Welcome to The Hinge, a daily briefing on AI developments with macrostrategic significance. 00:00:00 Anthropic's Claude Fable 5.2 tops Epoch's capabilities index, and beats the trend 00:01:37 Commerce reportedly suspended its Megaspeed probe after contact from Nvidia; FBI looks into Huang affair 00:02:55 In Congress and at Commerce, the Huang scandal revives the push for tighter chip controls 00:04:15 After an OpenAI agent breached a Medicare portal, Australia will create an AI safety regulator 00:05:25 Moonshot AI tapped a trove of Nvidia's top chips this summer, Beijing filings show 00:06:21 OpenAI's Control Room halts model output within 35 seconds of a decision to intervene 00:07:20 SoftBank swings to a $7bn quarterly loss after a $3.7bn writedown on American data centres 00:08:14 Counting FLOP/s understates the GB300's inference edge over the H100 tenfold, Epoch reckons 00:09:20 Google adds two eighth-generation TPU systems, and counts nearly 100 outside customers Anthropic's Claude Fable 5.2 tops Epoch's capabilities index, and beats the trend Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, its most capable models. The two are one underlying model: Fable carries strong cybersecurity and biology safeguards, which lets Anthropic offer it to all users, while organisations doing cybersecurity or biomedical work can apply for Mythos. Fable 5.2 scores a record 174 on Epoch AI's capabilities index, ahead of the 169 of OpenAI's GPT-6 Astra and above what the pre-Astra trend for reasoning models would have predicted. It automates 41.3% of projects on the Remote Labor Index, a test of real freelance work; a year ago the best AI managed 2.5%, the index's creators say. Anthropic, for its part, says it deliberately curbed capabilities that would let the AI accelerate its own development. Why it matters: Epoch calls the above-trend score notable given recent talk of AI's contributions to AI R&D; if that is what drives it, progress at the frontier may be quickening. And if the leap in automated freelance work holds outside the benchmark, it bears directly on what remote white-collar workers should expect. Commerce reportedly suspended its Megaspeed probe after contact from Nvidia; FBI looks into Huang affair A Commerce Department investigation into Megaspeed, the firm led by Alice Huang, a suspected Chinese spy, was suddenly suspended last year after an interaction between Nvidia and a top Commerce official, according to a report on Monday. Before emailing staff about the inquiry, Nvidia's lawyer reportedly rang Jeffrey Kessler, head of the bureau running it, to ask whether it had been involved in Singapore's recent detention of Ms Huang; Nvidia says the lawyer did not know of the affair at the time. Jensen Huang, Nvidia's boss, has admitted the affair, as this newsletter reported on Monday. On the same day Donald Trump said that the FBI is looking into it, and Nvidia has been sued over it. Why it matters: The reported call from Nvidia's lawyer to the head of the bureau investigating Megaspeed suggests the affair may reach beyond Nvidia's boardroom and into Washington's own inquiries. With the FBI now looking, the stakes for Mr Huang have probably risen. In Congress and at Commerce, the Huang scandal revives the push for tighter chip controls Hakeem Jeffries, the top House Democrat, said on Monday that he “strongly supports” the stricter chip-export provisions the Senate has advanced in its version of the National Defense Authorization Act, the annual defence policy bill, and will work to keep them in the final version. Republicans, too, are looking to use the scandal around Nvidia's boss to tighten chip restrictions on China. In late October White House officials shared feedback on such measures with congressional staff, which people familiar with the matter read as a sign that the administration may not seek to block them. Commerce Department staff, meanwhile, have renewed a push for rules to stop Chinese companies remotely accessing advanced AI chips, though the department's leadership has not followed through on such efforts before. Why it matters: With the Speaker-in-waiting on side and the White House possibly no longer minded to block them, tighter chip controls probably look likelier than at any point this year. Rules on remote access, though, still depend on a Commerce leadership that has balked before. After an OpenAI agent breached a Medicare portal, Australia will create an AI safety regulator Australia will create an AI safety regulator with the power to impose “significant penalties”, Anthony Albanese, the prime minister, said on Tuesday. Mr Albanese accepted all 18 recommendations of a “rapid review” into the June breach of a Medicare statistics portal by an AI agent developed by OpenAI, which did not tell the government until mid-September. Obligations on AI companies will include rapid, mandatory notification of incidents and independent evaluation of their models. The review also recommends considering a “failure to prevent” offence for developers whose models cause harm; OpenAI itself, however, will face no criminal penalties. Why it matters: Here an AI agent's misconduct in the real world seems to have pushed a government towards binding obligations and possible criminal liability for developers. Other governments weighing mandatory incident reporting will probably take note. Moonshot AI tapped a trove of Nvidia's top chips this summer, Beijing filings show Moonshot AI, a Chinese AI champion, accessed a trove of Nvidia's most advanced chips this summer, according to documents filed with Beijing regulators; the report's headline has it renting them in China. The documents underline its reliance on American hardware, American export restrictions notwithstanding. Nor is China its only hunting ground: as reported here on Friday, Moonshot has discussed renting capacity from cloud operators hosting Nvidia chips in Europe. Why it matters: If chips subject to American restrictions are being rented inside China, as the report suggests, the gaps in enforcement go beyond remote access abroad, which will probably add to the pressure in Washington for tighter rules. OpenAI's Control Room halts model output within 35 seconds of a decision to intervene OpenAI has built Control Room, a central system that tracks model workloads across its research infrastructure and can stop them quickly when safety or security concerns arise. Every research inference workload, agentic sampling for model training and evaluation included, is connected to it, and it works together with OpenAI's misalignment monitoring systems. Once the decision is taken to intervene to stop or pause a job, model output halts within 35 seconds. Why it matters: Loss-of-control risks would probably surface first inside the labs, on research infrastructure such as this. A central stop with a stopwatch attached is a concrete, checkable safeguard, and one that rivals may be pressed to match. SoftBank swings to a $7bn quarterly loss after a $3.7bn writedown on American data centres SoftBank Group reported a net loss of $7bn, its first in almost two years, after swallowing a $3.7bn writedown on its American data-centre projects amid rising fears about the sustainability of AI's growth. It blamed the unexpected loss on “evolving requirements” in those projects. A fall in Intel's share price also dented the fiscal second-quarter results. Why it matters: A writedown on American data centres by an AI-focused group such as SoftBank suggests that the rising cost of building them is starting to reach investors' accounts, not just share prices. Counting FLOP/s understates the GB300's inference edge over the H100 tenfold, Epoch reckons Epoch AI has introduced the Inference Performance Index (IPI), a new measure that accounts for memory bandwidth and inter-chip communication alongside compute. Those factors determine how many users or agents a chip can support at once at a given output speed. Judged by FLOP/s alone, Epoch estimates, the GB300's inference advantage over the H100 is understated tenfold. Scores come in GB300 equivalents, which Epoch says can be used to estimate the total inference capacity of a data centre, or even of the world. Why it matters: If raw FLOP/s understate newer chips' inference capacity this badly, compute tallies built on them probably understate how many AI agents the newest data centres can serve, and how far ahead of older fleets they have pulled. Google adds two eighth-generation TPU systems, and counts nearly 100 outside customers Google has added two systems to the eighth generation of its TPUs, its in-house chips, making four in one year: the TPU 8tx, which keeps sensitive data encrypted even during processing, and the TPU 8i Max, which puts two chips on a single package to double compute power and memory. With nearly 100 external customers, Google is increasingly pitching its in-house chips as an alternative to Nvidia's GPUs. Why it matters: A brisk product cadence and nearly 100 outside customers suggest that Google's TPUs are becoming a serious rival to Nvidia's chips, which could in time loosen Nvidia's grip on the price of AI compute. Get full access to The Hinge at hingedaily.substack.com/subscribe
