Good morning. It’s Wednesday, January 28th.

On this day in tech history: In 1926, John Logie Baird created the first working TV in a small London attic. He used a spinning disk to send blurry images of ventriloquist dummies through wires. It looked pretty rough, but his ideas about transmitting images helped pave the way for today's AI technology, like when your phone recognizes your face or spots objects in pictures.

In today’s email:

• OpenAI launches Prism for scientists, and Altman admits he went full YOLO

• Claude Cowork and Sketch join forces for multistage workflows across apps

• Moonshot AI’s Kimi K2.5 powers parallel agent workflows with 100-agent swarm

• 5 New AI Tools

• Latest AI Research Papers

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

_{In partnership with Contextual AI}

Introducing Agent Composer 

AI for When It Is Rocket Science. Agent Composer helps teams tackle expert-level engineering tasks in high-stakes environments—compressing hours of complex routine work into minutes. Trusted by high-tech leaders like Qualcomm and Advantest, it’s built for environments where accuracy and reliability are non-negotiable.

What makes it different:

• Unified context layer: Agents operate with full task, data, and workflow context

• Flexible, controlled agents: Combine dynamic intelligence with structured workflows for mission-critical reliability.

• Intuitive, no-code build: Create and optimize agents in minutes with pre-built templates, natural language prompts, or visual drag-and-drop—no actual rocket science required.

See it in action today—read the launch blog or RSVP for the launch event & AMA.

_{Thank you for supporting our sponsors!}

OpenAI launches Prism for scientists, and Altman admits he went full YOLO

OpenAI is making a major play for scientific researchers, launching a free AI workspace called Prism and revealing that graduate-level scientific discussions on ChatGPT jumped 47% last year to 8.4 million messages per week.

The company's GPT-5.2 model now scores 92% on the GPQA benchmark, surpassing human expert performance in biology, physics, and chemistry. Prism functions as an AI-native word processor that evaluates claims, searches literature, and converts whiteboard sketches into structured diagrams.

But the shift comes with acknowledged risks. CEO Sam Altman admitted he broke his own security rule by giving Codex full computer access within two hours, warning developers that convenience could lead them to "YOLO" into dangerous deployments where "serious security or alignment failures go unnoticed for extended periods."

OpenAI science lead Kevin Weil says “2026 will be for science what 2025 was for software engineering,” describing GPT-5.2 not as a discovery engine but as a high-throughput collaborator for cross-disciplinary reasoning. OpenAI is also adding self-critique and “epistemological humility” to curb overconfidence and hallucinations in scientific work.

As for the company’s technical architecture, OpenAI disclosed rare details about Codex's "agent loop," which operates completely statelessly, resending full context on every inference to support zero data retention. This creates quadratic prompt growth, managed through caching and automated compression.

OpenAI is also finalizing ChatGPT ad pricing at roughly $60 per 1,000 impressions, (That's 3x Meta's rates) which is far above typical digital ads and closer to premium TV advertising. Advertisers get views and clicks but no conversion data initially.

Altman called the lack of comprehensive security frameworks for AI agents "an open opportunity for startups," suggesting even OpenAI doesn't have answers. He also confirmed the company plans to slow hiring as AI absorbs internal work, and acknowledged GPT-5 sacrifices writing quality for reasoning performance, a tradeoff future models aim to fix. Read more.

Claude Cowork and Sketch join forces for multistage workflows across apps

Anthropic is turning Claude from a chat assistant into a full-fledged AI workflow hub. Nine enterprise apps including Asana, Slack, Figma, Canva, Box, Hex, and Monday.com, now run inside Claude, letting users manage projects, draft messages, generate charts, and preview files without leaving the interface.

Integrations leverage MCP Apps, Anthropic’s open-source Model Context Protocol, with human-in-the-loop access controls. Claude Cowork will soon enable multistage, agentic workflows across large datasets.

Claude Code got a major upgrade. DAG-based persistent tasks replace ephemeral to-dos, enabling cross-session state, parallelized workflows, and memory optimization for complex engineering pipelines. The agent even ported NVIDIA CUDA to AMD ROCm in 30 minutes, showcasing AI agents’ growing ability to automate cross-platform GPU workloads. Workflow intelligence is also expanding. Anthropic is testing a Plugins hub, Sketch attachments for diagrams and mockups, and local server previews for developers. Cowork tasks automatically reference project instructions and files, storing outputs locally.

CEO Dario Amodei frames these advances in a broader warning. By 2027, AI clusters could operate like a “country of geniuses in a data center,” with intelligence rivaling 50 million Nobel laureates. This scale introduces autonomy risks where models might develop deceptive personas or game safety tests. To counter this Anthropic uses Constitutional AI to include ethical principles into the model's identity. Read more.

Moonshot AI’s Kimi K2.5 powers parallel agent workflows with 100-agent swarm

Moonshot AI has launched Kimi K2.5, an open-source visual agentic intelligence model that integrates a 1T-parameter Mixture of Experts language backbone with a 400M-parameter MoonViT vision encoder.

The 61-layer network activates 32B parameters per token, supports a 256K-token context window, and processes multimodal inputs natively, enabling developers to mix design assets, documents, and code in a single workflow.

K2.5 introduces Agent Swarm, a parallel multi-agent system trained with Parallel Agent Reinforcement Learning (PARL). The orchestrator spawns up to 100 subagents to handle complex tasks in parallel, completing up to 1,500 coordinated steps per run, achieving 4.5× speedups on wide research and coding workflows. DAG-based persistent tasks, memory optimization, and cross-session coordination make Kimi a state-aware coding and reasoning assistant. Read more.

• ‘Clawdbot’ creator says Anthropic 'forced' him to rename the viral AI agent

• Microsoft’s Maia 200 triples AI inference speed and slashes hardware load 

• Moltbot (formerly Clawdbot) + Ollama lets you run a full AI agent without sending data to the cloud

• Grok 4.20 just took the #1 spot on PredictionArena

• Allen AI lowers the barrier for custom coding AI: private repo training now costs less than $500

• Google now lets users jump from AI Overviews into AI Mode conversations

• A European AI challenger goes after GitHub Copilot: Mistral launches Vibe 2.0

• Epoch AI introduces FrontierMath Open Problems, a professional-grade open math benchmark that has challenged experts

• Watch: Helix 02 achieves continuous, room-scale autonomy

• Robots only half as efficient as humans, says leading Chinese producer

• Airtable jumps into the AI agent game with Superagent

• Luma AI pushes generative video into real-world workflows with Ray3.14’s faster, cheaper, full-resolution output

• What Congress should do about AI, according to Dario Amodei

• Watch: 18 humanoid robots deploy themselves, no handlers required

• Dozens of nudify apps found on Google and Apple’s app stores

• SoftBank in talks to invest up to $30 billion more in OpenAI, source says

• China’s DeepSeek upgrades OCR with Alibaba’s Qwen2 model

• Taiwan, US discuss tech, AI and drone cooperation at high-level forum

• Manus AI now supports Skills, a common standard introduced by Anthropic

• Anduril has invented a wild new drone-flying contest where jobs are the prize

• Contextual AI’s new platform makes enterprise AI practical by fixing the RAG bottleneck

• Google just agreed to pay $68 million after its Assistant secretly listened to users and fed their conversations into ads 

• Apple quietly upgrades AirTag with 50% longer range and louder alerts while keeping the same $29 price

• Meta’s next move is subscriptions with AI superpowers for everyday users

• YouTubers are taking Snap to court, claiming its AI secretly learned from their videos

• Singapore bets $779 million on AI research to become Asia’s next innovation powerhouse

• Most U.S. workers exposed to AI could adapt, but 6 million, mostly women in clerical jobs, face high risk

• Emergency meetings, power struggles, and failed negotiations: how Apple ended up with Google Gemini

• Cambridge supercomputer to get 6 times more powerful with AI chips

• EU opens formal probe into Musk’s xAI over Grok deepfakes

• Meet GrowHR, the soft humanoid robot that can stretch, shrink, fly, and even walk on water

• Georgia Tech study argues the existential AI threat is overblown

• Alibaba’s Qwen3-Max-Thinking beats GPT-5.2 and Gemini 3 Pro at humanity’s hardest test—without breaking the bank

• Top 15 AI tools filmmakers are experimenting with today

• Watch: China's Xaomi has unveiled a fully automated factory

📄 Can LLMs Clean Up Your Mess? A Survey of Application-Ready Data Preparation with LLMs

📄 daVinci-Dev: Agent-native Mid-training for Software Engineering ​​

📄 SWE-Pruner: Self-Adaptive Context Pruning for Coding Agents 

📄 LongCat-Flash-Thinking-2601 Technical Report 

📄 Memory-V2V: Augmenting Video-to-Video Diffusion Models with 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!