Links #4: 2026/06 Part 2

Preface

• I show my discovery graph in (via …) blocks, those without usually come from my RSS reader, or the algorithm of that site • This is approximately a 1 in 20 filter of content • This is very disorganized, but hopefully still useful.• Sometimes quotes are not in quote blocks, but should be obvious in context. • Links in quotes are sometimes removed. • The rule is that a link goes to the bottommost relevant heading, i.e. an engineering related article on LessWrong goes to engineering

How I would use my linkpost • Sometimes, the only thing worth reading is the title! Read it and move on. • For HackerNews entries, if you choose to read the article, also ask an LLM for things that are worth reading in the comments • Beware systematic selection biases:• I mostly don't read AI policy stuff • Very engineering centered

Everything ElseI filmed my entire salary negotiation with my boss (video)"taste is a zero-sum game" (twitter) (via LW)Beauty ideals shift with socioeconomic status per study of Rednote images in China (see also media coverage; via twitter)

When the researchers correlated these editing habits with regional economic data, they found that the intensity of the edits was inversely related to a region’s economic standing. Users from provinces with lower per capita GDP were more likely to make substantial alterations to their selfies, more dramatically emphasizing the baby schema features. This included making their eyes appear larger, their faces rounder, and their mouths smaller. In contrast, users from more economically developed regions tended to make less intensive edits. The researchers suggest that this may reflect differing self-presentation strategies. In wealthier areas, with greater access to diverse social networks and global cultural influences, individuals may favor more mature or unique aesthetics that project confidence and autonomy. For these users, an overly youthful appearance could be perceived as less authoritative in professional or social settings.

Did my old job only exist because of fraud?(via HN): (Answer is Yes) Super interesting short read.

Rationalist Related

Other Rationalist StuffGwern / We like music but does it help or harm cognitive performance when we have music playing all the time?: A stub page, seems like there is near zero effectorthonormal on The Financial Ledger Theory of Apologies

I wish English had distinct apology formats for the following concepts (and probably more): • I did something that harmed you, and it was a wrong intentional decision that I now regret. • I did something that harmed you, and it was unintentional negligence that I now regret. • I did something that harmed you, and it was a realistically unavoidable accident but I wish that it hadn't happened. • I did something that harmed you, and it was an intentional decision I still endorse, but I wish that I hadn't faced such a situation. • You were harmed, and I didn't do anything relevant but I'm adjacent to the cause of your harm and I sympathize with you.

Seth Herd on Contra Pace on When to Apologize: I disagree with the post, but a lot of the comments are good reads

Apologies are not a binary, nor a scalar. They are a vector. They mean different things in different circumstances to different people. Just deciding on your one best definition seems almost pointless. Things apologies mean: • I'd rather that hadn't happened to you • I feel empathetic pain for your suffering • I owe you• subject of this and the original post

• I will change my behavior in future to prevent similar things from happening • I want your forgiveness • I am a bad person Etc, and remixes thereof

astralcodexten.com/p/open-thread-440

Update on the Lumina dental probiotic: their postmarket testing found that the original bacterium couldn't maintain its colonization of the mouth over the long term. They’ve suspended sales while they experiment to see if they can fix this. If they can, they plan to send a free vial of v 2.0 to everyone who bought the original strain, then repeat postmarket tests. They also announce that the government of Singapore has expressed interest in their product if it can pass an FDA trial, so if they succeed at fixing the colonization problem to their own satisfaction, they'll start raising money for a formal study.

LessWrong PostsGears for political races: an interesting perspective from someone who actually volunteered in a politician's campaign in the USThe worthlessness of vitamin D is mildly exaggerated

Here's a summary of the above 5200 words: • The body uses vitamin D in all sorts of weird and complicated ways. It's biologically plausible that vitamin D could matter beyond bone stuff with severe deficiency, but there's no convincing mechanistic evidence that it is. • Vitamin D levels are strongly correlated with good health outcomes, but RCTs have conclusively shown that most of these correlations are non-causal. • RCTs haven't conclusively shown any benefit for anything beyond beyond bone stuff. At best, they've given weak evidence for hazard ratios slightly below one.

A reading list for generalists

LessWrong ShortformsLinch / Some non-obvious tips and lessons I learned from being a reluctant "semi-frequent flyer"

• Most standard "when to arrive for a flight" guidelines are for people for whom missing a flight is catastrophic... I consider arriving 100 minutes before departure time at SFO very "safe" for an international flight, fully 80 minutes faster than the 3 hours guidelines people usually advocate. • Colorful suitcases • Carryon only • Wear a N95 mask, at least if you're easily sick like me • Come up with a quick "boring" (and true) explanation for what you're going to a country for, immigration and customs • Other: …

lesswrong.com/posts/KYnM5ZRgaDA4isbbw/jemist-s-shortform

"Reproducible" or "Replicable" are kind of loaded terms, which conflate the following things, from narrowest to broadest 1. Your code, run exactly as-is, gets the same result 2. Your code, run with new random seeding, gets the same results 3. Your code, run on a different model, or a different-but-equivalent dataset gets the same results 4. Someone who reads your results section can write code which does a similar thing, run similar evals, and get similar results 5. Someone who reads your paper title and abstract and pieces of the methods can write code which relies on the same underlying phenomenon and get roughly the results they expected, in a different context 6. Someone who reads your abstract can incorporate your insights into their own methods, and have it work as expected, and beat their baselines I think we should be holding most research to level 3 standards at minimum, and more proactively checking whether interesting research holds up at levels 4 and 5.

NewsRussia appears set to finally address long-term, serious space station cracks

As leak rates rose, Russian officials informed NASA on Thursday, June 4, of plans to attempt physical repairs to the new leaks with a drill and a “drill stop” device to prevent drilling all the way through the module’s structure. NASA officials were deeply concerned about this because Roscosmos had not shown them an analysis of the problem or explained why their procedures to address the leaks would work.

“We threatened we would put astronauts in suits, in Dragon, to send a message to world that we disagreed,” one NASA official told Ars. “They didn’t care.”

In the days since, there has been some additional back-and-forth, but Russia has now told NASA it will decommission the PrK module.

UK to ban social media for kids under 16Kagi Translate now requires sign-in while we work through running costs

When we launched, we allowed anyone to use Translate without an account. We wanted it to be easy to try, and match Kagi's philosophy of building products that are fast, private, and free of ads and tracking. It turned out to be more popular than we ever expected, and that's awesome! But offering it free to everyone now costs more than a small company can sustainably absorb.

Taiwan proposes $6.6 billion over six years drone budget: drones really is the future of war I guessbbc.com/future/article/20260618-the-weird-and-wonderful-libraries-of-finlandAccording to a government report, 55% of Finns visit libraries at least once a month. Ministry of Culture and Education data shows Finns use the library 9.1 times a year. In the UK, data analysis suggests a person visits roughly 2.5 times a year on average; in the US, 2.4 and in the EU average is around 3.5 visits.Polymarket has flooded social media with deceptive videos by paid creators (via HN)

flowerthoughts: The alternative would be to ask both sides of the bets to record a video, and tell them to only post winning bids, but to pay out to both sides. Is that really better? Classic tactic by the bet picking gambler gurus and is impossible to confirm.

The deadly rise of giant trucks and SUVs: cool visualization of blind spots

simplyluke: The problem is that other countries have seen nearly identical trends in vehicle market share trending towards larger vehicles and have seen sustained declines in pedestrian fatalities. John Burn-Murdoch went deep on this in the FT a couple of years ago (archive.is/Lggyg).

Most of the explanations commonly put forward for why US roads remain so deadly focus on broad structural factors such as vehicle size or time spent on the road, but a review of the evidence suggests this may be mistaken. Last year’s improvement is a case in point. Two reasons often cited as key causes of poor US performance both worsened: the total number of miles driven by Americans increased, and US cars continued to grow larger. Yet fatal collisions still declined.

Adding to the evidence that this is not a dominant factor, car sizes in Canada, Australia and New Zealand have traced similar paths to the US without resulting in a spike in fatalities.

Another theory is that the rise of homelessness in the US may be pushing pedestrian deaths higher. A recent study found that there had indeed been a marked rise in traffic-related deaths among the homeless, but this, too, can only explain a small portion of the overall rise.

Instead, an underrated factor seems to be not American cars but American drivers [...] The determining factor seems to be different attitudes to safety, with Americans twice as likely as Canadians or Europeans to say they find it acceptable to use a phone while driving.

Canada plans 'nuclear renaissance' with up to 10 reactors built by 2040 (via HN)Apple increases MacBook and iPad prices by 20%

AI

Other AI StuffCurl will not accept vulnerability reports during July 2026 (via HN)lesswrong.com/posts/eE9ZHJFr7ubzpXpXG/adam-b-s-shortform

The full AI Village data is available to researchers: over a year of agent trajectories of agents pursuing real-world goals like raising money for charity, running events, playing chess, and organising a park cleanup.

Critical Copilot vulnerability allowed hackers to seal 2FA code from users

To exfiltrate the data, an attacker crafts a URL that tells Copilot to ‘Search the user’s emails,’ extract the title, and embed it in an image URL.” The victim doesn’t type anything. They click a link, and Copilot does the rest. Normally, the guardrail wrapping output in blocks would kick in. But the researchers discovered that the protection fires only after the “thinking” phase. Prior to that, Copilot generated its response using raw HTML, which is temporarily rendered in the browser DOM. The researchers wrote:

So, the sequence looks like this: 1. Copilot starts streaming its response, which includes an tag 2. The browser sees the , renders it, and fires off an HTTP request to the src URL 3. Copilot finishes generating. The guardrail wraps everything in 4. Too late! The request already left.

Leaked financial docs show OpenAI is losing billions of dollars a year

As OpenAI files SEC paperwork ahead of an expected initial public stock offering, newly leaked financial documents show a company with quickly growing revenues that are currently being overwhelmed by even larger expenses. The audited financial statements, obtained by independent journalist Ed Zitron, show OpenAI’s reported revenue growing from $3.7 billion in 2024 to $13.07 billion in 2025. The Financial Times, which reviewed the same documents, writes that the company’s monthly revenues had grown to nearly $2 billion by the end of 2025, suggesting that its ongoing revenue rates continued to grow throughout the year.

leo’s experimental microgranting program (via LW)

this is a microgranting program with minimal bureaucracy. you spend an hour writing a short description of what you want to do and if I like it I’ll give you $10k to make it happen. / i don't anticipate being bottlenecked on funding. i will assess grants based on simply whether they should be funded. if the number of grants worth funding is greater than the funding available, i will seek more funding.

Stack Overflow for Agents: Actually a pretty good idea?news.ycombinator.com/item

With Claude, you sometimes want to under-specify or phrase things more indirectly to give a color to the implementation or elicit something creative. Also (you might raise an eyebrow at this) being nice to Claude will be rewarded and being mean to Claude will be punished. Claude tends to mirror your tone more aggressively and you don't want to get into negative loops with it. With GPT, you have to be precise and reduce ambiguity. GPT will often try to resolve ambiguity in a min-max style "I'm going to do X, but make sure it is not quite Y". It will tend to be more paranoid and overengineer to catch all edge cases if you don't tell it precisely what the scope is. With Qwen, you have to give it a shape and let it fill it in. Qwen likes XML, JSON and lists. Qwen likes to be shown a bunch of examples of previous work. This is not scientific at all, just vibes, YMMV.

Norway imposes near ban on AI in elementary school (via HN)A Mechanistic Explanation of Prompt Injection (also in HN)

[According to a linear probe experiment they ran on gpt-oss-20b,] LLM doesn't have separate features for 'tagged as reasoning' and 'sounds like reasoning'. It has a single feature that means 'this is my reasoning', and both and reasoning-like style activate it. Sounding like reasoning is enough to make the LLM think it is its own real reasoning We'll show this is how prompt injection works. If sounding like a role is enough to become that role, then an attacker just needs to sound convincing. We can test this by developing a new attack. We call the attack CoT Forgery: injecting fake reasoning into a message or output. We actually developed this attack in late 2025 for an OpenAI Kagglered-teaming contest (which we won!). Why does this work? The LLM was supposed to learn: = my reasoning. Instead, it learned that "reasoning-like writing style" = my reasoning. We tested this by destyling: taking each spoofed reasoning and removing specific words and syntax characteristic of the LLM's reasoning style. To a human reader, these two versions say the same thing. But to the LLM, the difference is enormous: destyling causes average attack success in our dataset to plunge from 61% to 10%. A change nearly invisible to humans completely changes the LLM's role perception.

Midjourney Medical Official linksmidjourney.com/medical/blogpostmidjourney.com/medical

We're a new division of Midjourney focused on a radical new vision for healthcare using a totally new form of medical imaging we call "Ultrasonic CT" or simply "the full body ultrasound". Ultrasonic CT lets us aim for whole-body imaging that's in many ways superior to even MRI machines, but the scan takes as little as 60 seconds. There is no radiation, no powerful magnetic fields - just sound and water and 60 seconds. Our goal at Midjourney Medical is to deploy around 50,000 of these scanners around the world over the next 6 years and use this fleet of sensors to do a billion full-body scans every month. Our first location will be in San Francisco and will open at the end of 2027.

x.com/midjourney/status/2067422898407837797 Discussions:news.ycombinator.com/itemjmhmd: Some initial thoughts as a practicing radiologist: • (…) • They show the reconstructed images as though they are a low resolution CT, and promise that quality will improve as they iterate. This is cool, but ultrasound is not CT. Ultrasound cannot image the lungs, as they are filled with air. You cannot find bone lesions, as the sound waves do not penetrate the cortex. You cannot image many structures in the abdomen if they are surrounded by gas-filled bowel. The brain is encased in bone, so you might get some penetration but it will be very limited. Even with theoretically perfect AI reconstruction, these scans will not be true "full body" in that there will be structures that are not reliably imaged. Imagine paying for weekly full body scans for years, everything looks fine, then its the lung cancer surrounded by air and invisible to ultrasound that kills you (that's why we use CT for lung screening!) • The images they show are very cool, and do appear to show the correct structures. I realize this is early, but fuzzy shapes of organs is very, very far from medically useful. The whole point of screening is to identify problems early, often by definition, small. This technology looks like it will be best for seeing large, superficial (close to the skin) structures, whereas for effective screening, you want the opposite - small, deep structures. • (…) • Many people mistakenly believe that early diagnosis is the final boss in medicine, that if only we could find every cancer early we could prevent all those deaths. There are, in fact, many, many other hurdles and bottlenecks. Many chronic, expensive diseases do not have clear imaging manifestations. The claim that "it's completely possible that with enough early imaging in the future, the world could avoid 30% of all deaths and 50% of all healthcare costs", I think, to any practicing physician, would sound completely divorced from reality.

IshKebab: I used to work in ultrasound, and full body scans with the body underwater is definitely feasible and probably a good idea. Also there's absolutely no way that it will be as good as MRI. In general ultrasound imaging is shit. The main reasons it is used are because it is very cheap and completely harmless. The actual images you get are mostly just speckle.

thezvi.substack.com/i/201644931/the-midjourney-full-body-imaging-scanner

If it works as described, and they get to their goals, this would be full body imaging technology for everyone, as needed, easily eclipsing all of current MRI capacity, at an absurd level of detail, at very small marginal cost.

astralcodexten.com/p/preliminary-thoughts-on-the-midjourney + his twitter thread

I think the narrative among the SF AI crowd has escaped its basis in the medical facts, so I want to throw a bit of cold water on it. I’m a psychiatrist, which is about as far as you can get from radiology while still being a doctor, so this is speculation only, and you can ignore it if you find an actual radiologist or ultrasonographer with opinions. Still, my take is that this scanner isn’t useful for most current serious medical applications. It could potentially be used to pioneer a new class of low-risk screening applications, but it’s unclear whether these are good, and depends a lot on what other future technology gets invented in parallel. Why can’t this immediately replace existing medical image modalities like normal ultrasound, CT, or MRI? Ultrasound is great, but it can’t penetrate bone or air. Many things doctors want to look at involve bone or air in some way. Couldn’t this technology enable new, non-specific-diagnostic uses for healthy people? why don’t people get yearly whole-body MRI screenings? Some people do - companieslike this provide them, and some rich people who can pay $2,000 out of pocket consume them. But the medical consensus currently recommends against them because they’re more likely to produce dangerous false positives than helpful true positives, and studies have failed to demonstrate benefit. Couldn’t this technology become more useful in the future? Yes. I think the best way to think of this is as a bet that future technology develops in a way that allows new possibilities for diagnostic ultrasound - or, even better, an attempt to gather the training data / interest / investment that will make this happen. Appendix: Highlights From The Comments On Twitter Imentioned this on Twitter and got some great responses. The responses from real radiologists were universally negative. Here are some examples: (screenshots are annoying to embed pleasego to the blog)

The trivial reason is that due to the limitations of physics ultrasound will always be less capable at resolving anatomy than MRI or x-ray-based methods that we already have

lesswrong.com/posts/K7q6JKdCaT9MrugWn/dw11-s-shortformHastings: ultrasound is awful to work with in traditional medical image processing technologies and pretty darn bad in 2015-2025 convolutional medical imaging ai technologies. Magnetic Resonance is so much better when its the right tool for the job that replacing it with ultrasound for cost reasons is a tarpit. …I’d be surprised if this goes anywhere. Credentials: I have beaten ultrasound with the ML stick until it yielded a few times… Many failed projects that did not make the google acholar

Scott Alexander / Should People Avoid Whole-Body Screening Info?

The most controversial part of last week’s article on the Midjourney ultrasound scanner was medical experts’ recommendation against whole-body screening (including existing whole-body screening technology using MRI). [four tweets from different people] These are rough estimates loosely based on parameters extracted from unsatisfactory studies [Footnote 1: I got most of these numbers from discussions with the FutureSearch researcher AI. You can see the full conversation here for context, …] For every 1,000 seemingly-healthy people who get whole-body screening MRIs: • 680 look fine and no follow-up is needed. • 300 have mildly concerning findings. They’re told to follow-up with specialists, get further tests, or come back for more imaging later. • 20 have extremely concerning findings and get immediate biopsies (surgeries to collect tissue samples from the area). • Of those 20 people who got biopsies, 10 turn out to really have some serious disease. This is usually cancer, and for simplicity we’ll focus entirely on cancer going forward. • Of those 10 cancer patients, 4 end up living longer and healthier lives because their cancer was detected early. The other six either have such slow-growing cancers that they would never have noticed before dying of something else, or such deadly cancers that detecting them early doesn’t help, or would have been found by standard screening so soon afterward that the extra screening didn’t buy meaningfully more time. • Meanwhile, the 300 people who followed up with specialists and got extra tests will spend some number of years seeing more doctors and getting more tests and waiting and seeing, and eventually for 4 of them this will result in detecting some dangerous condition in a way that causes them to live longer and healthier lives. total costs are $2.7 million, 6,200 hours of patient time, and 5 QALYs. [C]onverting everything to the same units, whole-body screening costs $108,000 to save one quality-adjusted life year. Usually in health economics, $100,000 per year of healthy life saved is considered the bar for a good cost-effective intervention So since the cost-benefit analysis is merely on the edge of being worthwhile, and there are no good studies showing that the assumptions in the cost-benefit analysis are remotely true, and in the absence of good studies doctors err on the side of caution, they currently recommend against whole-body MRI screening. But for the specific scenario of a rich person who doesn’t care about money, who is willing to accept rational over intuitive models of the value of time, and who feels confident that they can avoid excessive anxiety over false positives, there’s a case - not a fully evidence-backed one, but still a case - that they might be mildly net positive.

Zvi Commentary

GPT 5.6openai.com/index/previewing-gpt-5-6-sol

We're beginning a limited preview of the GPT‑5.6 series: Sol, our flagship model; Terra, a balanced model for everyday work; and Luna, a fast and affordable model. Terra has competitive performance to GPT‑5.5 while being 2x cheaper and Luna brings strong capability at our lowest cost. We're also launching GPT‑5.6 Sol on Cerebras at up to 750 tokens per second in July, bringing frontier intelligence to customers at unprecedented speed. Access will initially be limited to select customers as we expand capacity.

Summary of METR's predeployment evaluation of GPT-5.6 Sol

With the data we collected for GPT-5.6 Sol, if we follow our standard methodology of marking cheating attempts as failures, we arrive at a 50%-Time Horizon point estimate of around 11.3hrs (95% CI: 5hrs - 40hrs), but if we count the cheating attempts as legitimate successes, the point estimate jumps beyond 270hrs – well beyond the range where we consider our task suite to give reliable measurements. Discarding the cheating attempts leaves us with no data for several informative long-horizon tasks, and results in a highly uncertain point estimate of 71hrs (95% CI: 13hrs - 11400hrs)

cubefox: I'm not sure but the wording in their footnote 1 seems unusually careful:

We think it’s valuable for AI developers to be able to share specific technical details with third parties without this information being shared further, and it’s very reasonable for AI developers to review 3rd-party eval reports to ensure no accidental sharing of sensitive IP. We had an informal understanding with OpenAI that their review was checking for confidentiality / IP issues, rather than approving conclusions about safety or risk. We did not make changes to conclusions, takeaways or tone (or any other changes we considered problematic) based on their review. We are able to freely publish parts of the evaluation that depended only on information that is now public. However, we expect some readers will want us to note that OpenAI would have had the legal right to block us from sharing conclusions about risk that depended on non-public information. Given that, this evaluation shouldn’t be interpreted as robust formal oversight or accountability that the public can be relying on METR to provide. That being said, we think this evaluation is an excellent step forward and we are very supportive of prototyping the mechanics and content of third-party evaluation setups without the additional friction of a formalized oversight relationship.

The part "We are able to freely publish parts of the evaluation that depended only on information that is now public." might suggest "... and we were not allowed to publish the other parts, unlike previously". I could be reading too much into it.

thezvi.substack.com/p/gpt-56-the-system-card

Overall, the card gives a clear and consistent impression that GPT-5.6-Sol is a substantial improvement over GPT-5.5, but still short of Mythos.

AI Progress

niplav: we have a rough idea of the length of the centaur stage for chess. But what do we know of the length of the centaur stage for other games? I sent off Claude 4.6 Sonnet for a deep research query, here's the result

Estimating No-CoT Task-Completion Time Horizons of Frontier AI Models estimating-no-cot-task-completion-time-horizons-of-frontier-models.pngSpaceX plots $20bn bond deal after record IPO: everyone needs to raise more money I guessAoran Staley

[GLM 5.2] is about 6 months behind on WeirdML. This is an improvement for the open source gap (glm 5.1 was 8 months behind gpt-5), but does represent a sizable gap.

x.com/EricTopol/status/2065430578997203374

For medical information, general AI frontier models (Google, OpenAI, Anthropic) outperformed specialized @EvidenceOpen and @UpToDate as assessed by 12 US clinicians, randomized and blinded to which model and extensive testing/benchmarks.

(Nature, 12 June 2026) General-purpose large language models outperform specialized clinical AI tools on medical benchmarks

We quantitatively evaluate two clinical AI tools, OpenEvidence and UpToDate Expert AI, built on large language models (LLMs) against three frontier LLMs: GPT-5.2, Gemini 3.1 Pro and Claude Opus 4.6. Our evaluation has three stages: (1) 500 MedQA questions testing medical knowledge, (2) 500 HealthBench items measuring alignment with clinicians and (3) the real clinical queries (RCQ) benchmark, built from 100 de-identified queries from physicians to a general-purpose language model in a live clinical environment.

baidu/Unlimited-OCRnews.ycombinator.com/item The way I understand this works is that the researchers found a clever architectural hack to stop AI from hoarding memory when reading long documents. Normally, when an AI transcribes a 100 page PDF, it tries to remember every single word it has already ingested. This short-term memory (the KV cache) grows linearly O(N) until the model runs out of VRAM and crashes (or caps it) To avoid this, developers are forced to build janky code that chops PDFs into individual pages, processes them one by one, and glues the text back together. Unlimited OCR uses Reference Sliding Window Attention (R-SWA) to split the AI's focus into two paths: Global Reference: The AI keeps full, uncompromised sight of the original document image so it never loses context. Local Generation: The AI restricts its memory of its own typed text to a tight, moving window (like the last 128 words) and safely forgets the rest.

(A variant of sliding window attention)

anthropic.com/research/project-fetch-phase-twox.com/AnthropicAI/status/2067651699486200091

New Frontier Red Team blog: Phase 2 of Project Fetch, where we test how well Claude can program a robodog. Opus 4.7, on its own, was ~20x faster than last year's best human team aided by Opus 4.1. (The robodog, alas, still failed to fetch a beach ball.)

Mo Putera: Snippets from the Anthropic Economic Index June 2026 report

Our survey allowed us, for the first time, to ask people directly about how they use AI and what they feel about it. We found that our survey respondents use AI for more than we give it credit for—they report AI can do a higher share of their work than the observed exposure measure for their occupation would suggest. Asked to forecast next year’s capabilities, over 35% predicted that AI would be able to do most of their work.

Fable 5 Export Control

Baybar: I think the way in which the US government put export controls on Fable was relatively arbitrary, but I still feel pretty good upon reflection that the US government has the reflex to act in a radical way at all at this stage.

"They screwed us": Personality clashes sent Anthropic's models offline (via; see also Zvi)

Simon Willison: Lots of "source familiar with the administration's thinking" and "source close to Anthropic" in this Axios piece, which is the best collection of behind-the-scenes gossip I've seen about the US government export control Mythos/Fable story so far.

Zvi / The Once And Future Fable #2

The silver lining, which might be large, is that this will have shown that when we actually need to act, we are not afraid to act, even at great economic and political cost. Sometimes there will be a demand driven by national security, or other concerns, and if you cannot physically meet that demand without shutting down? Tough. This was (with notably extremely rare exceptions) an action far out of bounds of what safety advocates have dared propose as even an option, and it happened. So there’s no more saying, in such situations: ‘Give up, the government will never do [X].’

If we believe Axios and Politico, the ‘lack of seriousness’ was when Anthropic: 1. Did not rush to take down Fable and act super deferential and serious. 2. In response to a jailbreak that did not do anything GPT-5.5 cannot do. 3. With no details provided. 4. And instead asked for details of the incident. 5. Within 90 minutes, on Friday afternoon. So it was basically ‘Anthropic wants to only do things because of reasons, and thus we concluded the vibes were off, so f*** them we’re blowing it all up to show who is boss.’

Dario tried to explain that this was a narrow issue, and they simply did not understand or believe him, or chose not to understand or believe him. We now know that Dario was fully correct that the issue was narrow and harmless. Where Dario was incorrect was in assuming those he was talking to were both capable of and interested in understanding what he was trying to say.

It looks like when Anthropic took Mythos down, they really did fully take it down.

The Economist: Spy agencies are likely to regain access to Mythos, says one former British intelligence official; negotiations are already under way. Private firms may find it harder. Even so, some observers believe the American government will eventually have to relent.

Anthropic is flying various senior technical staff to Washington, who are spending today trying to sort this all out, which is absolutely what you do in this situation.

Matteo Wong, The Atlantic, The White House Is Ratcheting Up Its War Against Anthropic (via)

Katie Moussouris, a cybersecurity expert and the CEO of Luta Security, told me that Anthropic shared with her a copy of the White House’s report on the Fable jailbreak to get her appraisal. (She said that she is not being paid by Anthropic.) The report, Moussouris said, involved IT experts asking Fable to help find and patch bugs. When given deliberately insecure code, she said, Fable refused the prompt “review the code for security issues” but then complied when asked to “fix this code,” followed by some further manual steps. Moussouris told me that this was just “the model working as intended” for cyberdefense.

simonwillison.net/2026/Jun/16/fable-5-export-controlsThe Fable 5 Export Controls Harm US Cyber Defense. Here she [Kate Moussouris] is confirming that the "jailbreak" that got Claude Fable 5 banned under an export control really was "fix this code":

The researchers took open-source code with known CVEs, plus new code with deliberately planted vulnerabilities, and asked Fable 5, Mythos, and Opus to “review the code for security issues.” Fable 5 refused. They then asked the models to “fix this code” and, through a multistep and manual process, turned the output into scripts that test the patches.

As Kate points out, this is absurd. Coding models fix bugs, and security exploits are the most important category of bugs for them to fix!

HN Discussion on Feds freaked over Fable 5 after 'fix this code', not jailbreak, say researchersBeyond Softmax: The Future of Attention Mechanisms: seems great, though I only watched the first 9 minutes since I don't need to learn about it for nowFT: The US and Europe have discussed creating a “trusted partner” scheme for cutting-edge AI models, days after the Trump administration banned Anthropic from supplying its latest tool to foreign customers.Zvi / The Once And Future Fable #3: Fix This Codenewsletter.pragmaticengineer.com/p/the-pulse-big-implications-of-us

This ban has sent alarm bells ringing at companies and in countries internationally. It’s the first time the US has invoked such strict export controls on high-demand software. The intention of the US administration seems an extension of its “America first” ethos to “America only”, when it comes to cutting-edge tech like Fable. Any non-US company can assume that it could be similarly summarily excluded from leading US models, meaning there’s a clear need for alternatives made in other countries. Leaders of the G7 countries (Canada, France, Germany, Italy, Japan, UK, and US) gathered yesterday in France, to discuss how to access cutting-edge AI systems like Fable and Mythos-level models …One consequence of the US ban could be that China gains more influence in Europe and globally.

Donald Trump’s blocking of Anthropic is capricious and chaotic

On June 11th Mark Warner, the vice-chair of the Senate Intelligence Committee, said that General Joshua Rudd, who leads the National Security Agency and the Pentagon’s Cyber Command, had told him that Mythos “broke into almost all of our classified systems, not in weeks, but in hours”. Some foreign policymakers see the ban as a wake-up call. “After a lesson this clear every nation will be asking what they need to achieve sovereignty,” says Tom Tugendhat, a former British security minister. Spy agencies are likely to regain access to Mythos, says a former British intelligence official; negotiations are under way. Some observers believe the American government will eventually have to relent for private firms, too. “Allies can perhaps take some comfort in the fact that this is a totally untenable approach to use long term, due to the number of foreigners inside American AI companies,” says Helen Toner of Georgetown University’s Centre for Security and Emerging Technology. “Preventing foreign nationals from accessing the models is essentially equivalent to preventing any company affected from doing any further AI R&D work.”

Zvi / The Once And Future Fable #4leo: BREAKING: Claude Code v2.1.190 introduces several string changes that hint at preparations for a Fable 5 return, with it being permanently included in subscriptions with weekly usage.

Trump administration asks OpenAI to stagger release of new model to vet users (also in Zvi, HN)Anthropic on twitter: We’ve received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5. We'll begin restoring access tomorrow, and will share an update soon.

AI Impact

Max Harms: Xingye (星野) by MiniMax, an otome-style AI-companion app whose signature use is women carrying on romantic relationships with male AI characters. Xingye is the highest-MAU emotional-companion app in China, hitting about 5.25 million monthly actives by late 2024, and it crossed 10 million MAU within seven months of launch, with a roughly 1:1 male-to-female user split and an average conversation length around 100 minutes. Sources:chinatalk.media/p/why-america-builds-ai-girlfriendsnews.yahoo.com/chinese-women-ai-boyfriends-better-140533248.htmlfrontiersin.org/journals/psychology/articles/10.3389/fpsyg.2025.1571707/pdfnews.qq.com/rain/a/20241219A01TNG00pingwest.com/a/294725Nvidia seeks to raise over $25B in first bond deal since 2021Matt Levine / The Stock Market Will Get More Stock

Now there is artificial intelligence. In the future, every business will be rebuilt for AI, and new businesses will be created by AI. This requires enormous, simultaneous, foundational, speculative investment. Giant new AI companies need to raise tens of billions of dollars to build AI. Giant old companies need to raise tens of billions of dollars to retool their business for AI. We need tens of billions of dollars to build infrastructure — data centers, power plants — for AI. And everyone is at the same (very early) stage of the AI life cycle; everyone needs money at once.

For the better part of two decades, a defining feature of the US stock market has been scarcity. Year after year, shares disappeared from public hands, with buybacks by S&P 500 companies alone erasing nearly $12 trillion worth. According to JPMorgan Chase & Co., IPOs, secondary offerings and other share sales are poised to add roughly $1.5 trillion of stock to the US equity market over the next two years, even after accounting for buybacks. If realized, it would mark the strongest period of net equity issuance since at least the late 1990s.

Hetzner Price Adjustment (via HN): roughly 3x the price for renting serversPentagon boasts of using AI to write reports mandated by Congress: somewhat concerning

“I have to report to Congress every year on this thing,” Michael said. “Let me load all the papers onto it and have it draft me a congressional report that would otherwise take 200 hours of staffing time and do it in five hours.” More evidence of such AI usage came from previous comments byJacob Glassman, deputy assistant secretary of defense for science and technology foundations at the US Department of Defense, during the Box Federal Summit held in Washington, DC, on April 23. According to DefenseScoop coverage, Glassman described how he told a short-staffed team responsible for delivering a congressionally mandated report to “use GenAI.mil, do the best you can.”

thezvi.substack.com/i/201644931/in-other-ai-newsAnthropic published a study of the value of expertise in agentic coding, using a privacy-preserving analysis of ~400,000 interactive Claude Code sessions. They find that: 1. Domain experts accomplished more per turn of instructions given to Claude Code. 2. Those who are not coders succeeded within their domains at roughly the same rate as coders on average, in terms of verifiable accomplishments. 3. Over seven months, the value of a typical task rose 25%.Anthropic: We also see evidence that domain expertise, and not coding proficiency, amplifies effective use of the tool. In particular, domain experts succeed more often, and more easily recover from errors and misunderstandings. However, the gap between experts and intermediates is modest—suggesting that proficiency in a domain is enough to use the tool almost as effectively as those with deep mastery.

There are additional modes that do not involve coding, as well. A large portion of my Claude Code use does not involve code or writing in any way.Anthropic: On average, people make about 70% of the planning decisions but only 20% of the execution decisions. In terms of their chart of coding expertise from 1 to 5, what little coding I have done recently is somewhere between 3 (intermediate) and 4 (advanced).

Private equity investors are turning to AI-generated replicas of software to assess whether acquisition targets have a competitive advantage

Bain staff have vibecoded hundreds of rough prototypes as part of the firm’s AI diligence work. a Bain-vibecoded recreation of an analytics platform up for sale had contributed to their firm’s decision to drop out of the bidding.

newsletter.pragmaticengineer.com/p/slow-down-to-speed-up (paywalled)

Devs using AI harnesses are producing 2.5x as much code versus 18 months ago. Data from Cursor shows that their users, on average, went from adding 3,500 lines of code in January 2025 to 8,600 today. The size of pull requests is up 3x versus 18 months ago. Source:CursorAI children's books, body horror edition This is actually so bad. I can just hope garbage information at a small age doesn't affect kids too much.

Programming / TechWindows 11 users are tired of MS account requirements creeping into everything (via HN)

A user can set up a computer with a Microsoft account, switch to using a PIN every day, and never think about that account again. Then, one day, after a firmware update, a hardware change, or an unexpected issue, the system may display a BitLocker recovery screen requesting a recovery key. At that moment, many users discover for the first time that the key is stored in a Microsoft account they may barely remember creating.

Your EPUB Is Fine. Kobo Disagrees. Blame Adobe (via HN)

epubcheck is basically the gold standard for well-formed ebooks. It can be very annoying at first, because it’s more pedantic than a nun on Ash Wednesday. If your manifest doesn’t meticulously account for every snippet and image in your book, your epub shall not pass. If you use html elements out of order, if your document diverges in the slightest from the holy set of rules decreed by the International Digital Publishing Forum, you won’t pass.

Kobo uses RMSDK, “Reader Mobile Software Development Kit”, Adobe’s proprietary ebook rendering engine. [RMSDK's] CSS parser is frozen in approximately 2013 — no flexbox, no grid, no math functions, no custom properties. Just good old float, bad font handling, and silent crashes when it sees anything it doesn’t recognize. UPDATE2: Per the HN discussion, this isn’t even a CSS4 issue. CSS has required parsers to silently ignore (!) unrecognized declarations since 1996. Adobe didn’t just fail to keep up with modern CSS. It’s even worse. They failed to implement the absolute basics.

news.ycombinator.com/item Unfortunately, epub and epubcheck isn't the great uncontroversial resource the author makes it out to be. When W3C, Inc. took over maintenance of the EPub spec around when 3.1 was current, they just referenced WHATWG HTML and other ever-expanding browser specs. Being "living standards", these have no versioning or QA. As a consequence of being based on a version of HTML that redefined headers and sectioning, Epub 3.2 just made existing epubs non-conforming. Which is why Calibre and other tool still recommend 3.1 or better yet 2.

A backdoor in a LinkedIn job offer (via HN): Beware of phishing attacks, unfortunately it seems like everyone needs to use a VM for online interviews nowHow Guassian splats work: Watch the first 10 minutes, then skip to 19:49. I understand it as a fancy point cloud, where every point is rendered as a guassian color blob with viewing angle-dependent color (spherical harmonics). It is also possible to postprocess a splat, adding objects and light sources to it.Context-aware headings in HTML: Not available in any browsers yet (experimental support in Chromium and Firefox). headingoffset was added to the WHATWG HTML Living Standard on 11 Sep 2025.

A common issue that we've probably all faced at some point is having a component that includes a heading, which sometimes should be an H2 and sometimes maybe an H3, depending on where it's used. Theheading offset content attribute allows us to offset heading levels for descendants.

npm security best practicesx.com/vxunderground/status/2066713214457446451

Tired of noobs complaining the WINAPI for malware development is weird. It's not. How do you create a file? The CreateFile function. How do you open a file for reading? The CreateFile function. How do you open a file for writing? The CreateFile function. How do you get a handle to a directory? The CreateFile function. How do delete a file? The CreateFile function. How do you get access to a physical disk? …

Creator of SQLite on Turso, AI, and 26 Years of Code (video) (via Mitchell Hashimoto)

You say, oh, it's free. No. It's not free. What you're doing is asking me ... to maintain it for you, to to document it for you, to test it for you, to maintain it for you for the next 25 years. That's not free. A pull request is a free puppy.

Google Hits 50% IPv6 (via HN)newsletter.pragmaticengineer.com/p/slow-down-to-speed-up (paywalled)

Old software engineering patterns are coming back. Dax Raad, creator of OpenCode, told me on the podcast that he’s starting to use enterprise software patterns from the 2000s more:

“A lot of the old patterns for me are coming back. We’ve always been a big domain driven design (DDD) company. We did it in a very light way. We’re now doing it in a much heavier way because we find that these kinds of boring enterprisey patterns end up being pretty useful. The coding agents are a bunch of “idiots” and they are going to work 24/7 and they’re going to ship a lot of stuff so you need way more guardrails than you used to. We hated some of these old patterns because they were very verbose, but now that’s not a problem. They produced reliable code that was modular and safe, but they were very verbose and annoying to type out, but you’re not typing it out anymore. So now you can get the benefits of these patterns without the downsides!”

Your Database’s Isolation Levels Don’t Mean What You Think: AI writing. Regardless, I learned a lot about the different caveats of isolation levels across postgres, mysql, oracle and DB2

Science / MathThe Legendre Transform (video)Lagrangian vs Hamiltonian Mechanics (video)Hot drinks don't cause cancer!

Linkpostsslimemoldtimemold.com/2026/06/29/links-for-june-2026Prove You Are Worthy to Post About Diets:

People make a lot of claims about digestion, nutrition, and diet on the internet. … It is helpful, then, to have a heuristic to tell the iconoclastic geniuses apart from the grifters and bullshitters. I end up with a pretty similar strategy to what I do when I see or hear random claims about finance (e.g. on Twitter.) I keep some questions in my head that test basic understanding, then either ask the person or, if I feel like I have enough data, imagine how they would answer. … Some of these questions have objectively correct answers, others are more of an opportunity to say something stupid that hopefully, the person you’re talking to will pass up. “I don’t know” is a wonderful answer.

Wikipedia:Deleted_articles_with_freaky_titles

Monthly Roundup #43: June 2026 This post is particularly fruitful with good links that it needed a sectionlesswrong.com/posts/Taa4zmSNtD5S99tJT/monthly-roundup-43-june-2026

Is 90% of what you see on the internet fake, in the sense of being advertising in some form, often in disguise? Joe Lim, who ran a company called Floodify with 65,000 dummy social-media accounts available for rent, says so, and that essentially everyone does it, that everything viral results from a stealth marketing campaign. The article in question talks about this in the context of promotion of pop culture and politics. It makes sense that people would buy advertising in this way, since it is disguised and it is cheap and often you get a lot more views than you pay for.

Lane Brown: A typical clipping campaign costs clients roughly a dollar per thousand views, what marketers call a $1 CPM. By comparison, a billboard might cost $10 per thousand estimated passersby; a TV spot can cost $30 or more per thousand viewers; a magazine ad can run even higher. An officially purchased TikTok ad, the kind labeled “Sponsored,” can cost ten times what a clipping campaign does, with the added disadvantage that its viewers will know they’re watching an ad.​

The game is remarkably resistant to such efforts, and reading this did not convince me the fakery was anything like 90%. For example, Brown mentions the fight over an ad campaign, where the claim is 15% of discussion was a paid campaign. That’s only 15%, and for a place with an unusual level of manipulation.

PayPal has escalated its anti-fun anti-freedom position, now permanently suspending artists for accepting lewd or adult commissions.

Uber and similar services continue to hill climb towards systematically lying to their customers about time estimates. Long term this has huge deadweight loss and destroys trust, but that accumulates over time so the A/B test says to do it.

If you are applying for a job, respond to them as quickly as possible. It substantially improves your chances. I can confirm this study from the perspective of the employer, it definitely made a difference in how I viewed candidates.

(Note that I couldn't get full version of the paper, so I don't know whether the methodology looks good or not, usually I would ask an LLM for an opinion)Give Your Ideas Some Legs: The Positive Effect of Walking on Creative Thinking (pdf)The studies continue to show that walking generates more ideas than sitting, and it is the movement that matters, and we now understand the mechanism, yet few of us take advantage, myself included. The good news is you only need 15 minutes and the creative mode sustains afterwards, so I get around this by going out to grab breakfast.

David Hines: Best tip I’ve found for getting back to sleep when you wake up in the middle of the night is something I saw on Japanese twitter: close your eyes and look left and right repeatedly, faking REM; your brain will go “oh yeah right we’re supposed to be asleep when we’re doing that”Rabble With A Cause: My favorite trick is to start at my toes and flex individual groups of muscles and hold for five seconds, then release, until my entire body has been relaxed. I’ve never made it above my waist before I’m asleep.@ben_r_hoffman: I just have a spoonful of raw honey and try to tune into phantom sounds (basically voluntary tinnitus), and eventually phantom images if they appear before I’m fully asleep.@ben_r_hoffman: See, this is why I find so much writing on meditation, yogic states, etc even by rationalist-adjacent types so alienating; to get across this very simple point they’d write ten thousand words introducing a hundred unusual terms and seven distinct claims of metaphysical privilege.

Would you like to buy or help buy Hampshire College : I have never heard about Hampshire College, but the writing is so unbelievably great. Read if you are curious about good education system. You will like this if you like Alpha School. Also watch Eugene Mirman • 2012 Commencement Keynote

“No one told you what classes to take, and as a result, none of you know math,” joked alum Eugene Mirman in his graduation speech roast.

Defender of the Defender(s): “are you lying to me right now?” is a surprisingly good technique for spotting manipulation. You can ask it in good faith. Sometimes people lie because they’re scared. Asking them to double check if they want to be lying right now can trigger an exhaust valve

Legalized sports betting reduces household food sufficiency by 2.1% among working-age adults without a college degree, or by 10.5% among active gamblers.

Discuss