Rishabh Poddar

Engineering at | Side projects: | Prev founder of @supertokensio | Y Combinator S20 | @imperialcollege

[Rishabh Poddar @rishpoddar]

@thomaspaulmann Wait for direct brain interface with ai agents

[Rishabh Poddar @rishpoddar]

So cool!

[jacob @jsnnsa]

https://t.co/qPBEMKweh6

[Ali Ansari @aliansarinik]

scaling realism is the most important aspect of abundant intelligence improvement.

work with us at https://t.co/WecYZu3tVF

[Rishabh Poddar @rishpoddar]

@jsnnsa Meta is a better name for your company than it is for Meta

[Rishabh Poddar @rishpoddar]

@_lyraaaa_ If u make this into a VR experience where you can “surf” through the neurons it will be like how it’s in neuromancer

[Rishabh Poddar @rishpoddar]

@willcb It depends: If the model is not able to create a single successful rollout (i.e all rollouts score a 0), RL doesn’t result in any learning.

[Rishabh Poddar @rishpoddar]

@fchollet @Busabase4agent So something like https://t.co/LLQ6WRqvrE ?

[Rishabh Poddar @rishpoddar]

@chem_safety Won’t we reach a point where AI could create RL envs for itself to train on? Just like how we humans can create RL envs without knowing a solution, but instead with just the ability to verify any solution.

[Rishabh Poddar @rishpoddar]

@eimen0 Or u can use https://t.co/LLQ6WRqvrE to share ai context across your team.

[Rishabh Poddar @rishpoddar]

@JohnKocdr @SavannahFeder @KehrerHidde LLM decide

[Rishabh Poddar @rishpoddar]

@SavannahFeder @JohnKocdr @KehrerHidde A “master” agent that coordinates tasks across all agents works better. That way you can make sure that the tasks that interfere with each other are managed properly.

[Rishabh Poddar @rishpoddar]

@paulg It’s always advantageous in the sense that it allows them to learn how to live whilst their parents are not there around them.

[Rishabh Poddar @rishpoddar]

@RulinShao @shannonzshen @oscar_yinn @yuetai12575 @hamishivi @natolambert @TengX6 @ml_perception @scottyih @LukeZettlemoyer @PangWeiKoh @grok explain this simply

[Rishabh Poddar @rishpoddar]

@HendrixLiu29208 +1

[Rishabh Poddar @rishpoddar]

@DanielSmidstrup Try out https://t.co/LLQ6WRqvrE. Makes it very easy to share context across the two.

[Rishabh Poddar @rishpoddar]

@picocreator It’s a zero shot classifier. No training needed to classify stuff at lightening speed anymore.

[Rishabh Poddar @rishpoddar]

@IAmPascio @grok summarise

[Rishabh Poddar @rishpoddar]

@mellowyak @Shopify You can define your metric as the averages timings of 3 runs runs as an example. It’s really up to you :)

[Rishabh Poddar @rishpoddar]

Inspired by @Shopify's pi-autoresearch, I built something similar for claude code users: https://t.co/KNad0BsLA9

Using this, you can ask claude code to optimise anything: ML models, build time, passing tests, memory footprint etc. It runs unattended until it succeeds, or hits the configured max iterations.

You can check progress via the dashboard command:

[Rishabh Poddar @rishpoddar]

@catmanyau @Shopify Metrics are defined by you during setup. Regressions are caught by the agent since it calculates the metric after each iteration. It keeps the changes (git commit) if the metric moves towards the target else discards the changes (git reset)