Top Stories
Character.AI files for Chapter 11; Disney offers $20M for assets
Character.AI filed for Chapter 11 bankruptcy on October 13, five years after launch, with $50 million to $100 million in assets and liabilities, per TechCrunch. The chatbot startup has agreed to sell some assets to Disney for $20 million plus a non-exclusive perpetual tech license; the sale is expected to close in late November or early December, and if no qualified bid lands, the platform will be wound down. Character.AI had raised more than $193 million from Lightspeed, SV Angel, Greycroft, and a16z; co-founders Noam Shazeer and Daniel De Freitas returned to Google in August 2024 in a deal valued at $2.7 billion.
In a statement, CEO Colin Adams said the board determined that the court-supervised sale “provides the Company with the best opportunity to maximize value for all stakeholders,” and added that “if such a bid is unavailable, however, the platform will be wound down.” Disney will provide transition services through former Character.AI employees who joined it last month alongside former CEO Karandeep Singh Anand.
Bankruptcy filings show Character.AI’s unit economics did not close
The 404 Media read of Character.AI’s Chapter 11 filings lays out a months-long gap between revenue and compute, per 404 Media. Subscription revenue peaked at $2.8 million a month, sometimes dropping below $1 million; total revenue across ads, subscriptions, and in-app purchases peaked at $5.3 million a month. Compute costs ran $13.8 million a month in 2024 and $10 million a month in 2025; the filings show $56.1 million spent on compute from January to August 2026, plus $19.7 million in salaries and benefits, against monthly losses of $8 million to $9 million as of August 2026.
In an affidavit, CEO Colin Adams said “Character’s impressive technology is expensive to operate” and “at all times, Character has struggled with profitability and its operational expenditure has been high.” Disney’s reported $20 million offer to license some of the IP, per 404 Media, equals about two months of Character.AI’s 2025 compute spend.
Judge orders OpenAI to turn over Apple hiring records in trade secrets case
A federal judge granted Apple expedited access to OpenAI’s hiring documents in the trade secrets lawsuit Apple filed in July, with a five-day deadline for already-collected materials and 14 days for the rest, per TechCrunch. U.S. District Judge Edward Davila approved materials tied to OpenAI’s interviews with former Apple employees who applied to its consumer hardware and robotics groups, including interview checklists, guides, scripts, technical presentations, and records referencing Apple technology.
Apple’s lawsuit accuses OpenAI of a pattern of theft, including allegations that former employees used Apple laptops to download confidential material, and that former Apple employees at OpenAI coached others considering the move on what information to bring to their interviews, per TechCrunch. These are allegations, not findings. The judge also approved depositions of OpenAI employees Chang Liu and Alyssa Peng and chief hardware officer Tang Yew Tan, all former Apple employees, and access to private LINE messages between individuals named in the case. The judge denied some broader requests, including materials on metal finishes and power management, and left the door open for Apple to narrow and refile.
Google ships Gemini 4 Argon with a $2/$10 token price
Google launched Gemini 4 Argon on October 13 with introductory API pricing of $2 per million input tokens and $10 per million output tokens through November 30, then $4 and $20 afterward, per the Google AI blog. The model is available to developers and enterprises in AI Studio and the Gemini platform, to Google AI Ultra subscribers in Antigravity, and to Ultra and Pro subscribers in the Gemini app and AI Mode in Search. The post, under the name of Koray Kavukcuoglu, SVP at Google DeepMind, says Google is “giving developers, enterprises, scientists, and consumers access to our latest frontier model.”
The launch post claims Argon leads Gray Swan’s Indirect Prompt Injection benchmark at 0.7% attack success across 15 adaptive attempts and reduces jailbreak violations “nearly 2-fold” versus previous Google models; an external bio red team “could only unlock less than 1% of bio sub-tasks we test for.” Google is also offering 10,000 scientists six months of the highest Antigravity usage tier free of charge.
Liquid AI ships LFM2.5-24B-A2B for on-device agents
Liquid AI released LFM2.5-24B-A2B, a 24-billion-parameter mixture-of-experts reasoning model with 2 billion parameters active per token, designed to run private agent loops on consumer hardware, per the Hugging Face blog. Decoding runs at 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395, with throughput above 21,000 output tokens per second on a single NVIDIA B300 at high concurrency. The model has a 131K token context, up from 32K in its predecessor, and integrates with llama.cpp, MLX, vLLM, SGLang, and ONNX.
The model “thinks before it answers, calls tools, and works through multi-step tasks inside popular agent harnesses,” Liquid writes. The release is reasoning-only; LFM2.5-24B-A2B-Base ships alongside it for fine-tuners, and the team posted vendor-reported comparisons against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning on benchmarks including LiveClaw, BFCLv4, and τ³-Bench Banking.
Open vs closed: MIT Tech Review surveys the trade-offs
The October 13 MIT Technology Review explainer weighs the open-weights case for security, cost, and decentralization against filter-free deployment and the ease of stripping safety training, per MIT Technology Review. Elizabeth Seger of the Tony Blair Institute called avoiding “mass concentration of power” the strongest argument, while Hamish Hobbs at the Centre for Long-Term Resilience said that “if every individual had a loyal-to-them AI system that was safe [and secure], I think that would be a much less concerning world.”
The piece is skeptical on practice. Ghana NLP’s Paul Azunre said “the way it’s actually implemented in practice, [open source] tends to be a very extractive approach,” and David Gray Widder of UT Austin added that “openness is a floor, not a ceiling, for what we need to do to ensure that our technological futures benefit everyone.” The article notes Nvidia’s Jensen Huang has called for the US to release more open models, while Anthropic’s Dario Amodei has pushed for a more cautious approach.
AI labs are buying data from failed startups
A new resale market is routing code and workspace data from shuttered startups to AI labs and other model builders, per MIT Technology Review. SimpleClosure’s Asset Hub has handled more than 150 deals and expects to return more than $10 million to startups by year-end; 47% of those deals include code or data sales. Spirit Airlines’ data went to Google for $10 million; one founder, Keelan Miskell, told the outlet he netted roughly $25,000 for his code base, with a still-pending amount (likely around $15,000) for workspace data.
Sunset founder Brendan Mahony said “this operational exhaust data is like the wood shavings that everyone thought had no value, but it actually has a ton,” and Emory law professor Ifeoma Ajunwa said “employers should have express consent before selling workers’ data.” The piece notes that no company has yet released benchmarks comparing real-world training data with synthetic alternatives, and that data prices are likely to stabilize over the next year unless an AI bubble pops.
Ollama and Fly.io connect cloud models to sandboxed agent VMs
Ollama Cloud now runs inside Fly.io Sprites, full Linux virtual machines with their own persistent disks; the Ollama API key stays with a Fly.io gateway that adds it to each request, so the Sprite itself only knows a gateway URL and a connection ID, per the Ollama blog. “An agent in a Sprite can’t leak your key,” the post states; “new connectors block every Sprite by default” and “calls only work from inside a Sprite.” A preinstalled skill lets Claude Code, Cursor, Codex, and Gemini CLI use Ollama-hosted models without extra setup, and an Ollama Connector exposes kimi-k3, glm-5.3, and deepseek-v4.1-flash through the gateway.
A fuzzing example from Fly.io uses nemotron-3-nano:30b to generate malformed inputs, a Sprite to run them, and kimi-k3 to analyze crashes.
Quick Hits
- Pluto Security raises $42M for AI-tool monitoring. Out of stealth with an $8.5M seed and $34M Series A, the startup tracks employee AI tools, skills, extensions, MCP servers, and vibe-coded agents. CEO Shahar Bahat told TechCrunch a newly announced AI tool can show up inside companies’ environments within two hours of its release, and said “in general, I’ll say there was not one POC in which we entered and didn’t find malicious stuff.” TechCrunch.
- Amex publishes an “Intent Intelligence Framework” for agent purchases. The new API captures merchant, budget, and preference constraints before an AI agent checks out, and issues a tokenized credential rather than a raw card number. Amex VP Hannah Bozian said “our approach differs from the implementations of others in that it pulls intent earlier into the transaction flow.” TechCrunch.
- Cerebras CEO Andrew Feldman on AI risk rhetoric. Onstage at TechCrunch Disrupt, Feldman said “some of the leaders in the AI world maybe shouldn’t be allowed to talk in public” and added that “we have to be thoughtful about it. We have to regulate it reasonably.” TechCrunch.
- Feldman: a home AI assistant is not “very far away.” The Cerebras CEO sketched a model that runs “just like your cell phone, they’ll do some work on device and some work going back to the cloud” and that “will certainly have access to credit cards and be able to buy things for you.” TechCrunch.
- Industrial Next wants factory robots to switch tasks in minutes. Co-founders Lukas Pankau and Allen Pan, both Tesla alumni (Pankau later worked at Waymo), announced that BMW Group Technology Office will evaluate its dexterous manipulation for manufacturing. Pankau said “we can make robots do all these cool things like dance around, because there’s data for that, but we can’t make them assemble high-precision things.” TechCrunch.
- Federal court calls ALPRs mass surveillance; EFF loses in California. An Oklahoma federal district court ruled warrantless ALPR searches unconstitutional, calling them “a type of indiscriminate mass surveillance.” A California trial court separately dismissed EFF’s SIREN v. San Jose case; EFF said it will appeal. EFF.
Worth Watching
- Anthropic’s “Scion” source control system. Anthropic says it performs about 25 million CI code checkouts per day, and the main branch history is “almost 100 GB, and a compressed working copy is over 25 GB”; a team of seven built the Git replacement and completed the migration in five months, as Anthropic’s merged code changes rose from around 2,500 a day in May to over 30,000 a day by September. Simon Willison summarizes the engineering post.
- Science Corporation’s PRIMA chip in Stargardt patients. Max Hodak’s neural-interface company reports its first Stargardt disease implant, with a “next generation” implant expected in humans within the next year. Hodak told TechCrunch: “the long-term mission of the company is to understand how the brain creates the mind and to learn to build engineered brains.”
- Laplace Intelligence’s red-team pipeline for physical AI. The University of Michigan spinout says most of its revenue today comes from autonomous driving. CEO Howie Sun said the company is “essentially using simulation to predict the real world’s future” and is competing with Applied Intuition, which raised $1.2 billion at a $15 billion post-money valuation. TechCrunch.