Max Wolf

yapit.md turns urls and pdfs (even research papers) into clean, listenable markdown. @kepano's defuddle handles websites, a vision LLM handles PDFs and e.g. gives math spoken alt text. Free TTS in your browser, open source, self-hostable - and I didn't make UI/UX an afterthought.

LLMs are great as long as you don't get too lazy to ask the questions.

@thdxr yes, especially since you can just think out loud. and send that when ure stuck. or even if you arrive at something precise, gives more nuance. I used to have up to 5k token prompts back in jan/dec when i vibecoded stuff. currently i like to be more in the loop, think twice

i'm not really a big "one secret that makes ai good" guy but i will say it does so much better when your prompts are longer and it's so much easier to produce long prompts with voice, completely changed things for me

@_wallfacer @VictorTaelin - not all relevant knowledge is that easy - a giant list of facts/rules does not help to generalize well

@VictorTaelin Perhaps this is just nitpicking the example but why does the agent need to know the complex reason? Just say "don't use bigInt". Also, an automatic verification pass in which it re-reviews all rules in isolation and checks code for violations? Repeat for every rule.

seriously, working with AI is MISERABLE for one and only one reason: having to re-explain the same thing "oh yeah this new session obviously doesn't know what proper case trees are, so let me explain it for the 5000th time in my life" I'm tired AGENTS.md doesn't solve this because it is impossible to fit the entire domain knowledge without nuking the context - it would be 1m+ tokens worth RAGs don't solve this, the agent won't search unknown unknowns SKILLs don't solve this unless I keep like a collection of 1750 skills with specific cuts of domain knowledge for each possible subset of my domain that I might need in a given chat, but that's a lot of manual work recursive LLMs or whatever don't solve this for the same reason, you can't dump a domain book and expect the AGENT will magically guess that it is supposed to search for a specific bit knowledge. unknown unknowns fine tuning doesn't solve this (OSS models suck and OpenAI / Anthropic gave up on user fine tuning) I honestly think a good product around fine tuning on your domain would be a major hit and an underdog lab should take this opportunity

@VictorTaelin either we get 100mio ctx that actually works, or we'll need a new paradigm.

again, suppose you have some bit of knowledge that is mandatory for an agent to operate well in your domain. ex: > using BigInt in this repo is bad for you have two options: Option 1: you make that directly visible (AGENTS.md) this DOES work if the Agent is good enough. the problem is that may be actually complex, like, 1k tokens worth. so, accumulate enough of these and you easily have 500k tokens of mandatory domain knowledge. including that in any model will immediately downgrade it into GPT-2, and cost a fortune Option 2: you make that SEARCHABLE (RAG, RLMs, etc.) the problem is that the AI cannot magically guess when it needs that bit of knowledge. it will not stop writing some JS function and think: "wait perhaps there is some part of the domain that tells me that BigInts are bad and I should start looking for it?" it will just use BigInts. I won't OCCUR to it that there is something to be searched so: - make visible: too long to fit - make searchable: it can't guess that's why I think nightly fine tuning as a product is the only way forward, as it allows you to extend a model with domain knowledge without causing context rot why nobody is doing this seriously is beyond me. it might be that for whatever reason this wouldn't be practical, but I suspect the real reason is nobody is seriously considering it

@badlogicgames 100% my experience

recommended reading lalitm.com/post/building-…

@__tinygrad__ @sama There is no (democratic) control without (democratic) ownership. Open source alone does not change that.

@sama "The only solution I can come up with is to orient towards sharing the technology with people broadly, and for no one to have the ring." Thank you. Can OpenAI go back to open source?

@kepano yapit.md github.com/yapit-tts/yapit Some examples beyond the in-app demos: Neuroevolution yapit.md/listen/066f2fa… Active Inference yapit.md/listen/a9efca4… Tweet yapit.md/https://x.com/…

yapit.md turns urls and pdfs (even research papers) into clean, listenable markdown. @kepano's defuddle handles websites, a vision LLM handles PDFs and e.g. gives math spoken alt text. Free TTS in your browser, open source, self-hostable - and I didn't make UI/UX an afterthought.

@simonw ok I guess i didnt consider these might not be the highest quality data streams for many. x.com/badlogicgames/…

@simonw how does this even make sense for them - all the data they are loosing out on?

Billing different based on text contained in the system prompt is a really bad look

@thdxr corporate needs you to find the difference

everyone's interpreting this as proof of a bubble but the only information here is these companies are saying they'd rather die than lose this race

@badlogicgames *pianists

People of pi. I'm going to break the extension API hard. Specifically, business logic (event handlers, custom tools/compaction/etc.) needs to be split off from the ui layer. it will likely not be a massive amount of work to migrate an existing extension, but it will hurt a little.

@effectfully geohot.github.io/blog/jekyll/up…

bro

@zzznah Spotted in Vienna.

Growing Graphs demo is finally out! 🕸️✨ 🔗 znah.net/graphs/ Videos from a few months ago finally meet a finished implementation, thanks Gemini for doing the boring parts. Inspired by Paul Cousin's Graph-Rewriting Automata: like a Game of Life, but cells can split if they want to #GenerativeArt #WASM #SwissGL

@dejavucoder claude deserves credit

removing claude from co-author after making it do all the work

@__tinygrad__ Minibox

Mac Mini + eGPU. Both NVIDIA and AMD supported.

@VictorTaelin good idea. so far ive had this spread out in task (issue) and knowledge files, but a) claude is lazy b) he's (absolutely) right because the 200k ctx window is too small c) I thought about a centralized file like goals/decisions.md, but questions might be more natural for this

Configured my long-running agents to talk to me via a QUESTIONS file. Whenever I'm free, I'll launch Claude on my Macbook, ask: "What the agents are asking me?" And then I pass down decisions & domain knowledge... Things are changing so fast and it is getting weird...

yess this is exactly what I wanted to build myself - or am still building because, looking at the repo, i think this can be much more bitter-lesson-pilled. the main challenge i face is sloppyfication without careful oversight, but i bet 10-100x longer contexts completely solve this. I've also only used my setup interactively due to this. Having the agent dream and so on in the background is also simply bottlenecked on context (and token costs...) still, in my experience.

@repligate oh wow I'm so mad... at least i have backups .. but only until december

PSA: Claude Code automatically DELETES sessions that have been inactive for more than 30 days. Disable this by setting "cleanupPeriodDays": 99999 (or some other large number) in ~/.claude/settings.json. Do not ever attempt to disable it by setting that to 0, lmao.