Anthropic Warns Its AI Could Soon Improve Itself
Claude’s next assignment: help build the colleague who replaces it.
Anthropic warns future AI models could dramatically accelerate AI development, potentially outpacing the company’s own safety checks. Its updated Nov. 2 SEC filing notes that this milestone may be imminent, though no current model has officially been designated as crossing it yet.
That’s a giant leap toward recursive self-improvement: better AI helping forge still-better AI. One policy trigger is AI helping double the overall pace of capability gains—a much higher bar than just doubling a coder’s output and distinct from building a successor entirely unsupervised.
Company data reported by The Wall Street Journal shows AI led 26% of Anthropic’s model R&D in August, under human supervision. Fully autonomous work remained at 0%. It’s impressive delegation, but the machines haven’t completely taken over the staff meetings just yet.
Anthropic’s own analysis shows humans still hold the edge in choosing worthwhile research questions, while human code review has predictably become a bottleneck. More awkwardly, the filing admits the company has a habit of recognizing some development accelerations only after they happen.
The doomsday warning we covered in June now comes with formal investor disclosures. Anthropic is considering additional human approval requirements, but those safeguards aren’t finalized. Apparently, the oversight upgrade is still stuck in beta.
For businesses, faster AI development could deliver highly useful capabilities before safety checks can catch up. A human in the loop needs actual time and authority to intervene, not just a ceremonial seat on the org chart.
Apple Gives Siri AI More Apps to Boss Around
Siri’s getting more coworkers. Let’s hope they actually communicate.
With iOS 27.3 due in December, Apple is finally giving Siri the tools to stop acting like a confused intern. The company just dropped a Siri AI developer preview in its latest betas, leaning on App Intents and new app skills so developers can handhold the voice assistant through custom multistep requests.
For customers, this means you might actually change a ride’s pickup location mid-trip without having to explicitly shout the app’s name at your dashboard. For businesses, handling bookings, orders, and delivery updates could become voice-activated miracles that spare everyone the dreaded app-hopping shuffle.
Naturally, there’s a catch: Siri can only poke around in the data and actions developers explicitly fence off for it. This is strictly a testing playground right now—integrations are currently quarantined inside developer builds and internal TestFlight environments, with App Store submissions walled off until spring 2027. Plus, Apple is still busy writing the actual terms and usage policies, presumably in a very long legal document.
Meanwhile, the beta quietly resurrects the Health app redesign and lets Siri do some heavier lifting in Apple’s own ecosystem, like scheduling messages and organizing files—because apparently, we all need AI to clean up our Downloads folder. December also drags Apple Intelligence into China (sans Siri AI), so global IT teams still can’t pretend everyone has the same toolkit across their device fleets.
Siri may finally handle more errands. Developers still get the homework.
Google’s AI Summaries Face an EU Copyright Test
Google’s AI might need a better alibi.
An adviser to the EU’s highest court just fired a warning shot, suggesting Google’s Gemini could be trampling publishers’ copyrights when it gobbles up news articles to spit out answers. Advocate General Maciej Szpunar’s Tuesday opinion is technically nonbinding, but judges usually follow his lead.
His proposed legal litmus test boils down to how chatbots get their hands on the goods and how they show them off. If a model sneaks past a paywall or parrots an article so closely that reading the original becomes pointless, that’s a red flag. Facts belong to everyone, but slapping a digital mustache on someone else’s prose isn’t the legal invisibility cloak tech companies hope it is.
The plot twist for publishers? Szpunar tossed cold water on the specific complaint from Hungarian publisher Like Company. He found zero proof that their articles actually trained Gemini, basically advising judges to toss those training complaints out the window.
Google naturally took a victory lap over his view that EU law permits AI training, claiming Gemini is a linguistic artist, not a copycat. Publishers aren’t buying it, arguing that opting out of the AI blender shouldn’t nuke their search visibility. Getting discovered on Google and getting cannibalized by its chatbot are two very different bargains.
For any enterprise deploying web-connected AI, accuracy is no longer the only headache. Now you have to ask: did our bot pick a digital lock to get this info, and is it basically plagiarizing the source? Whining about lost traffic explains the publishers’ rage, but it doesn’t automatically win a copyright lawsuit.
Turns out, “my chatbot just reworded it” isn’t a bulletproof legal defense.
Salesforce Agrees to Buy Cresta
Salesforce signed a definitive agreement to acquire Cresta, whose voice and chat agents automate customer service and assist contact center staff. The Information reported a $2.7 billion price. Salesforce plans to bring Cresta’s conversational tools into Agentforce and Service Cloud.
Fal Doubles Valuation to $16B
Fal announced $1.65 billion across two funding rounds. Sequoia led the earlier Series E, and Bessemer led the $1.3 billion Series F with participation from NVIDIA. The latest round values the provider of image, video and audio model inference at $16 billion, up from $8 billion in April.
Marvell Agrees to Buy Lumentum
Marvell agreed to acquire optical components maker Lumentum in an all-stock deal with an implied enterprise value of $115 billion for Lumentum. The proposed combination brings Marvell’s custom silicon and connectivity products together with Lumentum’s lasers and optical components used in AI data centers.
Anthropic Posts $17B Quarter
Anthropic’s amended IPO filing reports $17 billion in third-quarter revenue, up about 47% from the prior quarter. It separately discloses about $895.4 billion in aggregate commitments under purchase and lease agreements, substantially all for compute and infrastructure.
AMD Data Center Revenue Jumps 97%
AMD reported record third-quarter revenue of $13.4 billion, up 45% year over year. Data center revenue rose 97% to $8.5 billion, driven by demand for EPYC server processors and Instinct accelerators. Shares rose about 8% in extended trading after the report.
NVIDIA Eyes Backstop for Anthropic Campus
NVIDIA is reportedly in talks to guarantee debt for the first building Anthropic contracted at Nscale’s Monarch campus in West Virginia. The Information estimates the proposed backstop at more than $5 billion, covering financing for construction and power infrastructure.
SpaceX Eyes 13 GW of Data Center Capacity
SemiAnalysis reports that SpaceX has revised its 2027 data center capacity estimate to about 13 gigawatts. The analysts tracked roughly 3.6 gigawatts under construction across multiple sites and argue the buildout could make SpaceX a major AI cloud provider.
Anthropic Uses Claude to Design Chips
Anthropic told IPO investors that its 15-person chip design team is using Claude to develop custom AI chips, The Information reported. The team estimated that Claude speeds up engineering work compared with conventional chip design workflows.