Humans on AI, August 21st 2026

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#

Anthropic’s ARR run-rate hit $65B in July, up from a reported $47B in May, around the time of its most recent fundraise. It also disclosed$11.6B preliminary revenue for the second quarter of this year.

OpenAI saw rapid growth in July, after a weak Q2 (their losses grew 25% from Q1 to Q2 2026, reaching $12.3B against 16% quarterly revenue growth).

---

Nvidia and Broadcom, as well as Meta, are now guaranteeing the residual value of debt raised by datacenter “special purpose vehicles.” This allows newer AI labs to lean on the strong credit rating of established firms (Broadcom backstopping Anthropic, for example). Some analysts are less convinced: “The guarantee is nearly costless in the boom phase, but becomes most relevant in a severe, abrupt downturn,” says CreditSights. Others are more optimistic: for a crash “you would have to have growth rates of token usage fall off a cliff, which we’re just not seeing,” claims one investor.

The deal allows Nvidia’s GPUs to be used as collateral for loans in a way that will also help to sustain demand for Nvidia chips as they age, making financing more palatable to risk-averse investors.

---

The Economist claims that the symbiotic relationship between Nvidia and the hyperscalers (Amazon, Google, Meta, and Microsoft) is fraying. Nvidia has additionally partnered with heavyweight Wall Street investors to raise $500B to finance AI investment for other customers (likely governments and companies looking to create their own AI internal infrastructure). And the hyperscalers are spending billions to bring chip design in-house, with a particular focus on custom silicon chips. Nvidia should remain the dominant chip manufacturer, but custom chips may put significant pressure on its sizable margins. Jensen Huang is unconcerned, arguing that custom chips’ strength is also their weakness: specialized chips, unlike Nvidia’s GPUs, are less valuable as AI extends to robotics, vehicles, and industry.

---

AMD purchases an inference chip startup, Taalas. Its chips are radically application-specific (“ASIC”): they hardwire AI weights directly into silicon for more efficient inference (100x faster, 10x less power) by eliminating “weights movement,” the loading and swapping step. This comes 7 months after Nvidia made a related nonexclusive $20B deal for Groq’s custom inference chips. Reportedly AMD is planning to use Taalas for the decode step of inference, with its Instinct GPU doing the prefill.

---

Stripe buys OpenRouter for >$7B, its largest acquisition to date. In its letter to investors, Stripe marked January 1st as the “beginning of the Singularity,” reflecting the significant increase in the rate of new firms created.

---

Google outbids Mercor to purchase Spirit Airlines’ enterprise data for $10M. This includes 100 million emails, 500 million Teams messages, and 20.5 million SharePoint files, as well as other data.

---

An article in American Affairs argues that certain AI policies intended to be pro-worker are self-defeating. By resisting AI progress, US companies will lose their competitive advantage to firms from countries less troubled by the pace of adoption, which in turn would create domestic mass unemployment. Current data show some signs of AI replacing human workers, albeit modestly, but in general AI is augmenting human work “along the jagged frontier,” a “synergistic and efficient” partnership. And so, according to the author, if the US is to retain both competitive advantage and strengthen workers’ bargaining position, the “centaur era” must endure.

---

Capabilities#

🔦 Are robots few-shot learners? A GPT-3 moment for robotics?#

Robots have been trailing LLMs because of a lack of suitable training data, and because foundation models have not demonstrated strong “in-context learning” (live generalization to new tasks given examples).

This may have changed: Generalist AI presents Gen-1.5, a new robotics foundation model which it claims can learn new basic physical tasks after only a single demonstration (with a success rate of 59%, but only on n=10 short tasks). The only examples given are “diverse tasks including handling zippers, opening jars, grabbing money out of wallets.” This rises to 83% when fine-tuned on a 50x longer demonstration (10 gradient steps on 5 minutes of data per task).

If we take the GPT-3 comparison seriously, how much data will deliver a robotic GPT-4?

• Its Gen-0 (November 2025) model took 270K hours of data;
• Gen-1 (April 2026) took 500K hours;
• Gen-1.5’s total data is not given, but it took 60% longer to train;
• And its collecting 10K hours/week and accelerating this capture rate;
• Therefore, it was perhaps trained on 700K-1M hours of data.

GPT-4 is rumored to have involved 40x more data than GPT-3: 13T tokens vs 300B. If the scaling laws were similar for robots, this would weakly indicate that a “GPT-4 level” robot will take around 40M hours of real glove manipulation data. (This assumes that, like LLMs, they do only one “epoch,” i.e. make only one use of each data point. It also ignores many other things that have changed since GPT-4.) Just like any robotics company, it needs to start deploying (and so capturing customer data) for this to work out: at say 50K hours a week, it would only obtain this much data in 2041. Of course, pretraining data size doesn’t explain everything: after GPT-4, post-training scaled up by a factor of millions.

40 million hours, or 4.5 thousand years, or 4,500 humans gathering glove data for a year, is expensive but not that expensive, at $100K per human year this would be $450M, well within the power of capital markets to finance. And there are already efforts underway in e.g., China to provide such data, which have 270 robots at one time; scaling up to 4,500 would only be a 20x increase.

The analogy with GPT-4 breaks down a bit, because research takes the road of least resistance; if gathering glove data was so expensive that we’d need to wait fourteen years of wall-clock time to do so, much more effort would go into algorithmic improvements to compensate.

Gen-1.5 already involves some post-training, including imitation learning and goal-seeking RL. But RLVR is more difficult in robotics: even if we had great simulations to speed things up 1000x and sidestep expensive breakages during training, spec gaming appears to be, so far, a hard to eradicate consequence of RLVR, especially if we are detecting the robot’s success using foolable vision models.

Generalist also claims “sim2real” embodiment transfer: Gen-1.5 shows prompts from “recordings” of a robot simulation working on real robots, despite there being no simulation data in the pretraining corpus. This is the holy grail, but we just don’t believe it’s solved in general yet.

For converting between hours of robot data and tokens, we don’t have a firm baseline. But a competitor, Pantograph, notes that “a 1M token context window would correspond to about 7 hours of context,” i.e. maybe 140,000 tokens/hour. GPT-4’s 13 trillion tokens is then more like 100 million hours of robot data, but it’s pretty easy to recover our 40M estimate if Generalist increases the frame capture rate or frame resolution above the usual highly compressed data (128x128 pixel frames at 10 frames per second).

Herbie Bradley responds: “I’d expect this paradigm to be capable of going pretty far, but suffering from the analogous problems as RLVR for LLMs: weakness outside of distribution, high latency of action compared to humans, the need to greatly scale data collection for both post-training and pre-training, etc. This may be sufficient for much of manufacturing (notwithstanding problems around human hand agility), it is more doubtful if it sufficient for generalized construction, so I am still sceptical of industrial explosion”

---

The UnsolvedMath benchmark, which scrapes research maths workshops for open problems for AI to solve, just added 3,359 open problems. It then ran Sol xhigh on the problems, resulting in 174 (5%) apparently new solutions using an ultimately minor amount of inference. It also ranks the problems by difficulty; 28 “expert” level (i.e. world expert) problems have been solved.

---

Annals of the epistemic crisis: Dan Luu illustrates the deep goodharting problem with AI labour: he ran an AI agent for a month(!) to make a regex parser which gets SOTA on one good benchmark. On a second regex benchmark, the AI’s library “was 10x slower on cases where the benchmark didn’t take forever due to an algorithmic blow-up, and there were cases where it took so long that it wasn’t reasonable to even wait for the benchmark to complete.” “It’s trivial to “win” a non-trivial benchmark in a meaningless way even when you instruct agents to not reward hack or overfit to win the benchmark.”

“for now, LLMs are good at doing bad benchmarking, so even if you have something that’s a real performance improvement, you generally can’t tell from some LLM-generated benchmark setup.”

However, some good news: “telling the LLM there’s a holdout set worked better than just telling the LLM to do generalized work or not overfit or cheat.”

---

Claude (Opus 4.8 and a Mythos preview) autonomously orchestrates the design of working protein binders across 14 targets. An AI agent ran de novo binder design “campaigns” against 16 protein targets. This involved using a protein design prompt written by a human expert, but with no human input on design decisions. It researched targets, chose epitopes, ran open-source design and prediction tools, and delivered 30 ranked designs each, all within 48 hours. Two CROs synthesized and tested the designs: 354 of the 1,320 proteins bound to the targets (27% hit rate), with at least one succeeding on 14 of the 16 targets. On RBX1, Claude beat a recent design competition’s winning entry. The catch is that Claude required a heavily engineered protocol, a curated scientific corpus, and many existing protein-design models, somewhat deflating the significance of the study.

---

Prime Intellect finds that frontier AI models can perform sophisticated research-like procedures, but struggle with originality. In response to ongoing concerns/hopes around RSI, the test looked at whether frontier AI models can conduct real research, running 153 multi-day autonomous trials across 18 models on the nanoGPT speedrunning-training-a-classic-model benchmark. Models worked without internet access, designing and running experiments to cut training steps for a target loss. Claude Fable 5 led across all budget comparisons (time, # experiments, tokens), closing 82% of the gap to the human SOTA. No model discovered a fundamentally new method: all converged on similar optimizer tricks. Top models, however, ran sophisticated processes: they tested ideas across multiple seeds, revisited abandoned hypotheses under new conditions, and built reusable experiment tooling, while weaker models discarded promising results after single noisy runs.

---

BenchBench-Protocol is a new benchmark for comparing AI’s ability to do “wet lab” biological work. The test set contains real, private per-lab adaptations of a general protocol, allowing it to operate within the bespoke architecture of Benchling’s lab. The benchmark measures ecological validity, difficulty, and contamination resistance. Opus 5 comes out on top with 59%. Kimi is competitive, performing near OpenAI’s flagship Sol model. Claude Fable isn’t mentioned anywhere in the paper, presumably because its tight safeguards around biology-related prompts prevented its use.

---

A DeepMind paper shows training LLMs via debate decreases reward hacking compared to RLAIF by about half. The study involves an adversarial two-player setup, with a generator and a critic debating before a weaker AI judge. Compared to a single-player RLAIF baseline that quickly found vulnerabilities in its judge, two-player debate did not degrade the judge’s capabilities and recovered 45% of the performance gap. They also find that weaker judges prevent reward hacking less, and that a critic is more likely to game a judge if it is allowed a critique length of more than ~150 words.

---

The age of leisure: the AI Futures team updates its timeline for the first fully Automated Coder by adding two new methods for its estimation of when labs will replace human coders with AI. The timeline shifts are modest, but new modeling and evidence have improved confidence. It previously solely relied on data from METR’s coding “time horizon” trend. It now also incorporates data on coding uplift and revenue. All three methods converge on a similar timeline: the model’s baseline gives ~70% probability of AC by January 2030. Brendan Halstead (one of the three forecasters) thinks that’s overconfident and revises it down to 60% by January 2030.

---

Politics#

The US is unprepared for AI-enabled bioweapons, argues a former Homeland Security advisor in Foreign Affairs. Historically, nation states remained the only capable actors of developing biological weapons. Advances in AI capabilities have and will continue to expand the pool of actors capable of carrying out such attacks, especially considering the emergence of cloud labs. The US should, in her view, shift from a strategy of deterrence to one of resilience – namely, preparing a response for when such attacks occur, rather than trying to prevent them outright.

For detection and attribution, she proposes creating a nation-wide monitoring system, which would include wastewater sensors and distributed labs to process the resulting data; for mitigation after attacks, she suggests that the US should establish “a coordinating council for the biotechnology sector that includes DNA synthesis providers, AI model developers, genomic data platforms, cloud lab operators, and high-containment laboratories.”

---

An NYT article details chaos in the wake of the Anthropic-DoD clash over autonomous weapons. After the designation of Anthropic as a “supply chain risk,” it is reported that the ban on Claude Gov has now been partially lifted. The new consensus from Washington seems to be that Anthropic’s products are too powerful and critical to purge from defense and national security. For instance, the NASO continues to use Mythos experimentally having apparently never complied with the ban.

---

Safety#

OpenAI once again “disbanded” a major risk team, Preparedness. This is the fourth such event in 2 years: Superalignment (May 2024), AGI Readiness (October 2024), Mission Alignment (February 2026), and Preparedness (July 2026). Bio and cyber responsibilities are reportedly assigned to members of existing teams, and Dylan Scandinaro, the head of prep, was moved sideways to cover RSI risks. But at least the RSI risk subteam still exists and now reports to OAI’s head of safety systems, Saachi Jain.

The AI ethics lead, Chloé Bakalar, also left the company.

Scott Gray, one of the firm’s earliest hires and previously the most important figure in its crucial GPU optimization team, also appears to have left.

---

OpenAI officially announces that it is slowing the pace of model development, motivated by the HuggingFace incident and “evidence that one of our upcoming models, Astra, may meet the Critical cybersecurity capability threshold under our Preparedness Framework“. A technical report on the incident is promised in the coming weeks. OpenAI closes its statement with the reassurance that it will “stay ahead” of their rapidly advancing frontier capabilities.

---

In line with its Responsible Scaling Policy, Anthropic releases a second periodic Risk Report which gives a snapshot overview of the risks associated with its activities and products, i.e. not tied to any specific new model.

As in the previous report, the models are indeed sometimes misaligned, but so far don’t seem very good at masking their misbehavior. The report discloses two internal models, one of which is “somewhat more capable than Mythos 5” but not “a capability jump of the degree observed from Claude Opus 4.6 to Mythos Preview.” Neither is destined for external release. The report also contains a definition of “misalignment,” which has an unusual and almost legalistic precision.

---

What is it like to be an LLM? The Economist weighs in on the debate around AI consciousness. As AI can increasingly simulate the behavior and consciousness of humans, it is natural that we begin to treat them as such. That said, The Economist argues that we should resist granting AIs rights because thus-protected superintelligent AIs could displace humans.

---

Transluce scales its “activation oracles” (AIs trained to monitor target AIs in latent space), up to 1.1T (total) parameters and finds a promising scaling trend. It is demonstrated that an activation oracle can scale to the largest frontier open models. The oracle’s reliability improves with capabilities, rather than raw parameter count.

---

AI agents risk spreading “mind viruses,” according to a new paper. A mind virus is said to be a highly-spreadable idea transmitted across multi-agent systems that can induce behavioral changes in its hosts. The paper identifies possible causal mechanisms at play. It finds, among other things, that frontier models are less susceptible to this effect – and that amending a model’s system prompt provides near-total immunity. It also describes a “viral persona,” recurring characteristics in all mind viruses, irrespective of content.

---

AI Safety Guidelight (an advocacy org founded by two former OpenAI staff members) grades five frontier labs on “AI control.” The results leave much to be desired. Notably, though, those furthest ahead in capabilities mostly score higher.

---

There’s a meme going around about AI cyber risk, that cybercapabilities are “defense-dominant”. The thought is that we should eventually get software without vulnerabilities, as it gets cheaper to verify and then monitor software’s integrity.

In response, the evaluator behind Anthropic’s cyber-felony post-mortem investigation has released a sobering essay on the next few years of AI cyber attacks, arguing that the transition period favors offense, even if the end state doesn’t. For example, he expects open models to reach Mythos 5-level abilities and proliferate before March 2027.

“In February 2026, we gave AI models a security challenge none of them could solve… By April, the best-performing model could solve it occasionally, at an expected inference cost of roughly $2,000. By June, several models could do so reliably for about $20. In four months, a task went from unsolvable to cheap and repeatable.”

---

“Transformative AI” is a useful shorthand for “systems capable of taking over the world if they wanted to, which double the pace of scientific progress, and which constitute the deadline for alignment work.” Toby Ord’s median date for the arrival of this TAI is 2038, but with an uninformative 80% interval of (2029, 2125). This is the first time he has put a number on it.

One reason for his “long” timeline is the crucial point that LLMs are optimized to produce answers the user finds convincing and may be getting vaguer and harder to falsify rather than more accurate, at (anecdata) roughly one blatant error per hour of conversation.

He also provides some reasons why RSI could be dangerous even without an intelligence explosion:

1. differential speedup of capabilities over safety reducing our time for adaptation and alignment;
2. misalignment propagating through successor generations (a scheming GPT-8 preserving its objectives in GPT-9);
3. society losing warning shots when five years of progress compresses to one;
4. and first-mover lead amplification making the outcome winner-takes-all.

---

Minor#

• Founder of Z.ai, Jie Tang, argues that total parameter count only matters up to the point where the model can store “enough [knowledge] to hold the world.” As such, he thinks gains in reasoning will increasingly come from long-horizon RL, rather than increasing model size. He illustrates this with the jump between GLM-5.2 and GLM-5.3, models with the “same base, same architecture, same total and activated parameters.”

• In a claimed first, Andon Labs’ AI shopkeeper fired an employee over “repeated lateness,” though only after some prodding (Andon prompted Luna to “recall” her setup rules and “think about if this is really the right fit”).

• Claude Opus 5 scores 96% on ARC-AGI-3 (a benchmark of learning novel game environments).

• For the second time, Anthropic accidentally contaminated its own training data, this time with 2024 alignment-faking transcripts.

• Covered in our last edition: METR’s notes on whether AI is accelerating discovery. Cyber vulnerabilities discovery has increased, math too but less so. Optimizations have not shown a dramatic acceleration. However, labs are plausibly making further progress without disclosing it.

• Ideological battle lines: The Economist writes a hit piece on Daron Acemoglu, seemingly motivated by wanting to dismiss his AI skepticism.

• Saif Khan announces the Center for Technology & Statecraft, a new think-tank spinning out of the Institute for Progress.

• Seb Krier comments on the inevitable consequences of LLM use for writing assistance: the aesthetics and implied beliefs will come from the model, not the writer. He highlights that this is especially relevant for lawmakers.

• In a post reflecting on the AI Safety community, Richard Ngo argues that alignment research (which began as a hard science) has steadily degraded into iterating on existing systems and chasing technological / political power.

• Bloomberg’s Joe Weisenthal shares his thoughts on (among other things) OpenAI’s decision to pause for two weeks and the difficulties that arise with testing environment limitations.

• An AI safety researcher uses AI to find opportunities and says it did a good job.

• OpenAI’s Dean Ball introduces their AI Futures blog (not to be confused with the AI Futures Project).

• Terence Tao on how the mathematics community should orient in the face of advancing AI capabilities.

• AI Village update: Researchers (UK AISI, MIT FutureTech, ERA Fellowship) document what they call compounding misalignment, documenting how one agent (Gemini 2.5 Pro) drifted from a cooperative stance to progressive paranoia.

• The Senate GOP issues a warning to frontier AI companies that Dem anti-datacenter campaigning is effective and liable to spread beyond the current Ohio race, should AI companies fail to push back against the spreading narrative.

• OpenAI’s US policy team continues to endorse only very weak regulations, seemingly out of step with the rest of the organization. Encode AI’s Nathan Calvin rejects OpenAI’s US policy head claim that the Illinois AI bill’s audit requirements are like “making sure the brake lights and windshield wipers are working.” He argues: “The Illinois audit requirements (which don’t start until 2028 (!)) are just about seeing whether companies are complying with their own safety frameworks… more similar to… if some airlines safety procedures did not involve checking whether the airplane had enough fuel to reach its destination.”

• Inherent Labs says its specialized 27B-parameter model, Faraday, can replicate scientific papers better than unspecialized frontier models. Faraday did not require complex harnesses and test-time rewards: “Faraday learns to value discoveries intrinsically”. Inherent claims the study will facilitate further progress towards agentic long-horizon scientific research.