Theo
Theo is a content creator focused on technology topics, primarily producing content and commentary in the areas of artificial intelligence, programming tools, and software engineering. His output covers AI-assisted programming, developer tool evaluations, and analysis of relevant technical research. He has established a notable presence in the technology community, particularly through in-depth discussions of cutting-edge AI programming practices.
Core Facts
The reviewer recommends letting Opus 5.5 handle implementation and long-horizon building, while assigning GPT-4.1 Sol to review, root cause analysis, codebase investigation, and task triage.
Sonnet 5.5 is on par with Opus 5.5 in offensive and defensive cybersecurity capabilities, making it the first model in the Sonnet series to adopt equivalent cybersecurity protection mechanisms.
Timeline (last 90 days)
Theo claims that T3 Code uses 100 times less data than Codex when switching threads and works smoothly even on poor mobile networks.
T3 Code is an open-source AI coding assistant maintained by three part-time developers that can manage multiple AI subscriptions and tools.
Replicating an existing project has clear success criteria and is easy to hill climb; whereas creating something nonexistent from scratch makes validating the design the real difficulty.
Theo maintains two threads: one endlessly advancing the TypeScript rewrite, and another periodically asking about the current bottleneck; through this dialogue he discovered that Claude had built its own queue to avoid excessive concurrency causing memory exhaustion, and found that insufficient compute was the main drag.
After adding machines and connecting to the cloud, wait times dropped from over a dozen hours per round to the hour level.
Theo considers TS Rust worthwhile because it can significantly shorten build times for real projects, while he describes the rest of the projects as "messing around."
Theo built projects on Claude including TypeScript rewritten in Rust (TS Rust), an alternative compiler that accelerates Rust compilation a hundredfold (Lust), and a skateboarding game Skate 3 Bench entirely generated by Claude, cumulatively running Claude for close to 4000 hours.
Theo first had TS Rust reach output completely identical to the official compiler, then required it to optimize to twice as fast as the official one; TS Rust ultimately produced about 1800 revisions and 7884 commits.
Developer Theo's codebase analysis benchmark shows that Argon is the only model with a false positive rate exceeding 50%.
After testing, programming blogger Theo pointed out that more than half of Gemini 4's judgments about his codebase were wrong, making it the only one of the 7 tested models with more than half errors.
40 more timeline events
All Facts (20)
Theo认为如果信息在代码库中能找到就不应放在AgentMD里，因为模型已经非常擅长通过bash命令和工具探索代码库
90%UnverifiedTheo建议在AgentMD中写入指令让Agent在遇到令其惊讶的事情时提醒开发者，目的是利用Agent反馈发现代码库中的架构问题
90%Unverified在AgentMD中告诉模型不要使用某技术（如tRPC），模型反而会更多地想到该技术，类似于'粉红大象'效应
70%Partially VerifiedThe reviewer recommends letting Opus 5.5 handle implementation and long-horizon building, while assigning GPT-4.1 Sol to review, root cause analysis, codebase investigation, and task triage.
65%Partially VerifiedSonnet 5.5 is on par with Opus 5.5 in offensive and defensive cybersecurity capabilities, making it the first model in the Sonnet series to adopt equivalent cybersecurity protection mechanisms.
65%UnverifiedAgent instructions in documentation decay as the codebase evolves, and if not updated in sync, they may mislead the Agent, producing worse results than having no instructions at all.
60%UnverifiedTheo claims that T3 Code uses 100 times less data than Codex when switching threads and works smoothly even on poor mobile networks.
50%UnverifiedT3 Code is an open-source AI coding assistant maintained by three part-time developers that can manage multiple AI subscriptions and tools.
50%UnverifiedTheo considers TS Rust worthwhile because it can significantly shorten build times for real projects, while he describes the rest of the projects as "messing around."
50%UnverifiedTheo maintains two threads: one endlessly advancing the TypeScript rewrite, and another periodically asking about the current bottleneck; through this dialogue he discovered that Claude had built its own queue to avoid excessive concurrency causing memory exhaustion, and found that insufficient compute was the main drag.
50%UnverifiedAfter adding machines and connecting to the cloud, wait times dropped from over a dozen hours per round to the hour level.
50%UnverifiedReplicating an existing project has clear success criteria and is easy to hill climb; whereas creating something nonexistent from scratch makes validating the design the real difficulty.
50%UnverifiedTheo built projects on Claude including TypeScript rewritten in Rust (TS Rust), an alternative compiler that accelerates Rust compilation a hundredfold (Lust), and a skateboarding game Skate 3 Bench entirely generated by Claude, cumulatively running Claude for close to 4000 hours.
50%UnverifiedTheo first had TS Rust reach output completely identical to the official compiler, then required it to optimize to twice as fast as the official one; TS Rust ultimately produced about 1800 revisions and 7884 commits.
50%UnverifiedThe sharp increase in tokens generated per prompt is a manifestation of people using models for larger and harder tasks after they became cheaper, better, and more capable, and rising bills and productivity gains are related.
50%UnverifiedMeasuring AI cost should consider three dimensions: how many tokens can be purchased per dollar, how many tokens are consumed to complete a task, and how much it costs to solve a problem at a specific level of intelligence.
50%UnverifiedThe author considers GLM 5.3 Flash to be a better model than GPT-5 in its own right, with stronger visual processing and better tool calling, particularly crushing GPT-5 in performance during hours-long agentic workflows.
50%UnverifiedTheo speculates that the fundamental divide between OpenAI and Anthropic is a difference in engineering versus research culture: OpenAI is engineer-led, while Anthropic places comparatively less emphasis on software engineers
50%UnverifiedMetrics-driven optimization carries risks: agents targeting reductions in React commits or V8 call counts tend to cut network requests that don't affect interactivity, leading to overly aggressive caching with insufficient invalidation
50%UnverifiedDuring the Sonnet 4 era, Anthropic heavily used models to generate code, accumulating a codebase full of slop, and is now using stronger models to go back and clean it up
50%