Nathan

@GoogleResearch I love me some compression

Introducing TurboQuant: Our new compression algorithm that reduces LLM key-value cache memory by at least 6x and delivers up to 8x speedup, all with zero accuracy loss, redefining AI efficiency. Read the blog to learn how it achieves these results: goo.gle/4bsq2qI

the male urge to build a GPU cluster at home

Has anyone actually used pageindex? I always see it labeled as a RAG killer but no use cases proving it.

@adcock_brett @richhomiecon Not a robotics expert or anything but I feel like all these “human-like” movements are so inefficient and clunky. Shouldn’t the bot just optimize the task for efficiency and precision rather than making it look like a human movement?

Today, Figure is showing another major milestone towards a robot in every home Running Helix 02, cleaning a living room fully autonomously

@0xSero Fair enough, I’ll have to test. The Claude code harness is extremely thorough in my experience. Hopefully droid will meet that expectation as well

Why do I recommend Droid? Look at the way it breaks down it's work, this is why Droid does better IMO. I have never seen it NOT use a plan, NOT check off the tasks, not run validation criteria. Even lower quality models do well in it because it forces them to just do what is told, in the right order, without over-complicating it. Yesterday I was seeing Claude, GPT, etc.. all make checklists, leave half of it unchecked, compact, and go on their own merry way.

@0xSero Why Droid?

Okay, so it goes like this best model provider: OpenAI best desktop app: Opencode best cli: Droid I am trying so hard to move out of Zed + Terminal but I struggle with it, tbh. This just looks amazing to me. I don't know how anyone will catch up to this.

♿️♿️♿️

March 12th is gonna be a dream

Prepped and ready for next thursday 💪🏼

Vibe coding taught me the basic stuff

With vibe coding, you accidentally learn: - how APIs connect everything. - why your .env file matters. - what localhost really means. - why deployments break but local works. - how auth actually works. - what really happens after npm install. - how backend logic flows. - how your database is structured. - why rate limits exist.

I love that I can just immediately vibe code a solution to any problem or annoying thing I run into

@rishimittal Hahaha true

@Submorf superior to clicking and scrolling for the first “grok is it true” comment

Actually really useful feature.

@ChrisLaubAI Most noticeable problem with Gemini models is the poor ability for them to use tools compared to other models

Gemini is now completely unusable I was a massive fan for a long time But it's gone so far downhill it's not even funny It hallucinates results while ignoring instructions, then when you call it out and attempt to correct it, will literally hallucinate more results and then deliver them as if it self corrected. Huge disappointment

@0xSero Interesting I’ve have very poor success with Gemini models using tools. It just wants to write a python script for everything rather than use an existing tool

I have been using Gemini-3.1-Pro for the last few days in Droid and my conclusion on it and Flash is very different than it use to be. I really disliked the 2.5-Pro variant because it was incapable in Cursor, the Gemini cli sucks as you all know and the models are unreliable outside of pure context crunching. I think there's too many issues around the service layer, including how you pay for it, and where you use it that has created a void around these models. There's many highly intelligent people daily driving Codex/Claude and they've shared enough about the model and where it shines/sucks that people tend to have better experiences. Gemini has been super consistent in Droid, it doesn't fall flat when calling tools as it does with most providers outside of Cursor. It's massive context is very useful for data crunching, as we know. Give it large datasets and let it build out plots and charts, recommend ways to refactor, etc.. The Flash version has been great as a Q&A model, it's very fast and works really well with summarize.sh as wall as in the Google search AI section. I think this model has tremendous potential to completely lead everything else. The writing style is the least cringe of any top lab. Very little negative contrasting "it's not x it's y", it uses more complex words often seems to produce very little em-dashes — I think this could be a staple.

@0xSero Thank you for sharing your perspective on this! Great info man

My most important video: youtube.com/watch?v=Adliws…

iykyk

Agent Clones > Agent Swarms trust me

You now have the ability to go from 0 cs formal education to getting Principal SWE jobs if you are a generalist, problem solver, and understand ai coding agents