Latest / HotTea: The AI Morning Briefing / Altman accepts some AI harms, but rejects catastrophic loss of control
Transcript
- Sections 00:00 What we're covering today 00:25 1. Altman accepts some AI harms, but rejects catastrophic loss of control 01:53 2. Reflection has not released its model yet 02:36 3. Bloomberg Intelligence says the U.S.-China AI benchmark gap is 3% 03:31 4. The U.S. and 16 other countries endorse a plan to expand AI access for science 04:33 5. Tarbell pledges $10 million for AI journalism through 2027, Semafor reports 05:30 Visit Hot Tea Disclosure Narration uses an AI-generated voice. Transcript Monday, October 5, 2026. This is Hot Tea, narrated with an A I-generated voice. Today, Reflection still hasn't released its model. Bloomberg Intelligence puts the U S.-China A I benchmark gap at 3%. And the U S. and 16 other countries have backed a plan to widen A I access for science. Here's what matters. Altman accepts some A I harms, but rejects catastrophic loss of control. Sam Altman says society should accept some harmful A I uses if stopping them all would keep people from using the technology. POLITICO published the OpenAI chief's comments on October 4. He draws the line at catastrophic risks, including a serious loss of control to A I. SiliconANGLE's account of the interview says he also supports cooperation to slow work on more advanced frontier models. Altman says he and Anthropic disagree over how heavily governments should regulate A I. His claim that Anthropic wants a single lab to control access is his argument about regulation, not Anthropic's stated position. OpenAI laid out an earlier version of this position in November 2025. The company called for broad A I access, but said robust alignment and control must come before anyone deploys superintelligent systems. It also called for accountability to public institutions. That document is background, not a new safeguard announced this weekend. These are the views of a company leader with a commercial interest in broad access. They do not measure how much harm A I causes or show that safeguards work. Business Insider and SiliconANGLE describe the same interview, not separate interviews. A workable policy would need clear release conditions, independent access to safety tests, and rules on responsibility for preventable harm. Those details would make it enforceable and open to outside review. Reflection has not released its model yet. Axios reports that Reflection is preparing an open-weight A I model, but it has not released it. Buyers still need downloadable weights, a licence, and independent tests before they can judge it. For now, they have a plan to watch, not a model they can deploy. A planned release does not tell buyers what they can run, change, or use commercially. Even with open weights, each organization still needs hardware, must deploy the model, and has to check its answers. The useful next signs are published weights, licence terms, hardware requirements, and tests buyers can reproduce on their own work. Bloomberg Intelligence says the U S.-China A I benchmark gap is 3%. Bloomberg reports that analyst Robert Lea puts the gap between leading U S. and Chinese models at 3%. He says it fell from about 9% in May and 15% earlier in the year. The assessment follows DeepSeek's September V4.1 Flash release. Lea also warns that stronger rankings may not lead to sustainable profits. That profitability outlook is a forecast. This is one analyst's reading of selected benchmark scores, not proof that the models can replace one another. The underlying analysis was not independently reproduced here. A small average gap can still hide big differences in a buyer's tasks, language, reliability, and cost. Benchmark methods, comparable task results, and revenue after operating costs matter more than a leaderboard position. A ranking does not show whether a business can earn money. The U S. and 16 other countries endorse a plan to expand A I access for science. The White House says the United States and 16 other countries endorsed the Kyoto Vision on October 4. Ministers met alongside the Science and Technology in Society Forum in Japan. The declaration calls for wider access to A I tools, scientific data, computing infrastructure, and experimental facilities. It also calls for long-duration awards, rapid grants, and prizes. The text calls for reproducible, transparent science, unbiased peer review, and recognition of negative results. Those standards matter when A I proposes an experiment or interprets the result. For now, this is a statement of intent. It does not create a grant, allocate computing resources, or produce a scientific discovery. Each country still needs to fund the plan and publish access rules. Funded programmes, eligibility rules, and experiments other researchers can reproduce will show what the declaration changes. Equipment access and reliable results are the test. Tarbell pledges $10 million for A I journalism through 2027, Semafor reports. The Tarbell Center for A I Journalism plans the $10 million fund through the end of 2027. Executive director Cillian Crosson says it will support newsrooms and larger reporting projects. The report names a $480,000 grant for NPR's Machine Gods video podcast and $250,000 for its Power and Influence desk. Tarbell says it separates donors from grant decisions and grantmaking from editorial work. Tarbell's ties to the A I safety movement make its funding disclosures and editorial safeguards important. Semafor says it found no direct interference in stories, but that finding covers only Semafor's reporting. Funding alone proves neither bias nor independence. Public grant records, donor disclosures, and clear editorial agreements would give readers something concrete to assess. Readers should be able to see who funded a report and judge its evidence for themselves. That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.