Prasad Chalasani
@pchalasani.bsky.social
Building:
Langroid - Multi-Agent LLM framework: https://github.com/langroid/langroid
Productivity tools for Claude-Code & other CLI agents:
https://github.com/pchalasani/claude-code-tools
IIT CS, CMU/PhD/ML.
Ex- ASU, Los Alamos, Goldman Sachs, Yahoo
Langroid - Multi-Agent LLM framework: https://github.com/langroid/langroid
Productivity tools for Claude-Code & other CLI agents:
https://github.com/pchalasani/claude-code-tools
IIT CS, CMU/PhD/ML.
Ex- ASU, Los Alamos, Goldman Sachs, Yahoo
Richard Sutton – Father of RL thinks LLMs are a dead end
YouTube video by Dwarkesh Patel
youtu.be
September 27, 2025 at 3:38 AM
Unsafe for whom and according to whom?
September 24, 2025 at 3:31 AM
Unsafe for whom and according to whom?
A while back I came across Claude-trace which shows all of Claude-code messages (including tool calls) in full detail in a nice foldable html view. Made a lot of things very clear to me.
github.com/badlogic/lem...
github.com/badlogic/lem...
lemmy/apps/claude-trace at main · badlogic/lemmy
Wrapper around tool using LLMs for agentic workflows - badlogic/lemmy
github.com
September 9, 2025 at 3:05 PM
A while back I came across Claude-trace which shows all of Claude-code messages (including tool calls) in full detail in a nice foldable html view. Made a lot of things very clear to me.
github.com/badlogic/lem...
github.com/badlogic/lem...
Truly bizarre situation they are in.
Curious why they can’t run some benchmarks with the model (if they suspect the issue is with the model itself) or some agentic coding benchmarks on Claude-code (if the issue might be with the scaffolding, prompts etc).
Curious why they can’t run some benchmarks with the model (if they suspect the issue is with the model itself) or some agentic coding benchmarks on Claude-code (if the issue might be with the scaffolding, prompts etc).
September 9, 2025 at 2:24 PM
Truly bizarre situation they are in.
Curious why they can’t run some benchmarks with the model (if they suspect the issue is with the model itself) or some agentic coding benchmarks on Claude-code (if the issue might be with the scaffolding, prompts etc).
Curious why they can’t run some benchmarks with the model (if they suspect the issue is with the model itself) or some agentic coding benchmarks on Claude-code (if the issue might be with the scaffolding, prompts etc).
My biggest concern now is — if the issue they have is as vague as “reports of degraded quality”, how do they even approach fixing it? And what measurable criteria will they use, to declare that it is fixed? Would they take a vibes-check opinion poll?
September 9, 2025 at 2:05 PM
My biggest concern now is — if the issue they have is as vague as “reports of degraded quality”, how do they even approach fixing it? And what measurable criteria will they use, to declare that it is fixed? Would they take a vibes-check opinion poll?
I admit I still haven’t weaned off of Claude-Code. I often have CC (Claude-code) consult/delegate with/to CC (codex-CLI) using my Tmux-cli tool, with CC1 watching/capturing CC2’s output or asking it to dump into markdown files:
github.com/pchalasani/c...
github.com/pchalasani/c...
GitHub - pchalasani/claude-code-tools: Collection of practical productivity tools for Claude Code
Collection of practical productivity tools for Claude Code - pchalasani/claude-code-tools
github.com
September 8, 2025 at 2:44 PM
I admit I still haven’t weaned off of Claude-Code. I often have CC (Claude-code) consult/delegate with/to CC (codex-CLI) using my Tmux-cli tool, with CC1 watching/capturing CC2’s output or asking it to dump into markdown files:
github.com/pchalasani/c...
github.com/pchalasani/c...
The excellent Claude Code scaffolding can be used very easily with GLM-4.5, Kimi-K2, Deepseek (all of which have Anthropic-compatible APIs) by setting a simple function in ~/.zshrc: github.com/pchalasani/c...
GitHub - pchalasani/claude-code-tools: Collection of practical productivity tools for Claude Code
Collection of practical productivity tools for Claude Code - pchalasani/claude-code-tools
github.com
September 8, 2025 at 2:39 PM
The excellent Claude Code scaffolding can be used very easily with GLM-4.5, Kimi-K2, Deepseek (all of which have Anthropic-compatible APIs) by setting a simple function in ~/.zshrc: github.com/pchalasani/c...
But context length half of Claude though
September 8, 2025 at 12:18 PM
But context length half of Claude though
E.g., see this recent anthropic status update. “Lower intelligence, malformed responses, issues with tool calling in Claude code” — quantized models is one way this could happen.
September 8, 2025 at 12:15 PM
E.g., see this recent anthropic status update. “Lower intelligence, malformed responses, issues with tool calling in Claude code” — quantized models is one way this could happen.
The single biggest missing feature in codex-cli was resume/continue; now they have it, and with GPT-5’s context length double that of Claude (I still haven’t got access to the 1m context model), and less sycophancy, I think it’s a very compelling alternative to Claude-code.
September 8, 2025 at 12:10 PM
The single biggest missing feature in codex-cli was resume/continue; now they have it, and with GPT-5’s context length double that of Claude (I still haven’t got access to the 1m context model), and less sycophancy, I think it’s a very compelling alternative to Claude-code.
There’s also Claude code router github.com/musistudio/c...
GitHub - musistudio/claude-code-router: Use Claude Code as the foundation for coding infrastructure, allowing you to decide how to interact with the model while enjoying updates from Anthropic.
Use Claude Code as the foundation for coding infrastructure, allowing you to decide how to interact with the model while enjoying updates from Anthropic. - musistudio/claude-code-router
github.com
August 25, 2025 at 11:51 AM
There’s also Claude code router github.com/musistudio/c...
I’m talking to code agents all day 😀 makes it easy to do a brain dump and then have them create a nice markdown
August 18, 2025 at 5:15 PM
I’m talking to code agents all day 😀 makes it easy to do a brain dump and then have them create a nice markdown
A speech to text (dictation) app is essential, to make it less tedious and less painful (literally, I mean to avoid RSI etc) to give code-agents all this context. My current favs are Wispr Flow (subscription) and VoiceInk (one time pay, fully local)
tryvoiceink.com/wispr-flow-a...
tryvoiceink.com/wispr-flow-a...
VoiceInk - Transform Speech to Text Instantly on Mac
Local AI-Powered Voice Recognition App for Mac. Convert Speech to Text Instantly with Offline Capability, 100+ Languages, and Complete Privacy. No Subscriptions, One-Time Purchase.
tryvoiceink.com
August 18, 2025 at 4:43 PM
A speech to text (dictation) app is essential, to make it less tedious and less painful (literally, I mean to avoid RSI etc) to give code-agents all this context. My current favs are Wispr Flow (subscription) and VoiceInk (one time pay, fully local)
tryvoiceink.com/wispr-flow-a...
tryvoiceink.com/wispr-flow-a...
Curious what you found to be the most effective way to use GPT-5 with CC. Some options I know of:
- have CC spawn another CLI agent running GPT-5
- have CC use an MCP server (e.g zen) that runs GPT-5
- Claude-code-router
- copy paste from ChatGPT
- have CC spawn another CLI agent running GPT-5
- have CC use an MCP server (e.g zen) that runs GPT-5
- Claude-code-router
- copy paste from ChatGPT
August 16, 2025 at 6:51 PM
Curious what you found to be the most effective way to use GPT-5 with CC. Some options I know of:
- have CC spawn another CLI agent running GPT-5
- have CC use an MCP server (e.g zen) that runs GPT-5
- Claude-code-router
- copy paste from ChatGPT
- have CC spawn another CLI agent running GPT-5
- have CC use an MCP server (e.g zen) that runs GPT-5
- Claude-code-router
- copy paste from ChatGPT
Not just pplx, It’s also beating o3 and Gemini 2.5 pro, which was very puzzling, until I saw a reddit discussion where they said they didn’t allow web-search on those other models. Yet another chart crime 😂
August 13, 2025 at 11:06 AM
Not just pplx, It’s also beating o3 and Gemini 2.5 pro, which was very puzzling, until I saw a reddit discussion where they said they didn’t allow web-search on those other models. Yet another chart crime 😂
Summary seems to be - People were expecting a step change in abilities, and got a router instead.
August 10, 2025 at 8:29 PM
Summary seems to be - People were expecting a step change in abilities, and got a router instead.
Claude-trace is fantastic. I was inspired by it to make nice html logs showing agent conversations in the Langroid agent framework: langroid.github.io/langroid/not...
Task Logs - langroid
Langroid LLM App Development Framework
langroid.github.io
August 3, 2025 at 11:35 AM
Claude-trace is fantastic. I was inspired by it to make nice html logs showing agent conversations in the Langroid agent framework: langroid.github.io/langroid/not...
There’s definitely a tradeoff of intelligence vs determinism
July 30, 2025 at 1:15 PM
There’s definitely a tradeoff of intelligence vs determinism
I am a CS person too, and I hear you. I guess we have to work at multiple levels
July 30, 2025 at 12:49 AM
I am a CS person too, and I hear you. I guess we have to work at multiple levels
So… basically a CustomGPT
July 30, 2025 at 12:47 AM
So… basically a CustomGPT