Anthropic's amended S-1 says AI-accelerated progress may be close

Anthropic released a revised S-1/A on Monday 2 November, with updated financials and new risk factors. The company says internal estimates for future models suggest it could soon reach the point its Responsible Scaling Policy defines as "dramatically accelerated progress attributable to AI." So far it hasn't designated any current model as past that point, and it's still assessing the data.

The filing also says Anthropic has, in the past, only spotted significant speed-ups in its own development after they'd already happened, and it could miss one again. Even if fully autonomous model building becomes technically possible, the company says it doesn't expect Claude to safely design and build later frontier generations without "meaningful human supervision or understanding." It's weighing RSP updates to keep people in control. Skeptics point out that an S-1 is a marketing document as well as a disclosure.

For builders: Plan for frequent Claude releases and possible policy changes. Any RSP revision could change which capabilities ship and when, so it's worth watching Anthropic's policy pages.

Marvell agrees an all-stock deal for Lumentum

The merger agreement is dated 1 November, and Marvell announced it on 2 November. Lumentum holders get 4.5 Marvell shares per share. That implies an enterprise value of $115 billion, about $1,132 per Lumentum share, or a premium of roughly 23.9% over unaffected closing prices on 28 October. On a fully diluted basis, current Marvell holders would own about 67% of the combined company and Lumentum holders about 33%.

The SEC filing sets an outside date of 1 November 2027, which can be pushed to dates in 2028 if regulators need longer. Termination fees run up to $5.8 billion payable by Marvell, plus a separate $3.2 billion antitrust fee. The companies' case for the deal is that AI data centres are moving from copper to optical links as bandwidth needs grow.

For builders: Optical interconnect is becoming a bottleneck for large clusters. One bigger supplier could affect networking costs and roadmaps at the clouds we rent from.

Newspapers sue Brave over search index and AI answers

Two suits were filed in federal court in Manhattan on 2 November. The New York Times brought one. The publishers of the New York Daily News, Chicago Tribune and Orlando Sentinel brought the other a few hours later. Both say Brave copies journalism into its index of more than 30 billion pages and sells access to AI companies through its Search API. The complaints list Anthropic, Perplexity, Mistral, xAI and Cohere as customers, though they aren't defendants.

The papers also object to Ask Brave, which they say reproduces passages word for word, and to a Brave setting promoted for getting past paywalls. Their claims include copyright infringement and DMCA circumvention. A court hasn't ruled on any of them.

For builders: If your grounding comes from a third-party search API, ask the vendor how it collects content and whether it indemnifies you. Provenance questions are reaching API customers now.

OpenAI reframes its policy pitch and backs mandatory testing

Semafor reported on 3 November that OpenAI now calls data centres "compute campuses" and is asking local communities for input on new sites. Chris Lehane, its top lobbyist, said OpenAI will push for a mandatory AI testing bill that it thinks could get 60 Senate votes. It has also recently endorsed two frontier AI safety bills in California.

Critics weren't convinced. Brad Carson of Americans for Responsible Innovation said OpenAI was "setting themselves up for failure" if it thinks "Orwellian name changes" will shift the debate.

For builders: A major lab now supports federal pre-deployment testing. If you ship capable models, keep evaluation records that would hold up under a formal review.

UK cyber chief puts a 2028 date on AI attacks against critical infrastructure

The NCSC released its Annual Review 2026 on 3 November, covering 1 September 2025 to 31 August 2026. In a speech that day, chief executive Richard Horne said the UK should expect AI-enabled attacks on critical infrastructure by 2028, including AI-powered malware in ransomware. He said AI lets attackers find and exploit "vulnerabilities that have been tolerated for far too long" at scale. The review expects AI mostly to speed up and scale tactics attackers already use. The NCSC estimates that malicious cyber activity costs UK organisations almost $20 billion a year.

For builders: Attackers and defenders have the same tools. Point AI-assisted scanning at your older code and dependencies, and shorten your patch cycle.

Meituan launches LongCat-2.5 for long agent tasks

Meituan released LongCat-2.5 on 2 November. It's a mixture-of-experts model with 1.6 trillion parameters, about 48 billion of them active per token. It handles text and images in one token framework, and Meituan trained it on long task sequences, including failures from real agent runs turned into repeatable tasks. It's live in Meituan's chat service and API, with GitHub and Hugging Face releases planned.

For builders: Wait for the published weights and licence before you plan to self-host it. Until then, the API is the only way to test it.

OpenAI hires former Army Secretary Dan Driscoll

OpenAI announced on 2 November that Dan Driscoll, who left the Army secretary job earlier in 2026, is joining in a new role: senior counselor for national security modernization. He'll advise leadership on AI for national security and cybersecurity, help US allies adopt AI, and build on OpenAI's Daybreak cyber-resilience programme.

For builders: OpenAI is putting more effort into government and allied work. Expect more sovereign deployments and the compliance features that come with them.

Sources

Talk about it with people building with AI

Our Auckland meetups are open to everyone. Members get a ForgeBox and the members' projects area for NZ$599 a year.

Become a member

Get the daily AI news by RSS