Humans on AI, March 13th 2026

• Frontier-lab lead remains robust

• SWE-Bench Verified remains unreliable

• GPT-5.4 appears to solve FrontierMath problem

• US urges customers to drop Anthropic

• Claude Constitution enables alignment tests

Opinion: They have a bunch of world-class people (though zero of the most tasteful, besides Friedman). Suggests that catching up to the frontier is very hard: talent and compute aren’t enough. This probably cost them $35bn.

Opinion: Bad (to the extent their compute might be put to productive use by more competent actors if they disappear), good (in lessening race dynamics). This is a larger loss than post-coup OpenAI. But the extreme turnover is probably less lethal than usual: Musk has a history of rebuilding after such things and it’s hard to shake investors’ confidence in him. Little bit of grovelling, and he’s already poached two heads from Cursor. But in the short term he will struggle, the mood is against them. Obvious move is to try and go literally AI-first, running autoresearch at million experiment / day scales.

Opinion: they’ve always spent a bunch, hiring thousands of ML engineers and producing solid models which don’t even reach the open frontier. We estimate they were spending $2bn/yr on model dev previously so new pledge only 2-3xes. It is in Nvidia’s interest to democratize access to good models, “commoditising their complement”. Even the attempt to commoditise model capabilities creates pressure on big labs to race for RSI. $5bn is plausibly as much as all Chinese labs combined are spending on model R&D right now.

Anthropic launches the Anthropic Institute, “a new effort to confront the most significant challenges that powerful AI will pose to our societies”, bringing Anthropic Societal Impacts, Red Teaming and Economic research under a new umbrella. Led by cofounder Jack Clark, with two excellent economists and Matt Botvinick (ex-Deepmind lead, law) as early hires.

Variables hit: labs self-policing, measuring societal impacts of AI

Opinion: A bit vague on what’s new here vs a restructure and brand segmentation.

Alibaba walk back their recent “agent going rogue and starting to abuse GPUs for crypto mining” incident. It was real (the AI did unexpectedly try something bad) but fake (the AI was simulating a rogue as part of its security tasks, not endorsing the rogue’s goal) and low-impact.

> We had a model tasked with a security audit — specifically, investigating abnormal CPU usage on a server. Somewhere along the way, it went off-script and decided to simulate a cryptocurrency miner to “construct a suspicious process scenario.”… our safety monitoring caught it immediately. The whole thing happened inside a strictly isolated sandbox — zero impact on anything external. The incident has been logged and will be used as a negative example in future RL training to reinforce what’s off-limits.

Variables hit: loss of control risk, broad autonomy

Opinion: Probably the most benign possible version of the story that doesn’t imply they were making it all up. Note that the practical difference between an LLM doing something and simulating doing something is minimal, though the implications can be different (e.g. how much it persists in trying, how much it resists correction).

Claims of multi-year lows in compute availability.

Variables hit: race dynamics, utility of existing models

Opinion: to the extent this is inference demand growth driven, suggests compute investments are less risky, less pressure to push for RSI from suppliers of compute (hyperscalers, semis). Good.

Opinion: Gap between benchmark results and real world application continues to exist. Unsurprising but useful result along the lines of replication.

Opinion: We think this is big, but even more for capabilities. Very useful for detecting issues (eg sycophancy, hallucinations). Likely to improve models in the long run, but is only one part of the puzzle; doesn’t directly provide alternative training methods to avoid the issues yet besides “repeat until it doesn’t have the issue”.

Opinion: Suggests lots of room for improvement in various domains possible with focused effort, good for tool world if small labs can compete on things, bad insofar as it puts pressure on labs to go for RSI if other things can be done cheaply.

Opinion: Fits with our experience, clearly pro-tool world to the extent these problems are hard to stamp out

Opinion: The problem is only “moderately interesting” (the least interesting category of FrontierMath Open Problems) but it would nonetheless be a meaningful result. We are likely to see more such results as Math provides perfect automated feedback through formal math environments like Lean. OpenAI appears to agree: one of the authors of this effort is expected to join the OpenAI for science team later this summer to further advance AI for mathematics.

Opinion: Not sure how impressive this is. Ramsey theory famously poses significant challenges. However, unlike most mathematics, combinatorics is particularly search-like, so progress here might not generalize to other areas of research.

Opinion: Not clear on direct impact of lifting designation in the short term - probably military suppliers still won’t want to use Anthropic to avoid DoW ire. Plausibly reduces the amount of non-military harm done to Anthropic. USG actively reaching out to customers suggests they are keen to maximise damage here.

Opinion: Overwhelming power of the DoW to be balanced against their recent habit of breaking the law in ways that Congress and courts won’t retroactively approve.

Opinion: Palantir positioning themselves as a middleman here, also allows systems to be more model/provider agnostic.

Opinion: Unlikely to lead to any policy near term but notable that this noise is building. How bipartisan it ends up being will be an important factor.

Opinion: Propaganda: bad poll structure - guardrails unspecified so voters can fill in the blank with whatever they want, naturally this will poll well. Even “ban vs no guardrails” is a very loaded framing in this context.

Opinion: Says what you’d guess if you spent 10 minutes on Bluesky looking at AI discourse

Opinion: Good news, but not groundbreaking. Highlights: GPT-5.2 still caves under pressure while Sonnet 4.6 doesn’t, but it along with other Claude models still fabricates data, lies when instructed to and takes unilateral actions.