Damodharan Jay

Web Developer, Typescript fanatic, Strong types for the win!

[Damodharan Jay @twitdamojay]

Thanks to @langchain , writing LLM apps is pretty easy with access to LLM memory, tools, agents and overcoming token length limitation. Was amazed at the quality of prompts used in the llm-math!

#langchain #ChatGPT #AI #Agent #

https://t.co/4UikVnZtsh

[Damodharan Jay @twitdamojay]

@shuding @_chenglou Isnt this the implementation of flexbox in CSS ? Why do we need to optimize layouting when it already formalized and meaningfully abstracted ?

[Damodharan Jay @twitdamojay]

@wasmerio https://t.co/xrzBXK07tP

I made the same last week! Also supported projects and multiplexer as well!

Thanks a lot for the awesome project @wasmerio @pi

[Damodharan Jay @twitdamojay]

Introducing webmux https://t.co/xrzBXK07tP - powered by @wasmerio - The super power edge sandbox running completely in the browser.
The pi (@pidotdev) runs in sdk and cli.
Give your agents the freedom it deserves in yolo mode - and multiplex as well!

[Damodharan Jay @twitdamojay]

@kirancodes Same thoughts. Its funny no one bats eye on z3 / liquid haskell sitting around for years

[Damodharan Jay @twitdamojay]

@wasmerio @pidotdev Absolutely!

The terminal multiplexer in browser is such a gamechanger.

Would love to open-source and get it reviewed from you

[Damodharan Jay @twitdamojay]

Damn! The unix sandbox in the browser built with @wasmerio is able to run @pidotdev

https://t.co/xrzBXK07tP (there are still some rough edges though - open sourcing soon...)

Sandboxes at the edge and that being in wasm is a huge boon for ephemeral apps

[Damodharan Jay @twitdamojay]

@mikeknoop Right now most rad harness is the cli with context

So ability to execute such commands in the latent space of the model with context is good enough ?

[Damodharan Jay @twitdamojay]

@mikeknoop New ARC-AGI benchmark loading..... ?

[Damodharan Jay @twitdamojay]

@wasmerio Can this run inside a browser like webcontainer ?

[Damodharan Jay @twitdamojay]

@saranormous claude code / gemini code / qwen code cli is the all the OS you need

[Damodharan Jay @twitdamojay]

https://t.co/pfU7HEMyXC
@karpathy introducing a new AI jargon - "People spirits" sitting on the internet: An amazing analogy for partial autonomous LLM agents.

Guess how many domain names gonna get registered similar to "vibe coding"....

[Damodharan Jay @twitdamojay]

@danielhanchen Wait for GRPO with unsloth on gemma3 with vision to try "aha" moment

[Damodharan Jay @twitdamojay]

@reach_vb qwen2.5-vl is GOATed for this usecase. It does structured generation amazingly well. Converts image to JSON / HTML easily.

[Damodharan Jay @twitdamojay]

@danielhanchen @UnslothAI Does it work with Qwen2.5-VL and other vision models ?

[Damodharan Jay @twitdamojay]

@danielhanchen @UnslothAI Amazing! Just when I was planning to spend money from pocket for 8xH100, unsloth saves me again!

[Damodharan Jay @twitdamojay]

@liangchen5518 Thanks a ton! Would love to extend your work to make it work for identification of elements.

Is it possible to share the info on your training infrastructure ? A terraform file or so ?

[Damodharan Jay @twitdamojay]

@VictorTaelin Wait should the LLM be able to reduce / execute the theorems to be useful ? Do you mean the private COT of the LLM should be able to execute in COT as well ?

[Damodharan Jay @twitdamojay]

@chheplo @rejaullah_40970 Just curious, how do we benchmark the finetune then ? Don't we some kind of benchmark to compare against ?