This is our weekly newsletter of AI developments. Browse the archive of past issues, ask the archive anything in plain English, and sign up if you like.

TL;DR#


Economics#

UK gov / Linkedin study of entry-level jobs: 80% of 38 major entry-level roles are shrinking. Worst: Accountant (–29%), Graphic Designer (–28%), Software Engineer (–27%), Data Analyst (–15%), and Legal Assistant (–14%), i.e. processing, organising, or producing structured information. At all seniority levels combined, these aren’t the weakest performers: it’s falling on juniors. In Engineering overall, entry-level is at -19% while senior is at -10%; this gap only opened in Jan 2026. “Entry-level hiring is not collapsing relative to the market. But within it, the composition is not uniform: information-processing roles are weakening while relational roles grow”

Opinion: Weak evidence for a labour market impact from AI vs other causes, but it is some evidence and we should expect to see replications across markets soon, to the extent that the effect is real.

Article arguing that a compute tax is bad policy, akin to taxing steel during the industrial revolution. Also claims that the compute tax base is too small to raise meaningful revenue and would instead push data centers overseas.

Opinion: We agree that there should be a presumption against taxes on intermediate goods. But he does not engage with the giant negative externalities we’re actually targeting, so this is an analysis with no sane denominator. Korinek persuasively argues that a per-GPU-sale tax isn’t best though.

SpaceX announces that they have entered into a $920M/mo agreement with Google to provide them compute.

Opinion: $11.60/hr for Blackwells is pretty much market rate for this kind of large cluster. If we expected this to pass through to SpaceX valuation in their imminent IPO – at even 1/3rd their predicted revenue multiple – it would more than pay for itself through Google’s ownership stake, so hard to rule out some skulduggery here with regard to the deal’s timing. But the deal likely makes sense on its own terms too.

Leopold’s fund hits $20B AUM after 270% surge in 2026

Opinion: Levered betting on AI has been a good business this year. Note this definitively tells us that the large put positions in the March 13F were not a significant part of their betting, since those puts did terribly.

Cognition introduces an AI productivity guarantee for customers: they try to measure the productive engineering output Devin is providing to enterprise customers, and if Devin is found to deliver less engineering value than what the customer is paying for, Cognition will fund the customer’s usage up to $10M until it does.

Opinion: Interesting incentive alignment experiment. Could help Cognition/Devin find more enterprise customers, though it’s unclear whether this will be enough unless they are also able to get their SWE line of coding models competitive with the pareto frontier.

Capabilities#

Annals of the data wall: Dwarkesh argues, with us, that recent AI progress has not come from improved sample efficiency but from widening and improving the data distribution at enormous compute cost (RL as ridiculously expensive synthetic data generation). “It is easy to forget how much data these models are trained on, and how much more it is than what we humans see in our lifetimes. We see these AIs as a galaxy glittering with capabilities, but at their center, invisible to the naked eye, holding all the constellations together, is an unimaginably massive black hole of data.”

The labs’ plan for the hard cases is to automate AI research first, then have the automated researchers solve sample efficiency.

Opinion: He is weirdly overrating open models based on the usual foolish benchmarks, and so overrating the power of distillation. But that’s a side point and otherwise we agree.

Against the above, OpenAI lead thinks we are still underrating inference scaling. His claim is that we can get an LLM to do most things now, if we just run it long enough with a good-enough harness. e.g. Mythos and GPT hacking did not plateau even after 100M tokens / task.

Opinion: Worth pondering: it is indeed totally fine to pay $20k in tokens to get an Annals-level proof, and totally fine to pay $2m in tokens to get a new ML optimizer. Strongly disagree that current LLMs are ASI in the limit of galactic token spend, but they are probably this for most coding tasks and some mathematical tasks.

This discussion is frustratingly qualitative. Is there a 2026 update for the inference scaling law? The old one is not that good! We could reconcile Dwarkesh and Brown quite easily.

One route to tool world: Gwern argues for more personalized AIs, “Guardian Angels”, with the intent to amplify existing humans instead of appealing to an economically-misaligned common “mode collapsed” denominator not immensely useful for any particular user. Building a prototype this summer, aiming for a 100x productivity boost.

“We increasingly lived in a world where LLMs were powerless to augment or help me, but ever more powerful to replace or hurt me… I, for example, have long struggled to get much use out of chatbot LLMs, because they are—surprisingly, given their pretraining and my extensive corpus—bad at imitating me, and their thoughts and insights invariably shallow and worth little… The chatbot paradigm has failed to augment knowledge workers. We keep hearing that the gains will “diffuse”, and we keep not seeing much in the way of benefits, and knowledge work remains a “weak link” O-ring mode where LLMs fail to improve the bottlenecks (while coming with their own drawbacks, like the externalized costs of forcing everyone to waste ever more time with CAPTCHAs and paywalls).”

See also: a spirited attempt by a brilliant former OpenAI scientist to deslop AI writing.

Opinion: Gwern post is also one of the clearest statements of LLM economics and shortcomings. He notes that Thinky is not (yet?) doing this properly because they need enterprise first. Market failure!

SWE-Marathon, yet another long-horizon autonomous coding bench, released. Actually interesting despite the scientific clickbait:

  1. Opus 4.8 and 4.7 beat GPT-5.5: they emphasise “26% vs 12%” but this is on the kinda silly strict binary “did all tests pass?” metric. It’s actually 71% vs 64% – but still!

  2. GPT is 4x cheaper than Opus 4.8

  3. trials used an average of 50M tokens per task, ran around 4 hours.

  4. Notable for being by an RL env company, releasing 320 GB of agent trajectories and all 1,300 rollout logs for inspection, and so probably quite insightful about that part of training.

  5. Changing the harness can halve success and double reward hacking.

  6. Kimi got 0% exact. The proprietary harness was worse than the Terminal Bench one(!)

Opinion: Despite the “SWE” name it’s not by the usual team and we like it. However, we should ignore the product clones category, because it’s both maximally contaminated and also nearly definitionally useless. SWE-Marathon looks much less hard without them.

Annals of the open source gap: How does GPT-4o (200B, May 2024) stack up against current-gen open models? Qwen3.6 35B and Gemma 4 26B punch way above their active param counts: both models get general chat quality that exceeds GPT-4o significantly (by winrate, as judged by Opus 4.6).

Opinion: Interesting study, made less interesting by the use of Opus 4.6 in place of human raters. I’m happy to use 2 years as an upper bound on the open source gap and don’t see any point in testing this further. We continue to deemphasize OSS.

Physics Nobelist Parisi using Claude Sonnet to help him with an analytic proof of a statistical mechanics problem.

Opinion: We’ve known they’re useful to open-minded thinkers for 18 months. This hundredth example just cements how undeniable this now is. Also a rare instance of Sonnet overhang.

Annals of LLMs not being AGI: Anthropic argument that biological data infrastructure is too messy for AI agents to navigate reliably because it relies on manual browser clicking. When asked to pull viral sequence data from the NCBI database, top models were extremely inconsistent, generating errors that could derail real-world outbreak tracking. The team built a deterministic retrieval software tool called “gget virus” to handle the navigation logic. Giving agents access to this tool pushed accuracy to near 100%, showing that scientific databases must be redesigned natively for machines.

Opinion: Surprisingly not bitter-lesson-pilled. Yes, biological data repositories can be an absolute nightmare to navigate properly, but then isn’t Anthropic expecting a country of geniuses in a data center some time soon? Surely a country of geniuses would be able to navigate and click around biological data repositories. Could be explained by the author (apparently?) not being an Anthropic researcher.

Anthropic demonstrates that Mythos can automatically reverse-engineer software security patches into working N-day exploits. Claude Mythos Preview built 8 working code-execution exploits for Firefox and 8 full privilege-escalation chains for closed-source Windows kernels. The Windows exploits took only a few hours and cost roughly $2,000 each in API credits to generate. The authors warn that standard patch rollout schedules are no longer fast enough to protect users: ‘“N-day” has become dangerously misleading. N-hour is closer to the reality we now operate in.’

Opinion: Mythos performing extremely well as expected. Also remarkable is Opus 4.7 and 4.8’s ability, given that Anthropic tries to limit the models’ cybersecurity-related abilities.

Contrarian review of the Mythos general access pilot: good security capabilities, middling general coding ability, expensive, bad/restrictive harness. “half of the guardrails don’t work”.

Opinion: Cool to see a grounded take from someone with access. It’s interesting to see how the model is seemingly deployed in a dedicated harness rather than Claude Code. In any case, at least if rumours are right, we’ll be able to test Mythos (“Claude Fable 5”) soon.

Tool worldism: an attempt to extract specific capabilities from models one at a time instead of using the full ensemble, to cut down costs. For very simple arithmetic tasks, you can cover 90% of behavior with just 5% of channels.

Opinion: Not really clear this adds much compared to distillation other than at small compute budgets. Also, most of their results are at small (<= 8B params) scale.

Politics#

OpenAI vision post makes noises towards cooperating with slowdown: “We have long believed there should ultimately be an international organization that helps coordinate leading AI efforts to reduce catastrophic risk… to make it possible for the world to take coordinated action, including slowing frontier development when needed, so societal resilience, safety, and alignment can keep pace.”

Opinion: Characteristically vague and deniable but it’s Altman saying it and it’s the right direction.

In some sense, everything we have so far – including Obernolte-Trahan – is a weak regulation of AI. We may not have time for strong ones. How good are these for x-risk? Michael Dickens adds up effects:

  • Direct risk reduction (small); building infra monitoring infra; making warning shots visible via mandated safety-test plus whistleblower protections; shifting the Overton window.
  • Opportunity cost (meh); regulation fatigue (maybe soon); slowing beneficial progress (ok); and displacing companies’ good self-regulation (meh).

Opinion: Kind of worrying that we don’t have good estimates of any of those effects (how often do weak policies promote later strong policies?), but kudos to Michael anyway.

Annals of Republican safety: Mitt Romney makes strong statement in favor of national AI safeguards, including the risk of extinction.

Opinion: Probably easier for him to take this position compared to the typical politician, given the lobbying capability of frontier AI labs. But the shift he is part of is making it easier for the typical politician.

Safety#

Formalizing the notion of “alignment tax”. Commentary on pushing past the pareto frontier of helpful-harmless via legislation/social pressure and warnings against safety theater. Argues that external safety pressure (advocacy, discourse) is worthwhile only when attached to enforceable demands.

Opinion: Worth asking to what extent Ant and OAI are actually sticking onto this notional frontier, as opposed to above it (raising money to) or below it. A full model would instead estimate how much a pause would buy us, as an input to the safety tech development inputs to the safety frontier, and what dual-use things like Constitutional training do to the frontier and the available budget for safety tech.

Incidents#

Epoch release Cyber Vulnerabilities explorer to check the extent to which AI models are getting better at finding software vulnerabilities. A sharp uptick around the release date of Mythos is observed.

Opinion: Pretty interesting data from Epoch as usual. Unfortunately, it’s hard to do proper credit assignment between Anthropic’s Project Glasswing, Mythos, GPT-5.5 and coding agents more generally.

Minor#

  • More thoughts on Obernolte-Trahan/GAIA from Brundage, one of the key IVO players.
  • Claude Fable 5 (Mythos, but with expanded guardrails) is rumoured to be slated for release either today or tomorrow.
  • Sriram Krishnan (thoughtful, honest AI advisor) leaves the White House. Huge loss. Not sure they have anyone good left who is thinking out loud.
  • OpenAI file confidential S-1, sounds conflicted about it.
  • Step towards good evals: AI code quality
  • France to field test AI battlefield system at NATO drill, weaning off Maven towards French-developed Arcadia.
  • Science article on AI automation of science focusing on astrophysics. Raises the usual concerns alongside the risk of a feedback loop: increase in volume of papers makes it increasingly difficult to avoid using AI to pore through it.
  • OLMo 3 32B’s representation of its own qualia is not affected by posttraining.
  • Large list of open model releases over the past ~week.
  • The UK CMA has ordered Google to provide clearer links to publishers in its AI search features and offer a page-level opt-out mechanism for AI scraping.
  • Yet another report of a model leaking others’ response data, this time of Claude, alongside outage. Unlikely to be true.
  • Harness-1, a 20B open search agent, released. Externalizes the search state, then train the model to use that, instead of a search-read-search loop.
  • Apple reveals Siri AI update running on Gemini.
  • Bun migration from Zig to Rust as a debatable case study of gradual disempowerment.
  • Analysis investigating how transformers internally measure the distance between tokens. Unsurprisingly finds that the model considers line breaks meaningful boundaries, much more so than punctuation.
  • Snarky commentary on the open-closed model capability gap and adjacent benchmaxxing. Nothing novel, but collates a variety of metrics.
  • Insight into fast modes: Xiaomi achieves 1000 tokens/s without Cerebras-like hardware – it’s “model-system co-design”, a mix of quantization and speculative decoding. Still the first time this is achieved without dedicated hardware on a 1T+ param model. 3x the price for 10x the inference speed. ##