How to Loop Your Transformer: Loop Scaling, Specialization, and Parallel Inference New paper from Meta. "Our experiments span models with up to 1B parameters, trained on up to 155B tokens using up to 1.07 × 10^21 training FLOPs, with up to 32 loops. We find that looping the full transformer generally gives the strongest quality at a fixed parameter budget, whereas looping a middle block of layers gives a better return when training FLOPs are constrained." "We compare exact weight sharing across 2 loops with small loop-specific low-rank adapters that allow each loop to learn a slightly different transformation. These adapters provide only modest gains when few loops are used, but become consistently more useful as the same weights are reused across more loops" "Naively applying Parallel Looped Transformers (PLT) in our setting does not preserve the gains from looping: quality can degrade as more loops are added. We find that preserving the current-token representation, separately normalizing it and the recurrent state, and controlling how they are combined substantially improves PLT."

Today we're launching Mouse-1. A fast decision model built specifically for agent harnesses. https://t.co/zbIMGwkztp

@charles_irl @xeophon @charles_irl I’ve been looking at public benchmark traces for 48 hours straight and I can tell you that benchmark quality is abysmal. I am on the floor in disbelief - Graders not setup correctly - Harnesses and environments that are slop coded and not really testing what you think they are, meaning the agent is doing the right thing but still getting dinged - Dumb tests like running in a half specified environment that doesn’t return feedback or errors if a tool call is made incorrectly so agent can’t even correct itself - Straight up incorrect assertions - Not giving agents sensible tools in the first place or not using the APIs properly @xeophon has spent the most time on this out of anyone I’ve talked to. I just started looking at it on Monday, he’s been at it for many months

Perplexity Flash takes the crown! ⚡ In our tests, it matched Wald's accuracy at nearly twice the speed on a 16 GB Mac mini. Synthetic agent decisions, real email threads, prompt injections: see why Flash is our new pick for email-completion checks. 👇 https://t.co/w7qK7mrGQN

Whether you're pushing LLM training & serving onto NPUs ⚙️, keeping accelerators in sync with PyTorch's release cycle , taming dynamic shapes in torch.compile , or exploring ultra-low-precision training. This agenda of #PyTorchCon has you covered 👇 See you in San Jose! #PyTorch https://t.co/HiMnrCCjsw

+1 more

DeepSeek v4.1 Flash-1017 is a good model for its very low price and open weights. I benchmarked it on induction, and it does significantly better than all the other DeepSeek models. Performance between GPT-5.6 Luna and GPT-6 Luna. Not bad, and not impressive either. GitHub: https://t.co/By8Ae7jH1H

Kimi K3.1 Pro does well on the induction benchmark: it ties with/surpasses Opus 5 and is well ahead of Qwen 3.8 Max and Muse Spark 1.3, making it the top open-weights model in this benchmark. Still well below GPT-5.6 Sol, let alone GPT-6.1 Sol or Opus 5.5, so it's not frontier yet and it is more expensive than GPT-6.1 Sol given its token usage. It is great to see open models advancing. Github: https://t.co/By8Ae7jH1H

@jurlycat shows Phone Harness, an open-source tool that lets Claude Code and Codex control a real phone by reading the screen, tapping, swiping and typing. Per the post it works on iPhone through iPhone Mirroring and on Android through ADB, with nothing installed on the phone and a Hermes plugin too. Existing mobile apps becoming a surface for agents is a big unlock. https://t.co/KnjJu4Gbfz

In agent driven startups, you can scale with vendors like Replit. I patched up 40+ apps < 2 hours! There is a page to view statuses of all of the vibe coded apps & prototypes. Where you can kickoff security patches and fixes and then deploy! https://t.co/7aFZeF5Bhw

@ArgusPlatformAI On the MATH (eval) dataset the tiny 0.6B student got 44% accuracy after distillation (up from 15% for the base model). The 0.6B reasoning reference model got around 48%. The 671B teacher model got 91.2%. But yeah, it's also 1000x larger.