AI News: November 3, 2026

1. Marvell Agrees to Buy Lumentum in a $115 Billion All-Stock Deal for AI Data Center Connectivity

Marvell. Marvell will acquire optical component maker Lumentum at a fixed 4.5 Marvell shares per Lumentum share, about $1,132 per share and a 23.9% premium to the October 28 unaffected price, for an implied enterprise value of $115 billion; Marvell holders will own about 67% of the combined company. The companies pitch the deal as one supplier for AI networking, from Marvell’s custom silicon to Lumentum’s lasers, optical transceivers, and optical circuit switches, and project more than $1 billion in annual run-rate cost synergies, with closing expected in the first half of 2028 pending shareholder and regulatory approval. Nvidia invested $2 billion in each company in March, and TechCrunch notes the deal ranks among the largest tech acquisitions ever, ahead of Microsoft’s $68.7 billion purchase of Activision Blizzard. Source TechCrunch

2. MIT Researchers Have GPT-6 Astra Edit 0.016% of Qwen3.5-0.8B’s Weights Directly, Lifting MATH-500 From 7.9% to 48.1%

MIT CSAIL. Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab describe “weight-space programming,” in which a general-purpose LLM reads and writes a network’s weights directly in a Python kernel, with no gradients, optimizer, or GPU. In the headline experiment GPT-6 Astra traced Qwen3.5-0.8B’s reasoning loops to one direction written by the MLPs of layers 15 and 19 and subtracted it from 64 output columns in each (131,072 weights); with a budget of 50 forward passes, MATH-500 rose from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval from 21.0% to 62.3%, while the share of MATH-500 answers hitting the 32,768-token limit fell from 93.0% to 33.1%. Other experiments had the model identify which game or language a small network was trained on, repair a control policy by changing 3.5% of its weights, and write a MinAtar Breakout network from scratch that outscored PPO baselines. The authors call the results “very early”: the same approach barely moved Gemma 3 270M (6.8% to 7.7%), which rarely loops, and they say editing “could not add an ability the model lacked”; full logs are on GitHub and the edited model is on Hugging Face. Source

3. The New York Times and Three Other Papers Sue Brave, Alleging It Sells Scraped Content to AI Companies

Brave. The New York Times sued Brave Software in federal court in Manhattan on Monday, and the Daily News, Chicago Tribune, and Orlando Sentinel filed a similar suit there the same day. The complaints allege that Brave built an index of more than 30 billion pages with an unidentified crawler that publishers cannot block, sells paid access to “real-time online content from the Brave Index” that AI companies including Anthropic, Perplexity, Mistral, xAI, and Cohere use for training or to answer user queries, and promotes a “Paywall blocker” in its browser. The Times calls Brave a “supplier, dealer, and distributor of stolen copyrighted content” and seeks an injunction and damages; Brave did not immediately comment. The suits extend publisher litigation beyond model makers (the Times has also sued OpenAI and Perplexity) to a search-index supplier that AI products use for grounding. Source The Verge

4. Salesforce Agrees to Buy Voice AI Company Cresta, Reportedly for $2.7 Billion

Salesforce. Salesforce signed a definitive agreement to acquire Cresta, whose voice agents handle complex customer calls end to end while its other tools coach human representatives in real time and analyze every interaction; customers include Marriott, Comcast, and Alaska Airlines. Salesforce did not disclose a price, but The Information reports $2.7 billion. Cresta’s conversation intelligence will connect to Service Cloud and Salesforce customer data so calls can be tied to outcomes such as resolution and retention, while Cresta keeps supporting the third-party contact center and telephony platforms customers already run. The deal is expected to close in the first quarter of Salesforce’s fiscal 2028 (February to April 2027) and follows Salesforce’s October 20 agreement to buy Braze. Source The Information

5. NorthCUA Releases Apache-Licensed Computer-Use Models From 9B to 397B, Reporting 88.6% on OSWorld-Verified

Shanghai Qiji Zhifeng. The NorthCUA team published a 121-page technical report and Apache 2.0 weights for three computer-use agent models built on Qwen3.5, Nano (9B), Mini (35B), and Pro (397B), along with the code for its NorthCUA Harness. Training combines an “experience flywheel” that generates executable tasks and keeps only verified trajectories, skill evolution for tasks the model cannot yet solve, mid-training and long-horizon SFT, and joint desktop and web RL. The report lists Pro at 88.64% on OSWorld-Verified (361 tasks, 100-step limit), 73.89% on WebArena-Verified, and 93.30% on ScreenSpot-Pro, with Nano at 78.90% and 67.16% on the first two; the team says Pro beats Qwen3.8-Max and Kimi-K3 on these tests, and an Ultrafast variant finishes matched WebArena tasks 2.3 times faster. Its OSWorld-G grounding score is a best of three runs, and the report acknowledges at least one grounding example that used a control seen in training. Source

6. Oracle Files to Sell $10 Billion in Stock and $10 Billion in Bonds to Fund Its AI Buildout

Oracle. Oracle filed with the SEC to issue $10 billion of common stock and $10 billion of senior notes, the financing it previewed at last week’s investor day, where it said this $20 billion would be its only additional capital-markets funding for the fiscal 2027 to 2030 investment cycle. The rest of the buildout is to come from vendor financing, customer-provided hardware, and customer prepayments, with positive free cash flow expected in the second half of 2028 and revenue of at least $225 billion targeted for fiscal 2030. Shares fell 8.8% on Monday and are down 33% this year; Oracle carried $125.3 billion of debt at the end of August, S&P cut it to BBB-, one notch above junk, in July, and its five-year credit default swap spread has risen 265% in a year to 292. Source

7. Investors Reportedly Push Anthropic Toward a $1.5 Trillion IPO Valuation as a Critic Urges the SEC to Hold the Listing

Anthropic IPO. Institutional investors have told Anthropic its IPO valuation should be around $1.5 trillion to $1.6 trillion, down from the roughly $2 trillion previously expected, citing the tech stock selloff, The Wall Street Journal reports. In Tech Policy Press, Matt Scherer of the Open Markets Institute argues that the SEC should use its authority to halt offerings whose registration statements are “on its face incomplete or inaccurate in any material respect” until Anthropic files two consecutive quarters of audited financials and fully discloses its roughly $895 billion in spending commitments, at least $582 billion of which he counts as non-cancelable, and he notes that the 2026 results in the filing are unaudited. WSJ Techmeme Tech Policy Press

8. Goldman Sachs Cuts Traders and Bankers Off From Claude Code and GitHub’s AI Coding Tools

Goldman Sachs. Goldman has limited advanced AI coding tools, including Anthropic’s Claude Code and Microsoft’s GitHub offering, to engineers, removing access for traders and bankers who had used them for months to monitor trading flows and create content, Bloomberg reports via The Decoder. A spokesperson cited “productivity and risk management,” said the experimentation phase is over, and said access now goes to the people best suited to use the tools long term; the restriction follows internal guidelines issued in October. The move reverses the approach CIO Marco Argenti promoted in July, in which non-technical staff “vibe coded” prototypes that engineers then scaled for production, and staff are reportedly frustrated by the rollback. Source

9. Wired Traces 400-Plus Disguised Apps That Switch to Deepfake Generators After Passing App Store Review

Investigation. Wired, with researchers from the nonprofit AI Forensics and security firm Infoblox, found a network of more than 400 apps and domains, linked to Chinese-language developers and a Hong Kong entity called KeLing Technology Limited, that present as pet-portrait, language-learning, or workout apps but query a server at install time (sending referrer, language, and an identifier) to decide whether to load a hidden web interface for generating sexually explicit and violent deepfakes from uploaded photos. Users often arrived through tens of thousands of explicit ads on Facebook and Instagram and paid through Apple and Google Pay, and Infoblox found almost 500 linked domains sharing code and encryption keys. AI Forensics counted 27 such apps live on Google Play in September with more than 350,000 downloads, many rated for all ages; Google had removed most before Wired asked, and Apple removed 19 apps, saying developers swapped content after approval and that it has suspended tens of thousands of “bait-and-switch” apps in the past year. Source

10. Palantir’s US Commercial Revenue Rises 148% and Overtakes Its US Government Business

Palantir. Palantir reported third-quarter revenue of $2.31 billion, nearly double a year earlier and above the $2.18 billion consensus, with net income of $1.13 billion. US commercial revenue grew 148% to $984 million and for the first time exceeded US government revenue ($953 million, up 96%), which CNBC attributes to businesses adopting more of Palantir’s AI tools. The company raised full-year revenue guidance to about $8.53 billion and its US commercial outlook to more than $3.56 billion, guided to about $2.65 billion for the fourth quarter, and its shares slipped about 3% after hours. Source

11. Tutor Intelligence Raises a $75 Million Series B Led by a16z for Robots Trained on Teleoperated Work

Tutor Intelligence. The MIT-founded robotics company announced a $75 million Series B led by Andreessen Horowitz, bringing total funding to $117 million according to Forbes, after a $34 million Series A in December 2025. Tutor puts robots into production before they are fully autonomous: human “tutors” teleoperate them to finish the work, and that data trains the robot foundation model shared by the single-arm Cassie (case picking, palletizing) and the two-arm Sonny (e-commerce fulfillment). a16z says Tutor charges hourly, can install in as little as 8 hours, plans more than 1,000 robots under contract by the end of 2027, and that nearly a third of customers who received robots in the past year were existing customers adding more. a16z Forbes

12. 42AI Emerges With $62.5 Million for a “Behavioral World Model” Sold to Brokerages

42AI. Tel Aviv-based 42AI came out of stealth with a $62.5 million seed round raised over three years from its founders and Infinity Group, which chairman and co-founder Amir Gal-Or founded, rather than mainly from outside investors. Its system turns behavioral observations into trait representations that feed a transformer-based world model, which outputs probabilities for a person’s decisions, while a separate language model generates explanations; training data spans psychological assessments, government and biobank datasets covering more than 2 million people, and customer data on about 30 million people. About 10 large brokerages use it to anticipate how traders act in volatile markets for liquidity, margin, and hedging decisions; the company claims 98% accuracy (measured as flows falling within a predicted range) and an F1 of 0.8 to 0.85 on trade direction and volume, and the founders did not explain what would stop brokerages from using predictions against their own customers. Source

13. Western Startups Are Stripping Censorship From Qwen and DeepSeek Models for US Companies

Rest of World. A Rest of World report details how startups post-edit Chinese open-weight models for Western customers: Israel’s Hirundo uses machine unlearning and weight editing to remove behaviors such as Qwen’s avoidance of Tiananmen, the 2019 Hong Kong protests, and Xinjiang camps, while Boston’s Lazarus AI realigned Qwen using topics from leaked censorship guidelines and answers generated by Gemini Flash, and says its “ReAligned” models run at about a dozen US companies. The piece cites CrowdStrike research finding DeepSeek R1 up to 50% more likely to produce code with severe vulnerabilities for politically sensitive prompts, notes that Airbnb, Uber, and Harvey have built proprietary models on Chinese bases and that OpenEuroLLM builds on Qwen, and quotes Hirundo’s co-founder saying many abliteration tools fail to fully remove censorship. The work comes after the administration barred civilian federal agencies from DeepSeek products in October. Source

14. Tsinghua and Berkeley’s uni-robot-agent Gets an Off-the-Shelf Qwen Model Past a Fine-Tuned VLA on Real Robots

Research. A paper from Tsinghua University and UC Berkeley (authors include Pieter Abbeel and Ning Ding) introduces uni-robot-agent, a five-layer harness that lets a general model control a robot through shared tools for multi-view triangulation, pose-targeted moves that report numerical errors, background jobs so the model can reason while the arm moves, and context compaction, using RGB images, calibration, and proprioception rather than simulator depth, CAD assets, or verifier feedback. Out-of-the-box Qwen3.8-Flash-Next scored 31.39 (25.00% success) on 42 RoboDojo simulation tasks versus 28.97 for GPT-6 Astra with RoboProbe, and 49.50 on 19 real-world tasks against 47.75 for a fine-tuned π0.5 and 35.50 for zero-shot Dream-1, while GPT-6 Astra in the same harness reached 75.50; the harness runs used a depth camera that the RGB-only baselines did not. The authors trace failures to four general skills (visual understanding, spatial grounding, geometric coding, and embodied reasoning), and the paper gives no code link. Source

15. SUSE Cuts About 10% of Staff, Saying “Automation and AI” Are Reshaping Software Companies

SUSE. The EQT-owned Linux vendor is laying off about 10% of its roughly 1,900 employees; CEO Dirk-Peter van Leeuwen told staff that “automation and AI are fundamentally shifting how software companies operate” and that “the requirement for some existing roles has greatly diminished.” A September voluntary separation offer to long-serving staff drew too few takers, affected employees will be told by November 4, and SUSE will present its fiscal 2027 strategy internally on November 16. Source

16. Garry Tan Says He Spends Up to $200,000 a Year on Tokens, and YC Runs an Internal Agent Called Stray Light

Y Combinator. On Walter Isaacson’s podcast The Builders, YC CEO Garry Tan said he spends “about maybe $100[,000] to $200,000 a year on tokens” on agents that handle research, coding, and personal administration with access to his email, Dropbox, Apple Notes, and 20 years of therapy notes, calling it “a crazy privilege that not everyone can do.” He said YC is “rebuilding the way we operate” around an internal agent called Stray Light that has context on every founder YC has backed, every application, and everyone who has emailed YC, and that personal agents will not go mainstream until costs fall. Tan, an early OpenClaw adopter, has published a set of 23 Claude Code skills called gstack. Source