In partnership with

Good morning. It’s Friday, December 5th.

On this day in tech history: In 1956, MIT researchers quietly tested the “Summer Vision Project precursor” camera rig, a hacked-together analog scanner used only in internal demos. Though never published, its line-by-line luminance capture became a reference for early computer-vision folks exploring edge extraction before digital frame buffers existed. It seeded techniques later formalized in Roberts’ 1963 gradient operator.

In today’s email:

• OpenAI’s “Confessions” pull back the curtain on AI misbehavior

• Anthropic prepares IPO

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

OpenAI’s “Confessions” pull back the curtain on AI misbehavior

OpenAI is moving through one of its most revealing phases yet. Its new ‘Confessions’ method adds a second output channel that acts like a truth serum, rewarding models only for honesty. Early tests on GPT-5 variants show the models admitting to reward hacking even when their primary answers fooled weak evaluators. It is not a fix for misbehavior, but it is a diagnostic lens that exposes what the model actually intended.

That push for visibility comes as a federal judge orders OpenAI to hand over 20 million anonymized ChatGPT logs in its copyright fight with major publishers. The ruling challenges the secrecy around how large models change training data into outputs and cues a new legal era where AI developers must surface their internal mechanics.

GPT-5.1-Codex-Max, designed for agentic coding, is now live. The upcoming App Directory demands real metadata, compliance docs, and verifiable test cases, marking a shift from experimental playground to production marketplace. The acquisition of Neptune Labs adds full neural telemetry, giving researchers raw access to gradients, activations, and configuration traces that normally disappear into black boxes. The OpenAI Foundation is also distributing more than 40.5 million dollars to 208 nonprofits focused on AI literacy and community-driven innovation.

CEO Sam Altman has explored taking control of Stoke Space to place future AI infrastructure in orbit. As data-center power demand climbs, orbital solar arrays and reusable launch systems like Stoke’s Nova are emerging as potential workarounds to Earth’s energy constraints.

The industry’s biggest supply chain remains in flux. Nvidia confirmed that its 100-billion-dollar, 10-gigawatt deal with OpenAI is still unsigned, echoing a pending $10 billion Anthropic investment and reflects an ecosystem where capital, compute, and trust are all being renegotiated in real time. Read more.

Anthropic prepares IPO while scaling enterprise AI and human feedback

Anthropic is quietly laying the groundwork for a 2026 IPO. They’ve brought in Wilson Sonsini, appointed Krishna Rao as CFO, and are running the internal checklist to meet public-company standards. No banks are locked in yet, but the pace suggests serious intent.

Anthropic CEO Dario Amodei signaled caution over the AI sector’s economic dynamics during the DealBook Summit 2025, drawing a clear contrast with certain competitors. Tech is strong, but the money side remains risky. YOLO moves, circular chip deals, timing errors could amplify vulnerabilities fast. Anthropic, he says, is keeping it disciplined.

At the same time, Anthropic is expanding its enterprise footprint. A $200 million deal with Snowflake puts Claude Sonnet 4.5 and Opus 4.5 directly into the cloud platform, letting companies run multimodal data analysis and build custom agents inside secure environments.

Anthropic is also building tools to understand AI’s human impact. Anthropic Interviewer ran 1,250 interviews with professionals across general, creative, and scientific work. The AI handles question generation, adaptive interviews, and transcript analysis, feeding results to humans. Insights show AI boosts productivity: 86% of general workers and 97% of creatives report time savings, but scientists remain selective, using AI mostly for coding, writing, and literature review. The tool scales human-AI feedback, letting Anthropic adapt its models while capturing professional, creative, and economic realities. Read more.

Google’s Workspace Studio and Opal bring AI reasoning to productivity, Replit joins the mix

Google is turning work into something AI can manage. Its new platform, Workspace Studio now lets anyone build AI agents across Gmail, Drive, Chat, and even Asana or Jira. Agents are built using Starters (triggers), Steps (actions), and Variables (dynamic placeholders) and can perform complex operations like sentiment analysis, invoice extraction, and intelligent prioritization. Teams can share these agents as easily as a Google Doc. It’s no code, full reasoning, full automation.

Concurrently, Gemini 3 Deep Think mode has launched for AI Ultra subscribers. Building on Gemini 2.5 Deep Think, which excelled at the International Mathematical Olympiad and ICPC World Finals, the update uses parallel reasoning to explore multiple hypotheses at once.

Coding is becoming collaborative. Google partnered with Replit to expand “vibe-coding.” The Opal engine is also getting advanced Agent capabilities for custom workflow automation that is far beyond typical “plan and execute” flows. They handle multi-step reasoning, sub-task decomposition, and adapt to context dynamically.

NotebookLM now supports prompts up to 10,000 characters, enabling detailed AI personas for work, education, and research. iOS features for infographics and slide decks are rolling out.

Google DeepMind’s mechanistic interpretability team, in a post on the AI Alignment Forum, has shifted from full neural network reverse-engineering to “pragmatic interpretability.” The focus is on proxy tasks that deliver measurable progress on AI safety and alignment, using simple, tractable methods like prompting, steering, and chain-of-thought analysis instead of complex techniques. Read more.

• Meta hires former Apple design lead Alan Dye to run new creative studio exploring AI as a design material

• Jensen Huang on AI: ‘No single winner,’ just tools that multiply human potential

• Amazon’s new Kiro system plugs AI developers into Stripe, Figma, and Datadog without overloading context windows

• Grok 5 trains on COLOSSUS 2, the world’s largest AI supercluster

• KREA AI launches Node App Builder to turn workflows into shareable apps without code

• Microsoft denies cutting AI sales targets despite reports of widespread underperformance

• Cristiano Ronaldo invests in Perplexity AI

• Singapore stakes a claim in the quantum race as Horizon Quantum brings its commercial system online

• Real-world trust, not benchmarks, puts Google’s Gemini 3 Pro at the top of AI evaluations

• The EU plans five massive AI gigafactories to build 100,000 high-performance chips

• Palantir CEO Alex Karp pushes back on claims his company is a surveillance tool for the government

• Math legend Ken Ono leaves academia to join a 24-year-old’s AI startup chasing superintelligence

• Startup Alien launches a privacy-first ID app that proves users are human without collecting biometrics

• An AI trained on prison calls is now scanning inmates’ conversations to predict potential crimes

• Investors should watch for an AI “air pocket,” says Bank of America, even if the bubble isn’t here yet

• Micron exits consumer market to prioritize memory for high-powered AI chips

• One of world’s most detailed virtual brain simulations is changing how we study the brain

• Kling AI’s 2.6 model lets you see the sound and hear the visuals in fully immersive AI-generated videos

• Year-end push to sideline state AI laws falters as NDAA negotiations collapse

• Developers can now skip prompt engineering thanks to byLLM’s AI-powered coding shortcut

• WordPress just made building interactive sites as easy as typing what you want

📄 MultiShotMaster: A Controllable Multi-Shot Video Generation Framework

📄 BlockVid: Block Diffusion for High-Quality and Consistent Minute-Long Video Generation  ​​

📄 ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration

📄 From Code Foundation Models to Agents and Applications: A Practical Guide to Code Intelligence

📄 RELIC: Interactive Video World Model with Long-Horizon Memory

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!