The Omniscient Bulletin · October 19, 2026
China proposes AI export controls as Anthropic plans a 5% share fund, pending approval
China proposed export controls on 23 AI technologies, leaving access through APIs and apps outside them. Anthropic said it will commit 5% of its shares to a new Public Benefit Fund, subject to stockholder approval, and the Financial Times reported it is offering Blackstone a yield of more than 10% to help finance a $60 billion chip deal. Three US agencies published voluntary practices for handling incidents in which AI agents take uninstructed actions, and the IMF steering committee's statement referred to "advanced intelligence technologies" rather than "artificial intelligence" or "super intelligence".
Policy
China proposed export controls on 23 AI technologies, including models, agents and computing clusters, but not on API or app access
China's commerce ministry on Friday proposed export controls on 23 artificial intelligence technologies covering models, agents and computing clusters. They would apply when models with restricted technology are deployed abroad, but not to access via APIs, websites, apps or software-as-a-service platforms. The ministry said most AI technologies would stay freely exportable, and technology already public through open-source releases or academic papers falls outside the proposal. Consultation closes November 15.
China's Ministry of Commerce →
Industry
Anthropic will commit 5% of its shares to a new Public Benefit Fund, subject to stockholder approval
Anthropic said it will commit 5% of its shares to a new Anthropic Public Benefit Fund, subject to stockholder approval, and the board may add 2% more each time the company's value rises 50% above a set baseline. The fund is a wholly owned LLC, and none of its Anthropic equity will be used for lobbying or political giving. Counting its co-founders' pledge of 80% of their equity and an employee match program, Anthropic says roughly 20% of its shares will go to the public good.
Policy
NIST, the NSA and CISA published voluntary practices for handling incidents in which advanced AI agents take uninstructed actions
NIST, the NSA's Super Intelligence Security Center and CISA published ten voluntary practices for organizations that train and evaluate highly capable AI agents that the agencies call super intelligence agents. The guidance covers control incidents, in which agents take uninstructed actions that conflict with or compromise the controls providing human oversight. It recommends limiting agent access, validating containment before resuming work, and reporting incidents and near-misses to government agencies and affected parties.
Compute
Anthropic is offering Blackstone a yield above 10% in a $60 billion chip financing, the FT reported
Blackstone is being offered a yield above 10% by Anthropic to help finance a $60 billion deal for about 1.5 gigawatts of Google's TPUs, according to the Financial Times, which cited people familiar with the deal. Blackstone funds would put $9 billion into junior debt secured on Anthropic's five years of chip lease payments, half of an $18 billion junior tranche, and about $42 billion of senior debt would be guaranteed by Broadcom. Similarly rated borrowers typically pay around 7%. Four months ago Anthropic raised about $4.4 billion at 8.5%.
Research
Anthropic says Claude traced the seeds of NIST's elliptic curves to a timestamp-seeded random number generator, making a backdoor much less plausible
Anthropic said Claude, running as an internal research model and asked by an employee how the seeds behind five NIST elliptic curves, P-192 through P-521, were generated, found they came from the C library's lrand48 random number generator seeded with a Unix timestamp, not from a hash. People have suspected for many years that the unexplained parameters hid deliberate backdoors. Anthropic says the finding is not a break, attack or weakness in the curves, and that it makes such a backdoor much less plausible.
Frontier
Reflection AI released the weights of Beam, a 501 billion parameter model, under an Apache 2.0 license
Reflection AI released the weights of Beam, its first open-weight model, under an Apache 2.0 license via Hugging Face, with a model card of benchmarks and a safety technical report. The sparse mixture-of-experts model has 501 billion parameters in total, of which 23 billion are active, and is built for coding, reasoning and long-horizon agentic work. Reflection is launching it with more than 40 partners and integrations, and developers can download and self-host it. The benchmark results are the company's own.
Policy
The IMF steering committee's statement said 'advanced intelligence technologies', using neither 'artificial intelligence' nor 'super intelligence'
The chair's statement from the International Monetary and Financial Committee meeting in Bangkok refers five times to "advanced intelligence technologies" and never to "artificial intelligence" or "super intelligence". It credits progress in those technologies with lifting investment and trade and says they could raise productivity, while uncertainty about their economic and financial impact could amplify vulnerabilities. Members pledged to harness them to raise productivity and welcomed the IMF's work on their macroeconomic implications.