171: 【AI季报 26Q2】从 coding 到 RSI，强者愈强的未来？

Summary

This LateTalk Q2 2026 AI review with Henry Yin of MOE Capital organizes the quarter around two loops: frontier intelligence pushing through coding, long-horizon agents, Auto Research, Recursive Self-Improvement, and Physical AI; and existing intelligence spreading into enterprise models, Open Source AI Models, Slack collaboration, Computer Use Agent workflows, voice, and image production. The source argues that OpenAI and Anthropic are no longer competing only on benchmark model quality, but on a full system of model capability, products, pricing, enterprise sales, ecosystem, safety posture, and organizational execution. Its main contribution to the wiki is to connect Codex, Claude Code, GPT-5.6, Fable 5, Claude Tag, Record and Replay, Enterprise Owned Models, GLM 5.2, and Interaction Model into one Q2 snapshot of AI capability and commercialization pressure.

Key Claims

• The episode frames Q2 2026 AI progress through two lines: pushing frontier capability forward, and diffusing current intelligence into workplaces, enterprises, collaboration tools, voice, computer use, and physical AI.
• OpenAI’s Codex is presented as regaining ground against Anthropic’s Claude Code after Claude quality/pricing complaints and OpenAI migration incentives.
• Fable 5 and GPT-5.6 are treated as high-end frontier models whose practical use is partly constrained by safety guardrails, access limits, or policy pressure.
• The source says Anthropic’s access limits on some frontier AI/ML tasks are a product and alignment problem, because silent degradation breaks user trust even when motivated by safety or policy pressure.
• Cursor is described as having lost independent-company status through a SpaceX/xAI-linked exit, making it a case of startup pressure under Model Provider Tool Competition.
• Recursive Self-Improvement is defined as a stronger loop than Auto Research: AI not only performs research tasks, but improves the next round of AI research capability.
• Anthropic’s internal code-generation and AI-safety-research examples are used as early evidence that coding agents may already be part of a self-improvement flywheel, while the source still treats full RSI as unresolved.
• Recursive and other RSI-oriented startups are presented as evidence that auto-research loops remain technically open enough for new teams to participate.
• OpenAI and Anthropic are both discussed as moving toward robotics or physical intelligence, but the source expects frontier model companies to focus more on robot brains, training, and world modeling than full-stack hardware.
• World Models and World Action Models are framed as the convergence of RL-style simulated learning and video-generation knowledge, especially when models become action-conditioned.
• Harvey and Applied Compute are used to introduce Enterprise Owned Models: enterprises may post-train or own domain models when frontier models are expensive, access is unstable, or the domain data is too valuable to leak.
• GLM 5.2, Kimi, and DeepSeek are treated as evidence that Chinese open models can become a practical enterprise-substitution layer when paired with U.S. post-training and application companies.
• Claude Tag moves Claude-style agents into Slack, making AI look less like a personal chatbot and more like a team coworker with tasks, context, permissions, and group visibility.
• Record and Replay is presented as an OpenAI computer-use route that records human GUI workflows into repeatable skills, but the source flags accuracy, latency, privacy, and permissions as constraints.
• Thinking Machines Lab’s Interaction Model is used to argue that voice is not just another modality; full-duplex speech could become a foundation for real-time AI interaction.
• MidJourney’s medical-imaging hardware turn is treated as a reminder that AI companies may reinvest software cash flow into unexpected physical-world bets rather than only model or app competition.

Key Quotes

“递归自进化” — the source’s Chinese rendering of RSI.

“AI 像研究员一样” — the source’s short definition of Auto Research.

“24 小时同事” — the episode’s frame for Claude in Slack.

Connections

• LateTalk, Henry Yin, and MOE Capital — show, guest, and investment context.
• OpenAI, Anthropic, Codex, Claude Code, Fable 5, and GPT-5.6 — frontier model and coding-agent competition.
• Cursor, SpaceX, xAI, and Model Provider Tool Competition — coding startup pressure and model-provider tool squeeze.
• Auto Research, Recursive Self-Improvement, Recursive, ML Coding, AI For Science, AI Verification, and Research Taste — AI research automation and RSI loop.
• Physical AI, World Models, World Action Models, World Model VLA Fusion, and Embodied AI — robotics and action-conditioned world-model branch.
• Enterprise Owned Models, Harvey, Applied Compute, GLM 5.2, Zhipu AI, and Open Source AI Models — enterprise post-training and open-model substitution branch.
• Claude Tag, Slack, Agent Harness, Agent Permission Boundaries, and Human-Agent Collaboration — team collaboration and permissions branch.
• Record and Replay, Computer Use Agent, AI Coding Verification, and Human Judgment Under AI — GUI workflow automation and verification branch.
• Thinking Machines Lab, Interaction Model, and Voice Interaction — real-time voice and full-duplex interaction branch.
• Meta, Google, Gemini, xAI, and MidJourney — other Q2 company updates in the source.
• AI Investment Metrics, AI Commercialization Pressure, and AI Economic Diffusion — business and investor interpretation of the quarter.

Contradictions

• No direct contradiction found. The source mostly updates existing AGI Three Acts, ML Coding, Recursive Self-Improvement, Model Provider Tool Competition, and Open Source AI Models pages with a later Q2 2026 snapshot.
• Timeline tension to preserve: earlier pages present Anthropic as the clearest coding-agent leader, while this source says OpenAI’s Codex regained ground in Q2 through product, pricing, and migration tactics. This is best treated as a sequential update rather than a contradiction.
• Source-local claim to verify later: the episode says Cursor was no longer independent after a SpaceX/xAI-linked acquisition. Existing pages only record provider pressure on Cursor, so this remains a claim from this source until corroborated by later ingests.
• Productive tension: Enterprise Owned Models and Open Source AI Models are framed here as ways to avoid frontier-lab dependence, while existing pages on AI Commercialization Pressure still warn that open or post-trained models must prove reliability, unit economics, and downstream business value.