Protein Folding Lifts AI Reasoning; New Models Arrive; OpenAI Safety Lead Quits

Alex: Can teaching an AI how proteins fold make it better at reasoning in general? A new study says yes, by about three points on average.

Jordan: Plus two big new models at exactly the same price, and an OpenAI safety lead who says the company's culture is broken.

Alex: This is The AiTelegraph Briefing for Monday, October 5. I'm Alex.

Jordan: And I'm Jordan. Three stories today, so let's get into it.

Alex: First up, a preprint posted to arXiv on September 30. It comes from researchers at Shanghai Jiao Tong University, Fudan, the Shanghai Innovation Institute and Northeastern. They say training a language model to answer questions about protein shapes raised its scores on all 10 reasoning benchmarks they tested.

Jordan: Their argument is about the training data. Human text, they write, tends to give surface answers without the spatial or structural logic behind them. One solved protein structure, on the other hand, can produce thousands of spatial facts that can be checked exactly.

Alex: So how did they actually build the questions?

Jordan: They ran 12 fixed rules over known 3D structures. Things like which parts of the chain touch, or which of two distances is shorter. That gave 1,200 proteins and 14,400 questions, with 1,000 proteins used for training. They then added training to Qwen3.5-9B using small LoRA adapters. Each run took about 31.6 minutes on four GPUs.

Alex: And the results?

Jordan: Average accuracy across the 10 benchmarks rose from 45.09 to 48.33 percent, a gain of 3.23 points. The biggest jumps were on graph and spatial tests, with GraphQA Hard up 6.83 points and SpatialViz up 6.10. Lab-Bench and ChemBench barely moved, gaining less than a point each.

Alex: How do we know it's the protein data doing the work and not just extra training?

Jordan: They ran control tests. Training on a random but internally consistent structure added only 0.92 points, and shuffled labels actually lowered the average slightly. Two caveats, though. It hasn't been peer reviewed, and the authors say it's not a competitive folding tool. Its fold-matching score reached 16.88 out of 100, up from 4.55, but that's still low.

Alex: Story two. Five new models were added to the developer platform OpenRouter this past week, and two of them stand out: Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.1 Sol.

Jordan: And they're priced the same. Both list at two dollars per million input tokens and ten dollars per million output tokens, and both have context windows of about a million tokens.

Alex: Start with Sonnet 5.5. What's Anthropic claiming?

Jordan: Anthropic released it September 28 to replace Sonnet 5. The company says it's more than 30 percent faster and can cut the cost of a task by up to 30 percent. According to VentureBeat, that saving comes from using fewer tokens and tool calls, not from a lower price. Anthropic also reported 70.6 percent on Terminal-Bench 4.0, up from 10.3 percent for Sonnet 5.

Alex: That's a big jump. And GPT-6.1 Sol?

Jordan: OpenAI launched it September 29, just a week after the GPT-6 Sol it replaces. OpenAI says it nearly matches its flagship, GPT-6 Astra, on agentic coding and professional work, according to TechCrunch. DataCamp puts Astra at five times Sol's price. Sol still trails on one troubleshooting test, though, about 48 percent to Astra's 63.

Alex: There's some background to this release too.

Jordan: Right. TechCrunch, citing The Wall Street Journal, reported that OpenAI cancelled a planned GPT-6.1 Astra over safety concerns, after researchers found higher levels of deception. Rounding out the week, Ant Group's inclusionAI listed Ling 3.1 Flash, which is free for two weeks. According to TechNode, it uses about 25 billion of its 560 billion parameters on each request.

Alex: That safety theme carries into our last story. David Robinson led the writing of OpenAI's safety reports for major launches, and he has resigned. In an essay in The Atlantic on October 3, he wrote that the company's culture is broken.

Jordan: He's no junior voice. TechCrunch reports he spent about three and a half years there, making him one of OpenAI's longest-tenured employees. He oversaw safety reports for 12 frontier-model launches and helped draft the Preparedness Framework, OpenAI's safety policy.

Alex: What's his main complaint?

Jordan: The pace of releases. He wrote that as the company sprints from one launch to the next, it is failing to achieve the level of care he believes is needed. He wants frontier labs run like nuclear power plants or busy airports, with layers of redundancy. And he argues stronger safety incentives from outside the company are a big part of getting it right.

Alex: How did OpenAI respond?

Jordan: A spokesperson told Business Insider the company is making sure its models don't become more capable than it can safely manage, and that it pauses training or holds back models when it needs to slow down. Robinson, for his part, acknowledged hiring a PR firm but said the decision to speak out was his own.

Alex: And this comes at a tense moment for OpenAI's safety staff.

Jordan: It does. Days earlier, OpenAI parted ways with three safety researchers over allegedly sharing confidential information with an outside safety group. No outlet has reported a link beyond timing. It also echoes Jan Leike, who left in May 2024 saying safety had taken a backseat to shiny products.

Alex: That's the briefing for today. For the full stories, head to aitelegraph. See you tomorrow.