Mann Patel

first day being noogler at @Google ✨

@jietang @yacineMTB @elonmusk @teortaxesTex AURA

@elonmusk @teortaxesTex won’t take that long

I think GLM 5.2 points to a 7 months gap currently It's around Opus 4.7-4.8 level, all told (modulo vision which in Opus's case is garbage anyway). Mythos reached Preview status (≥ Opus 4.8, functionally) by early Feb 2026. This means full PRC Mythos ("Fable") by Nov-Dec'26.

@SemiAnalysis_ dylan i like my job, let me have it 😭

Pretraining fundamentally does not make sense anymore for anyone other than frontier labs. Although there are a lot of people at enterprises & startups who have "Pretrainitis" to show “impact” and get promotions, fundamentally, it doesn’t make sense. There is probably higher ROI in partnering with a frontier lab to do prompt engineering, although it isn’t as “sexy” as pretraining.

@jmzoeee naur

my last day of tech week, back to the cave

@radixark raise @ly_h990 ‘s salary, she is a overworking herself! :3

guess who fumbled the bag hard…congrats!!

Codex with gpt-5.5 xhigh discovered a math trick for full vocabulary kl distillation jonathanc.net/blog/kl-cache-…

People talk, listen, watch, think, and collaborate at the same time, in real time. We've designed an AI that works with people the same way. We share our approach, early results, and a quick look at our model in action. thinkingmachines.ai/blog/interacti…

To train better open models, we need predictable scaling. Delphi is Marin’s first step: we pretrained many small models with one recipe, then extrapolated 300× to predict a 25B-param / 600B-token run with just 0.2% error. Getting there took some work 🧵

everyone benchmarks pass@k. everyone deploys pass^k.

@finbarrtimbers same

The best part of my job is I get to play with GPUs all day and someone else pays for it

1/ Deep learning is going to have a scientific theory. We can see the pieces starting to come together, and it's looking a lot like physics! We're releasing a paper pulling together these emerging threads and giving them a name: learning mechanics. 🔨 arxiv.org/pdf/2604.21691 🔧

Model shaping is still a craft of a few. That's what AI agents are for: learning it and doing it for everyone else. As a part of FrontierSWE benchmark we built a 20-hour post-training task on @tinkerapi and found the real bottleneck is research intuition.

@littewhite16806 most interesting part of this little project was writing a selenium verifier to 16personalities site and letting llm answer 5 scale questions haha

@punsbymann Nice work as a class project. Our method does not require training 8 different models. It looks like we generate several models from one, and your paper fuses several into one. Also, PEFT is different from pruning (e.g., pruning is not frozen). Welcome to stop by our poster! 😀

Your language model isn’t one person—it’s everyone. Check out Personality Subnetworks (ICLR 2026): a training-free framework to extract persona-specialized subnetworks. A step towards controllable and interpretable personalization in LLMs. Paper: arxiv.org/pdf/2602.07164

@littewhite16806 your work is totally unique and kind of opposite to mine! while you prune and prove personas already exist in your model (which sounds about right), i tried to train Lora’s that show composition of persona😄

the benchmark game has entered its IPO era

🧵 Decomposing the Delta: What Do Models Actually Learn from Preference Pairs? 1/n 💡 Why do methods like DPO and KTO actually improve reasoning? In standard alignment, we use preference pairs, but we don't fully understand what properties of the data drive downstream gains. We investigate two distinct notions of quality: Generator-level Delta - the capability gap between the models producing the chosen vs. rejected traces Sample-level Delta - The fine-grained quality difference within a single pair (Factuality, Coherence, Precision) 👇

@f14bertolotti @ChiahsuanL @MingyangKevinZh 🤩

This paper digs into what actually makes delta learning work. The authors show that it’s not just about the chosen trajectory being correct relative to the rejected one. What drives the gains is the coherence of reasoning across the trajectory. 🔗arxiv.org/abs/2604.08723