Viewpoints, commentary, analysis, interviews, and discussion of trends.
Addy Osmani, Morgan Westlee Lunt, and Qing Wang say the Claude Code team reviews changes continuously, with Claude reviewing and fixing most changes at least once before a human looks. They describe asking what could go wrong and which check—automated or human—is right to catch it, rather than using whether a human read the code as a proxy.
ClaudeDevs announced that Code Review is now generally available for Claude Team and Enterprise plans, with faster reviews, improved detection of consequential bugs, and per-repo Quick, Standard, or Deep review levels selected by admins. Boris Cherny says it runs on all his PRs and catches almost every bug before it gets in, helping his team move quickly while raising quality. https://twitter.com/ClaudeDevs/status/2110866210573386065
Replit CEO and co-founder Amjad Masad said the company’s work is increasingly driven by AI agents, which handle much of the initial support-ticket triage and can be deployed to submit fixes. He said people still provide judgment and taste, while spending more time deciding what to work on, discussing ideas, and planning architecture.
Live the kind of life where the suggestions from Codex Composer Text Predictions are always wrong.
@emollick posted: “Live the kind of life where the suggestions from Codex Composer Text Predictions are always wrong.” Ethan Mollick: I couldn't help myself. https://t.co/1D6DQqXvAC https://t.co/baQEzSDYII
@davis7’s review says Muse Wonder Full has good writing and “voice,” capabilities on par with 6.1 sol for non-specialized tasks, and orchestration. It lists the naming scheme and price and usage limits as cons, and says the full model is available in Muse Code through a subscription, with no API. Alexandr Wang called it a thoughtful review, said there is lots to improve, and was glad people enjoy using the model. https://twitter.com/davis7/status/2110533786895356118
@theo says Gemini 4 Argon has proven outright unusable for real code work in his codebases because it aggressively hallucinates about how systems work. Alexandr Wang says there is still room for Muse Wonder Full to improve and that they are working on it! https://t.co/D804UwQQQi
people are sleeping on muse wonder full :) https://t.co/dOhyW9f2UA
Try Muse Wonder Full on your hardest coding tasks—it will surprise you! “By my vibe testing, It's better than Astra. Not quite as good as Claude Code Fable and 5.5 Opus.” https://t.co/8RTGZ434kG
If an agent saves an engineer 2 hours implementing a feature but collectively costs the team 3 hours in alignment and review, the ROI is negative. Isolated chats are suboptimal in team environments because they lack context about what the team has already discussed and decided. The Kylon app addresses this by running Claude Code and Codex in a shared, multiplayer environment. Agents gain real-time visibility into team consensus and architecture decisions.
Muse wonders of the world
I think there should be companion fitness hardware for Codex and Claude Code. For example, reasoning intensity could be tied to your grip strength: the stronger your grip, the higher the reasoning intensity. That way, you could work out harder to get higher reasoning intensity.
with muse wonder, we aimed to build a model that is useful for real world tasks for real people, informed by our product adoption and how people use muse. we focused on things that might not look as glamorous on twitter (for example, not 3d demo games), which in retrospect we probably could have done more of.
The whole team feels dumb now, but it's true: we did not explicitly train our model on 3D demo games. Obviously, in retrospect, we probably should've done a bit of that... but no regrets. Please try Muse Wonder on your actual coding work!
MUSE WONDER 🍉 muse wonder is the best model in the world for personal agents and professional agents. we evaluated our model against the best models in the world across agentic and coding tasks, and we think you all will like it.
Andrej Karpathy describes four tiers of LLM exposure, arguing that public perceptions lag behind capabilities experienced by frontier-system users. He estimates roughly six billion people have barely encountered LLMs, one to two billion use casual chatbot products, 20 million use frontier models professionally in math and code, and 5,000 access internal frontier systems. He claims those internal systems coordinate thousands of agents on projects that would take experts years, while human review falls behind, and attributes progress in coding and math to reinforcement learning on verifiable rewards and commercial training priorities.
Code review is just one way to ensure quality. These days, some people act like the world will end if you don’t review AI-generated code, even though many people didn’t review code carefully before AI either. 😄
Ethan Mollick argues that many of the 5,000 people at labs with access to advanced models still haven’t used them to their full extent, with use concentrated in cybersecurity, math, and AI R&D. There are many other fields to advance and problems to solve, he adds, responding to @karpathy’s argument that the gap in shared understanding of LLM capabilities is widening.
Andrej Karpathy argues that understanding of LLM capability is diverging into a sharp funnel, with rough estimates of 6B people barely exposed, 1-2B casual users, 20M professionals using frontier LLMs in math and coding, and ~5,000 people accessing internal frontier systems. He describes increasingly ambitious agent workflows toward the narrow end, while saying human review and comprehension are starting to fall behind. He attributes the divide to problem size and duration, capability concentrated in digital and verifiable domains with economic value, and differences between free, paid, and internal access.
How AI shifted from a transductive paradigm (token-by-token answer completion) to an inductive paradigm (on-the-fly reasoning chain synthesis), in one chart. This is the big story of 2025 and 2026. This is what made coding agents work.
Satya Nadella says Microsoft aims to make Copilot a new operating system for work, connecting agents with business context, workflows, and systems of record. Microsoft announced 30+ new Copilot skills across Dynamics 365 Sales, Service, and Customer Insights, and introduced Copilot Managed Runtime for code hosted in a company’s environment under IT governance. Nadella describes a future where people can use Copilot’s Code experience to customize or build software connected to their existing business systems.
A QbitAI article complains that models such as Claude and GPT have become difficult to understand in coding contexts, using phrases such as 'ledger 15/15 all green' and 'one leg failed'. Online commenters suggest the main cause is chain-of-thought being compressed into 'caveman language' to save tokens, using 70% fewer tokens than plain language. The article says Opus 5.5 has improved its language ability, while Gemini is praised as one of the few models that still speaks naturally. CaveMan Skill, a GitHub project that teaches AI to speak in jargon, has 104K stars.
In an interview with Latent Space, Diogo Almeida, co-founder and CEO of TypeSafe AI, said Jev exceeded one trillion tokens processed per day less than a week after launch. Night-time traffic continues to rise, suggesting many calls come from automated machines.
Anthropic's Thariq Shihipar discusses the next phase of Claude Code on the Latent Space podcast, including Ask User Question, artifacts as persistent generative interfaces, multi-agent collaboration through Claude Tag and Projects, and Claude Mods for customizing the harness.
In his closing keynote at WeAreDevelopers World Congress North America, Simon Willison reviewed key LLM developments so far in 2026 in chronological order.
In a note dated September 24, 2026, Simon Willison said that the more time he spends working with coding agents, the more convinced he is that they make software engineering harder. Coding agents can produce astonishing results, he says, but getting their full potential requires exceptional discipline and knowledge.
A newly hired engineer at a large company says Claude Code generates everything on the team—specs, code, tests, PRDs, tickets and their handling, and reports. Engineers from L1 to L7 are all doing the same thing: chatting with Claude. No one reads the output, no one on the team likes this workflow, but they are told to produce as much as possible. Senior leaders repeatedly say pushing code is not the bottleneck and ask why progress is still slow; employees work 12 to 13 hours a day just to press Enter.
The latest GitHub Podcast examines five popular AI claims. AI-generated code still needs to be read and owned, with review effort scaled to risk; Skills and MCP address different needs, with Skills packaging team expertise and MCP standardizing connections to tools and data; and RAG is not dead, since it can coexist with agents, Skills, and MCP in one workflow. The episode also discusses judging AI fluency in hiring and how fine-tuning needs can signal codebase maintainability problems.
OpenAI has shared internal data showing how coding agents are reshaping AI research, including information on agent usage, experiment speed, task complexity, and research acceleration.
The most important skills for using AI coding agents effectively. An AI engineering skills map for coding agents. https://x.com/i/article/2095882148670832640
@thsottiaux says Codex is almost 100% reliable, has occasional resets, is open source, and “will have Astra.” Bubeck reacts: “Astra?! 👀”
Andrej Karpathy gave Opus 5 the opening passage of The Lord of the Rings and a 1M-token budget (about $10), asking it to render the story in Three.js. Opus ran for about two hours and wrote 5,500 lines of code to procedurally generate a rough but playable 3D scene.
Import AI 464 summarizes several research developments. Fable wrote a megakernel on KernelBench-Mega that achieved an 18.71× speedup with CUDA on an RTX PRO 6000 Blackwell, compared with 14.4× for Claude Opus 4.8, 11.14× for GLM-5.2, and 4.34× for GPT 5.5.
