Dhruv Saini ✈️ ICML

[Dhruv Saini ✈️ ICML @oofbaroomf]

just made PromptMaxx - it's for programming alongside @cline with Qwen Coder on
@cerebras when you need stuff thats only available on web LLM chat apps.

it uses a fast model to quickly apply pasted edits to your files and gathers llm-friendly context to copy

link in bio

[Dhruv Saini ✈️ ICML @oofbaroomf]

@natolambert Interested

[Dhruv Saini ✈️ ICML @oofbaroomf]

@jxnlco I want to pick myself because sometimes the model doesn't know how complex a question is until it reasons about it a lot.
I would trust it if I saw it and it was easy to change and it adapted to my preferences.
I use extra high most of the time, medium sometimes.

[Dhruv Saini ✈️ ICML @oofbaroomf]

@tszzl @runclepennybags this isn't consistent still.
distillation is training your model on a commercial model's OUTPUTS (which is just a bunch of text, sorta like a book). you aren't copying the model itself.
similarly, you are training on the copyrighted books, not copying the author's brain.

[Dhruv Saini ✈️ ICML @oofbaroomf]

@Sauers_ oof-baroomf

[Dhruv Saini ✈️ ICML @oofbaroomf]

@i2cjak you can also buy a bunch of random flcos fl or sony ecx microdisplays off of aliexpress for really cheap (they are designed for viewfinders, but work great for your use)

[Dhruv Saini ✈️ ICML @oofbaroomf]

imagine if all the ai labs teamed up.

would they make a model with anthropic's data recipe, google's tpus/long context/multimodal understanding, openai's rl infra
or
a model with anthropic's safety policies and api reliability, gemini's mental breakdowns, and 4o's sycophancy?

[Dhruv Saini ✈️ ICML @oofbaroomf]

@nilszeilon Probably Vivaldi.. - you can customize it with CSS themes, source-available, tons and tons of keyboard shortcuts, command chains, custom widgets, etc. But it's not open-source.

[Dhruv Saini ✈️ ICML @oofbaroomf]

a 15 year old takes ~12 million watt hours of food to raise.

deepseek v3 took ~1 billion watt hours of GPUs to train.

[Dhruv Saini ✈️ ICML @oofbaroomf]

i think the gap between humans and llms can be explained by this, and fixing it is necessary for agi

the ability of humans to continually learn with a sort of merged memory is very important; llms have jagged cliffs between pre-training, post-training, and in-context knowledge

[Dhruv Saini ✈️ ICML @oofbaroomf]

@FrameworkPuter does nobody else see the 13 dots?

[Dhruv Saini ✈️ ICML @oofbaroomf]

i think this metr benchmark may end similarly.

[Dhruv Saini ✈️ ICML @oofbaroomf]

why does it seem like ai progress is plateauing, yet we have graphs like this? i think it's because of three reasons

(1/4)

[Dhruv Saini ✈️ ICML @oofbaroomf]

arc-agi didn't flex the fact that there was no earlier progress on their benchmark, they flexed the fact that initial scores were low, and we already saw them fail (o3 is not agi despite getting 87.5% on arg-agi-1)

[Dhruv Saini ✈️ ICML @oofbaroomf]

awwww

[Dhruv Saini ✈️ ICML @oofbaroomf]

@pli_cachete except for cost though