Public · Free
AI competition is shifting from better answers to completed work
🔭 Main Thread
AI competition is shifting from better answers to completed work. The useful unit is increasingly a finished task, a coordinated deliverable, or a workflow with fewer handoffs—not another chat response.
🎯 Primary
- OpenAI: Terra emphasizes faster, cheaper cross-application computer use. The focus is everyday execution.
- Runway: Runway Agent reads files, runs code, and produces coordinated brand assets: a shift from individual images or clips to delivery packages.
- Anthropic: Claude's Max tiers become Max and Ultra. Verified Ultra accounts lose the five-hour cap; prices and weekly allowances stay unchanged. Longer uninterrupted work does not mean unlimited usage.
- Google DeepMind: Gemini 4 Carbon is previewed, not evidence of universal developer access. Genie 4 emphasizes real-time environments that respond to actions.
- Mistral: Large 4 brings a trillion-parameter multimodal architecture and open weights. Deployment choice expands; commercial-use terms still need checking.
💰 Investment
- a16z's Eve example cites a customer moving from roughly 100 to 500 clients per lawyer. The suggested vertical-AI advantage is fewer handoffs, not merely faster generation.
- Cohere announced an initial closing toward a round targeting up to C$7 billion. Its Canadian commitment includes C$4 billion in R&D and infrastructure and plans for more than 1,500 jobs through 2031—long-term industrial capacity, not just model launches.
🧠 Analysis
- Orchestration remains a differentiator. Greg Isenberg's Gamma conversation describes roughly 20 models and 200 calls behind a presentation. Taste, coordination, and feedback matter beyond the chat interface.
- Cheap answers are not learning. Li Xiaolai points to the harder question: whether capability and judgment actually accumulate in people and systems.
- Keep claims in their lane. Leaderboard leadership concerns particular evaluations; trusted-partner testing is not general availability.
🔨 Practice
- Measure cost per completed task. Cursor reports DeepSeek V4.1 Flash on its cost frontier at approximately $0.60–$0.98 per task across reasoning settings.
- Connect real work context—with authorization. Kimi's Gmail, Drive, and Calendar integrations make access to relevant material part of assistant utility. Its separate quota compensation announcement shows why billing reliability matters too.
- Treat data control as a deployment requirement. OpenAI Private Intelligence emphasizes control of sensitive enterprise data.
🔥 Professional Watchlist
- DeepSeek V4.1 Flash 1017 appears on the model trending list, alongside growing attention to task-level economics.
- MiniMax reports a 143.9GB 2-bit M3.1 Flash variant with 91.5% preferred-token agreement in specified evaluations. Smaller deployment footprints still involve quantization trade-offs.
- SambaNova reports M3.1 Flash serving above 200 tokens per second with lower costs than its predecessor. These are provider-reported figures.
👥 My Feeds
Anthropic's drug-development collaborations offer a quieter signal: scientific adoption depends on sustained research partnerships. The company describes early progress, not established outcomes.
🔭 Main Thread
Practical demonstrations are making AI legible: camera movement, complete learning paths, and assistants doing everyday tasks. The counterweight is ownership cost—who controls the system, and who keeps it working?
🔥 Viral Signals
- Controllable video: A Douyin camera-movement tutorial records 1,035,000 likes. Its 24-second demonstration turns unreliable prompting into a visible camera-path problem.
- Complete learning paths: A 53-minute programming tutorial records 882,000 likes. Long content can attract strong engagement when it takes viewers from setup to an application.
- Local control versus efficiency: A Reddit post relaying remarks about local inference records 5,329 points and 2,388 comments. This is evidence of discussion, not verification of the attributed remarks.
👥 Platform Feeds — China & Reddit
- Rednote: Lao Mai's nine Codex skills records 7,582 likes—a useful answer to “What can I do after installation?”
- Douyin: A 3D director's-console tutorial records 7,295 likes, pointing toward demand for visible composition, camera, and character controls.
- Reddit: An OpenClaw maintenance complaint records 141 points and 118 comments. One user's experience highlights repair and supervision costs; it is not a universal product verdict.