Avinash

AI, I Guess!

AI Research, Robotics, and mostly working on new frontier applied research.

[Avinash @a12348885]

In AI research, you cannot simply request LLMs to code or a plan a system. You must understand fundamentals to create something new and unique. X'ing this experience at ~3:45am lol

[Avinash @a12348885]

Very cool engineering work!

[Elon Musk @elonmusk]

Turns out the future does have flying cars

[Avinash @a12348885]

@FrancoisChauba1 @aiprof_mykel Great speakers, and loved the knowledge shared by both the Professors.

[Avinash @a12348885]

@pigflyingman @paulg Well said! I believe they already have these strategies in play internally, and agentic use pipeline for Amazon purchases. It’s just that they are holding their cards for now, I believe.

[Avinash @a12348885]

Time magazine will give the Person of the Year award to AI 🤣🍿

[Avinash @a12348885]

@FrancoisChauba1 @imrazaa AI slop mostly because of our post training strategies. The people who do RLHF and any other human driven feedback mostly don’t know what’s important in Research and frankly it’s not easy.

[Avinash @a12348885]

Sometimes the frontier LLMs are doing wild things if I give access to my system. I’m like “you should ask my permission”. It says yes and does the same thing that I told it not to do. 😅

[Avinash @a12348885]

@FrancoisChauba1 The confident tone of the nonintelligent ai claiming it does not glitch often should raise concern lol.

[Avinash @a12348885]

Same experience. I'm trying to do frontier research on a few important domains, and all I got from LLMs was gibberish most of the time.

[Omar Khattab @lateinteraction]

i don't see enough of this, so here again:

we're all wondering what the future of work looks like with AI progress, but I'm in constant agony trying to extract OK work from frontier models at "Ultra" settings

with levels of detailed context and elaborate feedback that i’ve never given a human in my life (… considering that my job is literally to give feedback)

at least up to Sept 2026, public frontiers models are - as they were in 2020 - systems that know a whole lot but are fundamentally quite dumb and unadaptive

[Avinash @a12348885]

Very funny. Bots are wrong about you lol. Psyops are expected because of the AI Race. When I talk to founders in the Bay Area, most of them are using Chinese open-source models since 2025 because of a lack of OSS models from the US in the past. Hope this adoption changes soon too.

[Avinash @a12348885]

@FrancoisChauba1 Probably, it also said “You’re right” too! 😂

[Avinash @a12348885]

@lateinteraction I didn’t find coding agents to be helpful in ai or frontier research. Have to stop using them. 😂

[Avinash @a12348885]

I bike to work now. Saving some money for the gpu costs that I incur because of ai experiments.

[Avinash @a12348885]

@pilatescapital below is the reason why I asked you to give access to your product for me lol. current LLMs from other providers are so bad and tend to do things a regular sane human being would never do. 🙃

[Avinash @a12348885]

I used frontier LLMs for an entire day today. It burned so many tokens and created 0 useful information for me to work with on an important RL problem. I would rather write all of this code from scratch instead of getting "You're right" every single time I correct an LLM.

[Avinash @a12348885]

I used frontier LLMs for an entire day today. It burned so many tokens and created 0 useful information for me to work with on an important RL problem. I would rather write all of this code from scratch instead of getting "You're right" every single time I correct an LLM.

[Avinash @a12348885]

Well… this is a major issue when someone is getting spammed. Also, the labs wants to make sure their data collection is classified well ie from human and agent data when they pretrain. 😅

We don’t need to be so strict about it when LLMs are used for a bit of writing help. Similar to the coding help we get from coding agents.

[Avinash @a12348885]

@FrancoisChauba1 @eve_bouff @mattkang Looking forward to meeting you in person.

[Avinash @a12348885]

@JeffDean It was pretty cool knowing you. Once an Uber driver was talking about you and how kind you were to him. It was a fun conversation. Love the way you approach life while still being humble.

[Avinash @a12348885]

open sourced a calendar agent that I use everyday for personal productivity: https://t.co/bQJdKJU0Nq