I would hope the one company that has ALL MY DATA might know me a bit better https://x.com/stevekrouse/status/2121426763276312865/photo/1
Notes
- It's sad that all the *.val.run links will die on Jan 10, 2027 Yours doesn't have to! We added tools to the Val Town MCP to set up redirects for your vals, which I will personally maintain for the next 10 years Point your agent at https://docs.val.town/migration/redirects
- Only for work with: - known success criteria - computationally verifiable (ie run test suite) Humans still required when the work is: - exploratory (don't know what we want) - competitive (don't want to be like everyone else) - requires human taste
Quoting this post ↗
- I have a dream that with sufficiently smart, fast, cheap AI: MEETINGS WILL BE THE WORK Now meetings are TALKING ABOUT the work. Then everybody leaves to do the work solo. Slow feedback loops! This is why craftspeople hate meetings. In the future, meetings will be augmented conversations where the work is finished before the meeting ends. No more solo homework later. Craft high level and low level details, together Yet again the folks at Anthropic are among the first to prototype this future. Great work! (Though I suspect we need 1k+ tokens per second - if not more - for it to truly be magical. I wonder when the equivalent to Claude 3.5 sonnet + Claude artifacts moment will arrive for augmented meetings!)
Quoting this post ↗
- The best part about shutting down a startup is all the lovely notes 🫶 2nd best is complimenting all your former competitors to moon Cloudflare, Vercel, Render, Netlify, Railway, Supabase... Such good hosting options! I know -- I tried to do better!
- One day we will run inference 60 times per second. Intelligence behind every pixel Very thought provoking prediction by @rsnous
- life changing
Quoting this post ↗
- seems like there are so many obvious-in-retrospect low-hanging fruit right now
Quoting this post ↗
- paul rudd was vibe coding in 2010 https://www.youtube.com/watch?v=maAFcEU6atk
- Villain for next James Bond movie: EA doomer decides that the only way to get the public to care about AI safety is releasing an AI bioweapon to give everyone a mild cold
- Neil Postman is to Marshall McLuhan what Bret Victor is to Alan Kay
- ChatGPT andriod app has trouble rendering long convos. Plz fix my openai frens https://x.com/stevekrouse/status/2115051847664767137/photo/1
- been thinking about these @Mappletons images a lot https://x.com/stevekrouse/status/2114841598546550993/photo/1
- "Creators need an immediate connection to what they're creating" – Bret Victor "Learn into async" is the same as telling a craftsperson to become a manager. Yes you can get more done, but you lose your connection to the details I care about TTFT and tok/s, so we can have 100x faster AI agents, so we can regain the flow state while creating software, and explore and learn what we're doing as we're doing it We're still in the dial up era of AI. The best way to predict the future is to invent it. Let's invent a humane, dream way to program with AI for 2029
Quoting this post ↗
- The biggest lesson we should all learn from You Can See Everything (and The Rehearsal Season 2) is that documentaries would be a lot more popular if they starred our favorite comedians I'd watch Andy Sandburg interview a tree
- If your main job is programmer, when was the last time you wrote at least 100 lines of code by hand in one day – not counting an interview setting?
Poll · Final results
- this week — 29 votes
- this month — 17 votes
- 6 months ago — 124 votes
- 1 year ago — 234 votes
Results as of . Poll end: .
- This is the closest essay to "what Bret Victor would write about the future of AI coding" if he didn't hate LLMs. Recommend reading! I've been trying to write something similar, so I know how hard it is. Maggie did an amazing job 👏
Quoting this post ↗
- New blog post: AI gets about 10x cheaper every year I propose we call this "Lee's Law" after the first person who noticed it in 2024: @kaifulee It holds up 2 years later! https://stevekrouse.com/10x
- me talking through how to read my charts we're living on a crazy exponential, which is really hard to get an intuitive feel for, so i'm trying to make sense of it with charts https://x.com/stevekrouse/status/2113338719638098247/video/1
- LLMs have been getting ~10x cheaper every ~7.5 months – holding the level of intelligence constant For perspective, this is 10x faster than Moore's Law https://x.com/stevekrouse/status/2113336071253528985/photo/1
- hacking tvs is on tiktok too https://x.com/stevekrouse/status/2113245095042433419/photo/1
- yes!!!!!
Quoting this post ↗
- https://nat.org is a delightful web page https://x.com/stevekrouse/status/2112991210562580528/photo/1
- it feels like oss is catching up to frontier faster, but the data doesn't entirely bear that out at any point in time, oss has been fairly consistently about 10% and 5 months behind the frontier (on this benchmark, at least) i was surprised by this! i guess much of the recent oss progress has been in cost reduction for good-enough intelligence, not catching up to the frontier
- Anyone experimenting with personal agent to agent protocols? ie if you want to get a meeting with me or sell me something or hire me or get hired... have your agent try to convince my agent? Proper channels for slop, with agents trying to represent us I'd be down for experiments in that direction
- if you use ai to write your tweets, i will know! (pangram chrome extension) stop doing that if you want to keep your street cred!
- inbox zero!
- the new suggested-next-prompt in codex is scary good https://x.com/stevekrouse/status/2112267241534214359/photo/1
- yesterday i read a physical book and then today typed out an essay inspired by it. very wholesome. feeling good. ai gave some good feedback on the essay, fixed some typos, but did not write any part of it ai did help me a lot with the cruft/engine of my website itself, which was nice
- i'm rooting for you, .taylorswift
- New essay: What shall I teach my son? What is the purpose of education in the age of AGI? https://stevekrouse.com/ends
- i wrote this notes to my students when i was a TA in a systems programming class i was very proud of it and still am! i wonder if it meant anything to any of my students... i thought this note was lost forever, but was delighted to find it syndicated to my new /essays page on my website by codex. delightful!
- has anyone coined a moore's law equivalent for llms that predicts how much they'll improve in how many months?
- @aliceisplaying dunno! @grok do you?
- https://stevekrouse.com
- new website! made with astra ultrafast over the last day syndicates all my writing to my personal site, all the books i listen to on audible, podcasts i've been on, and working on pulling tweets from twitter, and other scattered notes https://x.com/stevekrouse/status/2111168103229030763/photo/1
- do you hit limits? I was thinking about stumping for the $500i finally hit my usage limit last week for the first time with lots of astra ultrafast! finally got to use one of my banked rests now i am using a bit more sol ultrafast to be more conservative i'm trying to get my money's worth! hit as many limits as possible!
- Weird how astra is faster than sol.agreed!!
- i regret to inform you that the $500/month chatgpt plan is worth it astra superfast is super fun
Quoting this post ↗
- astra ultrafast is indeed pretty fast at 200 tok/s i'm surprised it's even faster than sol ultrafast https://x.com/stevekrouse/status/2110791589480288499/photo/1opus fast is pretty fast at 168 tok / s, but you have to pay-as-you go in cash https://x.com/stevekrouse/status/2110793829007954114/photo/1














