Shreyansh

Slope of the line is more important than the y-intercept.

Kimi K2.6 is the only thing at this point that’s stopping frontier companies from raising their prices. they’ve earned respect.

We spent millions building a wildly capable, human-like non-deterministic AI, and are now spending millions more trying to wrap it in guardrails and making it predictable and deterministic. Absolute cinema.

@VazeKshitij That's cool

@imbearguy Nope, I was there to sign with a startup

Ayy Bois. Career update soon. This shit is not a drill. And yes, I am in IIT-B rn to finalize the offer. Mode details incoming soon enough. 🫡🫡

@VazeKshitij Haha, last year for some event there, I reached 2 hours before, just to see the campus (and meet some ppl)

Btw - this is going to be me tomorrow at IIT-B.

@TheAhmadOsman Did you see the dwarkesh interview with Jensen huang? Whose words made more sense to you? Just curious coz from what I could understand, dwarkesh made pretty good points about generalizing ai but with odd references. I think u r the right person to ask this question

Let me make local AI easy for you Inference setup & optimization are simple now across pretty much any Inference Engine Just have an agent run evals overnight sweep a few quants per GPU / inference stack and generate bash scripts with best configs Wake up to tuned performance

@arpit_bhayani I feel good from the first sentence. Second is auto applied so no worries

It is okay not to make friends, but certainly do not make enemies. You never know where the person sitting next to you will be 20 years from now. The world is small, and it gets even smaller at the top.

@akshay_pachaar This is actually an art, always so excited to see breakdowns like this in simple words

Microsoft just mass-compressed LLM reasoning. their new paper introduces MEMENTO, a method that teaches reasoning models to manage their own context. instead of letting chain-of-thought grow into a flat 32K-token stream, the model learns to segment its reasoning into blocks, compress each into a dense summary (a "memento"), and mask the original block from future attention. the result is a sawtooth KV cache pattern where memory periodically drops instead of growing monotonically. training is a two-stage SFT recipe. stage 1: learn the block-memento format with full attention. stage 2: learn to reason with masked blocks. each block gets compressed 5-20x, and peak KV cache drops by 2-2.5x across model families. but the most surprising finding is the "dual information stream." when a memento is generated, the model can still see the full reasoning block. so the memento's KV entries get computed with full block context. after the block is masked, those KV entries stay. they carry implicit information that the memento text alone doesn't capture. recomputing memento KVs without block context drops accuracy by 15 percentage points. same text, different KV representations, significantly worse performance. this is what separates MEMENTO from prior work that rebuilds context from text alone and loses this implicit channel. they also showed the accuracy gap is a consistency problem, not a capability problem. the model can still solve the same problems, just less reliably. majority voting at k=3 recovers base accuracy, and RL closes most of the remaining gap. as reasoning traces get longer, models that compress their own intermediate state will serve more users on the same hardware. and the dual KV channel suggests in-place masking is fundamentally better than restart-based approaches. paper and dataset (228K traces) are public. link in the next tweet

I'm starting to like it #gym

@Yuchenj_UW given the impact and drama around it, It's probably 1

Seeing rumors that Claude Opus 4.6 got nerfed. Usually this boils down to 3 cases: - Unintentional. For example, a regression caused by changes in the inference stack or Claude Code. This is what evals are for before rolling out. - Intentional “optimizations” (quantization, reduced reasoning). If so, say it. If users pay for a model, they should get that model. - User psychology. The more you use a model, the dumber it feels.

Towards technological singularity

@ThePrimeagen ultrafix

I am such an idiot ultrathink ultraplan fix opus, no mistakes

@TheAhmadOsman you gave a 101 for LLMs that i love and follow can you give a 101 for LLM infra?

LLM infra right now is like Linux in the 90s We’re still early and there is a lot of work to do

@ayushunleashed whoever that arsenal fan is, i want to be friends with you

Priyam always coming up with fun ideas. What a night man.

@ayushunleashed Evals are just the test cases?

This is all going so meta with Claude code. I am using Claude to write a skill, then using claude to write evals for that skill, then claude is running subagents, providing skills to some agents while not to others to evaulate these agents on the evals ( test cases ), then noting down token usage and other stuff to evaluate weaknesses of the skill. Claudeception.

@ayushunleashed weights & biases have heard that its good

Any good content recommendations for building evals for long running Agents?

I questioned knowledge and saw that everything is somewhat biased, but some ideas still match reality better than others Finally, I realized even though I have values and think for myself, those same values can sometimes bias or distort how I see things Sasuke moment

I started by feeling life is unfair.. seeing some ppl earn easily while others grind Then I thought maybe some elite powerful ppl control what gets value and attention (manipulation), but I also realized it’s not fully controlled, just influenced (less facts to prove). ...contd

@ayushunleashed I'm seeing similar glitches, going till attempt 4 but no response.

Horrible experience with claude today. First sonnet 4.6 is hallucinating like it's GPT 3.5. Even when I tell it to do X, it reverts back to Y. Second, Are the rate limits decreased? A normal conversation hit the rate limit so fast.

@VazeKshitij Massive win, well deserved bhaii🙌

Ladies and Gentlemen Boys and Girls Lads and Lasses Today, since it's the end of the financial year, our increments were announced. If you didn't know, I work as a firmware engineer at an Industrial IoT company. I am from a tier-3 college, an electronics engineer, who works in a core company. 25% IT IS Y'ALL!!!! We are EXTREMELY close to the 7-figure mark. Lol, count in the twitter payout, and we are well past that too. I don't know what to say, I am still kinda processing it all. Unreal. Simply lovely.