OpenAI Adds Driscoll as Open-Weight Debate Sharpens at Wuzhen

OpenAI confirmed the hire on Monday 2 November, with coverage landing on 3 November. Dan Driscoll, the 26th Secretary of the US Army, takes a newly created role: senior counselor for national security modernization. He will advise on how advanced AI can strengthen national and cyber security. He will also work with OpenAI's leaders and outside partners, according to The Next Web.

The timing is not subtle. OpenAI has spent the year positioning itself as the default AI supplier to the US government. This hire buys it a person rather than a product.

Driscoll left the Pentagon in September after months of tension with Defense Secretary Pete Hegseth, The Wall Street Journal reported, as summarised by The Next Web. He is a former Army officer and a law school classmate of Vice President JD Vance. As Army secretary he oversaw nearly one million soldiers and more than 265,000 Army civilians, OpenAI said. He also served as acting director of the ATF, per his own account.

The Decoder adds a detail worth keeping. During his tenure Driscoll pushed AI modernization across the Army, including Project ARIA. That initiative put AI tools into budgeting, logistics and tactical model deployment. That is procurement experience, not just policy experience. The Decoder also notes he submitted his resignation in late August 2026 amid friction with Hegseth. Hegseth repeatedly removed senior officers or blocked promotions over Driscoll's objections, the report says. Driscoll keeps strong relationships with both parties in Congress, according to AP.

"National security requirements are changing faster than at any time in the past fifty years," Driscoll said in a statement.

Sam Altman framed it in export terms. "The US has an extraordinary opportunity to work with other countries to spread the adoption of democratic AI," the chief executive said. OpenAI said the hire builds on Daybreak, its cyber resilience programme with allies. The company has opened Daybreak to Ukraine and pitched it to utilities in the US.

The contract history underneath the announcement

The background matters more than the job title. In February 2026, OpenAI agreed to deploy its models on the Defense Department's classified network. The deal came after rival Anthropic's relationship with the Pentagon broke down, Rachel Metz reported for Bloomberg. The Decoder's account is blunter. The Pentagon classified Anthropic as a "supply chain risk" after chief executive Dario Amodei refused to drop restrictions on mass surveillance and autonomous weapons from his contracts. Altman announced his own deal the same day. Anthropic is still fighting its Pentagon case in court, per The Next Web.

So the national security fight between the two labs is not only about capability. It is about contract terms, and about who will write the rules for military AI use. Driscoll's congressional network is the asset OpenAI is buying.

None of this touches open weights directly. It sets the weather for everything else. The same government that picks a classified-network vendor also shapes the argument over whether model weights should be published at all.

Tsai makes the open case in Wuzhen

On 3 November, Alibaba chairman Joe Tsai championed open-source AI at the Wuzhen Summit, the South China Morning Post reported. He framed it against US pushback. That is the same day as the Driscoll news and the same week as a burst of open-weight releases from European and Chinese labs.

Public reporting does not give Tsai's specific arguments beyond that framing. Treat the SCMP piece as a signal of positioning rather than a detailed policy proposal. Alibaba's own open-weight record is the more concrete question, and public reporting does not settle it.

What is public shows an ecosystem moving without waiting for permission. GitHub's release notes for PSSA v0.5.0, a non-transformer language model project published 31 October, describe a parameter-matched transformer baseline. They also describe a Qwen3-to-PSSA converter with a pure-Rust teacher forward pass. The notes list cross-entropy plus temperature-KL loss verified by finite-difference tests, and native CUDA alongside WebGPU backends. This is a hobbyist-scale project, but the engineering checklist is the kind of thing that used to require a lab.

Anthropic's proof system, and why it matters to open weights

Anthropic published research on 3 November describing FlashProve, a zkSNARK-based system for proving that a response came from a specific model. The company's own numbers: for Llama 4 Maverick, a 400B-parameter mixture of experts, a single H100 generates proofs at around 110 tokens per second. Proof size is roughly 19 MB per 1,024-token request, using 4-bit quantized weights and integer arithmetic in prefill. That is about 70 to 95 times the GPU time of running the model itself on vLLM, Anthropic says.

Anthropic is explicit about the motivation. A user querying a model today cannot check that the responses came from the specific model a developer claims is responding. A routing service could secretly substitute another model for cost reasons. An attacker could tamper with the token stream. The system commits to a model and hides its weights; hiding the architecture is not a goal. Code, Lean proofs and documentation are released under Apache 2.0.

The open-weight angle is indirect but real. If inference can be verified cryptographically, the argument that closed weights are needed for trust weakens. If it cannot, every open checkpoint becomes harder to vouch for. Anthropic calls this intermediate progress, not a finished agenda.

What the benchmarks say about the gap

On the modelling side, the newest preprint is GSLM2, submitted to arXiv on 31 October by Yosuke Higuchi, Tom Fizycki and 15 co-authors. The work targets spoken language models. The authors say these lag text models in capability and data scaling efficiency. Their encoder, SpidR 2.0, induces discrete units at multiple levels of abstraction. A multi-stream model jointly predicts them.

The headline numbers: training data up to 6M hours and model scales up to 6.5B parameters. The gap to text-based models on sWuggy, sBLIMP and tSC shrank by about two orders of magnitude across read and conversational speech. Training code and checkpoints are open-sourced.

Elsewhere the picture is less flattering to scale alone. FrontierMat, submitted 30 October, is a benchmark of 330 problems drawn from frontier academic and industrial research. The fields are aerospace, semiconductors, embodied intelligence and nuclear fusion, each validated by experimentalists and industry experts. The strongest general-purpose model scores only 53.7%. On matched problems, cross-scale questions score 8.7 points below single-scale ones. Half the problems are solved by at most one model. Models fail chiefly by endorsing what is plausible rather than what was observed, the authors write.

A domain-adaptation study in the same paper raises an open-weight model from 36.7% to 55.4%. It uses meta-weighted continued pretraining and scale-modular self-distillation, matching the strongest proprietary model, with no decline on general benchmarks. That is one of the more useful data points for anyone arguing about open versus closed.

Efficiency work keeps chipping at the cost of running models

A 30 October preprint on mixture-of-experts pruning argues that expert outputs show structured redundancy. Across models and task categories, the stable rank and entropy effective rank of expert outputs sit far below the nominal number of experts. The authors propose two pruning methods, PEEP and DCEP. At 75% and 50% expert retention on DeepSeek-V2-Lite and Qwen3-30B-A3B they beat state-of-the-art baselines on most benchmarks.

Two more papers push in different directions. One, submitted 29 October, evaluates speech emotion recognition in audio-capable models. It finds that when audio and text disagree, the odds of error rise whenever audio is present, across eNTERFACE'05, MELD and CMU-MOSEI. Supervised classifiers on the same features behave oppositely. Another, also submitted 29 October, tests how language models anticipate the feeling of gambling or stock-trading outcomes. It finds what the authors call silicon gain amplification: models rated gains more intensely than equivalent losses, inverting canonical loss aversion.

None of these settle the open-weight argument. They do show that the research frontier is crowded with small, checkable claims rather than one big one.

The security ledger

OpenAI also moved on its own defences on 3 November. The Decoder reported that the company quadrupled bug bounty payouts. Critical vulnerabilities in core products such as ChatGPT or Codex now earn outside researchers up to $30,000. Exceptional findings earn up to $100,000, and with quality, loyalty and campaign bonuses the total can reach roughly $172,000. Anyone who reported a vulnerability in the last 30 days gets the difference retroactively. A security sprint will pair the firm Cathedral with OpenAI's own engineers. During the last sprint, 25% of OpenAI's product engineers worked on finding vulnerabilities, according to the company.

Whether open weights make that job easier or harder is exactly the question no number answers. Epoch AI's publication on open-weight models and the patch gap, dated 29 October, is not summarised here.

What is clear is the direction of travel. A former Army secretary is now advising a frontier lab. A Chinese executive is making the open-source case in public. A US lab is publishing cryptographic proofs that a given model produced a given token. And a benchmark built from real industrial research problems says the best general model still gets fewer than 54% right.

The open-weight question will be decided in procurement offices and courtrooms as much as in arXiv comments. On the evidence dated this week, both sides are staffing up.