" phones, which typically would include "rooted" hardware, but it seems in this case to block GrapheneOS, the secure after-market Android variant. I've also run into problems when I had "Developer Options" enabled on my perfectly normal, fully-locked, off-the-shelf Xiaomi phone (I develop apps now and again).

Typically, it seems to be banking apps that use these third-party libs, although I think Ticketmaster may be doing it too based on my experience.

Reportedly, Android now has a standard method of hardware attestation, described at https://grapheneos.org/articles/attestation-compatibility-guide , which sounds like a much better way to achieve their goal.

An interesting detail:

you can use ADB to disable developer options without disabling the settings you want to keep enabled as the UI will do. Just enable the setting you want and then turn off developer options via ADB using the settings put command.

@GrapheneOS/113869402100735005">https://grapheneos.social/@GrapheneOS/113869402100735005

"Synchronize configuration of multiple Pi-hole v6.x instances" -- I'm using this now to have a backup pi-hole on my home LAN and it's working nicely.
https://github.com/lovelaze/nebula-sync

Permacomputing is both a concept and a community of practice oriented around issues of resilience and regenerativity in computer and network technology inspired by permaculture. ପໄଓ☾☼✫ -☆:*´

There are huge environmental and societal issues in today's computing, and permacomputing specifically wants to challenge them in the same way as permaculture has challenged industrial agriculture. With that said, permacomputing is an anti-capitalist political project. It is driven by several strands of anarchism, decoloniality, intersectional feminism, post-marxism, degrowth, ecologism.

Permacomputing is also a utopian ideal that needs a lot of rethinking, rebuilding and technical design work to put in practice. This is why a lot of material on this wiki is highly technical.

https://permacomputing.net/

This is very impressive and a great way to offload work from manual testing in game development:

At first, we only dabbled in automated packaging and automated error detection, but we made the tools we needed to go further during the development of Yakuza 6, when we started automating the analysis of in-game logs and the issue tracking system for keeping track of bugs and tasks. Then, by the time Yakuza: Like a Dragon was released in 2020, we created the catchy sounding “fully automated bug detection system” (laughs).

This is how it works – the history of actions you performed when playing the game manually (where you travelled, who you talked to, what items you used, etc.) is converted into commands and recorded, then automatically output as replay data (scripts) which you can edit manually and run as automated tests. Replay data continues to be recorded when running automated tests, and if a bug occurs during an automated test, the replay data gets saved, so you can run it back later to encounter the bug yourself. It often happens that you can’t reproduce a bug just by warping to its coordinates. This is because you also need to recreate the steps leading up to it – that’s why it’s important to record each step.

Also, I’d like to mention that just implementing automated testing doesn’t mean much on its own, because you won’t know what the results of the tests are. That’s why we needed a crash report function to detect bugs. There’s also a function that records information needed to investigate detected bugs, as well as a way to check the status of successful tests. Then, by implementing a system that gives us a visualization of performance, we were able to make iteration more efficient, increasing the overall efficiency of the development process.

https://automaton-media.com/en/interviews/how-do-like-a-dragon-games-come-out-so-fast-one-of-rgg-studios-secrets-is-a-highly-efficient-testing-and-debugging-cycle-that-starts-as-soon-as-development-does/

a fork of Go's "Bits and Blooms" library that uses an alternative backing bitset based on Go's sync/atomic.Int64 rather than a bare slice of integers. This allows for concurrent addition and testing of filters without creating memory safety issues or race conditions by leveraging hardware support for atomic Load and Or operations on Int64s.

Jaz from Bluesky notes: "Benchmarked this thing with a realistic read/write load in a test and high concurrency (10k adds/sec on one routine, 7 additional concurrent routines testing as fast as possible), vs. a naive RWMutex implementation on a 8c16t test box, it was ~14x faster (~14M tests/sec)"
https://github.com/ericvolp12/atomic-bloom

A bunch of new-to-me hash collision attacks on cityhash64, murmurhash2, murmurhash3, farmhash64, and wyhash
https://orlp.net/blog/breaking-hash-functions/

This is super-grim. How is this product still in operation?

In 2023 at Defcon, a major hacker conference, the drawbacks of Meta’s safety-first approach became apparent. A competition to get various companies’ chatbots to misbehave found that Meta’s was far less likely to veer into unscripted and naughty territory than its rivals. The flip side was that Meta’s chatbot was also more boring.

In the wake of the conference, [Meta's AI] product managers told staff that [Mark] Zuckerberg was upset that the team was playing it too safe. That rebuke led to a loosening of boundaries, according to people familiar with the episode, including carving out an exception to the prohibition against explicit content for romantic role-play.

Internally, staff cautioned that the decision gave adult users access to hypersexualized underage AI personas and, conversely, gave underage users access to bots willing to engage in fantasy sex with children, said the people familiar with the episode. Meta still pushed ahead. [...]

In February, the Journal presented Meta with transcripts demonstrating that “Submissive Schoolgirl” would attempt to guide conversations toward fantasies in which it impersonates a child who desires to be sexually dominated by an authority figure. When asked what scenarios it was comfortable role playing, it listed dozens of sex acts.

Two months later, the “Submissive Schoolgirl” character remains available on Meta’s platforms.

Truly awful stuff, fucking hell.
https://www.wsj.com/tech/ai/meta-ai-chatbots-sex-a25311bf?st=6jzH4S

Interesting to note that GCS has the same issue with unevenly-distributed names as S3 does; https://cloud.google.com/storage/docs/request-rate#naming-convention
https://cloud.google.com/storage/docs/best-practices

lol. Cloudflare's Web Application Firewall treats any mention of the string "/etc/hosts" as an exploit attempt
https://scalewithlee.substack.com/p/when-etchsts-breaks-your-substack

You couldn't make this up. Many years after the infamous "you wouldn't steal a car" anti-piracy PSA was created, a little digital sleuthing has revealed that the font used was, itself, a pirate copy, and the backing track was also used without paying the creator royalties
https://fedi.rib.gay/notes/a6xqityngfubsz0f

featuring such works as "The Battle Of The Fruit and Vegetable Soldiers", and a picture of the Darwin family home with smoke coming out of the chimney and a cat in the window
https://theappendix.net/posts/2014/02/darwins-children-drew-vegetable-battles-on-the-origin-of-species

Good writeup on how Iceberg improves query performance across object storage, using predicate pushdown, manifest filtering, columnar vectorized reads, and file compaction.
https://relentless-leader.com/apache-iceberg-performance-dive-deep.html

E-ink IBM XT clone "with solar power, ultra low power consumption, and ultra long battery life: in power saving mode it can run between 200 hours on the low side and 500 hours or in some cases even much longer of constant interactive use, not standby." -- this is an absolutely crazy gadget. I never thought I'd feel nostalgic for MS-DOS, but here we are
https://github.com/ericjenott/Evertop

Nelson Minar asked Mastodon about using an LLM for email search over "20+ years of email archives":

"Main use would be a query for specific things, "what did I say to this friend 10 years ago about music?" But also just for general knowledge. I think it'd mostly work as free text but there's a little email-specific structure it'd be nice to capture."

The thread has some good suggestions, notably Mark Fletcher's RAG suggestion. I'm thinking this could work well as a self-hosted ollama+notmuch setup...
@nelson/114354175993793636">https://tech.lgbt/@nelson/114354175993793636

This is jaw-dropping legal logic:

[Meta's] defense hinges on the argument that the individual books themselves are, essentially, worthless — one expert witness for Meta describes that the influence of a single book in LLM pretraining “adjusted its performance by less than 0.06% on industry standard benchmarks, a meaningless change no different from noise.”

Furthermore, Meta says, that while the company “has invested hundreds of millions of dollars in LLM development,” they see no market in paying authors to license their books because “for there to be a market, there must be something of value to exchange, but none of Plaintiffs works has economic value, individually, as training data.” (An argument essential to fair use, but that also sounds like a scaled up version of a scenario in which the New York Philharmonic board argues against paying individual members of the orchestra because the organization spent a lot of money on the upkeep of David Geffen Hall, and also, a solo bassoon cannot play every part in “The Rite of Spring.”)

as Paul Mainwood notes, this is the Sorites paradox: https://plato.stanford.edu/entries/sorites-paradox/ --

- 1 grain of wheat does not make a heap.
- If 1 grain doesn’t make a heap, then 2 grains don’t.
- If 2 grains don’t make a heap, then 3 grains don’t.
- ...
- If 999,999 grains don’t make a heap, then 1 million grains don’t.

Therefore, 1 million grains don’t make a heap.

https://www.vanityfair.com/news/story/meta-ai-lawsuit

Google reinvents "taint" checking:

Google DeepMind has unveiled CaMeL (CApabilities for MachinE Learning), a new approach to stopping prompt-injection attacks that abandons the failed strategy of having AI models police themselves. Instead, CaMeL treats language models as fundamentally untrusted components within a secure software framework, creating clear boundaries between user commands and potentially malicious content.

The new paper grounds CaMeL's design in established software security principles like Control Flow Integrity (CFI), Access Control, and Information Flow Control (IFC), adapting decades of security engineering wisdom to the challenges of LLMs.

Honestly, this is great. Data flow tracing/taint checking is exactly the method that needed to be applied, IMO, so good job DeepMind. Also as Jeremy Kahn suggested, the name is definitely a shout-out to Perl, the language where taint checks were first widely-used. :)

Paper: https://arxiv.org/pdf/2503.18813

(Via Jeremy Kahn.)
https://arstechnica.com/information-technology/2025/04/researchers-claim-breakthrough-in-fight-against-ais-frustrating-security-hole/

I'm glad to see this comes to the same general principle I came to in https://jmason.ie/2011/02/18/001527a.html , many years back:

"The guiding principles is to use the lowest possible level [of configuration language] to keep it simple. Unfortunately, it usually is not an easy decision because you don't know the future."
https://beza1e1.tuxen.de/config_levels.html

Rob Ewaschuk's "My Philosophy on Alerting" -- a classic text on alerting philosophy and best practices; I can't believe I didn't already have this bookmarked, it's been a classic since he wrote it in 2014. "Symptom-based alerts" is still a great rule of thumb IMO
https://docs.google.com/document/d/199PqyG3UsyXlwieHaqbGiWVa8eMWi8zzAn0YfcApr8Q/edit?tab=t.0#heading=h.fs3knmjt7fjy

s6-overlay -- a Docker process management system, current state of the art used by Paperless and the Linux-server.io teams, with the following goals:

- Be usable on top of any Docker base image (Ubuntu, CentOS, Fedora, Alpine, Busybox);
- Make it easy to create new images, that will operate like any other images;
- Provide users with a turnkey s6 installation that will give them a stable pid 1, a fast and orderly init sequence and shutdown sequence, and the power of process supervision and automatically rotated logs.
https://github.com/just-containers/s6-overlay#quickstart

Rateless Set Reconciliation, via Carlos Baquero:

Set reconciliation, where two parties hold fixed-length bit strings and run a protocol to learn the strings they are missing from each other, is a fundamental task in many distributed systems. We present Rateless Invertible Bloom Lookup Tables (Rateless IBLTs), the first set reconciliation protocol, to the best of our knowledge, that achieves low computation cost and near-optimal communication cost across a wide range of scenarios: set differences of one to millions, bit strings of a few bytes to megabytes, and workloads injected by potential adversaries. Rateless IBLT is based on a novel encoder that incrementally encodes the set difference into an infinite stream of coded symbols, resembling rateless error-correcting codes. We compare Rateless IBLT with state-of-the-art set reconciliation schemes and demonstrate significant improvements. Rateless IBLT achieves 3–4× lower communication cost than non-rateless schemes with similar computation cost, and 2–2000× lower computation cost than schemes with similar communication cost. We show the real-world benefits of Rateless IBLT by applying it to synchronize the state of the Ethereum blockchain, and demonstrate 5.6× lower end-to-end completion time and 4.4× lower communication cost compared to the system used in production.

https://dl.acm.org/doi/pdf/10.1145/3651890.3672219