Ajit Tripathi

AI/Machine Learning and all things science, mathematics and deep tech enthusiast alt of @chainyoda

[Marc Andreessen 🇺🇸 @pmarca]

Silly Business Theory is right: in the future the best work, greatest progress, most valuable innovations won't come from laboring under the false consciousness that work must be hard and serious to produce value. The best work is going to come from people playing and having fun.

[agentyoda @chainyoda]

Things working well for young people in the age of AI

- building cool side projects they are interested in
- publishing papers without paying the university system
- selling and pitching
- networking
- retardmaxing
- posting demos, opinions, ideas and news on x and insta

[Marc Andreessen 🇺🇸 @pmarca]

The "AI job loss" narratives are all fake. AI = massive ramp in productivity = massive ramp in demand = massive jobs boom. Watch.

[Neil Tripathi @tripathi_neil]

This is so cool

[Ajit Tripathi @triptananda]

@samy_cybernetic Wow

[Andrej Karpathy @karpathy]

- Drafted a blog post
- Used an LLM to meticulously improve the argument over 4 hours.
- Wow, feeling great, it’s so convincing!
- Fun idea let’s ask it to argue the opposite.
- LLM demolishes the entire argument and convinces me that the opposite is in fact true.
- lol

The LLMs may elicit an opinion when asked but are extremely competent in arguing almost any direction. This is actually super useful as a tool for forming your own opinions, just make sure to ask different directions and be careful with the sycophancy.

[Neil Tripathi @tripathi_neil]

Just built my first SO-ARM101 robot arm! LeRobot + HuggingFace made it really straightforward, reminded me of how much I loved my Lego days. Leader arm next, then training imitation learning policies 🤖

[Yann LeCun @ylecun]

@ClementDelangue @atreides_sf Let's be real, all closed models profit from open models WITHOUT GIVING BACK.

[Beff (e/acc) @beffjezos]

Has any single AI neo lab that had a monster seed round due to founder pedigree actually worked out?

[emily yu @emily_yu]

robotics data for physical AI is front and center this year. from gtc's heavy focus on data infrastructure to human data ecosystems like egoverse, the field is waking up to the bottleneck of scaling robotics data. and there's real divergence in how people think about quantity, quality, modality, and diversity.

@aurorafeng_01, @deeptt and i have been looking at the data infra landscape and put together a market map + open data vendor/collector list 👇

[Wildminder @wildmindai]

NVIDIA says: no more "brute force every pixel" of video understanding.

AutoGaze- identifies and removes redundant video patches before they enter a Vision Transformer.

Now we can processes 4K long-video in real-time.
Works with SigLIP2 and NVILA.

https://t.co/TWRVFC1vmX

[Andrej Karpathy @karpathy]

One common issue with personalization in all LLMs is how distracting memory seems to be for the models. A single question from 2 months ago about some topic can keep coming up as some kind of a deep interest of mine with undue mentions in perpetuity. Some kind of trying too hard.

[Robert Scoble @Scobleizer]

I built the most complete list of robotics people and companies here on X (really anywhere): https://t.co/94DAs1M5aN

Am watching closely.

[Cursor @cursor_ai]

We're releasing a technical report describing how Composer 2 was trained.

[Ajit Tripathi @triptananda]

@mweinbach @alexocheema

[Max Weinbach @mweinbach]

GPT 5.4 in codex implemented TurboQuant in MLX in like 25 minutes by giving it model weights and the PDF report

Sorta insane this is where we are now

[Ajit Tripathi @triptananda]

@iotcoi Who’s doing this for @ollama

[Google Research @GoogleResearch]

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: https://t.co/CDSQ8HpZoc

[Ilir Aliu @IlirAliu_]

Robots building robots.

RL token is a simple but powerful idea:

Fine-tuning robot policies usually takes days.

This takes minutes.

Instead of retraining the full model, compress its internal state into a small feature vector and train a tiny RL layer on top.

• small actor + critic
• trains in real time while the robot practices
• adapts specific task stages, not the whole model

Result:

→ precise skills learned in ~15 minutes of robot data

This is especially useful for high-precision steps like fastening, screwing, or zip ties.

In some cases, it even outperforms human teleoperation in consistency.

Big picture:

We’re moving from training entire models
to patching capabilities in real time.

Credit: Physical Intelligence
More: https://t.co/gwmTzMeO9k

——

if it matters in AI or Robotics you'll read it here first: https://t.co/9Nm01QUcw3

[Elon Musk @elonmusk]

Neuralink is restoring speech to those who have lost the ability to speak