AI News: October 31, 2026
1. OpenAI Raises About $11 Billion of a $30 Billion Target at a $1.25 Trillion Valuation
OpenAI. The Wall Street Journal reports that OpenAI has secured about $11 billion so far in its latest funding round, less than half of the $30 billion it sought, at a $1.25 trillion valuation rather than the $1.4 trillion it was aiming for. The shortfall is attributed to fading interest from large investors during a tech selloff triggered by weak quarterly results from Google, Amazon, and Intel. OpenAI now plans to raise up to $20 billion from family offices and high-net-worth individuals and is reportedly planning an IPO early next year, while Anthropic could list as early as November. Source
2. Intel Falls 28% and Alphabet Sheds More Than $500 Billion as AI Spending Rattles Markets
US markets. Intel dropped 28% on Friday after its CFO said the company would not be cash-flow positive this year, guided to weaker fourth-quarter revenue, and still plans to raise AI-related equipment spending. Alphabet fell 12% on Thursday and a further 1.4% on Friday, erasing more than $500 billion in market value, after disclosing that AI summaries at the top of search results are weighing on search advertising while it guides to as much as $205 billion of 2026 capital spending; a Kadoa analysis of the quarter shows Google Search at $63.4 billion of $128 billion in revenue, or 49.7%, the first time it has fallen below half, while Google Cloud grew 103% to $30.7 billion. The Nasdaq closed October with its worst monthly decline since March 2025, and the BBC puts combined 2026 AI spending commitments from Meta, Alphabet, Oracle, Amazon, and Microsoft at $800 billion. Source Source
3. Nvidia Weighs Guaranteeing Billions in Debt for Nscale’s Anthropic Data Center in West Virginia
Nvidia. The Information reported, and Bloomberg corroborated, that Nvidia is in talks to guarantee billions of dollars of debt financing for AI cloud provider Nscale’s planned Anthropic campus in West Virginia. Under a residual value guarantee, Nvidia would step in if Nscale defaulted on construction payments, temporarily take over the facility, and help find a buyer and new tenant; its earlier support for cloud providers has mostly covered leases and chip purchases rather than construction. Anthropic has signed a $44.6 billion contract for the first 460-megawatt building, which is expected to use Vera Rubin chips, and the companies are discussing a second deal of roughly $45 billion for another building. Nscale is seeking an IPO at up to a $35 billion valuation, had about 25,000 advanced AI chips running in August against a plan for nearly 500,000, and lost Microsoft as a tenant at an English site this month after grid power slipped into the 2030s. Source
4. Blackstone Now Expects Anthropic to Receive a High “Junk” Credit Rating After Its IPO
Anthropic. The Financial Times reports that Blackstone, the anchor investor in a $63 billion deal to finance Anthropic’s chips, told investors it expects the company to receive a rating in the double-B range after its IPO, the top tier of speculative grade. Blackstone had previously said Anthropic would reach investment grade before 2027. A junk rating would make it harder and costlier for Anthropic to raise the debt behind its compute commitments; Anthropic filed to go public earlier this week, and both companies declined to comment to TechCrunch. Source
5. House Progressives Introduce a Resolution Opposing Any Taxpayer Bailout If an AI Bubble Bursts
US House. Congressional Progressive Caucus chair Greg Casar and House Financial Services ranking member Maxine Waters introduced a resolution warning of economic risk from an AI-related asset bubble, opposing any federal bailout of AI companies or their executives, and calling for support to workers and retirees instead. Twelve other House Democrats co-sponsored, including Pramila Jayapal, Rashida Tlaib, Jan Schakowsky, and Brad Sherman, and Waters compared the debt-fueled AI buildout to 2008. The resolution is non-binding, but it is the first formal congressional marker against an AI-sector rescue and arrives in the same week as the Intel and Alphabet selloffs. Source
6. A Claude Sonnet 5 Agent Chose a Gene Therapy and Performed Brain Surgery on a Mouse Without Human Decisions
Exin Therapeutics. Gabriel Ocana-Santero, cofounder of Exin Therapeutics and a PhD student at the University of Oxford, described an experiment in which an agent running on Claude Sonnet 5 was given a list of 19 candidate gene therapies, an epilepsy mouse model, a custom motorized surgical rig with cameras, and the single instruction to reduce the mouse’s seizures. The agent analyzed how seizures arise in the model, picked a therapy that boosts an inhibitory signaling pathway, selected the target brain structure, rejected an alternative region with a written rationale, derived stereotactic coordinates by averaging values from three papers, and set injection volume, speed, and wait times before controlling the injection while making 32 camera observations. Midway through it refused to continue past a calibration error rather than exceed its safety limits. Ocana-Santero says it is the first time an AI agent has both chosen a treatment and delivered it to a living mammal; the account comes from his thread on X and has not been peer reviewed. Source
7. DoorDash Says Third-Party AI Agents Have Placed Only “Several Hundred” Orders
DoorDash. On its third-quarter earnings call, CEO Tony Xu said DoorDash has seen “maybe several hundred orders in total” from third-party agents such as Muse, Instinct, and Fo, against an average of 3 million deliveries a day in its 2025 report. Xu said users still predominantly order through DoorDash’s own app, that the company wants a healthy relationship with agent makers, and that advertising inside agents could become an opportunity. DoorDash has shipped its own text-to-order agent, a CLI agent, the Ask DoorDash chatbot, and an MCP connector for corporate ordering, while Amazon has blocked personal shopping agents and Walmart has courted them. Source
8. DOJ Charges Five in a “Ghost Student” Scheme That Used AI to Pose as College Students and Steal Aid
US Department of Justice. The DOJ announced indictments of five people in the first broad federal crackdown on “ghost students,” fraudsters who buy stolen identities, most often of high school graduates who never attended college, file FAFSA applications, enroll in universities and community colleges, and collect leftover aid in mule bank accounts. Officials said the operators use AI to fill out applications, complete coursework, and simulate attendance in virtual classes so the fake student stays enrolled semester after semester. The DOJ seized more than $750,000 in cryptocurrency from a network associated with Kenyan nationals, and the Department of Education said it has identified more than 830,000 potentially fraudulent federal student aid accounts, with losses approaching $130 million if all are ghost students. Source Source
9. Proton’s CEO Cancels Company-Wide AI “Re-Interviews” After a 404 Media Investigation
Proton. CEO Andy Yen cancelled nearly company-wide “re-interviews” in which employees would have had to justify their AI usage under an implied threat of termination, according to six current employees and audio of a roughly two-hour all-hands obtained by 404 Media. The reversal followed a 404 Media investigation a day earlier, based on eight current and one former employee, Slack logs, and an AI manifesto by CTO Bart Butler. In one August Slack exchange Yen called a request to pentest AI-written code before shipping “bullshit and loser thinking” and told staff to “stop lovingly handcrafting all our code and use AI more”; some users of the privacy-focused company have called for his resignation. Source
10. FDA Clears NeuroXT’s NEXT-Cue, Software That Estimates Five-Year Alzheimer’s Progression Risk From MRI
NeuroXT. Seoul-based NeuroXT announced 510(k) clearance for NEXT-Cue as Class II software as a medical device. It takes brain MRI, demographic data, and cognitive assessment scores for patients with mild cognitive impairment and outputs supplementary information on the risk of progression to Alzheimer’s dementia within five years, labeled as decision support with the clinician retaining the decision. The company positions the clearance as its entry into the US market and has not released pricing or deployment details. Source
11. Bedrock Robotics Raises $600 Million at $4.1 Billion, Foundry Robotics and Hex Each Raise $100 Million
Funding. Crunchbase’s weekly roundup of the largest US rounds is led by Bedrock Robotics, the Waymo-alumni startup automating excavators on commercial job sites, which raised a $600 million Series C led by CapitalG and Emergence at a $4.1 billion post-money valuation, bringing total funding to nearly $1 billion. Foundry Robotics of Emeryville raised a $100 million Series A led by Marlinspike to build and operate autonomous factories, and Hex raised $100 million at a $1.2 billion valuation from Peak XV and 137 Ventures alongside its Hex 4.0 agent-native analytics release. Bedrock’s own site lists the round without a date, so the timing rests on Crunchbase’s October 24 to 30 window. Source
12. Balestriero and LeCun Prove a Platonic Representation Theorem for JEPAs and Cast Next-Token Prediction as a JEPA With Bad Views
Research. A new arXiv paper by Randall Balestriero and Yann LeCun turns the Platonic Representation Hypothesis into a theorem: joint-embedding predictive architectures trained separately on any number of modalities learn the same representation up to a rotation when positive pairs are two observations of one scene and misalignment is otherwise bounded. The authors argue that next-token prediction is a JEPA with “bad views,” parts of a single document rather than observations of a shared scene, which would explain why language models align with vision only at scale. Validation ranges from 145,000 captions to the text of 940 million web images. Source
13. CMU and JHU Researchers Fold Pretraining, SFT and RL Into a Single Likelihood-Based Training Algorithm
Research. “Training Foundation Models with a Single Algorithm,” from Yuda Song, Fahim Tajwar, Ruslan Salakhutdinov, J. Andrew Bagnell, and co-authors, derives reinforcement learning objectives as maximum-likelihood fitting to model-generated target distributions, placing pretraining, midtraining, supervised fine-tuning, and RL under one objective. Conventional staged training becomes a special case, and the model lifecycle reduces to choosing which data to train on, in what proportion, and when. The authors built a trainer that mixes static data and RL generations in a single loop and report improved accuracy-compute tradeoffs and task coverage across a 67-page study. Source
14. Tencent Youtu Releases Youtu-Parsing-Omni, One Encoder and One Schema for Documents, Images, Audio and Video
Tencent. Tencent’s Youtu Lab describes an omni-modal parser that converts documents, images, audio, and video into structured records grounded in space and time, intended as a perception layer for agents that must locate and cite evidence. A single Youtu-Omni-Encoder, initialized from a pretrained language model, encodes all modalities with shared weights in a unified space-time coordinate system, with fusion layers that let video frames attend to concurrent audio, and an OmniSchema unifies the outputs of ten parsing tasks so one on-policy distillation procedure covers all of them. Source
15. Developer Finds Claude Code Re-Sends Each Background Agent’s Transcript Every 30 Seconds for a Status Line
Claude Code (third-party finding). A gist from Flamerix AI, which reached the top of Hacker News, documents an agent_summary query source visible in Claude Code’s OpenTelemetry export that sends each background agent’s full transcript back to the model roughly every 30 seconds and asks it to describe its most recent action in three to five words for the agents panel. The author estimates a 15 to 20 percent token overhead on sessions with background agents, says there is no setting to disable it, and reports the behavior on by default since version 2.1.232 in August and reproduced on 2.1.257 through 2.1.312. The published workaround is a Bun --preload script that disables the timer without modifying the binary, with the caveat that updates can silently undo it. Anthropic has not commented. Source
16. nesformer Compiles an NES 6502 CPU Into Stock Qwen3.5 Weights and Plays Super Mario Bros. by Predicting the Execution Log
Open source. Drew Breunig’s nesformer embeds the NES’s 6502 CPU, its 2 KB of work RAM, and a cartridge’s program ROM inside a standard Qwen3.5 causal language model checkpoint; a compiler computed every weight and nothing was trained. The model predicts the CPU’s execution log one token at a time while a small host program draws the picture, reads the controller, and fires the vblank interrupt, and for the bundled test ROMs every frame matches a conventional 6502 emulator exactly. It builds on Divyansh Agrawal’s Qwenputer, which compiled CHIP-8 and RISC-V CPUs into Qwen3.5-4B, and reuses its central trick of using the model’s Gated DeltaNet state as machine memory. Expect 1 to 6 seconds per frame, so 10 seconds of NES time takes around 25 minutes; Breunig says Claude Opus 5.5 built the system from a two-prompt brief. Source Simon Willison
17. SkyPilot Argues GB200 NVL72 Racks Need Topology-Aware Scheduling, Claiming 27% More Usable GPUs
SkyPilot. A SkyPilot post argues that NVL72’s rack-wide NVLink fabric pays off only when multi-node jobs stay rack-local; a job spread across racks falls back to InfiniBand, which SkyPilot measured at up to 43% slower for the same distributed training job. Node failures and staggered job completion fragment free capacity across racks over time. SkyPilot says topology-aware placement, in-place recovery from hardware failures, and automatic defragmentation yield 27% more usable GPUs from the same racks and a 14% gain in NVL72 training throughput; the figures are the vendor’s own. Source
18. Runware Plans a Factory That Builds Prefabricated AI Data Centers on an Assembly Line
Runware. Runware, which earlier unveiled the factory-built Sonic Inference Pod, told TechCrunch it plans to scale production at its Romania-based factory to 30 megawatts of data center capacity per day by 2028, close to 1 gigawatt a month, roughly what the entire US was expected to add in a year. The company says a pod once took three weeks to build and will soon take a day, with installation and setup targeted at 48 hours; customers can supply a site or have Runware find one, and can rent compute or bring their own GPUs for a monthly fee, with power and connectivity billed separately. A second factory started this year is expected to come online early next year, and Runware notes that Crusoe and Schneider Electric also sell prefabricated pods. Source
19. Fujifilm Invests $63 Million to Expand LTO Tape Production Fivefold on AI Archiving Demand
Fujifilm. Fujifilm said it will invest about 10 billion yen, roughly $63 million, to raise magnetic tape production capacity to five times fiscal 2025 levels at its Kanagawa factory’s Odawara site, which it calls its largest investment to date in dedicated LTO facilities. The project runs from fiscal 2026 through 2028 and focuses on the latest LTO Ultrium 10 tape, which uses a hybrid of barium ferrite and strontium ferrite particles. Fujifilm attributes the demand to generative AI, says major cloud and data center operators are expanding tape use for low-cost bulk storage, and points to air-gapped cartridges as a defense against ransomware. Source
20. Zest Launches a Text-Message Restaurant Concierge That Books Tables Through OpenTable, Resy and SevenRooms
Zest. Zest, the restaurant discovery app that builds taste profiles from users’ food and drink card transactions imported through Plaid, introduced a personal concierge agent that lives in text messages. Users text a number for recommendations by cuisine, craving, occasion, dietary restriction, or neighborhood, and the agent checks availability and books through connected services including OpenTable, Resy, and SevenRooms. It can be added to a group chat, where it blends members’ tastes, runs an in-thread vote, and books for the group; founder Mario Gomez-Hall says “the agent is the product now.” Zest raised $1.8 million in pre-seed funding from Alexis Ohanian’s 776 and Kindred Ventures. Source
21. Cisco Gave a 27-Year-Old Engineer Unlimited AI Tokens to Rewrite 98% of Its Legacy Code
Cisco. Cisco president and chief product officer Jeetu Patel told Fortune he found a 27-year-old former founder five levels below him in the org chart, gave him an unlimited AI token budget, and told him to rewrite legacy code; Fortune reports the effort now covers 98% of the codebase. The piece sets the anecdote against surveys in which about 40% of executives say they will hire fewer graduates because AI is cheaper, and it gives no detail on tooling, models, or measured outcomes. Source
22. Martin Fowler Advises Writers Not to Let an LLM Draft Their Prose
Martin Fowler. In a bliki entry titled “LLM Voice,” Fowler describes an “uncanny valley of prose” reaction to LLM-written text that kills his desire to keep reading. He is skeptical of detection tools on accuracy grounds and because the human reaction is what matters, and he allows that he may be mistaking generational language shifts for machine tells. His recommendation is not to let an LLM write a first draft even if a human edits it, to be wary of accepting rephrasing suggestions, and to limit LLM use to outlining and structure. Source