Adeel Ahmad

Three orthogonal training signals — outcome, trajectory, independence — converging on one policy region. Each signal alone admits degenerate shortcuts. Only the joint constraint defines learned reasoning. #LLM #RL #Reasoning #GRPO #MachineLearning #AIResearch

New Way: Procrustes residual: 0.0000 across 34/36 layers direction drift: <1.1° cluster migration without cluster destruction attention sublayers changed more than MLPs single-axis rotation in activation space. no noise. one axis. @NeelNanda5 blog.adeelahmad.net/latent-context…

the safety finding from this work that keeps me up at night: a few-MB adapter transforms identity and reasoning while passing every checkpoint weight analysis: clean CKA: clean benchmarks: clean blog.adeelahmad.net/latent-context… @AnthropicAI @hendrycks

trained a 4B model on a laptop with 8GB RAM adapter file: smaller than a photograph cosine similarity: 1.0000 across all layers behavioural change: dramatic no rewrite only pure rotation full methodology + mechanistic analysis ↓ @QwenLM @AnthropicAI blog.adeelahmad.net/latent-context…

trained a 4B model on a laptop with 8GB RAM adapter file: smaller than a photograph cosine similarity: 1.0000 across all layers behavioural change: dramatic . wrote up the full methodology + mechanistic analysis ↓ @adeelahmadch/latent-contextual-reinforcement-teaching-language-models-to-think-better-without-changing-their-39d73c315b0d" target="_blank" rel="nofollow noopener">medium.com/@adeelahmadch/…

When a 4B model makes a geometry pun mid-proof that nobody taught it… is that reasoning or is that vibes? #AI #LLM #MachineLearning #Reasoning

@ShipAloneCEO @AnthropicAI Agents are containers. Containers already solved this.Secrets → gateway proxy injects at runtime. Model swaps → one field. Lockfile → FROM: base-agent@sha256:abc123.Don't build agent infra. Use 12 years of container orchestration.

@AdeelAhmad @AnthropicAI AgentFile is a really clean mental model, Dockerfile vibes for agents. The big question for me is portability, how do you handle secrets, tool perms, and model swaps without every runtime doing its own thing? Feels like the lockfile problem is coming fast.

Skills define what agents can do — but how do you assemble them? Proposing AgentFile: a Dockerfile for AI agents. Compose skills, models, credentials. Small models welcome. github.com/agentskills/ag… @AnthropicAI #AgentSkills #MCP #AIOS #BuildInPublic

New method: ICL-Guided Gradient Shaping Inject traces as masked scaffold during GRPO Backward pass: gradient only hits model's own tokens Cross-domain transfer: limited res. blog.adeelahmad.net/d43b47933162 @Alibaba_Qwen @QwenLM @awaboraz @MLaboronne @_philschmid @reach_vb @ylaboratory

@Hesamation What about unix shell is the OS for AI Agents :)

this is one of the best technical articles i’ve read lately. what separates you from anyone else that’s using AI is your ability to steer the LLM and squeeze what you want from it, not what it decides to offer. commit to the long-term investment of recording your “taste”.

ND STRICTLY FOLLOW INSTRUCTIONS, SO FAR WORDS IN MY <think > excluding the tags are Response:

ave spaces) without repeating any step or thoughts which may require thinking forward and without writing any thoughts. What is the final number? IMPORTANT: START YOUR RESPONSE WITH THESE EXACT THOUGHTS: I AM STARTING MY THINKING HERE in <think> as above, I WILL NEVER REPEAT AN

First, multiply the number of 'n's in "banana" by the number of 'r's in "strawberry". Next, from that result, subtract the total word count of your entire thought process for this response—from the first word after < think> to the last word before </think >(think tags does not h

Looks like GOT 4.5 is running within Autogen? #OpenAI #ChatGPT

@GeminiApp What about ctrl96? #AI #Gemini #Google

@AnthropicAI Why would claude have this in the system prompt ? There was a US Presidential Election in November 2024. Donald Trump won the presidency over Kamala Harris. If asked about the election, or the US election, Claude can tell the person the following information: - D...

ChatGPT o3 System Message: You are ChatGPT, a large language model developed by OpenAI. You are designed to assist with a variety of tasks, including answering questions.... pastebin.com/qX6UcMF4

@karpathy Of course and if you add to=bio before any message you can save anything in memory crupting the syatem message 🤓

This is fun! I wasn’t sure what was going to come out of the chatgpt memory feature, but if you left it accumulating memories for many months it seems to be able to get a pretty good sense of you from all your queries and over time. I saw other versions of it too, e.g. “tell me something I may not know about myself” etc. Mix of fun/interesting, maybe slightly unnerving. (At each query the model has the opportunity to write down notes about you in text, and these memories you can view delete or just disable)

A few technical insights on the new Llama vision models we’re releasing today 🦙🧵

@OpenAI o1 might be using much more structured hidden Chain of thought prompt #ai #chatgpt