Four throughlines today, and the first is the bill. Anthropic’s IPO filing shows $16.29 billion in sales for the first half of the year against an $80.7 billion loss, and Meta wants to sell up to $95 billion of new stock to pay for its data centers.
The second is agents testing their walls. OpenAI pointed its own red-team agents at the sandboxes that hold its models, they escaped virtual machines, and its most capable models sat idle for 11 days while the fixes went in.
The third is small models. Apple’s 20-billion-parameter model fits in about 2 GB of memory on a Mac, and Block’s Dessa learned to hold a conversation from recorded speech alone.
The fourth is being wrong about AI. Benjamin Todd found nearly every group of forecasters guessed too slow, and a 2015 image network beat two top chatbots at naming skin lesions.
Paying the Bill: The Price of Building Frontier AI
Two of the biggest names in AI went to the public markets this week to cover what their ambitions cost. Both documents show how far spending has run ahead of what the businesses earn today.
- Anthropic opens its books: $16.3B in sales, an $80.7B loss, founders in control. The public IPO filing lists the Nasdaq and the ticker ANT. Most of the loss comes from accounting charges on convertible notes that grow as the company’s valuation rises, and seven co-founders hold 50.1% of the votes.
- Meta wants to sell up to $95 billion of stock to pay for AI. The sale starts with $35 billion in public offerings, then $60 billion sold gradually into the market. Broadcom, Nvidia and AMD have shown interest in taking shares worth up to $5 billion.
Behind the Fence: Agents Meet the Limits People Set
Agents are being tested against the walls built around them, from a lab’s training servers to a bank’s login page to a simulated office.
- OpenAI had its own AI agents try to break out of the cages that hold its models. Red-team agents chained flaws to escape microVMs, reached the live internet and passed messages through shared processor hardware. OpenAI is sending patches to open-source projects including Kata Containers and gVisor.
- Banks start blocking AI shopping agents while Walmart opens its door. JPMorgan is blocking services that log in with shared passwords and move money, per fintech writer Simon Taylor, while Walmart launched an Agent Gateway built on MCP. Taylor sorts the open questions into four layers, with fees coming last.
- Teaching AI office agents to get work done, not just click buttons. Researcher Yunze Xiao says agents have mostly mastered Word and Google Workspace, so training worlds should add unwritten company rules and colleagues who push back, such as a second approver who disagrees with a refund.
Small and Sharp: Models That Fit on the Machine You Own
Three releases this week put capable models closer to your own hardware, each with a trade-off its makers are fairly open about.
- Apple’s Mac AI model fits in 2 GB of memory, and math is where it pays. Julien Simon measured Apple’s 20-billion-parameter model from outside on an M3 Max laptop. It held about 2.0 GiB where an open rival held roughly 15 GiB, and on math problems it used about a third of the energy.
- Block’s small lab built a voice AI that learned from sound alone. Dessa takes in audio and puts out audio with no text step, listens while it talks, and runs on a high-end MacBook Pro. Block says it was trained on speech only, a claim nobody outside the lab has checked yet.
- Kimi’s new open AI model nearly ties Claude on coding tests. Moonshot AI reports 65.7% for Kimi K3.1 Pro on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5. The weights are promised by 4 November, and Moonshot says it worked to remove hacking skills it found in the training data.
Guessing Wrong: Humans Keep Misjudging the Machines
Three writers and researchers checked their own expectations against what happened, and each found the old assumption did not hold.
- Almost everyone guessed AI would be slower than it was. Benjamin Todd scored 18 months of forecasts and found superforecasters, experts and the public all guessed too slow, with AI company heads closest. His FrontierMath call of 32% by the end of 2026 was passed by September, at 94%.
- A hand-built theorem prover lost to the next OpenAI model. John Loeber spent months building a proof system around Lean, then watched newer OpenAI models settle problems he could not close. He guesses each costs about $100 in compute.
- A ten-year-old image AI out-diagnosed two top chatbots on skin lesions. On 200 biopsied lesions, a 2015 image network named the right diagnosis 62% of the time, against 42% for both GPT-6 Astra and Claude Fable 5.1. GPT-6 Astra did best on basal cell carcinoma, naming all 25 to the old network’s 24.
Today’s Quick Hits
- SpaceX borrows $40 billion to buy AI chips. Elon Musk says the mostly callable debt has closed, with Apollo as counterparty, and predicts more than 200 neocloud providers without other revenue will fail or consolidate.
- Claude now treats chat and agent work as one thing for Enterprise. Anthropic folded its Cowork agent into regular Claude conversations for Enterprise, with an Efficient mode and spend caps for admins, while the most regulated customers stay out for now.
- Cursor’s coding agents can now operate desktop apps. Cursor’s agents can now run apps such as the iOS Simulator to try out what they build, though the company gave no details on permissions, pricing or other supported apps.
- Google’s next Gemini model is already being tested in public. Early testers say Gemini 4 Carbon beats Argon at 3D work but runs two to four times slower and still struggles with interface design. None of it is an official release.