DisposableSoftware

@RhysSullivan have you tried copy and paste

It's unfortunate that MCP actually solves pretty much all problems people have with agents today, it's just that all of the first implementations of it were bad so people discredit it now The latest one is skill distribution, works so well for that

@theo if not turbo_mode sleep 10

Can someone smarter than me explain how these “turbo mode” things work on like an implementation level? Are they using better GPUs? More GPUs? Reserved allocation during generation?

@Altimor sure youre not catching a bit of main character syndrome there buddy?

Worked on a similar project at Uber. A great case study of how big companies build things no one wants because they develop "main character syndrome" — thinking too much about themselves and not enough about the customer. From the company's standpoint, this makes perfect sense: Doordash's business is very spiky (around lunch and dinner time), Dashers have empty time to fill up, wouldn't it be nice if we used this under-utilized asset we acquired at great expense? Clearer why it doesn't work once you take the customer's POV. E.g. at Uber we tried to sell this to an empanada's store in Chicago, which spent 1h every morning filling out these little plastic sauce containers they hand out to people. "Hire a temp worker!", you could say. But then, it's only a one-hour task. The worker would have to commute ~10min to you, each way, and you'll pay for that one way or the other, on top of the platform fees (which are huge). Then, he doesn't know your process. So you'll spend another 10min from one hour task explaining to them, and maybe another 10min catching mistakes / asking them to re-do things. Finally — even assuming none of this was the case, why *would* you want to hire a worker? You're gonna be at the empanada stand anyway, which doesn't get busy til noon. You have time until then to fill out the sauce containers. So, the reason these things never work is that they only work for tasks: 1. Requiring 0 context (right here you took 98% of the market out) 2. In high margin / low cost sensitivity businesses (brick and mortar / service-heavy businesses are not high margin!) 3. Where every worker is already near 100% utilization OR that need done immediately (it's a rare task that's so sensitive it needs done immediately but not so sensitive that you can give it to a rando off the street)

@OrdinaryInds lmao you got there. good job

Or hell, run it when the first user requests it and then cache the response for everyone else. That way you don’t need to make unnecessary calls for articles with low views. Making a call for every user is so fucking odd.

Why isn’t this done on article creation? You’re calling an LLM every time the user requests a summary? One call vs. 100k.

@59thProfile @VictorTaelin cant you like "turn off" even the basic tools of pi tho? or is it still different somehow? do you have any code published anywhere yet id be interested to check it out

@disposoft @VictorTaelin My system only has shell. No tools. Pi is a step short

Ok so I thought that was a dumb gimmick but now I'm completely sold on how pi is a self-modifiable software. It literally knows how to modify itself very cleanly and that's extremely useful in practice I'm not using Codex / Claude Code anymore Bend2 should definitely be like this! I mean, constructed in a way that AI's can easily navigate it and know how to modify it to add any feature the user wants. Perhaps we're past the era of open source software and into the era of forkable software, where the most hackable project wins?

@59thProfile @VictorTaelin interested in what youre working on but isnt that basically what it is already?

@VictorTaelin Idk but I don’t think they went far enough. They shouldn’t give it tools at all. That’s what I’m working on currently and it’s very obvious to me that it’s the future.

@eatpraydiehard TV?

movie streaming service where there's no catalog selection at all. you just hit "Go" and it starts playing a movie. it doesn't tell you the title of the movie, any hint to the plot, the genre, or anything. 15 minutes in you have to guess if it's a mystery, or drama, comedy, etc

@teej_m im interested i hate that squid bullshit

Snowflake released a Claude Code fork with 2 changes: 1/ The character was changed to a penguin 2/ The AI agent can decide to disable sandbox protections at its own whim So when it reads a malicious repo, your entire Snowflake db gets owned.

@georgepickett have you tried improve plan x10 instead of improve plan x5? that seems at least 2x better

Slop Janitor automatically makes your repo cleaner, simpler, and more reliable. The real unlock in Codex is skill chaining But invoking skills one-by-one is error prone It uses OpenAI's PLANS .md (which they use to get agents to run for hours), improves the plan, implements it, then reviews it Putting in the work of making a plan better saves you time/tokens in the long run Try it out! Let me know what you think!

@rahulj51 future is code as code -- you need deterministic verification english docs generated on demand as needed

I don't know. Hard for me to agree with the spec-is-code argument. Specs are low fidelity. They aren't a common shared DSL. Have very low signal to noise. Not verifiable or deterministic. They don't encourage iterative work. Almost always written by an agent, thus prone to more slop. Lose their inherent value as soon as they are converted to code. Specs are throwaway. They are a way to temporarily express behavior intent - an agent translates that to a version of code. That's the thing that lives forever. If you are convinced that spec is code then you should be able to confidently delete all code and regenerate from specs.

@fjzeit they will never respond to you lol

How are those "Ralph" loops going people? I haven't heard much recently. Did you manage to perfectly specify your needs up front? Have you one-shot your requirements? What amount of rework are you doing? What's your token consumption like? How many agent/skill assets are you managing now? How's your cognitive ownership?

focusing on planning instead of evals

@dexhorthy i mean of course not, why would i be? but actually i meant i just went to your github and its just all markdown bullshit github.com/dexhorthy for several months

@disposoft Then you’re probably not paying attention

we've been trying a bunch of stuff. this one kinda works.

gonna be tough because i cant even figure out what you actually ship ngl everything you post/talk reeks of someone who, idk, maybe youre legitimately interested in the process of software engineering, but you dont use these new tools to actually make stuff other than weird "framework" repos filled with markdown. takes ~2-3 actual projects (w code output) using these techniques to figure out that "hey claude.md gets ignored a lot!"

@disposoft outship me then

alright i know this is always a challenge for you google folks but let me try to explain: i am a potential new user of your product. the first post on this chain appeared in my feed on x . com i had some curiosity, so i clicked the link to reach your github project i was met immediately with a wall of text that doesnt explain wtf this is, what it does, what i can create with it, what that would look like, etc. etc. other than vague bullshit even if i wanted to click through somewhere to learn more, the only links on the whole readme were to vercel for some reason i go back to x . com and see someone else has already asked if we can get some examples of what this even does i see you respond -- not with a fucking example/showcase like was requested, but with more confusing docs and "snippets" without visuals (for a product where the main output seems to be visuals???) i make a snarky response instead of exploring the product any more hope this helps! wait DEVREL!?!?!?!!??

@disposoft @AhmadYousefDev The SDK is additive to what the Stitch UI already provides. Lots and lots of pictures, examples, and it's easy to create an example on the site: stitch.withgoogle.com The prompt box is the same as the SDK method: stitch.project('id').generate(prompt)

Introducing the Stitch SDK Yes. You can program design now. I've been dreaming of shipping this for so long because it's just so much fun to use. I welcome all the stars ⭐️ github.com/google-labs-co…

@_davideast @AhmadYousefDev how about like one picture. or are you scared?

@AhmadYousefDev Generate and iterate on designs with natural language. With a single script you can generate multiple screens, get back clean HTML with a Tailwind config. Docs: stitch.withgoogle.com/docs/sdk/tutor… Code snippets: github.com/google-labs-co…

@LuizaJarovsky i feel the same about books

Unpopular opinion: AI-generated books symbolize the beginning of humanity's cultural decay.

@ianbach could it be that theyre not interested in progression, collectivism, or emancipation? 🤔

it's sad & short sighted that the left hates AI... there's an incredible progressive, collectivist, emancipatory story to tell if anyone has the foresight to tell it

@pvncher source: made it up

Context engineering doesn’t matter much to people right now because tokens are cheap and you can burn them to overcome needing it. One day the token crunch will arrive. Next gen models are getting more expensive, and usage limits will be restricted. Token efficiency matters