@pete_millspaugh @NASA @ICANN @UofMaryland in other words, this is a TIL TLD?
Notes
- @JustJake @robcobbable 🥹 https://youtu.be/1WC8dxMC4Xw?si=WhNY98UYAS4ijC5D
- @jmduke How do you do something with some frontend interactivity? Client side react?
- @jmduke link to code or docs?
- I'm constantly falling into this pattern: 1. Start with Request / Response function 2. Upgrade to Hono for routing 3. Hono JSX for server-side routing 4. Ugh, need frontend stuff, rearchitect to client-side React app The JS world really needs a minimalistic fullstack framework. Very excited for Remix 3
- @steveruizok @samwhoo come again!
- @simpsoka great jules crossover episode - google's gotta do a party there for sure
- @MatijaSosic i want to be notion too though 😝
- @modal_labs i did think of you this morning
- @juecd__ @OpenAI @sama @onkernel shoot your shot!
- if lovable is canva, val town would like to be figma https://x.com/antonosika/status/1945899595957162104
- So exciting! The world needs more thoughtful technical folks like Pete to become writers 🤩 https://x.com/pete_millspaugh/status/1945850547354947684
- RT @charmaine_klee: PostHog uses @ValDotTown to pipe Octolens insights into PostHog - tracking brand perception and product sentiment!! Th…
Reposted by Steve
Text may be partial · Read original
- @anshnanda_ @ghumare64 @motiadev @vibecodeapp Then why do you need a typescript god?
- @anshnanda_ @ghumare64 @motiadev @vibecodeapp I've got some bad news for you about vibe coding
- i am fanatical that everyone needs a clipboard manager it takes 30 min to install & get familiar with and then your life is better forever do yourself a favor and download maccy! https://x.com/stevekrouse/status/1846257984873095637
- @mjackson I'll give it 5 hours
- @nicoalbanese10 well said! i wasn't able to find this in the ai sdk v5 docs, but makes sense given that you're still working on em
- yesterday i prototyped kimi k2 on townie, val town's coding assistant it seemed reasonably good at making val town apps, which exciting because it's ~4x cheaper & faster than claude 4 sonnet (on groq)
- @nicoalbanese10 why stopWhen vs maxSteps? What are other kinds of conditions?
- @zlumer_eth There's a dynamic interplay between top-down and bottom-up, and I find that if you create a library boundary too early, you end up annoyed having to rewrite the abstractions again & again If there are pre-existing good libraries, that's ideal Otherwise, I typically keep…
Text may be partial · Read original
- Amazon's Kiro looks awesome and seems to get a lot right (spec driven development), BUT still seems to be missing one big piece that I haven't seen any coding agent address yet: Helping you with the dynamic interplay between top-down and bottom-up design. Going up and down the ladder of abstraction. Allowing the details at the bottom bubble up to the top to change your goals, and sometimes, if you let it, even change you as a person There's an implicit bias in all these tools towards your original high level plan. This opens you up to the very real risk that your plan is misguided, and that instead of operating at the high level, the tool should be exposing the lowest level details to you, and how those constraints may impact your vision For example, I want to add persistent chat history to Townie, Val Town's AI assistant. Do I want a high level implementation plan? No! I want to understand at the lowest level the shape of the data that the frontend and backend need, and how we'll encode and decode it for storage, and which storage system is best. The intricacies at the lowest level are the important bits: which version of ai SDK are we on? Do we first want to upgrade to v5 before we lock in this schema in the database? Does the AI SDK have a guide for persistent storage? (Yes.) What does it say? What are the tradeoffs between Val Town storage options, sqlite and blob storage? Before we lock in the schema, can we test hydrating the frontend and backend with test saved data to validate it works end to end? Can we test the serialization to sqlite and back to ensure that step is bug free? This is how I as a human programmer would approach this problem. Most agent system approach is as if the human is the PM and the agent is the programmer. You give the agent high level direction, and it figures out the low level details. That's fine if you want to be a PM, but I want to *program* it with an assistant and that means doing lots of exploration at the bottom and have those concerns bubble up to top level designs Another thing the tool should help you with is understanding the true cost, now and in future maintenance, of a feature. For example, adding DMs to an app sounds conceptually simple, but any application engineer knows otherwise and should talk you out of it, not help you spec it
- @isidentical say more! can you triangulate it amongst other models?
- must-read, so good https://x.com/calvinfo/status/1945155574414397542
- @walterra @ValDotTown You can pause a cron on val town. Is that what you mean?
- No-code tools threw out the baby (coding) with the bathwater (all the complexity to run & deploy code) Val Town is, as @tylerangert phrased it, "just code"
- RT @brettberson: After 12+ years of publishing tactical company building playbooks on The Review, we’ve decided to start a brand-new public…
Reposted by Steve
Text may be partial · Read original
- @simpsoka Oh, I used Townie! I should've said that instead of "Claude 4 Sonnet"
- @yescynfria ah I was hoping you were making a deep cut reference. he's the love interest in the fault in our stars and he pretended to smoke for laughs
- @yescynfria very augustus waters of you
- @simpsoka Nope! Do you think it'd do better? And if so, why? :)
- @samwhoo Great question! In this case, the true source of the bug was that the underlying data was html. I figured by printing out an actual value of the string from the database. If it tried to write a unit test, it would've used a non-html value
- https://letsjelly.com/
- i'm immediately 2x more likely to use software with arrested development lorem ipsum https://x.com/stevekrouse/status/1944795835436249593/photo/1
- Tried to fix a bug this morning in a vibe coded project with more vibe coding Claude 4 Sonnet spun out writing a bunch more code The problem was that we were trying to html-escape content that we thought was text but was actually html. The solution was to stop escaping it. The actual diff was deleting a function call, which was less than a single-line change. It would've taken a lot of time & money for LLM to figure out what took my human brain 10 minutes How could an LLM have figured it out? It would've needed to do a lot better job instrumenting the code and actually running it to see the problem. I think this is one of the biggest bottlenecks in AI-assisted programming
- Apparently there's FIQL, RSQL, OData, CEL, according to chat, link below, but I've never seen those before and they don't seem super elegant https://chatgpt.com/c/68750946-debc-800a-98fc-d1a57cbbf695
- I want to add filters and sorts to Val Town's search (of vals, logs) but it seems like there's no common filter language spec. How can this be?? The github issues one, ie `is:issue state:open author:stevekrouse sort:updated-desc`, is awesome but custom How have y'all solved this problem?
- @MaineFrameworks @normonics Cool!
- Oh wait maybe this is @karpathy's https://eurekalabs.ai/
- I have a proposal for an AI benchmark: teaching humans, validated by a test for the human without the AI present This would solve Bloom's 2 Sigma Problem! As far as I can tell, today's AI can't help beginner students truly learn stuff Anyone working on this?
