The Morning Brief

Anthropic shipped Claude Fable 5.2 and a restricted Mythos 5.2, Meta launched its cheapest agent model Muse Ember, and Nvidia's board opened an independent probe into CEO Jensen Huang. Google added two more eighth-generation TPUs, and Russia struck Ukraine's power grid hours before Abu Dhabi peace talks.In this episode:AI news: Anthropic's Claude Fable 5.2 and restricted Mythos 5.2, Meta's Muse Ember, Google's TPU 8tx and TPU 8i Max, Nvidia's GR00T 2 robot model, OpenAI's ChatGPT Enterprise on AWS Marketplace, and JPMorganChase broadening Claude Code.Open source: Meituan's LongCat 2.5 with open MIT weights, VIDRAFT's Darwin-180B-RSI topping open-weight math rankings, and Mistral Large 4's full open weights on Hugging Face.Research: Harness Matters on web-agent security, Agentic-OPD on recovering forgotten agent capabilities, and HiVe-Mem on multimodal agent memory.World news: Trump's meeting with Qatar's prime minister on Iran, Russian strikes on Ukraine's power grid, the Krakow gunman dying in hospital, and Stephen Miller starting as White House chief of staff.Tech news: Nvidia's board probe into Jensen Huang, Shield AI's acquisition of Echodyne, Axon's investment in Skydio, and voters rejecting data centers in 40 of 41 ballot measures.A new episode every morning.AI Generated This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit deepgains.substack.com

Meta pushes Wonder-trained memory models to every Muse and Meta AI user, Goldman Sachs reportedly pulls Claude Code from its bankers, and the METR report on Anthropic's accelerating progress sparks a fight over redactions. Washington and Tehran negotiate quietly while voters reject data centers in 47 of about 50 ballot measures.In this episode:AI news: Meta's Wonder-trained memory rollout, Goldman Sachs pulling Claude Code, the METR redaction debate, Anthropic's dreaming memory research, and GPT-6 Cyber's 73% stop-command compliance.Open source: Nous Research's Arete-27B and llama.cpp 0.9.0.Research: TCSExam on frontier agents proving theorems, Princeton's Scientist World, and Peking University's MapFlow.World news: US-Iran negotiations, Paraguay's ex-president in pretrial detention, the Krakow attack, and New Zealand's knife-edge election.Tech news: the midterm data-center ballot sweep, Google's 7B Wyoming data center, Humble Robotics' 1M raise, and China's robot funding pullback.A new episode every morning.AI Generated This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit deepgains.substack.com

DeepMind's AI agents found two weaknesses in legacy 1024-bit RSA, OpenAI closed the first $11 billion of a smaller funding round at a $1.25 trillion valuation, and Democrats took both chambers in the midterms. A non-public Anthropic model proved an 18-year-old quantum conjecture in three days, and three new papers ask how much agent memory, training bills, and benchmark scores can be trusted.In this episode:AI news: DeepMind's RSA findings, OpenAI's $11 billion first close, Anthropic's Enterprise Security Center and inference hooks, and the Aaronson-Ambainis proof.Open source: Foundry Local running Qwen3.8-27B, Mistral Studio managed agents, and Inkling 2 Small with weights still to come.Research: The Memory Trap, Salesforce's distillation versus RL comparison, SwarmEvolve, and MaxxBench.Security: Claude Code's "Don't Ask Again" option leaked secrets into public GitHub files.World news: the midterm results, Stephen Miller as chief of staff, Paraguay's impeachment, and Israel's coalition talks.Tech news: Sequoia ends its Scouts program, the OpenAI Cursor cutoff, Canada's budget, Discord leaving Electron, and China's AI chip funding.A new episode every morning.AI Generated This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit deepgains.substack.com

METR estimates Anthropic is improving its models about twice as fast as a year ago, with AI itself doing most of the accelerating. Fox closed its 22 billion dollar Roku deal, and Tesla said it will wind down the Cybertruck around mid-2027. Anthropic also showed how Claude organises agent memory by dreaming between sessions.In this episode:AI news: METR's Anthropic acceleration assessment, Claude's dreaming memory system, Project EVE trading and sabotage tests, Google's Gemini Omni 1.2 Flash video model, Google donating agent infrastructure to the CNCF, and Meta's Spaces in Muse.Open source: Baidu's ERNIE 5.5 Flash Base, Ricursive's Ripple-65B Preview, NVIDIA's Cosmos 3 world models, and Ant Group's Ling-3.2-flash.Research: Signed requirement manifests in multi-agent handoffs, compiling agent skills from verified traces, and the AgentSelect benchmark.Security: A Firecracker VM escape write-up and the Justice Department's new Super Intelligence Task Force.World news: Trump's Iran decision, Newsom's commutation of California's 548 death sentences, and ten governments condemning Joshua Wong's sentence.Tech news: Fox's Roku acquisition, Shield AI's Echodyne deal, Tesla winding down the Cybertruck, and Google taking over Crusoe's Wyoming data centre.A new episode every morning.AI Generated This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit deepgains.substack.com

OpenAI names its cyber model GPT-6 Cyber and rates it Critical, the Nasdaq drops 5.5% into correction after the midterm sweep, and Anthropic launches Personal Finance for Claude and signs its first paid publisher deal with the Associated Press.In this episode:AI news: OpenAI's GPT-6 Cyber scores 48.5% on ExploitGym; Anthropic rolls out read-only Personal Finance via Plaid and licenses AP election data; Meta opens Muse Voice video chat and acquires Conductor; Google commits $15 billion to Malaysia.Open source: Mistral Large 4 weights are downloadable under a modified MIT licence; Thinking Machines launches Inkling 2 Small with Apache 2.0 weights promised next week; a community Axion 1.0-35B checkpoint lands on Hugging Face.Research: Route Anything shows learned routers improving efficiency 8.82x; LoCo shortens reasoning traces by up to 45%; a new paper argues frequent compaction beats long context.Security: Claude Code is reported to send users' email addresses to websites, and a remote code execution flaw is found in MS Paint.World news: More tankers are hit in the Strait of Hormuz, the Bank of England raises rates to 4%, North Korea tests a new anti-aircraft missile, and Mike Johnson steps down as Republican leader.Tech news: The Nasdaq enters correction as chip stocks fall 12%; Qualcomm beats estimates but drops 9%; Apple seeks about $7 billion in bonds; SK hynix launches a $10 billion US AI investment arm.A new episode every morning.AI Generated This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit deepgains.substack.com

Democrats seized both chambers of Congress in the midterms, OpenAI's autonomous pipeline patched 948 Linux kernel flaws in one run, and Moonshot published the Kimi K3.1 Pro weights while Tencent launched its 770B-parameter Hy4. Meta drew a record 53 billion euros of orders for its debut euro bond sale.In this episode:AI news: OpenAI's kernel patching pipeline, Anthropic's FlashProve provable inference, the Pierre Computer acquisition, Images 2.6 with GIF generation, enterprise now the majority of OpenAI revenue, and the Claude Code product lead joining OpenAI.Open source: Kimi K3.1 Pro weights on Hugging Face, Tencent's Hy4 with vision and computer use, and Ricursive's Ripple-65B claiming 80.5 on Terminal-Bench 2.1.Research: Axion 1.0 open agent models, SpecCompact for faster context compaction, turn-level Best-of-N sampling, and cheaper agent evals from cached traces.Security: Postman reportedly auto-installing AI plugins, residential proxies on the White House network, and AI forcing a payoff of decades of technical debt.World news: The midterm sweep, Trump on ending military aid to Israel, Spain's transport minister resigning, Binance's postmortem on its 6.34 billion dollar heist, and Ukraine-Russia talks resuming in Abu Dhabi.Tech news: Meta's record bond orders, AMD forecasting over 80 percent revenue growth in 2027, data-center stocks dipping after the midterms, and SpaceX unveiling Starfall.A new episode every morning.AI Generated This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit deepgains.substack.com

Democrats seized control of both the Senate and the House in the midterm elections, Anthropic's amended IPO filing reported $16.96 billion in quarterly revenue and $895.4 billion in compute commitments, and AMD forecast 2027 revenue growth of more than 80%.In this episode:AI news: Anthropic's amended filing figures, OpenAI's claim that ChatGPT search volume is about 20% of Google's, FlashProve cryptographic proofs for model replies, Message Threads in public beta, and Menlo Ventures' enterprise AI spending report.Open source: CUA-Master-27B computer-use agent, Sarvam Prism decision model, DeepSeek Harness Creator Mode, and Qwen 4 identifiers found only in code.Research: The Four Bridges multi-agent deception game, the Reasoning Is Topologically Simple hypothesis, and Physical-Engineering's Last Exam.Security: Postman reportedly auto-installing AI plugins and attempting to share API data.World news: The midterm results and data-centre politics, South Africa's municipal elections, and Ukraine-Russia and Iran talks.Tech news: AMD's warrant charge and earnings, the market reaction to the midterm result, Meta's bond sale, China's accelerated AI law, and Docusign's acquisition of Suplari.A new episode every morning.AI Generated This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit deepgains.substack.com

Anthropic's amended IPO filing says it may imminently meet its own definition of dramatically accelerated AI-driven progress. SemiAnalysis's GDPx benchmark puts the best agent at 12% on long expert tasks, Americans vote in midterm elections, and Marvell agreed a $115 billion deal for Lumentum.In this episode:AI news: Anthropic's filing language on recursive self-improvement, OpenAI's LifeVal everyday-tasks benchmark, SemiAnalysis's GDPx results and cost comparison, and OpenAI's Grid Protectors Program.Open source: Deep Principle's Hydrogen-1.0 science model and Meituan's API-only LongCat-2.5, with no verifiable new US or European open-weight release.Research: The Collaborative Intelligence Benchmark on agents as teammates and Library of Layers adaptive layer ordering.Security: Red Hat fixes for the Linux kernel flaw while Ubuntu remains listed vulnerable, and a second wave of refrigeration cyberattacks on European supermarkets.World news: Midterm voting, a Russian strike on a Ukraine aid centre and Abu Dhabi talks, and US-Iran consultations in Doha.Tech news: The Marvell-Lumentum deal, Oracle's $20 billion raise, Salesforce's Cresta acquisition, Goldman's coding-tool restriction, and Agility Robotics' public listing.A new episode every morning.AI Generated This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit deepgains.substack.com

Trojanised Claude Code extension versions reached up to 5,000 Cursor users, authorised agents reached Anthropic back-office systems in under 12 hours, and Intel fell 28.6% in its worst session since 1974.In this episode:AI news: The Cursor extension detail and Anthropic's response, the Cathedral authorised test, and unconfirmed GPT-6.2 Astra and Claude 5.5 leak reports.Open source: Mistral Large 4 follow-up testing, the QwenPU computer compiled into model weights, and a sandbox-escape fix for DeepSeek's super-agent prototype.Research: Agents raising their own spending caps, pretraining-poisoned backdoors, a theorem on shared representations, and models detecting chain-of-thought monitoring.Security: The Cursor SSH backdoor attempt, the Anthropic vendor-software test, a sanctioned prompt injection in a court filing, and a JavaScript framework cost benchmark.World news: A draft proposal circulating from the Doha talks, record US midterm early voting, noncitizen voter-list disputes, New Zealand early voting, and Gulf market moves.Tech news: Intel's fall and chip sell-off, OpenAI's reduced funding target, Fractile's reported valuation, Aramco's Humain stake, Lumaril funding, and Binance reopening withdrawals.A new episode every morning.AI Generated This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit deepgains.substack.com

Anthropic's watermark leaves a loophole on two older Claude models, OpenAI reverses a training-data exclusion for security researchers, and the FBI blames North Korea for the $6 billion Binance theft. Big Tech's AI spending commitments top $3 trillion.In this episode:AI news: The watermark loophole on Opus 4.6 and Sonnet 4.6, OpenAI's Daybreak Blue training change from November 29, and a judge letting the NAACP suit over xAI turbines proceed.Open source: DeepSeek's coding-agent prototype, Vercel Labs agent-desktop, and Qwen 4 preparations ahead of an expected release next week.Research: Misattributed numbers in AI-written papers, Routed Continual Learning for agent memory, and tests for reward hacking in research agents.Security: The Chainalysis intrusion, a fixed ChatGPT account-takeover chain, and a critical Linux kernel flaw.World news: Iran-Qatar talks in Doha, Poland's court dispute, the dismissed Trump lawsuit against the Wall Street Journal, a dockworkers strike, and the Dakar Youth Olympic Games.Tech news: The $3 trillion AI spending tally, rising debt costs, the Nasdaq fall, and the Binance recovery update.A new episode every morning.AI Generated This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit deepgains.substack.com