MitraAI

Our story...

Their UX is poor [no desktop client with MCP support to connect to the filesystem, WTH] and their availability still needs to be improved. But, man the model quality has improved substantially. It is like flicking the 0 --> 1 switch.

For the past couple of years we had written off Google. They were too conservative to enter LLMs and their Gemini models sucked for a long time. The new kids ran the show. Now, Gemini 2.5 feels such a powerful thing. #google #GoogleCloudNext

The way many showcase vibe coding projects is similar to using a shovel to dig shallow pits. There's very little value in this. Unless in a rare case when you have always needed a hole and without this modern shovel you could not do that. #vibecoding #coding

Just shared a new article on "The State of Reinforcement Learning for LLM Reasoning"! If you are new to reinforcement learning, this article has a pretty generous intro section (PPO, GRPO, etc) I addition, I cover 15 recent articles focused on RL and Reasoning Models,

Our development processes will transform radically as we generate technical debt within a single day, forcing us to implement rigorous quality standards like those the transportation industry adopted mid-20th century. Are you preparing for this shift, or will you be left behind?

Engineers will deliver 10x value for the same pay and hours, with most benefits flowing to end users. Those still working at 1x will be purged from the industry.

We have surveyed over 500 people on vibe coding and interviewed dozens and here is what the future of AI generated coding is. Most people are unable to grasp the level of change. Let us cut through the hype. A thread...

One of the “holy grails” of robotic learning is to perform generalizable everyday household mobile manipulation tasks. With a new bimanual mobile robot, our latest work BEHAVIOR Robot Suite (BRS) takes a stab at this very very hard, still yet-to-be-solved problem!🤖🦾🤩

@oscarle_x This is amount of compute, which is correlated with but not exactly cost. In general, you can think of inference-time compute as a slider, with the ability to expend more compute to improve the quality of results (for some classes of problems more than others, & up to a point).

Announcing Cold Compress 1.0 with @answerdotai A hackable toolkit for using and creating KV cache compression methods. Built on top of @cHHillee and Team’s GPT-Fast for torch.compilable, light-weight performance. Develop novel methods in as little as 1 line of new code.

What are the top 10 ranks from IITJEE, the hardest exam in India, 1998 doing today? Abhinav Kumar, rank 1, didn't even go to IIT. — IMO Gold medalist — went to MIT to do 3 majors — PhD at Harvard — MIT Professor in Math — Probably made $100M+ at Renaissance Tech 1/10

"Code is there to explain the comments to the computer." — Andy Harris

this is amazing, but more importantly, liberating! @kyutai_labs leading the charge on real time voice assistants and as a true open-science non-profit, will release the code and details. congrats to a bunch of my former FAIR colleagues at Kyutai on the launch!

HSR layout is like the Hayes Valley of SF here. Once you're tapped into the startup ecosystem, you start bumping into people in popular places often. And the talent? - just insane. Idk how VCs aren't as heavily investing here. In 1 month, I have met some insane builders here.

I got an early demo of this when I visited @uwcse a couple months ago and the ability to isolate sounds in your environment was pretty great. Nice work, @b_veluri, Malek Itani, Tuochao Chen, Takuya Yoshioka, and @ShyamGollakota!

@kunalb11 Walking talking robots.

Who’s building cool AI applications in India. Comment below with some brief description. (Serious replies only)