Character.AI's bankruptcy, Disney's $20M discount bin

Good morning 👋 Character.AI just became the first AI-native bankruptcy of this cycle, and Disney picked through the wreckage for $20M. The headcount deal buried underneath the headline number is the one worth your attention.

In today's issue:

• 🔭 Character.AI's bankruptcy, Disney's $20M discount bin
• 🧠 RTX Spark laptops hit shelves, DeepSeek's V4.1-Flash gets faster and cheaper
• 🔬 Agent swarms that train coordination and skill together, plus why LLMs won't drop "not X, but Y"
• 🛠️ A weekly trap for the agent regression nobody notices until it's too late
• 🔗 Microsoft walks from a UK data center, OpenCode's breach, NVIDIA's EU pledge

🔭 THE ONE THING

🪦 Character.AI's bankruptcy, Disney's $20M discount bin

Character Technologies, the company behind Character.AI, filed Chapter 11 in Delaware on Oct 12 (case 26-11603). Disney's stalking-horse bid is $20M, and it's a license, not a buyout: a non-exclusive, perpetual grant to the training, safety, and serving code plus nine patents. The real price tag is in the headcount. Disney is hiring 66 of the roughly 86 staff who made it this far, with about $180M in comp over three years plus a $30M retention pool, per Bloomberg. Character.AI was still burning $8-9M a month after cost cuts, which says $20M buys Disney the safety plumbing for companion AI cheaper than building it in-house, and the auction isn't even final until late November. If you're building consumer AI on venture money, the unit economics that killed this company didn't go away just because a bigger brand bought the carcass.

🧠 MODELS & RELEASES

• 💻 NVIDIA RTX Spark laptops are on shelves now, with ASUS, Dell, HP, Lenovo, Microsoft and MSI all shipping units starting at $2,399. Blackwell RTX graphics and up to 128GB of unified memory on a laptop means real local inference, not a demo. NVIDIA
• 🔎 DeepSeek shipped V4.1-Flash-1017: better benchmarks, faster completions, shorter outputs, and cheaper to run than the last version. It also opened its Search and Fetch APIs at $0.75 per 1,000 calls each, free through Claude Code and Codex until Oct 31, when the meter starts. DeepSeek

🔬 RESEARCH HIGHLIGHTS

• Market-based post-training New work lets agents in a swarm bid for the right to act, instead of hard-coding who does what. Winning a bid earns the training credit, so coordination and individual skill learn together instead of in separate passes. Swarms trained this way scale to larger populations and lose less performance when agents drop out. The old split, train the agent then bolt on orchestration, looks like the wrong one. arXiv:2610.22322

• Librarian Agents A builder agent reorganizes a raw pile of documents into a hierarchical filesystem of synthetic files and links before anyone asks a question, then hands it to a separate agent to answer from. Trained with RL to directly optimize downstream accuracy, it beats retrieval and knowledge-graph baselines by 33 points on BrowseComp-Plus, with a 31-point jump on domains it never saw. The authors name the asymmetry outright: agents built on something like Claude Code are good at navigating a filesystem you hand them, bad at building one worth reusing later. arXiv:2610.22273

• The "not X, but Y" tic A diagnostic paper explains why LLMs lean on "not X, but Y" constructions 2.5 to 9x more than humans do, and it's not that models are getting it wrong. Only 1.3% of instances were actual errors; 52.8% were just unneeded. Reward models and LLM judges score the construction higher even when context doesn't call for it, which means RLHF is teaching the habit, not failing to catch it. Explains a lot about why AI prose converges on the same tics, and why editing this newsletter to dodge that reflex is real work. arXiv:2610.22249

🛠️ TRY THIS

Set a trap for the regression nobody notices until the product is gone

1. Pull 15 to 20 real queries from your agent's logs or support tickets, the ones users actually ask, not the synthetic test cases you wrote once and forgot.

2. Write a pass/fail rubric for each, grounded in what a correct answer looks like for your product, not a generic "sounds reasonable" check.

3. Replay that set against production on a schedule (a cron job, a GitHub Action, whatever you've already got) and diff the pass rate against last week's run.

4. Alert on any drop. AWS's AgentCore write-up cut mean time to detect from weeks to under 4 hours running this exact loop on a 30-minute cycle.

Prompt: Given this list of user queries and expected answers, score each production response as pass or fail against the expected answer, name the failure mode in one line, and flag anything that wasn't a failure mode last week.

Worth a look: zeron (open source, 3,200+ stars) gives you one control plane across Claude Code, Codex, Cursor and Devin. If you're already juggling more than one coding agent, it's the dashboard that tells you when one of them quietly goes sideways.

🔗 QUICK LINKS

• Microsoft walked away as anchor tenant at Nscale's Loughton, Essex site (23,000 Nvidia chips, still waiting on a grid connection that might not land until the 2030s), leaving the facility customerless right as Nscale chases a ~$3B IPO. Telegraph
• OpenCode says a breach through a legacy system hit 1.4% of Console users, exposing stored OpenAI, Anthropic, and Google keys. Some users are already staring at stray charges in the hundreds. X
• NVIDIA is pledging open learning resources and compute access to European research labs. No dollar figure attached, so judge it by what labs actually get, not the press release. NVIDIA
• ChronoScale Holdings posted Q1 FY27 revenue up 148% to $41.5M, riding a new 50MW Microsoft deployment built on Nvidia's GB300 NVL72 gear. SEC filing

See you tomorrow.

Pradeep Perugu