Nikkei Asia reports that the off-balance-sheet debt of US tech giants tied to AI obligations has ballooned to about $3.2 trillion, nearly doubling in six months from the roughly $1.65 trillion the paper estimated in July. The figure covers Alphabet, Microsoft, Amazon, Meta and Oracle, and includes purchase commitments for chips and data centers plus leases that have been signed but have not yet started, meaning the obligations sit outside reported balance-sheet liabilities.

• Nikkei estimates US tech giants' off-balance-sheet AI obligations at about $3.2 trillion, almost double its July figure of $1.65 trillion, citing purchase commitments and leases that have not yet commenced.
• The obligations are roughly triple the leases and long-term borrowings the same companies report on their balance sheets, and the gap is widening faster than traditional capex, according to a Wall Street Journal analysis of the same structures.
• Much of the exposure is concentrated in long-duration supplier contracts and signed leases that land on the balance sheet all at once when a data center goes live, creating a deferred-depreciation and refinancing cliff.

Read the full story on AiLookout.news

xAI is powering part of its Colossus 2 data center near Memphis with eight converted jet-engine turbines installed at a site in Mississippi, according to reporting by the New York Times that TechCrunch and Tom's Hardware summarised on 3 November. The units are rebuilt by Missouri-based ProEnergy from GE engines normally fitted to Boeing aircraft. Mississippi permits allow a 25% higher limit for volatile organic compounds than the standard, and the Southern Environmental Law Center has challenged the approval.

• Converted aeroderivative turbines give AI builders power in roughly 24 months, against about seven years for a 100MW grid connection in Northern Virginia, according to figures cited by TechCrunch.
• xAI's eight units in Mississippi each produce up to 50MW, and the state permits allow a 25% higher limit for volatile organic compounds, which environmental groups are contesting.
• The trade-off is efficiency and emissions: converted jet engines burn more fuel and emit more per megawatt than turbines built for stationary power plants.

Read the full story on AiLookout.news

Palantir reported third-quarter 2026 revenue of $2.313 billion, up 96% year over year, with US commercial revenue of $984 million, up 148%, surpassing US government revenue for the first time. Adjusted earnings per share were $0.44 against consensus of about $0.40. The company raised full-year revenue guidance to between $8.530 billion and $8.534 billion.

• US commercial revenue reached $984 million, a 148% year-over-year increase, and passed US government revenue for the first time in the company's history.
• Total revenue of $2.313 billion beat consensus, but guidance implying 15% sequential growth in the fourth quarter, down from 20% a year earlier, weighed on the shares.
• Palantir raised full-year revenue guidance to $8.530-$8.534 billion and lifted its US commercial target above $3.559 billion.

Read the full story on AiLookout.news

Apple released a Developer Preview in the iOS 27.3, iPadOS 27.3, macOS 27.3, watchOS 27.3 and visionOS 27.3 betas that lets any developer integrate an app with Siri AI using App Intents. Apps can expose custom actions and content, chain several actions into a single skill, and surface entities from the Spotlight semantic index. Apple says the preview is early access before these experiences ship to customers in a few months, with App Store submission opening in the spring.

• Any app can now integrate with Siri AI through App Intents, replacing the smaller set of supported schemas that shipped with the first version.
• The preview adds skills that combine several app actions into one request, plus personal context, onscreen awareness and natural language handling.
• Apple says App Store submission for these experiences opens in the spring, Siri AI remains unavailable in the EU on iPhone, iPad and Apple Watch, and several terms are still unpublished.

Read the full story on AiLookout.news

Anthropic's Frontier Red Team released FlashProve, an open-source prototype for what the lab calls provable inference. A model provider commits to a set of weights and then attaches a zero-knowledge proof to each response showing it came from those weights. The proof hides the weights but not the architecture, and verifying it is cheap enough to run on a local machine. Anthropic published the code, a technical report, a protocol specification and Lean proofs under Apache 2.0, and said the work is ongoing research with no deployment planned at Anthropic or anywhere else.

• FlashProve lets a user check that a response came from the model they asked for, even when the provider is the lab itself, using a zkSNARK proof attached to the output.
• On a 400B-parameter Llama 4 Maverick in 4-bit quantization, one H100 generated proofs at around 110 tokens per second, with roughly 19 MB per 1,024-token request, about 70 to 95 times the GPU time vLLM needs to serve the same prompt.
• Anthropic released five artifacts under Apache 2.0 including executable code and Lean proofs, and states plainly that the system hides model weights but not model architecture, and that nothing is being deployed.

Read the full story on AiLookout.news

Adam Fry, OpenAI's consumer product lead for ChatGPT, said at a press event that the volume of search queries run through ChatGPT now equals about 20 percent of the volume run through Google. The comparison is based on OpenAI's internal measurement and its own view of Google's number, so the exact ratio is a claim, not an audited figure.

• ChatGPT consumer product lead Adam Fry said ChatGPT's search-like query volume is now roughly one-fifth of Google's, while noting the growth is based on OpenAI's analysis.
• OpenAI also said ChatGPT grew from 1 billion to 1.2 billion weekly active users in two months, added 250 million weekly voice users, and now generates 1.5 billion images per week.
• The comparison is apples-to-oranges in one direction: OpenAI counts any query that triggers a web search, and it did not specify the source of the Google volume number.

Read the full story on AiLookout.news

Fal, a five-year-old company that hosts and runs open-weights image, video and audio models for developers, raised a $1.3 billion round at a $16 billion valuation, double its April valuation. The round is one of the largest in the AI inference sector, and the company says its biggest problem in the summer was not demand but having to rate-limit paying customers.

• Fal's $1.3 billion round was led by Bessemer Venture Partners, with participation from Nvidia, Sound Ventures and T. Rowe Price, and values the company at $16 billion.
• The company announced its Series E and F at the same time, saying a combined $1.65 billion was raised across both rounds, and that it passed $800 million in annualized revenue earlier this year.
• Fal's chief executive said the company had to rate-limit customers in the summer because it could not serve demand, and the funding is earmarked for more compute and a bigger team.

Read the full story on AiLookout.news

AI Discover is a new tab in the Netflix TV app that suggests films and series by mood, audience or time available instead of by viewing history alone. It rolls out first to US members watching in English on supported TVs, and Netflix describes it as a test kitchen rather than a finished product.

• Netflix's AI Discover tab rolls out first in the US to members using English on supported TVs, marked by a red gem icon and a "New" label.
• Themes include "Something Light", "Scare Me", "Family Movie Night" and "Quick Watches"; each title has a Tune Your Recommendations button to keep or block it.
• Netflix is not publishing accuracy or engagement metrics, and the feature is unavailable on kids profiles and outside the TV app at launch.

Read the full story on AiLookout.news

StateBench is a four-tier benchmark for long-context state tracking from the Allen Institute for AI and Hugging Face. It finds that recurrent and hybrid architectures like Gated DeltaNet generalize much further than transformers when trained directly, and that none of the frontier models tested robustly solves the harder levels as inputs grow beyond their training lengths.

• StateBench spans four difficulty levels rooted in Krohn-Rhodes theory: r-trivial, aperiodic, soluble, and general finite state tracking, with up to 64,000 state changes for length-generalization tests.
• Frontier models - Gemini 3.8 Flash, Claude Opus 5.5, and GPT-6 Sol and Luna - handled r-trivial and aperiodic tasks but dropped to 13% or lower accuracy on the longest general finite problems.
• 275M-parameter Gated DeltaNet and hybrid models trained directly on StateBench generalized almost perfectly out to tens of thousands of assignments, while transformers never learned the hardest level even when trained directly.

Read the full story on AiLookout.news

Anthropic's amended IPO filing shows about $895 billion in purchase and lease commitments, nearly all tied to computing and data centers. Q3 revenue reached about $17 billion, but the company swung to a $567 million operating loss after a $974 million operating profit in Q2, with compute rising to 58% of revenue.

• Anthropic's amended S-1 lists about $895 billion of aggregate spending commitments, nearly all for computing and data centers, with about $70 billion due within a year and roughly $450 billion due by the end of 2030.
• Third-quarter revenue hit about $17 billion, but the company posted a $567 million operating loss after its first operating profit of $974 million in the previous quarter; compute costs rose to 58% of revenue from 48%.
• Top commitments include about $314 billion in purchase commitments (up from $55 billion at end-2025), Broadcom at roughly $161 billion in equipment leases, Google at about $121 billion, SpaceX up to $120.5 billion, and Amazon at $110 billion.

Read the full story on AiLookout.news

Google plans to disable AI tools for K-12 students next year unless a school administrator turns them on, affecting Gemini, NotebookLM, Gemini in Classroom and Google Search. Because AI features are integrated into Search, schools that restrict AI entirely will lose search access and students will see a message from their school, Google education VP Chris Phillips told The New York Times.

• Google will turn AI features off by default for K-12 students next year unless a school administrator explicitly enables them.
• The change covers Gemini, NotebookLM, Gemini in Classroom and Google Search.
• Because AI Overviews are built into Search, schools that restrict AI lose access to Google Search entirely.
• Google announced the policy alongside the launch of Guided Learning in schools and says it will share specifics in the coming months.
• Schools that want search without AI summaries have no setting for that, because the two features ship together.

Read the full story on AiLookout.news

The Financial Times reports that Nvidia has told its data center partners they must sell at least 51% of their services to non-Chinese companies. The instructions were communicated verbally through sales and regional teams, affect only future contracts, and come as Washington tightens controls on Chinese access to advanced AI computing.

• Nvidia is requiring cloud computing clients to sell at least 51% of their services to non-Chinese companies, per the FT's report citing five people with knowledge of the matter.
• The guidance was delivered verbally through Nvidia's sales force and regional teams rather than in writing, and only new contracts are affected — existing agreements are not.
• The move extends US restrictions on Chinese access to advanced AI computing beyond chip exports to overseas cloud leasing, an area Chinese AI developers have relied on.

Read the full story on AiLookout.news

Apple is adding two content descriptors to the App Store Connect age rating questionnaire — Shared Location and Dating and Relationships. Apps that include Dating and Relationships get a minimum 18+ rating. Developers must declare them from November 16, and all apps must be updated by January 12, 2027.

• Apps with dating or relationship features — including a virtual companion that engages in sexual or romantic content — will carry a minimum 18+ App Store rating.
• Declarations start November 16, 2026, and every app must complete them by January 12, 2027, even if the app does not update its age rating otherwise.
• In Brazil, Shared Location maps to an A16 rating and Dating and Relationships to A18, unless Apple grants an official MJSP rating for apps like parent-child location sharing.

Read the full story on AiLookout.news

OpenAI built Pando, a new backend for Git repositories, to handle the massive parallel development load that Codex generates. It uses causal consistency instead of repository-wide consistency, reuses packfiles across fetches, and cut median push latency by nearly 5x. Pando now powers all development in OpenAI's monorepo and delivers nearly a petabyte of Git data per day.

• Pando uses causal consistency instead of repository-wide consistency, allowing pushes to unrelated branches to proceed in parallel without contending on a single coordination point.
• The system reuses packfiles across fetches instead of generating an individual packfile for each client, enabling many readers to fetch the same history without per-client CPU cost.
• Pando cut median push latency by nearly 5x and now serves all development in OpenAI's monorepo, with reliability that had fallen below 90% before the migration.

Read the full story on AiLookout.news

HuggingFace introduced Model Integration Contracts (MICs) for its Transformers library, a testing framework that shrinks large checkpoints into small, CPU-runnable fixtures with recorded baselines. The approach caught a UMT5 regression that had gone undetected for two months and is now running in CI, though it does not yet block PRs.

• Model Integration Contracts shrink real checkpoints into small fixtures (e.g., 66 MB vs. 1,227 MB) that run on CPU and preserve every layer type, tokenizer, and processor file.
• The UMT5 SDPA bug introduced on July 30, 2026 went undetected for two months because standard forward-pass tests did not expose it; the MIC fixture flagged it on every commit replay.
• A CLI tool called tmic drafts contracts with LLM assistance but requires human review for baseline approval; coverage is still partial across model families.

Read the full story on AiLookout.news

CrowdStrike announced two acquisitions on November 3, 2026: Dreadnode, which sells automated AI security testing, and Fiddler AI, which sells agent observability and explainability. Dreadnode's testing will be generally available on the Falcon platform in Q4 of CrowdStrike's fiscal 2027, and Fiddler will be sold as Falcon Observe on the same timeline. Both deals are expected to close in that quarter, subject to customary conditions, and neither press release disclosed a purchase price.

• CrowdStrike signed definitive agreements to acquire Dreadnode for automated AI security testing and Fiddler AI for agent observability, both announced at Fal.Con Europe in Barcelona on November 3.
• Dreadnode runs more than 100 known AI attack strategies, including jailbreaks and prompt injection, and maps findings to OWASP Top 10 for LLMs, OWASP Agentic Security, MITRE ATLAS and NIST AI RMF.
• Fiddler will be sold as Falcon Observe and will evaluate agents before deployment, trace their decisions and tool calls, and flag behavioural drift in production.

Read the full story on AiLookout.news

OpenAI appointed former US Army Secretary Dan Driscoll as senior counselor for national security modernization, a newly created role at the company. Driscoll will focus on how artificial intelligence can strengthen U.S. national security, cybersecurity, and help allied nations adopt AI responsibly. The hire signals OpenAI's deepening engagement with defense and intelligence agencies as the military pushes to expand its use of AI systems.

• Former US Army Secretary Dan Driscoll joins OpenAI in a new senior counselor role focused on national security modernization and cybersecurity.
• The hire comes as the U.S. military and intelligence agencies accelerate AI adoption, with OpenAI positioning itself as a defense contractor alongside Microsoft and Palantir.
• Driscoll resigned as Army secretary earlier this year following reported tensions with Defense Secretary Pete Hegseth over drone and modernization policy.

Read the full story on AiLookout.news

Anthropic added two sentences to the Claude Code system prompt on Opus 5.5 to steer the AI tool away from recreating specific existing video games. The update follows weeks of Steam being inundated with AI-generated game clones that decompile and repackage classic titles. Addy Osmani at Anthropic confirmed the change, saying the update targets IP infringement while keeping original game development and genre games working. Developers have reported the filter also blocks legitimate projects modding or enhancing existing games.

• Anthropic added two sentences to Claude Code's system prompt barring recreation of specific existing games, while allowing original games and genre games.
• The update targets a wave of AI-made game clones flooding Steam, particularly retro game decompilations and ports generated by AI coding tools.
• Developers report the filter is inconsistently enforced and sometimes blocks legitimate game-development work, and can be bypassed with trivial prompt rewording.

Read the full story on AiLookout.news

Squarespace is reducing its global workforce by approximately 19%, around 300 employees, according to a statement provided to TechCrunch by representative Jeff Seedman. The Verge first reported the layoffs on November 2, 2026. The company called the move a restructuring meant to simplify operations and move faster, and it comes as prompt-driven AI site builders take market share from traditional website platforms.

• Squarespace told TechCrunch it is cutting roughly 19% of its global workforce, about 300 employees, in a restructuring.
• The layoffs land as prompt-driven builders from Wix, Lovable and Google change how sites get made.
• Squarespace's own AI bets, the 2024 Blueprint tool and the unreleased Forma platform, have not yet changed its position.
• The company already went private once, after struggling to hold public-market interest following its 2021 listing.

Read the full story on AiLookout.news

ADNOC announced on November 2, 2026 at ADIPEC in Abu Dhabi a multi-year partnership with OpenAI and G42 that makes it the first energy company to deploy OpenAI's advanced AI platform. The first two applications are integrated field development master planning and a system that values every barrel of oil before it leaves the ground. The models run on G42's sovereign infrastructure in the UAE, with human experts retaining oversight of consequential decisions.

• ADNOC says it is the first energy company to deploy OpenAI Frontier, an agent platform, across its business rather than in a single application.
• The partnership starts with whole-field development planning and in-ground barrel valuation across production, refining, logistics and trading.
• OpenAI's models run on G42's sovereign AI infrastructure in the UAE, and ADNOC says human experts keep oversight of consequential decisions.
• OpenAI's Daybreak Blue security work is folded in for cyber defence of energy infrastructure, but no contract value, start date or performance figures were disclosed.

Read the full story on AiLookout.news

Goldman Sachs has restricted Anthropic's Claude Code and Microsoft's GitHub Copilot to its engineering staff only, following an internal review of productivity and risk. The move reverses months of broader access and has frustrated traders who used the tools for client trading flow monitoring and market simulation. The bank's own custom AI assistant, which draws on models from OpenAI, Anthropic and Google, remains available firm-wide.

• Goldman Sachs limited Claude Code and GitHub Copilot to engineers after a productivity and risk review, reversing months of broader access.
• Traders had used the tools for monitoring client trading flows, simulating market shifts, and drafting text and visuals.
• The restriction follows similar moves across Wall Street as JPMorgan, Citi and Bank of America set their own AI boundaries.

Read the full story on AiLookout.news

Y Combinator CEO Garry Tan revealed on Walter Isaacson's podcast The Builders that he spends between $100,000 and $200,000 a year on AI tokens. Tan uses AI agents for coding, research, and personal life management, and has open-sourced his 23-skill Claude Code toolkit called gstack. He also disclosed that Y Combinator itself runs an internal AI agent named Stray Light that has context on every founder the accelerator has ever backed.

• Garry Tan spends $100K to $200K per year on AI tokens and has open-sourced his 23-skill Claude Code toolkit gstack.
• YC runs an internal AI agent called Stray Light with access to every founder application and email in the accelerator's history.
• Tan says AI agent costs are 'too expensive' for mainstream adoption and warns against free Big Tech-controlled alternatives.

Read the full story on AiLookout.news

Anthropic disclosed in its IPO filing that the retail-price value of its internal Claude usage exceeded its monthly employee compensation costs in April 2026. By September, internal AI usage had surged more than 250-fold since January. More than 90% of the code added to Anthropic's software in September was written by Claude. The company treats AI tokens as a growing component of how work gets done, alongside traditional salary, bonuses and stock awards.

• The list-price value of Anthropic's internal Claude usage first exceeded monthly personnel costs in April 2026.
• Internal AI usage value surged more than 250-fold from January to September 2026.
• More than 90% of Anthropic's added code in September was written by Claude, not human engineers.

Read the full story on AiLookout.news

OpenAI announced on November 2 that it will run another internal security sprint with an outside partner, Cathedral Systems, whose researchers will use OpenAI's own models to challenge its defenses. The company is also revamping its Security Bug Bounty program, more than quadrupling rewards for critical findings on average: critical vulnerabilities in core products now earn up to $30,000, with exceptional findings eligible for up to $100,000 and up to $172,000 with bonuses.

• OpenAI's next security sprint brings in Cathedral Systems, whose researchers will use OpenAI models alongside company teams to challenge its defenses.
• Critical vulnerabilities in core products, including ChatGPT, Codex and the API Platform, now earn up to $30,000, up from $6,500 in OpenAI's illustrative example for a one-click account takeover.
• The bug bounty program adds report-quality, loyalty-tier and priority-campaign bonuses, lifting exceptional critical payouts to as much as $172,000.

Read the full story on AiLookout.news

Apple started delivering the Audio Intelligence suite for Apple Watch Series 12 and Ultra 4 in the iOS 27.3 developer beta on November 2. Two conversation features are the headline: Live Rewind, which shows the last 15 seconds of speech as a text snippet when you double-press the digital crown, and Siri Recaps, which ambiently summarises conversations into high-level notes. Apple says raw audio is never stored or accessible to the operating system, apps, or Apple itself, and that summaries auto-delete after seven days unless saved.

• Live Rewind replays the last 15 seconds of a conversation as text when the wearer double-presses the digital crown; unsaved snippets disappear permanently after about 30 seconds.
• Siri Recaps generates high-level summaries, not transcripts, omitting speaker attribution and sensitive data like financial or authentication information.
• The features require Apple Watch Series 12 or Ultra 4 paired with an iPhone 16 or newer, are English-only at first, and use the S11 chip's Secure Exclave plus Private Cloud Compute.

Read the full story on AiLookout.news

DeepSeek released creator mode for DeepSeek Harness on November 2, 2026. Users describe a plugin in chat and Harness builds it, extending tools, skills and the interface. DeepSeek's own Harness page describes the app as in public preview and open source, with composable plugins. The company published the prompts behind seven demos, including a 3D landmarks gallery, a Transformer visualisation, a poster workbench, a video editing workbench and a paper reader, while cautioning that some plugins needed several rounds of refinement.

• DeepSeek Harness now has a creator mode where a chat prompt produces an installed plugin with its own sidebar entry or overlay.
• DeepSeek published seven demo prompts, from a 3D landmarks gallery to a video editing workbench and a conversation statistics panel.
• Harness is in public preview and open sourced, installed with Node.js and npx or by cloning the repository.
• DeepSeek warns that some demos needed multiple conversation rounds, so a single prompt is not a guarantee of the result shown in the video.

Read the full story on AiLookout.news

Marvell Technology and Lumentum announced on November 2, 2026 a definitive agreement for Marvell to acquire Lumentum in an all-stock transaction. Lumentum stockholders receive a fixed 4.500 Marvell shares per Lumentum share, an implied enterprise value of $115 billion, or roughly $1,132 per Lumentum share, a premium of about 23.9% to the unaffected closing price on October 28. Marvell holders will own about 67% of the combined company and Lumentum holders about 33%. The deal is expected to close in the first half of calendar 2028.

• Marvell is acquiring Lumentum in an all-stock deal with an implied enterprise value of $115 billion and a fixed 4.500 exchange ratio.
• Marvell holders keep about 67% of the combined company; Lumentum holders get about 33% and three board seats, including CEO Michael Hurlston.
• Both stocks fell on announcement day, with Marvell down about 13% and Lumentum down nearly 19%.
• Closing is guided for the first half of calendar 2028, so the deal faces roughly 18 months of stockholder and regulatory review.

Read the full story on AiLookout.news

Google expanded Dreambeans to India on November 2, the app's first market outside the United States, with support for 13 local languages including Hindi, Punjabi, Tamil, Telugu and Bengali. The app generates daily illustrated stories from connected Google services without a prompt. TechCrunch reported that Sensor Tower estimates put US downloads at about 55,000 since June, with a steep drop in October.

• Google launched Dreambeans in India on November 2 with 13 local languages, its first market outside the US, on Android and iOS.
• Sensor Tower estimates cited by TechCrunch put US downloads at roughly 55,000 since June, with about 5,500 in October after an 84 percent month-over-month fall.
• Stories can only be generated in one selected language at a time, so the Hindi-English mixing common in India is not supported, according to Google product manager Gozde Oznur.

Read the full story on AiLookout.news

Apple reported fiscal 2026 fourth-quarter revenue of $114.5 billion on November 2, up 12 percent year over year, with net income of $30.7 billion and diluted earnings per share of $2.10. Services reached a record $31.6 billion. On the call, chief financial officer Kevan Parekh said Apple sees AI as a reason for customers to upgrade and pointed to opportunities across third-party apps and partnerships, without saying what those opportunities are.

• September-quarter revenue was $114.5 billion, up 12 percent, with iPhone revenue of $57.6 billion and Mac revenue of $10.8 billion; iPad fell 10 percent and wearables fell 8 percent.
• Services hit an all-time high of $31.6 billion, up about 10 percent, while Apple guided December-quarter services growth to 8 to 9 percent, citing third-party licensing headwinds.
• CFO Kevan Parekh said AI offers "opportunities across third-party apps and partnerships as well as first-party opportunities" but did not describe them; the call was John Ternus's first as chief executive after the September 1 transition from Tim Cook.

Read the full story on AiLookout.news

Trajectory raised a $45 million Series A at a $300 million valuation, led by Sequoia with Nvidia, Bessemer Venture Partners and Bain Capital Ventures participating. The company, founded about a year ago by three Stanford classmates who worked at Google DeepMind and Apple, sells a platform for continual learning: it collects feedback on how a model performs inside a customer's application and retrains an open source model on it. Harvey, Rogo, Decagon, Mercor and Nvidia are named users, and the platform is now publicly available.

• Trajectory raised $45 million at a $300 million valuation led by Sequoia, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures participating.
• The product connects to a customer's application, collects feedback on model performance and retrains an open source model on it, so specialists can improve the model themselves.
• Harvey, Rogo, Decagon, Mercor and Nvidia are named as users, but no revenue, margins or independent benchmark results have been published.

Read the full story on AiLookout.news

Omen AI, a San Francisco startup building real-time fluid sensors for AI data center cooling, acquired Aquatic Labs for $16 million in cash and stock on November 2, 2026. Aquatic Labs founder Allan Adams joins as chief science officer and leads research and development. The deal folds alkalinity sensing into Omen's elemental and biological monitoring, while Omen reports $46 million in total funding and work inside data centers with more than 10 gigawatts of combined capacity.

• Omen AI paid $16 million in cash and stock for Aquatic Labs, announced November 2, 2026, and made founder Allan Adams its chief science officer.
• The two sensors are complementary: Omen reads what is dissolved in the fluid, Aquatic Labs measured alkalinity, the property that predicts scale and corrosion.
• Omen reports $46 million in total funding and work in data centers with more than 10 GW of combined capacity, but has not disclosed named customers, revenue or independent uptime results.

Read the full story on AiLookout.news

42AI, an Israeli startup co-founded by Amit Gal-Or and Ehud Ezra, exited stealth on November 2, 2026 with $62.5 million in seed funding for a "behavioral world model" that predicts how specific individuals will act. It is deployed at 10 large brokerages for liquidity, margin and hedging decisions, is generating seven-figure revenue, and claims an F1 score of 0.8 to 0.85 for trading direction and volume. The funding came largely from the founders and Infinity Group over three years, and the company plans to raise more.

• 42AI exited stealth with $62.5 million, raised largely from its founders and Infinity Group over three years rather than from a conventional venture round.
• The model combines psychological assessments, government and biobank datasets covering more than 2 million people, and customer records covering almost 30 million people into per-person probability profiles.
• It runs at 10 large brokerages and claims 98% range accuracy for trade flows and an F1 score of 0.8 to 0.85 for trading direction and volume, figures the CEO qualifies as range measures rather than per-trade predictions.

Read the full story on AiLookout.news

In the amended S-1A Anthropic filed on November 2, 2026, the company added risk-factor language saying that its "internal estimates and our expectations regarding future models, rather than our current models" indicate it "may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI'". Anthropic said it has not designated any current model as exceeding that threshold, that it has previously failed to identify accelerations in its own development rate until after they occurred, and that it is considering updates to the policy to keep humans in control of frontier model development.

• The amended S-1A says Anthropic's internal forecasts, not its shipped models, point at the company soon meeting its Responsible Scaling Policy definition of dramatically accelerated AI-driven progress.
• Anthropic wrote that it has in the past not identified a significant acceleration in its own development rate until after it happened.
• No current model has been designated as crossing the threshold; Anthropic says it is weighing policy changes to keep humans in frontier model development.

Read the full story on AiLookout.news

Oracle filed plans on November 2 to raise $10 billion in common stock and $10 billion in senior notes, using the full $20 billion external funding cap it described at its investor day one week earlier. Shares fell 8.8% on the news, extending a yearlong decline that has cut the stock by a third in 2026. The company plans to use the proceeds for capital expenditures, debt repayment and investments tied to its AI data-centre buildout, and forecasts at least $225 billion in revenue for fiscal 2030.

• Oracle filed to sell $10 billion in common stock and $10 billion in senior notes on Monday, matching the cap it previewed at its October 27 investor day.
• Shares fell 8.8% and are now down 33% year to date; credit-default-swap spreads have widened 265% over the past 12 months.
• Oracle expects positive free cash flow in the second half of 2028 and projects at least $225 billion in revenue by fiscal 2030 to justify the spending.

Read the full story on AiLookout.news

OpenAI introduced LifeVal on November 2, 2026, a benchmark that measures how well its AI models complete everyday tasks people want delegated, such as finding childcare, filing government forms and managing a constrained grocery budget. The evaluation draws on American Time Use Survey data and a demographically representative survey of 3,995 respondents. GPT-6 Astra achieved a best-recorded score of 80.6% on the benchmark, with perfect marks on several specific errands including toddler sensory-class booking, pre-trip vehicle-inspection booking and budgeting a halal grocery order.

• LifeVal measures end-to-end task completion rather than question answering, using everyday scenarios from time-use research and a 3,995-person survey.
• GPT-6 Astra recorded a best score of 80.6% overall with three tasks scoring 100% — though OpenAI has not disclosed the run count, scoring denominator or typical variation.
• The benchmark consciously excludes consumer-facing features; OpenAI describes it as an internal evaluation tool developed over the past year.

Read the full story on AiLookout.news

Anthropic filed an amended Form S-1 with the SEC on Monday showing Q3 2026 revenue of $17 billion, up 1,127% year over year. The filing also revealed $895 billion in compute and infrastructure commitments and a net loss of $23 billion for the quarter. The numbers mark an official update to earlier investor communications that pegged Q3 revenue at $16.8 billion.

• Anthropic's Q3 2026 revenue hit $17 billion, up from roughly $1.4 billion a year earlier, according to the amended S-1 filed Monday with the SEC — slightly above the $16.8 billion figure reported from earlier investor briefings.
• The company reported $895 billion in compute and infrastructure commitments, up significantly from earlier disclosed figures, and a net loss of $23 billion for the quarter as it continues heavy investment in datacenter capacity.
• The amended filing covers nine-month revenue of $33.24 billion and sets the stage for what would be one of the largest technology IPOs in recent years, expected to price before the end of 2026.

Read the full story on AiLookout.news

SpaceX CEO Elon Musk confirmed that Grok Bot's usage limits were intentionally throttled after the cost of running xAI's agentic AI agent became 'unsustainably high relative to revenue.' Users reported hitting 100% usage caps over Halloween weekend even when they had barely used the service, and Musk acknowledged the caps were deliberate. The move comes as SpaceX's Super Intelligence division reported a $431 million profit last month, with AI now accounting for 52% of company income.

• Users unexpectedly hit 100% Grok Bot usage limits over Halloween weekend, with some reporting their allocation jumped from 50% to full without apparent usage — Musk confirmed it was a deliberate throttling, not a glitch.
• Musk stated the cost of serving Grok Bot was 'unsustainably high relative to revenue' in an X reply, though he has not specified what pricing or cap changes subscribers can expect going forward.
• The throttling comes despite SpaceX's Super Intelligence division reporting a $431 million profit, with AI making up 52% of company income — the profit would have been higher without $1.2 billion in one-time costs related to the Cursor acquisition.

Read the full story on AiLookout.news

The New York Times sued Brave Software in Manhattan federal court on November 2, 2026, alleging the browser maker scrapes paywalled journalism, sells access to its 30-billion-page index to AI companies including Anthropic, Perplexity, Mistral, xAI and Cohere, and markets a paywall blocker. A second suit filed the same day by the New York Daily News, Chicago Tribune and Orlando Sentinel makes similar claims. Brave did not immediately comment; none of the AI companies were named as defendants.

• The Times and three Tribune-linked newspapers filed separate federal suits in Manhattan on Monday, both targeting Brave Software rather than the AI companies that buy its index.
• The complaints say Brave crawled more than 30 billion pages with a crawler that does not identify itself, leaving publishers no way to block it, then sold paid access to that content for AI training and answers.
• The Times suit also attacks Brave's browser paywall blocker and its practice of blocking publishers' ads while selling its own.
• Brave, founded in 2015 and backed by roughly $250 million in funding, did not immediately respond; the five AI companies named as recipients of its index were not defendants.

Read the full story on AiLookout.news

OpenAI announced the Grid Protectors Program on November 2, 2026, committing its future data center campuses to voluntarily reduce electricity demand during grid emergencies. The first enrolled site, the Effingham County campus in Georgia, pledges up to 1,000 megawatts of flexible demand reduction — enough to power roughly 700,000 homes — which Georgia Power estimates will save customers about $950 million a year from 2029.

• OpenAI launched the Grid Protectors Program, a framework in which future data center campuses commit to cutting electricity demand when the grid is under strain.
• The first campus is Effingham County, Georgia, pledging up to 1,000 megawatts of flexible demand — equivalent to about 700,000 homes — with Georgia Power projecting $950 million in annual customer savings from 2029.
• OpenAI said families and small businesses should not subsidize compute campuses, positioning the program against regulators and lawmakers pushing for data center demand flexibility.
• FERC has moved to reform large-load grid connections, and a Senate permitting bill would let states condition grid service on peak demand limits.

Read the full story on AiLookout.news

Meta's advertising business grew 26% in Q3 2026 to $63.1 billion, more than doubling Google Search's 12% growth. The 14-point gap is the widest since 2021, putting the two businesses within $300 million of each other by quarterly revenue. CNBC reports the divergence reflects AI's double-edged effect: Meta's AI targeting tools boosted ad performance, while Google's AI Overviews and AI Mode are cannibalizing the search clicks that fuel its ad business.

• Meta ad revenue hit $63.1B in Q3, up 26% YoY, while Google Search grew just 12% to $63.4B — the widest growth gap since 2021 and just $300M separating the two.
• CNBC reports Meta CFO Susan Li told investors that AI business assistant users drove 5% more conversions and 9% higher ad spend; nearly 10 million advertisers used at least one AI tool in Q3.
• Google's AI Overviews and AI Mode surface answers instead of links, reducing click-through rates and slowing search ad growth even as query volumes hit record highs.
• Both companies face investor skepticism: Alphabet lost $500B in market cap on its worst day ever after Q3 earnings, while Meta's stock fell 14% in October.

Read the full story on AiLookout.news

In a quarter that CNBC called the first where AI benefits showed up in measurable earnings, Santander reported $216 million in business value from AI initiatives in the first nine months of 2026, Novartis CEO Vas Narasimhan said AI "started to improve our P&L on a tangible basis," and Standard Chartered CFO said AI is appearing "everywhere across the business." The reports mark a shift from years of vague AI promises to specific, quantified financial outcomes from major non-tech companies.

• Santander quantified $216M in business value from AI in the first nine months of 2026, with its CFO saying AI is now directly linked to positive contributions in revenue and costs.
• Novartis CEO Narasimhan said Q3 2026 was the first quarter AI tangibly improved the P&L, with core operating income up 10% YoY at constant currencies.
• Standard Chartered reported AI "everywhere across the business," driving 14% operating income growth and visible operating leverage as income outpaces cost growth.
• Meta said its AI investments delivered "substantial incremental advertising revenue" with 26% total revenue growth, while Google Cloud revenue more than doubled to $31B driven by AI infrastructure demand.

Read the full story on AiLookout.news

On November 1, 2026, Proton CTO Bart Butler published the "AI @ Proton, Engineering Manifesto," an internal memo he wrote about a month earlier. It says "'Programming' is dead. Engineering is not, nor is expertise," and tells engineers their job is now to help LLMs write better code faster. It was published two days after a leaked story about Proton re-interviewing staff on AI use, which Proton has since said it will not do.

• Proton CTO Bart Butler published his internal "AI @ Proton, Engineering Manifesto" on November 1, 2026.
• He writes that an engineer's job "is no longer to write code" but "to help the LLM write better code faster," with models and agents as direct reports.
• He gives three reasons Proton hasn't yet seen big productivity gains: engineering culture, being stuck at "phase 1" of speeding up existing workflows, and a collective action problem.
• The memo appeared two days after a leaked story about staff re-interviews on AI use; Proton has since said it will not re-interview staff and that AI use will not cost jobs.

Read the full story on AiLookout.news

Apple CEO John Ternus told Wired's Steven Levy that Apple doesn't "want to ship AI for AI's sake." He said Apple is "absolutely going to leverage AI" in new products, but that people shouldn't have to think about whether a feature came from AI or traditional software. TechCrunch reported the comments on October 31, 2026.

• John Ternus, Apple's CEO since the beginning of September, said Apple doesn't "want to ship AI for AI's sake."
• He said Apple is "absolutely going to leverage AI" in its new products.
• Asked about AI agents making apps unnecessary, Ternus said creating products is "a deeply human and personal thing."
• Apple has often been described as behind on AI after delaying a more powerful Siri, which launched this summer.

Read the full story on AiLookout.news

On November 1, 2026, a Hugging Face repository called burner22/numbat appeared with 772 safetensors shards totaling 7.72 TB, claiming to be an Anthropic Mythos-class checkpoint. Within hours, CellCog's forensic analysis proved the tensor data was generated from name-seeded random noise — not trained weights. Separately, a 'Claude Mythos 5.5' staging reference briefly appeared on Anthropic's official page, unrelated to the Hugging Face repository.

• A 7.72 TB Hugging Face repository claimed to be an Anthropic Mythos-class checkpoint; forensic analysis proved the tensor data was seeded random noise, not trained weights.
• CellCog verified the debunk by matching each shard's random seed to its tensor name using Philox RNG, confirming zero training loss reduction.
• A separate 'Claude Mythos 5.5' staging reference briefly appeared on Anthropic's official Mythos page on November 1, unrelated to the Hugging Face repository.

Read the full story on AiLookout.news

On October 29, 2026, Y Combinator Labs and Olam Labs published 'Agents in Games', detailing experiments where 50 GPT-6.1 Sol agents ran in Minecraft for three hours. The agents sent roughly 5,900 messages, formed committees, appointed a single fertilizer operator, and never built a functioning society. In competitive tests, GPT-6.1 Sol beat Claude Opus 5.5 teams 6-1 across Bed Wars and Mindustry.

• YC Labs and Olam Labs ran 50 GPT-6.1 Sol agents in Minecraft for three hours; they formed bureaucracy rather than civilization, spending 11 minutes appointing a single fertilizer operator.
• In competitive tests, GPT-6.1 Sol beat Claude Opus 5.5 teams 6-1 across Bed Wars and Mindustry, with GPT developing more aggressive hive-like strategies.
• The researchers open-sourced their Minecraft agent harness, noting the patterns resemble dynamics seen in multi-agent coding workflows.

Read the full story on AiLookout.news

Anthropic has quietly updated Claude's safeguards to block video game reverse-engineering, modding, and decompilation tasks. The restrictions affect both malicious cloning attempts and legitimate development work, drawing criticism from the modding community and highlighting the irony of AI companies policing the same kind of copyright-adjacent activity that their training data relied on.

• Claude now blocks reverse-engineering, decompilation, and modding of video game code, affecting both malicious cloning and legitimate development.
• The restrictions come after a wave of AI-powered game decompilations and ports that let users run Mario in Elden Ring, WoW in browsers, and PS5 jailbreaks.
• Critics note the irony of Anthropic — which trained Claude on copyrighted web content — now blocking users from similar copyright-adjacent activities.
• The broad restrictions may push users toward less restrictive AI models, creating a race to the bottom on ethical safeguards.

Read the full story on AiLookout.news

Axios reports that OpenAI will launch a coordinated policy and PR campaign immediately after Tuesday's November 5 midterm elections, when Democrats are expected to retake the House of Representatives. The campaign has three pillars: proposing mandatory "speed limits" on the most powerful AI systems, rebranding OpenAI's data centers as "AI compute campuses," and backing state legislation that gives local communities more input on data-center buildouts. OpenAI chief global affairs officer Chris Lehane told Axios the midterms will be "a bit of a movie trailer for what you're going to see in 2028," when he expects AI to be the defining issue of the presidential election.

• OpenAI plans a policy and PR campaign right after the Nov 5 midterms, proposing mandatory AI "speed limits" and rebranding data centers as "AI compute campuses."
• The move anticipates Democratic victories that would give the party control of the House, shifting AI regulation from a friendly White House to a more skeptical Congress.
• Lehane says the campaign is about helping communities see AI as empowering and holding companies accountable for transformative local benefits.

Read the full story on AiLookout.news

On October 31, 2026, Anthropic renamed its two highest Claude subscription tiers. The $100 Max 5x plan is now simply Max. The $200 Max 20x plan is now Ultra. Prices stay the same, and existing subscribers migrate automatically. The substantive change is inside Ultra: verified accounts no longer have a rolling five-hour session limit. Weekly usage limits stay unchanged. New Ultra accounts keep a temporary five-hour cap until verification clears, which Anthropic says helps prevent fraud and abuse. A separate weekly limit for Claude Fable also remains.

• Max 5x ($100/month) is now Max; Max 20x ($200/month) is now Ultra. Prices unchanged, existing subscribers migrate automatically.
• Verified Ultra accounts no longer face a rolling five-hour session limit. Weekly usage caps and the separate Fable weekly limit are unchanged.
• New Ultra accounts keep a temporary five-hour cap until verification clears. Verification runs on account and payment history.

Read the full story on AiLookout.news

Memdoor published an open-source multi-agent coding workflow on October 31 that uses five Qwen3.8-27B model instances — each in a specialized role — to solve competitive programming problems. On the first ten LiveCodeBench Hard problems, the system scored a mean of 9.8 out of 10 across five passes, beating Claude Fable 5's single-model score of 8.0 and edging past the 9.6 of the GVS5H harness it reproduces. Fifty runs cost $35.79 total, or about $0.39 per run at the median.

• Five Qwen3.8-27B agents scored 9.8/10 on ten LiveCodeBench Hard problems, beating Claude Fable 5's 8.0 and the GVS5H harness's 9.6.
• The workflow uses specialized roles — planner, two ideators, solver, manager, plus a repair worker — communicating only through files on disk.
• Fifty runs cost $35.79 total; a median run took 40 minutes and $0.39 on Qwen3.8-27B with a 128K reply cap.

Read the full story on AiLookout.news

OpenAI updated its Help Center on October 31 to say that starting in December, Plus, Pro, and Business subscribers will share one token-based usage allowance across ChatGPT and Codex. Only Instant chat is exempt. OpenAI published no allowance sizes and no precise start date beyond the month. The change merges the chat pool with the existing Work-and-Codex pool as OpenAI brings agentic capabilities into the main ChatGPT interface.

• OpenAI will merge ChatGPT and Codex usage into one shared token-based allowance on Plus, Pro, and Business plans starting in December — only Instant chat stays outside the meter.
• OpenAI has published no allowance sizes and no specific start date beyond 'December'; the only planning tool is your own measured token burn.
• At the limit, subscribers can wait for a reset, upgrade their plan, or buy credits. The change reflects OpenAI merging chat with Work into a single product experience.

Read the full story on AiLookout.news