Today in AI — October 6, 2026

A day of second-order moves: the labs are buying task data from software vendors instead of scraping the web, Waymo is borrowing to fund the robot world, and the biggest bank in the US just put a number on what Anthropic's latest model cost it in risk.

Models & Research

• OpenAI is training GPT-6 Astra on Ironclad's real contracting work. Ironclad staff helped turn 11 tasks across legal, commercial and procurement work — setting up NDAs, approval workflows, clauses that change by jurisdiction — into research problems, each graded against 8 to 50 criteria. On OpenAI's own evaluation, Astra averaged 55.0 percent against 41.6 percent for GPT-5.6 Sol, while estimated time per attempt fell from 37 minutes to 19.2; the training tasks were synthesized from public SEC filings, not customer contracts, and the time figures are simulated. OpenAI is now inviting a small number of software companies to hand it the same kind of messy, high-value work — if the pitch lands, the next frontier models get trained on the software you already use.
• A Microsoft page briefly confirmed looped transformers in the GPT-6 family — then got edited. A LocalLLaMA thread documented the publicly accessible page saying GPT-6.1 Sol runs two inference passes through its base model "instead of three," which would line up with earlier reporting that OpenAI's recent models use recurrent-depth training; the poster says Microsoft has since removed the passage. This is an architecture detail that leaked via a support page and vanished hours later — screenshots, not a press release, are the evidence, so treat it as strong but unconfirmed.
• EmpirioLabs open-sourced Aplomb 1, a 5.3-billion-parameter "decision model" with a 1-million-token context window. It accepts text, JSON, images, video and audio in a single request and returns calibrated probabilities for tool choices and their arguments, built on Qwen3.5-4B plus a Qwen3-Omni audio encoder. The headline benchmark claims come from the lab's own runs, and the license is free only for organizations under $1 million in annual revenue — but the weights, the full run logs and the benchmark kit are all public, which is more than most "decision layer" launches offer.
• A hobbyist matched a 114M dense model with a 21M one — by giving it a 6.4-billion-parameter lookup table that lives on an SSD. Following product-key memory and Meta's memory-layers work, the table is 4-bit and memory-mapped from NVMe, so the model writes about 140 tokens per second on a Radeon RX 9070 while using 0.4 GB of VRAM; cold reads cost roughly 8.5 milliseconds per token, and bolting the table onto an already-trained model didn't work. The author published the repo, the model, the failed runs and the success criteria he wrote down before each experiment, with the big runs costing about $70. The interesting long-term idea: knowledge sitting in inspectable, individually editable rows instead of opaque weights.
• Qwen 4 reportedly lands at the end of October. A LocalLLaMA poster says an Alibaba day-zero partner let the launch window slip, which would compress the gap between versions considerably. Alibaba did confirm at last month's Yunqi Conference that Qwen 4 is in training on a new architecture, with later releases scaling toward 5-to-10-trillion parameters — the specific date, though, is hearsay until Alibaba says it.

Industry

• Waymo has grown its first-ever debt raise from more than $3 billion to $5 billion. Bloomberg reports the enlarged size reflects rising AI costs and a robotaxi fleet expanding fast enough to need the capital now; the company was first revealed to be courting Pimco and Blackstone for the initial sum in early September. Borrowing rather than selling equity is a bet that per-ride economics arrive before the interest bill does — and a signal that Alphabet prefers debt off its own balance sheet for the robot business.
• Flai raised a $27 million Series A led by Base10 Partners for AI that answers car dealerships' phones. The software picks up calls, emails and texts — booking service appointments, transferring shoppers to the right salesperson — and TechCrunch reports it now books roughly 50,000 appointments a month. Vertical AI keeps winning the boring way: own one high-volume workflow completely, then charge for the minutes it saves.

Policy

• JPMorgan CEO Jamie Dimon says cyber risk went up "tenfold after Mythos." Speaking with Bloomberg TV at the bank's technology summit in London, Dimon said cyber was already JPMorgan's biggest risk before Anthropic published Mythos, and that AI "created vulnerabilities that we didn't know about." He explicitly avoided existential framing — the fix is "rolling up our sleeves and going to work," including a new 50-company coalition for critical infrastructure — but the operative line for the industry is that a frontier model release now moves a bank's risk calculus by an order of magnitude.

Tools

• Hark Pro, a privacy-pitched personal assistant from serial founder Brett Adcock, ships in wide release. Its model is trained specifically for computer use: it drives your machine, surfaces its browser navigation in a small window so you can watch it work, and the pitch against Meta and OpenAI is blunt — "we're not here to like sell you ads and steal your data." There's a free tier plus a subscription for heavy users, and an AI-native device is promised for 2027; the open question is how much general capability a computer-use-specialist model trades away.
• 54 GB of VRAM for $35: a dormant crypto-mining rig bought off a Russian marketplace. A LocalLLaMA poster negotiated a dusty box of nine P106 6 GB mining cards — GTX 1060-class parts — down to 3,000 rubles, confirmed all nine worked, and booted it from a USB stick. Used mining hardware is the discount aisle of local inference, and aggregate VRAM is the only spec that matters when the model has to fit somewhere.

Would you trust a bank's risk numbers that move tenfold on one model release — or a $35 mining rig with your workload? Tell us in the comments.