Theo's Rust tsc, and the $400k that got nowhere
Five months and about $400k of Codex tokens got Theo nowhere on a Rust rewrite of the TypeScript compiler. About $20k of Opus finished it in two weeks, and he says he has not read a line of it.
The big disagreement
Thorsten Ball: preaching craft right now is the wrong advice
Thorsten Ball says senior engineers who present themselves as Stewards of the Craft of Programming are telling young engineers the wrong thing during a period of real structural change. His argument is not that deep knowledge and good architecture stopped mattering, it is that nobody truly knows what will matter, and the sermon lets secure people feel comfortable when they should not.
But it's the wrong advice to not rethink everything right now.
Receipts
Theo ships tsc-rs, the compiler Opus finished after Codex burned $400k
Theo released tsc-rs, an open source drop-in replacement for tsc, its type checker and LSP, written entirely in Rust by agents. He says five months and roughly $400k of Codex tokens got nowhere, then roughly $20k of Opus got there in two weeks. He also notes `bun check` landed the same day and is probably the right choice for most Bun apps, with tsc-rs faster only on Effect ts checks.
I have not read a single line of the code.
Theo runs his own usage numbers: Sol on the $500 plan just barely fits
Theo pulled a week of his own prompt counts and says he can probably main GPT-6.1 Sol Ultrafast on the $500 plan while just barely hitting the limits. He excluded the ts-rust work as a long running YOLO mission, says 583 a week is his total usage, and ran the same comparison against Opus on his $200 plan.
Astra feels pretty close to unusable.
How they actually work
Pocock: tell the agent to care about the codebase, not just the task
Matt Pocock argues agents are already good at strategic programming and are simply not trained to care about it, so the fix is instructing them to care about the codebase itself rather than only the task at hand. He adds that most people do not care much either, as long as the code works.
So all you need to do is to tell your agent to care about its own codebase.
Model and agent watch
Haiku 5.5 lands at roughly 75% less to run, with 100k context
Anthropic shipped Claude Haiku 5.5 and says it costs around 75% less to run than Haiku 4.5 on average. Boris Cherny says it carries a 100k token context and puts the cost gap at 10x cheaper than Haiku 4.5. Theo's reaction to the release was that nobody had mentioned it to him.
Haiku 5.5 is a good Haiku.
Open tabs
Most Vercel traffic is now bots
Guillermo Rauch reports 58.18% of Vercel network traffic is bot-originated over the last 30 days, up from 32% in January 2024. Over 60% of deployments are now agentic.
Karpathy says the capability gap is now a funnel
Andrej Karpathy estimates about 75% of people have barely touched LLMs, while roughly 20M see agents finish projects that used to take weeks. He says that gap is widening.
Frequently asked questions
Why did Theo switch from Codex to Opus for the tsc-rs project?
Codex spent roughly $400k over five months without finishing the TypeScript compiler replacement, while Opus completed it in two weeks for around $20k. The same setup with a different model moved from stalled to shipped at a twentieth of the spend.
What does Matt Pocock say about making AI agents write better code?
Pocock argues that agents are already good at strategic programming but are not trained to care about codebase quality. The fix is instructing them to care about the codebase itself rather than only the immediate task at hand.
How much cheaper is Claude Haiku 5.5 compared to the previous version?
Claude Haiku 5.5 costs around 75% less to run than Haiku 4.5 on average, with one estimate putting it at 10x cheaper. It includes a 100k token context window.
What percentage of Vercel traffic now comes from bots?
58.18% of Vercel network traffic is bot-originated over the last 30 days, up from 32% in January 2024. Over 60% of deployments are now agentic (controlled by AI agents).
Why shouldn't junior engineers focus on craft and fundamentals right now?
Thorsten Ball argues that during a period of real structural change, nobody truly knows what skills will matter. Senior engineers preaching traditional craft practices may be giving advice that makes sense for them but wrong for juniors deciding what to learn next.