C S Krishna

Artificial Intelligence Researcher & Practitioner; Author: UnReal Elections

[Mihail Eric @mihail_eric]

My Stanford course 𝗧𝗵𝗲 𝗠𝗼𝗱𝗲𝗿𝗻 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿 starts next Tuesday.

This Github repository will hold all the assignments (and has the ones from last year).

Bookmark it. See you next week.

https://t.co/b6o254kwon

[Eric Xing @ericxing]

Glad to share my perspective on the recent AI-pacing and humanity risk narratives on @CNBC. AI, however powerful, is a piece of software residing inside a chip; to do harm in physical world and on nature, in addition to be intelligent enough to know how (e.g., to design a potent enough virus that is biologically viable not just in theory), it needs to further command supply chain, facilities, people, and additional experimentation to overcome the idea-to-physical gap, which can be and already are subject to intensive guardrail. I am less worried an immediate or even a remote risk from that direction just because of all the missing scientific and engineering links are hard to bypass; I am more worried what the AI frontier labs are making and talking about are not subject to academic investigation, 3rd party audit, and objective and quantitative forensic analysis because of their close-source or mere “open”-weight nature, which is not helpful for trust, safety, and transparency. At @IFM_AI we open source the full house (data, weight, code, checkpoints…) of our models in our recent K2-horizon release. I say NO to slow down AI, I say YES to a more open, faster stronger, and more cost effective AI, which means even faster research and development, because there are too many important problems for AI to solve, such as cure cancer.

[Zhengyao Jiang @zhengyaojiang]

So they created an action native model.

Conventional LLMs take actions by first generating JSON text sequentially, which is then parsed into actions.
This model can only output actions, more precisely, probabilities of discrete actions.

I don't know the exact architecture, but apparently this specialization makes the model very cheap and fast to run.

This is especially true for outputs, since it doesn't need autoregressive decoding.

[Alex Smola @smolix]

Here are a few scenarios of what might have happened with the Navier Stokes drama, none of which look good for @OpenAI.

1. Attempted scoop

OAI found out that a solution of Navier Stokes existed and they piled an army of agents on the problem to genuinely solve it and scoop Tristan and Levent.

2. Unethical employee

An OAI engineer got wind of the solution and used GPT's responses to Tristan and Levent to prime their own solution approach. This would explain why the agents suddenly struck gold.

[Eric Xing @ericxing]

Agree. Not just in LLM-biz, industry in general, even in academia, university presidents, provosts, are better served if they themselves are down-to-earth doers and active practitioners/researchers, l learned this first hand in building @mbzuai .

[C S Krishna @CSkrishna]

@gneubig great was waiting for this. Thank you, Prof. Neubig 🙏

[Graham Neubig @gneubig]

We started to post videos for CMU 11-768, AI Agents!

All of the videos will be posted to this playlist, so please bookmark/follow it if you want to be notified of new ones! I'll also try to post them on this thread too.

https://t.co/HAJgKXve0x

[Omar Khattab @lateinteraction]

reminds me of when Shopify folks (@kshetrajna) saved $5M using DSPy - the company seems to be really good at this kind of thing!

https://t.co/dLdvqd9Vo9

[Lakshya A Agrawal @LakshyAAAgrawal]

The GEPA team went through all feedback we received about prompt/harness optimization runs and used HCI principles to develop deep observability and steerability tools into GEPA/optimize_anything including re-architecting GEPA itself, allowing both you and your agents to know exactly what is going on in the run.

The new interface allows you to understand what candidates are being proposed, which datapoints are useful and if any datapoints are misleading or failing repeatedly?

With the built-in GEPA agent, you can now chat about the run, inject new candidates on the fly, as well as use @cursor_ai-style brushes to give very targeted feedback to various components like the candidates and evaluators, and see the changes immediately.

We are very excited to iterate further on this, and invite you to please sign up for the user study to provide us feedback!

[alex zhang @a1zhang]

Read full blog, condensed summary below: https://t.co/wgNFJctc9x

[Fascinating History @Fascinate_Hist]

A 2,000-year-old statue of the Greek goddess Aphrodite unearthed from the mud during excavations in the ancient city of Aizanoi, Türkiye.

[Graham Neubig @gneubig]

This Fall at CMU we're teaching a new course on AI Agents!

The goal is that you learn how to create a scaffold, build evals, and train an agentic LLM using RL.

We'll try to balance theory and practice, and introduce modern frameworks and best practices.

[C S Krishna @CSkrishna]

https://t.co/a5JsHkADwc

[C S Krishna @CSkrishna]

@KanwalSibal All these are fringe to core establishment, no? Out of favour, even

[C S Krishna @CSkrishna]

@Empedoclesss So from india pov, asim dasgupta n power good for India?

[C S Krishna @CSkrishna]

https://t.co/iDQ7z9SXr5

[Matei Zaharia @matei_zaharia]

Really excited to open source a new project: Omnigent, a meta-harness for AI agents.

It lets you build multi-agent coding and custom agents, sitting above Claude Code, Codex, Pi, and agent SDKs to let you compose them. It also adds live collaboration and rich control policies.

[Kevin Patrick Murphy @sirbayes]

My talk at MIT, on "Agentic AI systems: from scruffy to neat", is now available. I cover 3 examples of agentic systems - Bayesian linguistic forecaster, autoharness, and code world models - which combine LLMs, code and planners in different ways. Links below.

[Omar Khattab @lateinteraction]

Claude Code is finally an RLM (oct 2025), congrats to Anthropic :-)

[Iceland Cricket @icelandcricket]

Our Prime Minister @KristrunFrosta meeting her Indian counterpart today, sorting out our first five match Test series. No doubt about it, you can see a cricketing look in her eye.