Yash

The highlight of my Twitter profile up until now has been getting a reply from the OG yay.... @karpathy the new intern at @AnthropicAI xD

@_avichawla Valuable points mentioned 💫

As an AI Engineer shipping agents to production, please learn: - Not every intent needs an agent - Early stopping over indefinite retries - Fallback parsers for structured output - Evals for agent behavior not just output - Delivery infra that's framework-agnostic - Provider diversity as a reliability decision - Model portfolios over single-model stacks - One agent with good tools over multi-agent - Cost attribution per feature, not per invoice - Full-chain tracing, not just endpoint logging - Deterministic signals before LLM-as-a-judge - Production traffic repeats. Cache accordingly - Guardrails as middleware, not per-agent code - Human-in-the-loop is a design pattern, not a fallback Most of what blocks agents from going to production isn't the core logic. It's the plumbing around it. That's why most points above focus on agent ops, not just agent dev. Plano is a 100% open-source infrastructure layer that handles routing, orchestration, guardrails, and observability for agentic apps. I have shared the GitHub repo in the replies. 👉 Over to you: What else would you add here?

@sbincx Why aren't you ?

If you’re so smart why aren’t you working on something that makes you feel like a moron on a daily basis?

@zhyncs42 Nice, added to the reading list

Correctness is critical for LLM inference engines. Recently, I found TRT-LLM’s work on Hypothesis Testing Methodology to be extremely professional. #hypothesis-testing-methodology" target="_blank" rel="nofollow noopener">github.com/NVIDIA/TensorR…

@jino_rohit It's just a ton of hacks

reading the awq quant paper today. also realized theres a tons of research work you can still do in quantization once you have a decent grasp of the general direction of the space

@brookeleblanc What were their intentions to test or to humiliate if the other person didn't know something about?

When I was switching roles late last year I interviewed for a company that wanted me to put everyone I knew in a spreadsheet And I immediately stopped process w/ them. My career/life, and all the people in it, is so much bigger than a spreadsheet. For a nonpartner, non cofounder role, never ever do this. Unless you have a lifechanging amount of skin in the game and your Rolodex will be completely confidential. Even then. Probably don’t do this. Who you know is your IP.

@elonmusk Certainly

The 400’s in Rome were brutal

@samiramanabi Why are you being so offensive

Hmm… did people forget to take the undergraduate computer architecture course?

@ariG23498 Nice

That is what I want to cover in the first part. I am about to complete my write up, and submit it to my colleagues for a review. Hope they like it! 🤞

@elonmusk How much % of the grok users use it for difficult coding tasks ?

Grok foundation model V9-Medium (1.5T) has finished training. Evals look good. A lot of Cursor data was added in supplementary training and there is more to come. Fine-tuning is underway and reinforcement learning begins in a few days. 2 to 3 weeks to public release. This will be a major improvement over the 0.5T v8-small that currently serves all Grok production traffic, especially for difficult coding tasks.

@glcst Got the point but Not so great an analogy ig cause the one refuses to adopt ai would really be paranoid when the models become much powerful and fast.

the one guy in the team who refuses to adopt AI and still ships more than everyone

@ziqi_huang_ It's very useful for the robotics I believe

An interesting work on Physical AI: PhysX-Omni. First unified sim-ready generation framework for rigid, deformable, and articulated objects, with a diverse dataset and new benchmark. 🌐 physx-omni.github.io 💻 github.com/physx-omni/Phy… 📦 huggingface.co/datasets/PhysX…

@eliebakouch The performance degradation is so much that it's better to use search

asked 2 questions about the claude desktop app, defaults to haiku 4.5, both wrong answers :(

@teslaownersSV Pre-optimization is always bad. It's better to build a bad version of something and then optimize it as @elonmusk said.

Engineering is magic

@docmilanfar Curious to know what you said to the barista then ? xD

can confirm. Google had no category for visiting faculty back then, so we all got (green badge) "interns" status. a barista once asked me if I was too old for an internship 😅

@luke_metro Significantly very less, not on priority definitely cause the one which is on priority can help solve that problem possible which is a powerful AI

sometimes I wonder how many TPUs Google Deepmind is currently siccing on solving P vs NP

@elliotarledge That's bad for the grass though xD

Touching grass

@s_batzoglou Surely one of the important ones till date !

Probably the most important AI-in-math paper to date

@leafs_s It's really a great paper, I read it the day after it was published.

From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence This paper argues that classical information measures such as entropy and Kolmogorov complexity are insufficient for understanding modern AI learning. Entropy mainly measures randomness, while Kolmogorov complexity measures the shortest possible description of data. However, neither fully explains why some datasets are far more useful for training AI systems than others. To address this, the authors introduce epiplexity, a new concept intended to measure the amount of learnable structured information available to a computationally limited learner. The key idea is that information is not absolute. Its value depends on: the learner’s computational limits, the structure and ordering of the data, and the transformations used to generate or present the data. The paper suggests that epiplexity could provide a theoretical framework for: understanding why certain training data are more useful, designing better datasets, improving data generation and augmentation, and studying learning efficiency in AI systems. arxiv.org/abs/2601.03220

@_aidan_clark_ @ssi Also meanwhile xD

@_aidan_clark_ 3B-4B just for the research compute is enough and almost every big lab has it, I would prefer all of them and also @ssi xD

If you want to work on pretraining-for-AGI, join OpenAI, Google, Meta or the Anthropic/XAI/Cursor supergroup. The bitter truth of the widening compute gap is that all the problems which are actually on the critical path to AGI now demand that level of compute.

@AJakkli It's a perfect example of "narrow ai" as @ilyasut said

I really liked Eric's take on why alpha go is profound: A 10-layer network can only do 10 sequential steps of thinking, by construction. And yet those 10 steps can "amortize and approximate to very high fidelity a nearly intractable search problem."