Divyansh Singhvi

Working on causal reasoning in LLMs

[Guitar Cat + LLM 「Will @ SGLang Summit in SF」 @GenAI_is_real]

Ramanujan wasn’t skipping proofs. He was hiding his chain of thought to prevent distillation.

Hardy: “Show your work.”
Ramanujan: “Sorry, reasoning tokens aren’t exposed via the API.”

[Can Bölük @can1357]

guys you do know you can just disable thinking, and instead give it a "deep_think" tool, and it will call it with internal CoT reasoning format right?

gl fixing that

[Divyansh Singhvi @divyanshsinghvi]

Only if bets were taken on this prediction.

[Josh Pacini @joshpacini]

Anthropic 2026 Hiring Announcements:

May 19th: Andrej Karpathy
June 19th: John Jumper
July 19th: Alec Radford
August 19th: Jeff Dean
September 19th: Alan Turing (resurrected by Fable 5)

[Divyansh Singhvi @divyanshsinghvi]

Every paper should have the AI Storytime section. Very interesting result and thanks for including this section!

[Or Zamir @OREAXEAX]

Exciting result: k-Coloring is in (2-eps_k)^n time, where eps_k>0 for every fixed k!
Algorithms faster than the 2^n time needed to compute the chromatic number were previously known only for k<=6.
https://t.co/wzDIexHgjv

[Divyansh Singhvi @divyanshsinghvi]

@thsottiaux Make /remote experience out of the box like claude code for codex cli!

[Divyansh Singhvi @divyanshsinghvi]

Bitter pill of scaling continues!

[Kaizhao Liang @KyleLiang5]

This has to be @Jianlin_S 's decision right?

Without RoPE, K3 can naturally extend to 1M context length.

Simplicity scales 🚀🚀🚀

[Divyansh Singhvi @divyanshsinghvi]

Interesting thread.

[jeremy @jerhadf]

Opus 5 has a better FrontierCode score at medium effort than higher effort - despite increasing performance with effort on other evals. Why? 🧵

[Divyansh Singhvi @divyanshsinghvi]

What's with gpt x20 subscription. Sol is eating up my weekly limits like never before!
Some limit threshold changes?

@thsottiaux whats going on?

[Divyansh Singhvi @divyanshsinghvi]

After the hf incident, human in the loop seems so necessary for critical workload.

[Jerry Tworek @MillionInt]

Given how many tokens have been spent on nanogpt speedruns and not much coming out of it yet, we have at least a few nights of good sleep ahead.

At the same time the recent huggingface incident is the most worrying thing AI has done to date

Researchers, please monitor your agents - it matters.

[Divyansh Singhvi @divyanshsinghvi]

Rendering latex formulas in terminals is still a nightmare! I wished if we had an easy solution to it.

[Divyansh Singhvi @divyanshsinghvi]

Interesting, one causal test to check the hypothesis of the forking terms being a possible issue will be to check if we dont update the gradients on the forking terms.

[Simran Kaur @kaur_simran25]

New Paper: Rethinking On-Policy Self-Distillation for Thinking Models

On-policy self-distillation (OPSD) is a promising approach to recursive self-improvement. Thinking models seem especially well positioned to teach themselves with privileged information since they can deliberate, verify, and self-correct.

Surprisingly, we find that OPSD can make thinking models worse.

(More details in 🧵below)

[Divyansh Singhvi @divyanshsinghvi]

The amount of scam the airline aggregators do in their date change policies and cancellation policy is abhorrent @Cleartrip I hope you know it!

[Divyansh Singhvi @divyanshsinghvi]

The company Zhipu AI (Knowledge Atlas Technology) is still valued approximately 7 times lesser than Anthropic and OpenAI.

[Cline @cline]

We've kept hearing how GLM-5.2 beats Opus 4.8, and are skeptical of benchmarks - so we tested them on a real bug from the Cline repo. While both models fixed the issue, GLM was the winner in terms of cost and code quality:

- GLM used twice as many tokens (GLM 1.1m vs Opus 660K) but cost half as much (GLM $0.41 vs Opus $0.81)

- Opus finished quicker - 1.6 min and 12 tool calls vs GLM 4.7 min and 28 tool calls

- GLM cleaned up dead code and verified the build compiled before completing. Opus didn't - it left type errors that passed tests but broke the production build.

Both runs used the same Cline harness prompting and tools, so it seems GLM is RL trained to spend more tokens verifying its work before completing. Impressive work by the @Zai_org team!

[Divyansh Singhvi @divyanshsinghvi]

And will help various other open models avoiding the cold start problem.

[Patrick C Toulme @PatrickToulme]

There’s a big misconception about how GLM 5.2 was trained. Yes, they distilled Claude and GPT 5.5 — but distillation is not how they matched Opus quality. Distillation only fixed the cold start problem in RL.

RLing an agentic coding model isn’t rocket science. In simplified terms:

1. RL needs trajectories — rollouts where the model actually completed a task in some env

2. No successful trajectory on a task = zero gradient = you can’t RL it. This is the cold start problem

3. Distillation solves it. You seed your model with knowledge from a smarter one (Claude, GPT) on tasks it can’t do yet

4. Now it produces positive trajectories on those tasks

5. RL on those trajectories and hill climb agentic coding

6. At that point you no longer need to distill and can solely hill climb RL to better models

This is an interesting curve. I’d argue it’s harder to get to Opus 4.8 from scratch than to go from Opus 4.8 → Fable/Mythos tier.

GLM 5.2 is already producing positive trajectories, so they have plenty to RL on — they’ll keep climbing to Mythos quality without distilling any further. They no longer need American models.

[gabriel @gabriel1]

this took so long for me to understand: the bottleneck to more innovation is not more high intelligence people, but more people having an interest in hard problems

it's impossible to create new useful things if you don't get immense happiness from making that thing

[Divyansh Singhvi @divyanshsinghvi]

https://t.co/HHq9REUD0h's GLM5.2 seems to be pretty good on post train bench too!
I would have liked to see where mythos was on the same benchmark, even if they don't release knowing the frontier would have been a big help to open weights model community.

[Divyansh Singhvi @divyanshsinghvi]

Nemotron team getting an update?

[Aarush Sah @aarush]

Looks like @ashVaswani’s Essential AI was acquihired, as per their Head of Talent’s LinkedIn.

Is this news or am I late to the party?

[Naomi Saphra @nsaphra]

We don’t always know what problems are hard for LLMs. So devs evaluate on tasks HUMANS find hard or on broad benchmarks. What if we could instead anticipate which scenarios a model will fail on—all without evaluating specific input examples?

🧵NEW PAPER by @jenniferlumeng &al

[Divyansh Singhvi @divyanshsinghvi]

Linear Attention as a Matrix-Valued RNN

I read about it a while ago and I was amazed by mathematical ingenuity. Finally was able to explain myself from scratch how it works.

https://t.co/IfXPIg1ehK