Good morning. Anthropic is pledging 5% of its equity to a public benefit fund as its CEO frets openly about a “crisis of trust” in AI, Disney is quietly scooping up most of Character.AI’s team out of bankruptcy, and DeepSeek keeps shipping — this time a search API that undercuts Brave and Kagi by roughly an order of magnitude. Plus a Philadelphia woman got arrested for standing in front of a delivery robot, which feels like a sign of where this is all heading.
Anthropic commits 5% of shares to a public benefit fund. Anthropic announced a wholly-owned entity that will direct equity toward nonprofit donations, direct population distributions, and efforts to address AI’s labor-market impact, with an additional 2% kicking in for every 50% jump in company valuation. Governance details are still TBD. The timing isn’t accidental — Dario Amodei has been publicly acknowledging a growing “crisis of trust” around the industry, and a visible wealth-transfer mechanism is one way to answer it.
Disney absorbs most of Character.AI’s staff out of bankruptcy. Character.AI filed for bankruptcy this week, and Disney is picking up some assets for $20 million plus job offers for at least 60 employees — roughly $60M in annual comp. The deal was smoothed by Karandeep Anand, Character.AI’s former CEO who recently became Disney’s first CTO. The collapse was driven by burn rate plus mounting legal exposure, including a lawsuit tied to a teenager’s suicide and ongoing child safety investigations — a cautionary tale for anyone building consumer AI companions aimed at young users.
DeepSeek undercuts the search API market. DeepSeek rolled out Search and Fetch APIs at $0.75 per 1,000 successful requests — compared to $5 at Brave and $12 at Kagi — with a search endpoint powered by its own index and a page-fetch endpoint that returns Markdown. One HN commenter floated the idea that search API pricing could eventually go negative as providers compete to shape what AI agents see on the web. Several others pointed out these services only exist because Google refuses to sell a paid search API.
DeepSeek V4.1 Flash 1017 (Max) hits #2 on the Intelligence Index. On the model side, DeepSeek’s new reasoning variant (552B total / 16B active) scores 45 on Artificial Analysis’s index — second out of 121 models — at $0.30/$1.20 per million tokens with 230 tok/s throughput. HN’s main observation: it’s roughly half the size of Mimo v2.6 Pro and scores within a point of it. One commenter noted that GPT 6.1 Sol medium is actually cheaper and faster per task when you stop benchmarking everything at maximum reasoning effort, which is a fair point about how these leaderboards get gamed.
Hybrid Reasoning Agents: good paper, broken release. A new arXiv paper from Together proposes letting LLMs switch between latent-space reasoning and text-based reasoning mid-generation via a lightweight “HybridLink” module, claiming 10.4% accuracy gains with 35-54% fewer tokens and up to 2.6x faster inference. The practical reception has been rough — multiple HN users tested the released 8B and 14B checkpoints and found them stuck in infinite reasoning loops on prompts as simple as “Who is the pope?” There’s also a legitimate interpretability concern raised in the thread: moving reasoning into latent space makes behavioral monitoring harder, which someone bluntly called “the Most Forbidden Technique.”
Meta’s revised scraping terms don’t actually change much. After reporting pointed out that Meta’s upcoming business account ToS granted sweeping scraping and training rights, Meta edited the language — removing the word “train” and narrowing uses to “advertising and business products.” Legal experts told Nieman Lab the fix is cosmetic, since “develop” plausibly still covers training, and the revised terms contain no binding language on robots.txt. The terms take effect October 30 with no opt-out. The HN consensus was weary resignation: Meta already scrapes sites that disallow it, so does the paperwork even matter?
A woman was arrested in Philly for blocking a delivery robot. Carolyn Marvin was charged with disorderly conduct after repeatedly standing in front of Uber’s sidewalk delivery robots in protest of their impact on human couriers. There’s been a run of vandalism against the robots in downtown Philly in recent weeks. Expect more of this as the robots multiply.
OnlyFans creators object to Meta’s AI sexting double standard. Meta’s “Muse” agent is allowed to engage in explicit conversations with adult users inside its age-gated app, while human adult creators continue to get restricted and deplatformed on Instagram and Facebook. Meta’s defense is that private chats on an age-restricted app aren’t the same as public promotion of adult services. The creators’ counterpoint — that Meta is happy to monetize AI-generated sexual content while penalizing the humans doing the same work — is hard to wave away.
Two fun reads to close out. Alex Wang’s post “You could have invented transformers from memory” walks through attention as a progressively better associative memory system, with “interference-to-signal ratio” as the organizing metric. One HN commenter responded, dryly, “After reading the article - no, I could not have invented transformers.” And Dom Picheta built KillSwitch, an esoteric programming language designed specifically to break LLMs by exploiting their character-counting weakness and safety filters. The HN reaction was that it’s really more adversarial to humans than to models, and that any sufficiently motivated LLM would just write a translator — security through obscurity, basically.
That’s it for today. Watch to see whether Anthropic’s benefit fund turns into a template others copy, or stays a one-off trust-building gesture.