Disney hires most of Character.AI’s workforce as the company breaks up in bankruptcy

Character.AI filed for Chapter 11 earlier this week (company announcement), and Disney has now agreed to hire at least 60 of its employees, a majority of the staff at an estimated $60 million in annual compensation, plus $20 million for a non-exclusive perpetual license to some of its technology; the bridge is former Character.AI CEO Karandeep Anand, who recently became Disney’s first CTO. Disney is the stalking-horse bidder: its offer sets the court-supervised floor, others can still outbid it, and the company says that if no qualified buyer takes the platform itself, it will be wound down. Disney wanted people and a tech license, not the product: a companion-AI leader with a large, engaged user base could not carry what its own filing calls “significant legal, regulatory, and operational headwinds”. My read, and it rests on this one case: that is a warning sign for the rest of the companion category.

Anthropic commits 5% of its shares to a public benefit fund, with a no-lobbying rule

Anthropic is moving 5% of its equity into a wholly-owned LLC called the Public Benefit Fund, and the board can add 2% more each time the company’s valuation rises 50% above a set baseline; the plan still needs stockholder approval. The detail worth remembering: none of this equity can be used for lobbying or political giving, in the US or abroad, at a moment when AI-industry political spending is climbing fast. Counting founder pledges and employee donation matching, Anthropic says roughly 20% of its shares will go to public-benefit purposes, including work on AI-driven labor market disruption on top of its earlier $200 million economic-futures commitment.

DeepSeek ships V4.1-Flash-1017: faster, with shorter outputs

DeepSeek released V4.1-Flash-1017 on Saturday as a minor update to September’s V4.1-Flash, claiming gains across reasoning, coding, agentic and visual benchmarks, with weights on Hugging Face under an MIT license. Its model card puts Humanity’s Last Exam at 42.6, up from September’s 36.8. The more telling claim is “shorter outputs”: token volume is where agent costs live, and DeepSeek keeps selling efficiency alongside capability, consistent with its low-cost playbook.

Research radar

Hybrid Reasoning Agents: think in latent space, act through text

The paper proposes agents that reason in latent space, carrying thought as continuous hidden states instead of writing it out token by token, and switch to text only to call tools or answer; a lightweight HybridLink module handles the switching, trained with dedicated SFT and RL stages. Reported results: about 10.4% average improvement across 11 benchmarks, 35%–55% fewer tokens, and roughly 2x faster inference. Agent-efficiency researchers should read it, and safety researchers even more so: readable chains of thought are one of the few handles we have for monitoring model reasoning, and that monitorability is fragile to exactly this kind of architectural choice. The authors argue their continuous hidden states remain directly inspectable; I’m not convinced that a vector you can probe is the same thing as reasoning you can read.

One line to remember: Character.AI had the users and still went under, by its own account because of legal, regulatory, and operational weight. Whether that bill comes due for the rest of the companion category is the thing to watch.