The UK invests in controlling AI risks; law firms need to measure learning

Other news

1. Moonshot cuts Kimi K3.1 Pro's offensive cyber capabilities before releasing its weights

Models · Reuters report on 24 October.

Moonshot says it removed cybersecurity data and tasks from training to reduce offensive uses. On its internal xdev-bench-hard test, the final model solves 81.0% of tasks, against 90.5% for a version without those measures. The test allows up to three attempts. These are company results, without independent verification of that comparison. [3]

Reuters reports a commitment to release the weights—the files for running and modifying the model—by 4 November. The reduction is measured against an unrestricted version from the same training pipeline, rather than its predecessor: the new model still outperforms its predecessor on this test. [3][4]

Why it matters. Releasing those files changes the control the provider retains. Buyers of a tool need to know which safeguards belong to the model and which depend on its hosting service. A lower score does not eliminate offensive capability either.

2. Rights holders in 274 compositions recorded by Elvis sue Anthropic

Copyright · Complaint filed on 23 October.

ABG EPE IP LLC and other rights holders have sued Anthropic PBC in the federal court for the Northern District of California. The complaint, case 5:26-cv-12473, alleges unlicensed copying to train Claude and reproduction or transformation of lyrics in its responses. It identifies 274 compositions; the subject is musical works, rather than an allegation of cloning the singer's voice. [5]

Paragraph 89 seeks $41.1 million in maximum statutory damages for wilful infringement: 274 works at $150,000 each. The complaint also includes claims involving circumvention of protection measures and removal of rights information. That figure neither captures the entire lawsuit nor represents an award. [5]

Why it matters. The dispute requires a distinction between material used in training and what the system subsequently delivers. The allegations are the plaintiffs'. The copy consulted reproduces the filed document; it does not establish a ruling on the merits.

3. A trial with lawyers separates AI-assisted performance from unaided judgement

Legal training · September study, revised in October; coverage on the 24th.

The Decoder highlighted a randomised trial involving 133 patent attorneys at eleven US firms. The original NBER study specifies that 91 completed the protocol. One group had access to a Google Labs assistant for three months. Assisted drafts improved in quality, especially among junior lawyers, according to assessors who did not know the participants' assignments. [6][7]

In a subsequent review of an application without AI, the advantage was concentrated among senior lawyers. Juniors showed no average improvement: there were more good and more poor results. The research is a working paper without peer review; Google funded the trial's direct costs. [6]

Why it matters. I would assess the delivered document and the lawyer's reasoning separately. Asking a junior to explain a correction and then complete an exercise without an assistant can test something that delivery speed conceals. This is a training proposal, rather than an experimental finding about the best teaching method.

4. Spanish travel businesses open another path to booking: talking to an AI

Business · Report published on 24 October; earlier initiatives.

Lighthouse attributes 1.4% of its measured hotel traffic to AI, twice the share six months earlier, according to El País. [8]

Hotusa's Eurostars uses Guido to guide travellers and book within a conversation, according to Eurostars information reported by Revista Gran Hotel. Amadeus has described its collaboration with Anthropic since 3 September. eDreams announced its Alexa+ agreement on 7 October. These predate the report; they were not launched yesterday. [9][10][11]

Why it matters. The customer relationship may begin in an assistant outside the hotel or agency. When reviewing that channel, I would start with who confirms the price, who completes the booking and who handles complaints. The traffic figure does not measure completed purchases.

LLM vs. NCM: a new AI architecture?

Research · arXiv: 9 October; blog: 12 October 2026. Two weeks ago.

Stanford and Together AI propose NCM: it generates learned codes from structure to detail, in parallel within each level. It remains a language model. MIT-licensed code and experimental weights are available. [19][20][21]

Against diffusion, it reports 21× on LM1B with better generative perplexity, 10× on OpenWebText with competitive quality and 31× on XSum matching ROUGE-L. In Sudoku, it beats the tested autoregressive and diffusion models with fewer evaluations. This does not establish that advantage over production LLMs. [19]

Why it matters. My hypothesis is hybrid: NCM for generation, LLMs for reasoning and tools, symbolic systems and graphs for verification. For Hermes, I would explore NCM as a specialised engine for bulk document generation and transformation, without replacing its main reasoner yet.

NCM has not demonstrated greater intelligence than today's LLMs. Its promise is questioning sequential generation. If it scales while retaining its efficiency, it could change inference economics and the design of future AI systems.