AGI will be achieved when an LLM gets 100% on SWE Bench the way a human would: by deleting all the tests it can't get to pass
Notes
- RT @anselm_io: Rare opportunity to personally influence a future Alan Kay talk! He tends to revisit similar themes so I’m always excited f…
Reposted by Steve
Text may be partial · Read original
- @redblobgames @steveruizok Clearly!
- @SalehOfTomorrow Yeah maybe we can ease folks into it
- @SalehOfTomorrow We build for engineers We should probably teach debugging for aspiring engineers We are not the product for anyone who is allergic to code
- One possible path forward for Val Town is to be 2 and 3 of the list above, and provide good observability for 4, and let someone else (bolt, lovable, cursor, devin, etc) hook it all up. We can be just the development and deploy platform
- I think we're not that far from agentic coding with very little human involvement We need: 1. Smart LLM 2. Good sandbox 3. Constrained stack 4. Good feedback loops 5. Good tools And I think all the pieces are there. In 2025-2026 I think someone will put it together
- BUT they were onto something! Most of the world doesn't know how to debug And there really is more to automate here We've bottled up intelligence. Let's get it in a tight feedback loop with the outputs of its actions, in a safe sandbox
- Those users (for the most part) were non programmers who wouldn't even read the error messages They and AI would go down some really stupid rabbit holes and waste a ton of money. Adding rate limits seems to have put a stop to this silly behavior
- Yet immediately our users begged us to automatically send the errors back. They complained that making them hit the "ask townie to fix it button" was stupid drudgery we were forcing them to do Kinda like how Tesla autopilot keeps checking your hands are on the wheel
- Right now in Townie, we detect errors. We then surface those errors to you, and you can optionally decide to send them back to the AI We did this to minimize cost and also to ensure it doesn't go off the rails and do something scary
- That might sound depressing (to have less human involvement) especially if you (like me) love the activity of programming But it's not as bad as you think if you unpack it
- I think it's important to acknowledge that full agent coding is not there yet. Fast feedback loops with humans are the way, today But we shouldn't forget that agentic loops with less and less human involvement is the goal https://x.com/HamelHusain/status/1880129026435756153
- RT @noampomsky: Hello!! Reminder that I'm doing another round of my matchmaking experiment where I get 1000 people to answer 70 questions a…
Reposted by Steve
Text may be partial · Read original
- the problem with effective altruism is that they thought about it too much, unfortunately
- @holtbt now the competition's getting serious!
- On Monday – MLK Day, Jan 20 – Cerebras & Val Town are cohosting a virtual hackathon The 60-second Hackathon! Inspired by http://cerebrascoder.com, which creates apps in seconds There's also a track to create Cerebras AI apps hosted on Val Town Come join! Details below... https://x.com/cerebras/status/1880367618386702748
- @rough__sea miss you too!
- @moXplOd @brianships @vjisherenow neat! was just looking for something like this
- unexpected townie use-case: creating illustrations! https://x.com/charmaine_klee/status/1880307658013466712
- RT @appodeepa: Chatting with AI via email is a neat idea so built it with @ValDotTown. Super simple to use: 1. Add your OpenAI API key 2…
Reposted by Steve
Text may be partial · Read original
- @TheFlowAgency @alexiglesias_me Thanks Glenn!!
- @alexiglesias_me been a while since i played with @webflow! excited! 👋 @callmevlad
- RT @alexiglesias_me: Going live with @stevekrouse to explore https://val.town in 5 min, see you there! https://www.youtube.com/watch?v=Cm-yqc8Hve8
Reposted by Steve
- Doing a livestream tomorrow morning at 9am ET. Come join! https://x.com/alexiglesias_me/status/1880006123157352767
- RT @charmaine_klee: 👇 receive reddit keyword alerts via sms (twilio) https://www.val.town/v/charmaine/redditKeywordSMS
Reposted by Steve
- @simonw @mattpocockuk @ValDotTown @rough__sea is cooking up something new & awesome here!
- RT @charmaine_klee: 🔔 get a discord notification every time someone mentions your product / company / favorite keyword (like "taylor swift…
Reposted by Steve
Text may be partial · Read original
- @DormRoomFund @cursor_ai no way - is cursor a dorm room fund company? 🤯
- @yawnxyz @alfongj @raw_works @pomdtrr @tmcw @iamseeley @mxmcd @neverstew @jdan @Willthereader @vawogbemi i think the 'bug' in this case is that we default to vals as unlisted, but many of the vals alfonso creates (and others too) could be public, but folks just forget to flip the switch
- @alfongj @raw_works @pomdtrr @tmcw @iamseeley @yawnxyz @mxmcd @neverstew @jdan @Willthereader @vawogbemi haha so sad - yeah i think the way that datasource works, it can't pick up unlisted vals and if i were to do this query on the backend, it'd kinda break the contract of what an 'unlisted' val means, ie that we don't tell other people you have it
- @wasauce @ValDotTown @raw_works @pomdtrr @tmcw @iamseeley @yawnxyz @mxmcd @neverstew @jdan @Willthereader @vawogbemi Great question! Sadly, I think that sort of data is probably private. It'd be weird if your hosting company bragged about your usage analytics, right?
- The most prolific Val Town users, by # of public vals @raw_works recently asked me which power users he should check out. This list has many of the best ones :) https://x.com/stevekrouse/status/1879282475823952138/photo/1
- https://stevekrouse-prolific.web.val.run/
- RT @charmaine_klee: i joined @valdottown! ever needed to run quick scripts, demo your APIs, automate internal workflows, or test backend…
Reposted by Steve
Text may be partial · Read original
- @daniel_levine @bradgessler @github Yes! https://docs.val.town/integrations/github/receiving-a-github-webhook/ Shoot me a DM if you get stuck - also happy to jump on a quick call to onboard you / pair program
- @anselm_io Maybe he'd have no interest in this, but I'd love to hear him riff on the new LLMs tools. Does he use any of them for anything?
- RT @arjunattam: the thing I like about @valdottown is how it enables “logs driven programming” logs from all executions (incoming requests…
Reposted by Steve
Text may be partial · Read original
- RT @prashhtrivedi: You know the moment when you write a love letter to service you adore, and the CEO gives the shoutout? That's my kind of…
Reposted by Steve
Text may be partial · Read original
- @jdan Yeah I misunderstood how it used *since*
