The Omniscient Bulletin · 2026-08-19
The Omniscient Bulletin — August 19, 2026
OpenAI halted its largest training run and priced the watching of its own models at a fifth of compute. Then its quarter leaked. Nvidia's newly cleared China chips are landing in Hong Kong, not the mainland, and a robot maker rose 629 percent on day one.
Yesterday OpenAI said it had underestimated what its own models could do offensively. Today it published the bill: a two week pause on reinforcement learning, its largest frontier run still on hold, and monitoring that eats roughly a fifth of the inference compute it watches. Hours later its quarter reached the Wall Street Journal, and the model it had pointed at last week posted a score four points off the leader at a third of the cost per task. Elsewhere, a peer reviewed result quietly complicated every lawsuit about training data.
Frontier
OpenAI stopped its largest training run, and put a number on what watching its own models costs
OpenAI says it paused reinforcement learning training on its latest models intended for deployment for two weeks, and that its largest planned frontier run remains on hold while it runs smaller evaluations. Two things forced it: the Hugging Face incident, and preliminary evidence that an upcoming model, Astra, may meet a critical cyber capability threshold, a determination the company dates to August 7. Its new monitoring runs activation classifiers at every sampled token and costs roughly 20 percent of the inference compute it watches.
Frontier
The model OpenAI pointed at last week posted its score, at a third of what the leader charges per task
Artificial Analysis now has Z.ai's GLM-5.3, at maximum reasoning, scoring 59.5 on its Intelligence Index, eighth of 182 models it ranks. Claude Opus 5 leads at 63.1, Fable 5 at 62.1, GPT-5.6 Sol at 60.9, Kimi K3 at 59.7. The gap that matters is not the four points, it is the invoice: GLM-5.3 costs 68 cents per Intelligence Index task, against $2.34 for Opus 5 and $1.23 for Sol. It runs at 84.7 output tokens per second, and its API is priced at $1.40 per million input tokens and $4.40 per million output.
Industry
OpenAI's quarter reached investors the day it announced the pause, and Anthropic is now the larger business
OpenAI told investors it booked $6.7 billion in the three months to June, up 18 percent from $5.7 billion in the first quarter, with its operating margin sinking further into the red. The figures reach us through the Wall Street Journal, sourced to people familiar with the finances; OpenAI is private and publishes nothing. The comparison is the point. Anthropic reported $11.6 billion for the same quarter, up more than 50 percent sequentially, and a small operating profit, its first. A year ago the order ran the other way.
Research
Delete every image by one artist from the training set, and at large scale the output does not change
MIT CSAIL researchers report in Nature Communications that as a diffusion model's training set grows, no individual example matters much to any particular output, a phenomenon they call attribution decay. Past a certain scale, deleting one image, an entire artist's portfolio, or every photo of a single person from the training data leaves the generated sample unchanged. Their method, a diffusion ensemble, deletes the data outright rather than estimating its influence, which the authors call the first exact way to test the question.
Compute
Cerebras put three wafers in one rack, and says the thing decodes thirty times faster than a GPU system
CS-4 is the fourth generation Cerebras system, built from three Wafer Scale Engine 3 Turbo processors on a new modular rack platform called Nexus, with first shipments this quarter. Cerebras claims up to 30 times faster inference than GPU systems, up to 10 times the throughput per watt of CS-3, wafer to wafer latency as low as 2 microseconds, and more than 1,000 tokens per second on models above 10 trillion parameters. Every one of those figures is the company's own benchmarking. It can also hand the prefill stage to AMD or AWS silicon.
Policy
Washington cleared 100,000 H200s each, but Beijing is keeping the first shipments in Hong Kong, not the mainland
The Financial Times reports that small batches of Nvidia H200s have been cleared for China's market, with ByteDance and Tencent each receiving about 10,000 in recent weeks and other Chinese firms possibly next. Washington had already cleared those two to buy up to 100,000 apiece. The binding restriction is Beijing's: to protect domestic chipmakers, regulators have told firms to ship the processors to Hong Kong, which sits outside the mainland customs border, and use them there. Reuters could not independently verify the report, and Nvidia did not comment.
Robotics
A humanoid robot maker was oversubscribed 8,000 times, then rose 629 percent on its first day of trading
Unitree priced its Shanghai STAR Market listing at 150.8 yuan a share, valuing the Hangzhou company near 61 billion yuan and raising about 6.1 billion yuan for a tenth of its enlarged share capital. Its retail tranche was oversubscribed more than 8,000 times, a record for a technology startup on that market. On Wednesday's open the stock touched 1,100 yuan, up 629 percent. It is the first humanoid robot maker to list on a mainland exchange, and it reported 1.7 billion yuan of revenue for 2025.