Matt, is this true?

you should be able to hail waymo’s like a regular taxi. you should also be able to talk to the waymo to tell it where to go. the waymo should also speed and weave through traffic. the waymo should also tell you sexually deviant stories from its past.

@garrytan it’s easier to enter flow state with 0.618 miller lites i’ve found

I'm giving up drinking because of Claude Code. I need my brain to be maximally pristine so I can sling 10k LOC a day

@johnloeber Good post but I don’t think this is Jevon’s paradox. You have described the setup for the paradox (the cost of producing software is going down) but not the contradiction (no evidence that the demand for software is increasing).

Jevons Paradox At the start of this year I thought I was going to try to read two books a month. I am now reading zero books for the rest of the year. Why? AI is getting so much better that I have so much more leverage on my own time that it is becoming way too expensive to spend time on reading. Projects that previously would have taken me 1,000 hours I can now do in 20. Everything that I've ever dreamed of creating, writing, etc., that was previously out of reach, is suddenly within reach: just another 10 hours here, another 10 hours there, having and being able to use all of this AI leverage when other people aren't, feels like a huge advantage. That feels like the hottest possible iron that must be struck immediately. I've always been a standard-issue Silicon Valley nerd who's sometimes going to stay in and write software instead of going out, but now I'm feeling this stronger than ever before. The opportunity cost on my time has risen wildly and, funnily enough, this totally locks in my behavior. Instead of relaxing more, doing the same amount of work, and having more free time, I am busier than ever before. Jevons Paradox is very, very real.

@celestepoasts grok it

"asking a language model" should be a verb, like "googling" is to "looking something up on the internet with a search engine"

overheard in an sf coffee shop yeah man software engineering isn’t even worth learning these days. hell, Claude code is already dead. the only way you’re gonna make it is if you can connect to the universal consciousness

software engineering is dead... long live software engineering

@arvidkahl how big do you think a context window is? How do you think the context window works?

I don't really understand devs who jump back and forth between LLM providers. Claude, Codex, Claude, Gemini, back to Codex, then Claude. What marginal improvement can be worth the massive context shift every single time? Pick one and build. You won't have time to switch.

@pvncher if you are CEO of a company that is losing money, how do you convince investors you are worth $380B?

It’s so hard to contrast Boris’ take with Dario’s claim that: “software engineering is fully automated by 2027. You might have the 0.01% engineer left over, but that's it, the job is finished.” Why does Dario keep making such insane claims?

@_chenglou what does LLM stand for?

@_chenglou what evidence do you have that LLMs can reason outside of language? Bitter lesson strikes again

Stupidly late realization on why LLMs are so good at reasoning: human’s reasoning capability is bottlenecked by language! It’s not that languages are good at reasoning; reasoning ended up being defined by language first and foremost. The medium truly shapes the message

@DavidOndrej1 this is already known to practitioners (both that all models are overfit on these problems and that the frontier models are still clearly ahead) but it’s almost deceitful to not mention the relative cost: Minimax 2.5 is not on par with Opus but it is ~15x cheaper

@kimmonismus What’s “big”? The idea that an agent uses multiple tool calls? That search should be fast?

This is big. To quote their CEO „Why the AI ecosystem needs this: AI Agents now use multiple tool calls within their tasks. When the end-to-end task needs to be fast (like seconds), then any underlying web search tool calls need to be near instant.“

@OfirPress Difference is two-fold: 1) if you don’t trust the car, you fear a violent death vs. fearing sloppy code and 2) programming is verifiable before production in a way that driving a model 3 on the interstate is not

Autonomous coding today is at the same place autonomous driving was in 2016. You kind of feel like 'holy shit, this is incredible, none of us will code in 3 years' but in reality it'll actually take at least 20 more years till people stop programming.

@elonmusk @robustus Probably not. In the medium term, the bitter lesson will bite. In the long term, Shannon entropy remains.

@robustus Code itself will go away in favor of just making the binary directly. The next step after that is direct, real-time pixel generation by the neural net.

Turns out with claude code, my decades long strategy of NOT deeply learning: - regexs - sql - nginx confs - elaborate shell commands - advanced shell scripting - any javascript framework - perf optimization - webpack, cdns, bundlers - 1000 other things ...was entirely correct.

@emmettscully learning? yes. earning? tbd.

@mmcbrien4 worth it?

"The world needs much more compute"

what it felt like walking out of Marty Supreme (2025)

chase brain activation

my girl asked me why I’m not responding to her texts… had to let her know I’m going stealth mode

i go to the fast casual restaurant and order the same “seasonal pick” until the season changes