Vishal Misra(@vishalmisra) 's Twitter Profile Photo

LLMs cannot “recursively self improve”

This falls out from the conceptual matrix described in section 2.1 of our paper below. Any LLM can only approximate this matrix, so it has rows missing. For “improvement” it needs to fill out missing rows (1/n)

arxiv.org/abs/2402.03175

account_circle
Vrigu(@vrigu) 's Twitter Profile Photo

Vishal Misra ESPNcricinfo There’s more to it Vishal.
Dravid was the one who volunteered on the behalf of all three. Sachin and Ganguly had to be “persuaded” by Dravid. Sachin definitely wanted to play.
Here’s part of it directly from the horse’s mouth.
hindustantimes.com/cricket/rahul-…

@vishalmisra @ESPNcricinfo There’s more to it Vishal. 
Dravid was the one who volunteered on the behalf of all three. Sachin and Ganguly had to be “persuaded” by Dravid. Sachin definitely wanted to play. 
Here’s part of it directly from the horse’s mouth. 
hindustantimes.com/cricket/rahul-…
account_circle
Kody Technolab(@kody_technolab) 's Twitter Profile Photo

Catch Vishal Mishra live on May 12th, Tuneland 2024, . 📍
👉 kodyrobots.com/events/b-praak…
With his heart-touching melodies like Pehle bhi Main and , experience the futuristic touch of our surveillance robot and serving robot !

account_circle
Vishal Misra(@vishalmisra) 's Twitter Profile Photo

Revisionist history. All the senior players voluntarily dropped out of the T20 World Cup and in fact it was Sachin who suggested Dhoni be made captain.

Revisionist history. All the senior players voluntarily dropped out of the T20 World Cup and in fact it was Sachin who suggested Dhoni be made captain.
account_circle
Cedric Yau(@ctyau) 's Twitter Profile Photo

Vishal Misra martin_casado How did you generate the token probabilities for the input prompt? This looks a lot like the OpenAI Playground for Completions output but in the playground, I can only get logprobs for the generated tokens. Was this slide an overlay?

@vishalmisra @martin_casado How did you generate the token probabilities for the input prompt?  This looks a lot like the OpenAI Playground for Completions output but in the playground, I can only get logprobs for the generated tokens.  Was this slide an overlay?
account_circle
Cedric Yau(@ctyau) 's Twitter Profile Photo

This talk Vishal Misra was an eye-opener to how few shot prompting works under the hood to change output token probability distributions. I forgot LLMs don't produce tokens directly. They produce embeddings from which nearest neighbors tokens are chosen.

youtu.be/sLodkyHlQhY

This talk @vishalmisra was an eye-opener to how few shot prompting works under the hood to change output token probability distributions. I forgot LLMs don't produce tokens directly. They produce embeddings from which nearest neighbors tokens are chosen.

youtu.be/sLodkyHlQhY
account_circle
Kody Robots(@kody_robots) 's Twitter Profile Photo

Experience the soulful tunes of Vishal Mishra live at Tuneland 2024 on May 12th in GIFT City! 🎶
kodyrobots.com/events/b-praak…
Don't miss this blend of great and innovative ! 🤖🎤

account_circle
Raghav Bikhchandani(@raghav_bikh) 's Twitter Profile Photo

Meanwhile careerist faculty on US campuses like Cricinfo co-founder Vishal Misra continue to defend tyrannical police state activity and mock brave students opposing the genocide. What 'concrete investments' are possible when Israel will bomb Rafah & controls aid transportation?

Meanwhile careerist faculty on US campuses like Cricinfo co-founder @vishalmisra continue to defend tyrannical police state activity and mock brave students opposing the genocide. What 'concrete investments' are possible when Israel will bomb Rafah & controls aid transportation?
account_circle
martin_casado(@martin_casado) 's Twitter Profile Photo

tl;dr LLMS as Bayesian learning : Given a prompt, looks for something close in training set, then uses prompt for new evidence. Then computes a posterior using this new evidence. This posterior distribution is what is used to generate the new text. (Vishal Misra)

It's so crazy…

tl;dr LLMS as Bayesian learning : Given a prompt, looks for something close in training set, then uses prompt for new evidence. Then computes a posterior using this new evidence. This posterior distribution is what is used to generate the new text.  (@vishalmisra)

It's so crazy…
account_circle
dmitriy(@DmitriyLeybel) 's Twitter Profile Photo

Vishal Misra Sus.

What do you mean by 'recursively self-improve'?
Are you referring to synthetic data not being useful? They're given external guidance, are they not?

@vishalmisra Sus. 

What do you mean by 'recursively self-improve'?
Are you referring to synthetic data not being useful? They're given external guidance, are they not?
account_circle