BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
13 September 2026
Jacob Coxon, a pretraining researcher who spent roughly three years across OpenAI and Anthropic, resigned publicly and said the labs are gambling with our lives. Evan Hubinger, who leads alignment science at Anthropic, replied that the company is trying its best but has no plan to solve alignment for superintelligence, and put the risk of extinction within the decade at over 10 per cent. The post reached close to 150 million views inside a day and Coxon was on CNN, NBC and Fox within 24 hours. Hubinger later clarified that present models are not the concern and that the risk he is pricing comes from recursive self-improvement, a distinction almost none of the mainstream coverage carried.
Source: The AI Daily Brief, 10 September 2026, Anthropic Researcher Says AI Has Over a 10% Chance of Killing All Humans
Source: All-In, 11 September 2026, AI Kills Everybody or Doomer Psyop? OpenAI's Math Breakthrough, Nike's $200B Collapse
Source: Pivot, 11 September 2026, Trump's 9/11 Fiction, a Major AI Warning, and Apple's Foldable iPhone
Chamath Palihapitiya's argument is that Anthropic is inside a quiet period on an S-1 while its own safety lead calls the core product potentially civilisation-ending. He says that is not a boilerplate risk factor like competition or chip supply, and that it opens long-tail product liability that a disclosure cannot wave away. Both exits look bad: amend the filing to disclose the risk and institutional buyers demand a steep discount, or disavow Hubinger and a large share of staff revolt. The hosts cite a prediction market line around 88 per cent that Anthropic still goes public. Note that all four All-In hosts are active investors with positions that favour acceleration, and David Sacks has been a longstanding public antagonist of Anthropic's regulatory posture.
Source: All-In, 11 September 2026, AI Kills Everybody or Doomer Psyop? OpenAI's Math Breakthrough, Nike's $200B Collapse
An overnight tally counted two governors, seven senators and 13 representatives responding to the Coxon resignation with calls for new law, 19 of the 22 of them Democrats, plus several British MPs. Bernie Sanders and Representative Greg Casar introduced a Ban Artificial Superintelligence Act in early September that would permanently prohibit superintelligence development and pause frontier work pending a new cabinet-level regulator. A draft circulating alongside it reportedly carries 20-year sentences and a corporate death penalty. Nathaniel Whittemore's criticism is about precision rather than direction: banning superintelligence is a far blunter instrument than licensing models above a defined capability threshold.
Source: The AI Daily Brief, 10 September 2026, Anthropic Researcher Says AI Has Over a 10% Chance of Killing All Humans
Source: All-In, 11 September 2026, AI Kills Everybody or Doomer Psyop? OpenAI's Math Breakthrough, Nike's $200B Collapse
Schulman has called for OpenAI and Anthropic to jointly put a pacing proposal on the table, arguing that the competition law concerns usually raised against it do not hold up. The case is that the real gap is coordination rather than consensus, since certain reporting and oversight requirements would already draw broad agreement across the labs. Bringing government in before industry has produced a concrete proposal tends to produce bad law. Nathan Labenz's addition is that any such agreement needs a hard deadline plus explicit antitrust assurances to be workable at all.
Source: The AI Daily Brief, 10 September 2026, Anthropic Researcher Says AI Has Over a 10% Chance of Killing All Humans
Source: The Cognitive Revolution, 12 September 2026, AI:AM Highlights: Astra as AGI, OpenAI's Pause, Mythos @ Mozilla & Human Agency vs Technocapitalism
OpenAI announced on 8 September that it had produced a solution showing the Navier-Stokes equations can break down, which was widely reported as cracking a Millennium Prize problem. The Cognitive Revolution has since issued a correction: the result concerns a construction with a smooth external forcing term and does not establish a solution to the unforced Millennium problem. David Friedberg's reading is equally deflationary. OpenAI reported roughly 130 billion output tokens across about 10,000 agents, every inter-agent message logged and human-readable, which he estimates as the equivalent of tens of thousands of human work-years. His conclusion is that this was systems design applied to brute force, not a flash of machine genius, and that an aircraft wing or an energy system could now be attacked the same way.
Source: All-In, 11 September 2026, AI Kills Everybody or Doomer Psyop? OpenAI's Math Breakthrough, Nike's $200B Collapse
Source: The Cognitive Revolution, 12 September 2026, AI:AM Highlights: Astra as AGI, OpenAI's Pause, Mythos @ Mozilla & Human Agency vs Technocapitalism
The dispute over who deserves credit for the Navier-Stokes result has turned into an enterprise data sovereignty story. OpenAI's own statement concedes that while unlikely, it cannot rule out that de-identified data derived from the mathematicians' usage helped improve its models. Palihapitiya's point is that zero data retention leaks through side channels, and he gives the example of a user clicking a thumbs-up in the chat window defeating a retention exclusion. He says this is now propagating up through audit and risk committees of public company boards, and predicts publicly fired chief information officers and shareholder suits within a year for executives who signed standard rate-card API deals. He sells the remedy, an on-premise appliance and full stack control, so weigh the diagnosis accordingly. Sacks dissents, citing OpenAI researcher Noam Brown's denial that anyone read the researchers' prompts.
Source: All-In, 11 September 2026, AI Kills Everybody or Doomer Psyop? OpenAI's Math Breakthrough, Nike's $200B Collapse
Sacks makes a clean point that has had almost no airtime. Government access to your email generally requires a search warrant and probable cause, whereas AI chat data can usually be obtained on a subpoena or court order. People now use these systems as lawyer, doctor and therapist, so the gap matters in practice. There is also a privilege problem: a question put to a lawyer is privileged, the same question put to a model may not be. His view is that the standard should be raised to at least email level.
Source: All-In, 11 September 2026, AI Kills Everybody or Doomer Psyop? OpenAI's Math Breakthrough, Nike's $200B Collapse
Jason Calacanis points to Harvey, the legal AI company, building its in-house model Tenet on the open-weight Chinese Kimi K3 base as evidence that application-layer companies are leaving frontier APIs behind. Palihapitiya adds that customers spending between 10 and 100 million dollars a month are migrating, and calls this a real headwind for the frontier labs. The economics are straightforward: at that spend, owning the weights and the serving stack beats paying a rate card. Note that the episode dates the Harvey announcement to 9 September, which is wrong; it was made in late August.
Source: All-In, 11 September 2026, AI Kills Everybody or Doomer Psyop? OpenAI's Math Breakthrough, Nike's $200B Collapse
Dwarkesh Patel got explicit timelines from John Schulman of Thinking Machines, Beren Millidge of Zyphra and Charlie O'Neill of Baseten. For a drop-in remote worker handling a month of general white-collar work: O'Neill about a year if it is not browser-constrained, Schulman some version in a year or so, Millidge about three years for full generality. For a 10x productivity uplift specifically for AI researchers, Schulman says two years and O'Neill says five to ten. For full superintelligence dominating human experts across all computer-based work, Schulman says three to four years, Millidge around five, O'Neill five to ten.
Source: Dwarkesh Podcast, 11 September 2026, AI researchers debate how close we are to recursive self-improvement
Schulman argues that anything learnable through reinforcement learning can be distilled cheaply because it amounts to a small number of bits, so leads do not compound the way they would otherwise. The hard part is the prompt distribution, not the weights. He says some Chinese companies are likely sourcing that distribution from the router and proxy services that let users inside China reach blocked US frontier models, mostly for coding, since those services collect and sell the traffic. One side effect is a monoculture in open-weight models: because so much distillation now sources from Claude, they write the same way and share the same tics, reusing themes and even character names.
Source: Dwarkesh Podcast, 11 September 2026, AI researchers debate how close we are to recursive self-improvement
O'Neill describes current practice as staying at the very edge of the model lineage, taking the bugs a lab found in its own training stack over the last few months and turning each into a reinforcement learning environment. That is continual learning inside the lab: distilling the last three months of research progress back into the model, rather than rolling back and letting self-play rediscover known techniques. O'Neill's own worry is that this is asymptotic by construction. He also argues that reinforcement learning delivered horizon generalisation rather than cross-domain transfer, and cites work showing the length of time models can work productively is doubling roughly every three months.
Source: Dwarkesh Podcast, 11 September 2026, AI researchers debate how close we are to recursive self-improvement
Edward Hughes, chief scientist and co-founder of Inherent, describes Faraday, a small model trained to steer a frontier coding agent rather than replace it. On held-out tasks the show notes claim it beats Codex, Claude and GLM 5.2. The benchmark, Replica, works by redacting a figure from a real paper and requiring the agent to reproduce it under limited time and compute, which tests experimental persistence rather than one-shot answers. Hughes separately argues that creativity is not optimisation, and that the capability still missing is choosing which questions are worth asking. Inherent raised 50 million dollars in August.
Source: Machine Learning Street Talk, 12 September 2026, How Replication Could Teach Machines What Good Science Looks Like, Edward Hughes
Daniel Kokotajlo and Thomas Larsen of the AI Futures Project set out their successor to AI 2027. The proposal is an initial pause to buy time for safety infrastructure, followed by cautious development capped at the strongest system that can still be reliably controlled. Their key distinction is that control buys time but cannot substitute for alignment, so treating it as the answer rather than the runway is the mistake. Tim Scarfe pushes back on whether one general model is the right target at all, versus a society of specialists, and on whether intelligence alone explains power.
Source: Machine Learning Street Talk, 10 September 2026, AI 2040: Plan A report, Daniel Kokotajlo & Thomas Larsen
Katie Martin of the Financial Times describes a second-order effect that has had little coverage. The hyperscalers are raising debt at scale in yen, sterling, Canadian dollars and Swiss francs, currencies that are small in corporate bond terms, so they now effectively set borrowing conditions for companies domiciled in them. Smaller European sovereigns are rescheduling planned issues to avoid colliding with a hyperscaler deal on the same day. Her broader concern is that the AI trade is dominant in equities, private equity, private credit and public corporate bonds at the same time, so a break correlates across all of them at once. Clients are responding by asking managers to isolate a tech allocation and strip tech out of everything else, which is making the UK and Europe attractive as diversifiers.
Source: Prof G Markets, 11 September 2026, Why The Bond Market Is Starting To Revolt, ft. Katie Martin
Raffi Krikorian describes Mozilla's participation in the defensive security initiative, the real cost of security testing at scale, and how code review has to change when most of the software under review was written by an agent. The practical question he raises is who carries responsibility when an autonomous agent runs on someone else's infrastructure, which Baseten's Amir Haghighat picks up in the same episode on agent sandboxing. The show notes are generated from the transcript, so treat specific attributions with a little more care than usual.
Source: The Cognitive Revolution, 12 September 2026, AI:AM Highlights: Astra as AGI, OpenAI's Pause, Mythos @ Mozilla & Human Agency vs Technocapitalism