In partnership with
In Today’s Issue:
🫠 Altman's three endgames for superintelligence
💰 OpenAI races to raise money before Anthropic's IPO
🧠 Anthropic's next model reportedly reads your intentions
🧮 Berkeley mathematicians react to OpenAI's proof
🤖 Dream-1: a robot brain for $3 an hour
✨ And more AI goodness…
⚡ The Signal
The AI race is accelerating on every front at once, and its best-known runner has told his staff that the industry still has to pick an endgame.
OpenAI is rushing to raise about $30 billion before Anthropic goes public, Anthropic is readying a model that reportedly reads what users really want, and MiniMax's new model gets within a few points of OpenAI's GPT-6 Sol on agent work at a third of the cost. Sam Altman's options, shared behind closed doors: a global watchdog, one merged super-lab, or the scenario he calls terrible, with OpenAI due to pick its side within a year. At Berkeley, mathematicians are already arguing about what their field is for after OpenAI's claimed proof, and Dream Labs now rents out a robot brain by the hour. Measuring the speed is easy. Steering it is the part nobody has worked out.
All the best,
Kim Isenberg
(REUTERS)
💰 OpenAI Races to Raise $30 Billion Before Anthropic's IPO
OpenAI is racing to raise about $30 billion before Anthropic's expected stock-market debut in mid-November, in a round that could value it at more than $1.4 trillion. Reuters reports that SoftBank and Abu Dhabi's MGX are in talks to join, and that OpenAI is pitching $50 billion in annualized revenue, rising to $70 billion by year-end. The rush is about comparison: Anthropic's IPO disclosures will give investors a detailed look at its business, after a $16.8 billion third quarter that left some backers arguing for a $1.5 trillion valuation instead of the hoped-for $2 trillion.
👉 tl;dr: OpenAI wants its money raised before investors can line up its numbers against Anthropic's public filings.
The Information's report on Anthropic's unreleased model (The Information)
🧠 Anthropic's Next Model Reportedly Reads Between the Lines
Anthropic is preparing a new model that reportedly works out what users actually want, even when they barely say it. According to The Information, one person who has used the unreleased model says it reasons and intuits better than Anthropic's current commercial models: asked to fix a single bug, it may track down the whole class of problems behind it and ask whether to fix all of them. It is not yet clear whether it joins the Fable or the Opus family. Google is reportedly readying Gemini 4 Carbon, and OpenAI a cheaper new Terra model for the coming weeks.
👉 tl;dr: Anthropic's next model reportedly infers the goal behind a request, so it can fix the root cause instead of only the symptom you named.
A UC Berkeley math lecture (The Daily Californian)
🧮 Berkeley Mathematicians Split Over OpenAI's Proof
Six weeks after OpenAI claimed a solution to Navier-Stokes, one of math's seven $1 million Millennium Prize Problems, UC Berkeley's mathematicians are split on what it means for their field. Fields Medalist Richard Borcherds says he had expected this capability in about ten years, thinks AI may outperform beginning graduate students and reports that some students are dropping out or worrying about their futures. Ph.D. student Robin Huang calls today's AI "very stupid and very inefficient" for research, NYU's Tristan Buckmaster questions how the result was produced, and the proof is still drawing scrutiny.
👉 tl;dr: OpenAI's proof has split Berkeley's mathematicians: a Fields Medalist sees AI catching up with young researchers, while a Ph.D. student calls today's AI too weak for real research.
🎬 Watch This
Nick Bostrom, the philosopher behind the book Superintelligence, tells AI-safety researcher Roman Yampolskiy why he doubts anyone can keep a radically superintelligent system "in a bottle" forever: sandboxes and access limits buy time, but in the end the system has to want what we want. He also says it "may be less misleading to say that it is an AGI than to say that it is not" about GPT-6, which still lacks continual learning, research taste and physical skill. A thoughtful companion to today's lead: Altman asks who should control superintelligence, Bostrom asks whether control can last.
"We need to constantly differentiate between the things we can do and the things we should do."
– Jack Clark, Anthropic co-founder, in his IMF Per Jacobsson Lecture
Clark, who told the IMF on Saturday that the singularity has begun, argued that getting through it will take "unusual and serious rules" and restraint: the kind of cooperation Altman's preferred endgames depend on.
Washington is throwing DeepSeek out just as DeepSeek reportedly digs deeper into China's chip industry. Huawei engineers have been working from DeepSeek's offices to build its requests into AI chips due next year and beyond, The Information reports, a possible route to frontier training without Nvidia. Meanwhile, DHS has ordered federal civilian agencies to remove DeepSeek products, including third-party tools built on it.
Altman Maps Three Endgames for Superintelligence
The Takeaway
👉 Altman told OpenAI staff about three proposals for developing superintelligence safely and said OpenAI must decide within a year which one to push for.
👉 He appeared to favor an IAEA-style international agency; the "Demis plan" would merge the major labs into one CERN-like body.
👉 He called the "Eliezer plan", in which one lab uses self-improving AI to cut rivals off from computing power, "terrible".
👉 He also said OpenAI cannot afford to give up ground now, as it and Anthropic head toward public listings.
Sam Altman has told OpenAI staff that the race to superintelligence needs an agreed endgame, and that OpenAI must decide which one to back within the next year. According to The Information, which spoke to people with knowledge of the remarks he made last week, Altman laid out three paths the industry has discussed for developing the technology safely. He appeared to favor the first: an international agency modeled on the International Atomic Energy Agency (IAEA) that would enforce safety standards, an idea he has promoted publicly since a 2023 blog post with Greg Brockman and Ilya Sutskever.
OpenAI CEO Sam Altman (Heather Diehl/Getty Images via The Information)
The second, which he called the "Demis plan" after Google DeepMind's Demis Hassabis, would fold the major labs into a single CERN-like body that supplies new models to partner countries. The third, the "Eliezer plan", is named after AI-risk theorist Eliezer Yudkowsky: one lab reaches recursive self-improvement, AI that improves itself, and uses the resulting economic might to "melt everyone else's GPUs," leaving rivals unable to catch up. "Obviously I think that one's terrible," Altman said, adding that he included it for "completeness." One person with knowledge of the comment said it was seen as a subtle dig at Anthropic, given Yudkowsky's ties to people there.
Altman on the third scenario, as reported by The Information (The Information)
The catch is that both paths Altman prefers need rivals and governments to cooperate, while he also told staff that OpenAI cannot afford to give up ground right now, with OpenAI and Anthropic both moving toward stock-market listings. Cooperation so far is thin: OpenAI, Google and Anthropic still disagree over who should test new models in their planned Standards Authority for Frontier AI, and Washington mainly relies on voluntary reviews of models before release. The whole account comes secondhand, from people with knowledge of the internal remarks.
Why it matters: OpenAI's pick, due within a year, will show whether the leading labs back outside oversight or keep racing toward the one-winner scenario Altman himself calls terrible.
Sources:
🔗 https://www.theinformation.com/articles/altman-describes-three-endgames-ai-including-one-single-firm-melts-everyone-elses-gpus?rc=bfliih
🔗 https://www.theinformation.com/articles/google-openai-anthropic-ai-safety-group-takes-shape
The AI notetaker that gets the hard words right
Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.
Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.
Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.
MiniMax M3.1 Flash on Agents' Last Exam: pass rate vs. total cost (MiniMax, using Agents' Last Exam leaderboard data)
The chart: MiniMax's launch chart for M3.1 Flash, released yesterday, plots pass rate on Agents' Last Exam, a 152-task test for AI agents, against the estimated cost of the full run. M3.1 Flash at max effort passes 29.6% for about $111; Claude Opus 5.5 leads at roughly 38% for well over $1,000, and GPT-6 Sol gets about 32% for roughly $320.
The lesson: A model whose weights MiniMax promises to publish on October 23 now reaches almost four-fifths of the leading score on agent work for under a tenth of the cost, a big deal for anyone paying to run agents around the clock. Developers noticed before the launch: as the anonymous "Space Bunny", an early build sat at #1 on OpenRouter for ten straight days, according to MiniMax.
The caveat: This is MiniMax's own chart. Its model ran in the Claude Code harness at max effort while rivals are shown with their best published setup, and all costs are estimates. Opus 5.5 still leads by about nine points.
🤖 Dream Labs Rents Out a Robot Brain for $3 an Hour
⚡ Bottom line
Dream Labs launched Dream-1, a model that runs two-armed robots it has never seen, rented out by the hour.
💡 Why it matters
Robot makers can rent intelligence as a cloud service instead of training a separate brain for every machine.
🔎 What it means
If outside tests keep confirming it, cheap general robot brains could open robotics to far more builders.
A startup that came out of stealth yesterday wants to sell robot intelligence the way AI labs sell chatbots: by the hour, to anyone. Dream Labs, led by CEO Joel Jang and backed by a $90 million seed round led by B Capital and Kindred Ventures, launched Dream-1, a model it says controls two-armed robots "it has never seen, on tasks it was never shown, in places it has never been." Access costs $3 an hour, or 75 cents for a lighter version, and the API is open to everyone.
Dream Labs' public Robot Playground, where anyone can type a task for a live robot (Dream Labs via SiliconANGLE)
The bet is on video. Dream-1 learns mostly from footage of people doing ordinary work, such as stacking shelves and packing boxes, predicts what should happen next and then works out the robot movement that makes it happen. Jang compares it to driving: "When you get into a rental car, you don't relearn how to drive." For now it handles robots with two arms and simple grippers, mainly picking, placing and folding clothes; wheeled robots and humanlike hands come later.
The strongest evidence so far comes from RoboCurve, a benchmark group that ran Dream-1 through 120 real robot runs on the first tier of its RoboDojo-RC test and published every video. It scored 59, the first model above 50, ahead of Claude Opus 5.5 and Gemini 4 Argon at 48 each.
Dream-1 tops RoboCurve's RoboDojo-RC Tier 1 leaderboard (RoboCurve, CC BY 4.0)
Cheap robot hardware already exists: Dream Labs works with more than 20 makers whose machines cost about $2,000 to $10,000. What was missing, in Jang's words, was "a brain any robot can use." One benchmark tier is not proof that Dream-1 will hold up in a messy warehouse, but a general robot brain that costs less per hour than a coffee changes who can afford to build useful robots.
Take the prompts our creative team actually uses
The key to maintaining brand quality at scale? Mastering how to communicate with AI models and embedding them into your creative strategy.
Join our honest discussion with Ari Murray, Chief Digital Officer at Salt and Stone, and get 5 tips for expert LLM prompting with ready to run prompts.
