Tensors

Codex Open Sourcing AI models 👀 huggingface.co/blog/hf-skills…

@Zoom Zoom has an AI division 😂

Zoom achieved a new state-of-the-art (SOTA) result on Humanity’s Last Exam (HLE): 48.1% — outperforming other AI models with a 2.3% jump over the previous SOTA. ✨ HLE is one of the most rigorous tests in AI, built to measure real expert-level knowledge and deep reasoning across complex problems. What that means for you: ✅ More accurate summaries ✅ Better reasoning ✅ More powerful automation in AI Companion 3.0 Click the link to learn more. 🔗 zm.me/3MxVbyS

@OpenAIDevs These are impressive. Looking for the codex models

GPT-5.2 evals

@xiangyue96 If the model is task specific it does without a doubt

There are competing views on whether RL can genuinely improve base model's performance (e.g., pass@128). The answer is both yes and no, largely depending on the interplay between pre-training, mid-training, and RL. We trained a few hundreds of GPT-2 scale LMs on synthetic GSM-like reasoning data from scratch. Here are what we found: 🧵

🚨 China's tech dominance: Leads in 7/8 AI categories per ASPI's 2025 Tracker. $56B invested in AI this year alone. #ChinaAI #TechRace

🚨 AI research crisis: One author claims 113 papers in 2025; conferences like NeurIPS hit 21K submissions. Review quality tanks amid AI slop. #AI #Academia

@DeItaone I don’t feel like this should be that big of a surprise though

DEEPSEEK USING BLACKWELL CHIPS TO BUILD NEXT MODEL: INFORMATION BLACKWELL CHIPS WERE SMUGGLED INTO CHINA: INFORMATION

@CNBC “The company said it plans to release audio-only glasses with its Gemini AI assistant and glasses that will include an in-lens display.”

Google to launch first of its AI glasses in 2026 cnbc.com/2025/12/08/goo…

@ns123abc Realistically wonder what percent of humans sleep less than him

>he doesn't know

@Alibaba_Qwen New RL dropped 🔥

🚀 We introduce Soft Adaptive Policy Optimization (SAPO) — a smooth, stable, and highly effective RL method for training large language models. Why SAPO? 🔹 Hard clipping is brittle — gradients vanish or explode 🔹 MoE models amplify variance, making training even more unstable SAPO replaces hard boundaries with a continuous, temperature‑controlled gate that: ✨ Smooth trust‑region behavior → no abrupt gradient drop ✨ Sequence-level coherence → align sequence‑level behavior ✨ Token-level adaptivity → preserves useful gradients & boosts sample efficiency ✨ Asymmetric temperatures → significantly improved stability, esp. in MoE models What does this mean in practice? 📈 Longer stable RL runs 📈 Higher Pass@1 📈 Stronger performance on Qwen3‑VL across math, coding & multimodal tasks SAPO offers a more scalable and reliable foundation for RL-tuning large language & multimodal models. 📄 Paper: arxiv.org/abs/2511.20347 📚 Blog: qwen.ai/blog?id=sapo

@maximelabonne @asapzzhou I see, only .5 B are available. Are there benchmark results?

@maximelabonne @asapzzhou Woah this is sick. Are all diffusion based qwen variants published to hugging face?

Open recipe to turn Qwen3 into a diffusion LLM 👀👀 > Swap the causal mask for bidirectional attention > Source model matters a lot for performance > Block diffusion (BD3LM) >> masked diffusion (MDLM) > Light SFT with masking Great work from @asapzzhou with his dLLM library!

@akshay_pachaar @ClementDelangue Early concept is good but all of my workflows will run into issues without it being an iterative process. Looking forward to feedback loops to further improve reward functions, prompting, etc. in the future.

HuggingFace just made fine-tuning 10x easier! One line of English to fine-tune any open-source LLM. They released a new "skill" you can plug into Claude or any coding agent. It doesn't just write training scripts, but actually submits jobs to cloud GPUs, monitors progress, and pushes finished models to the Hub. Here's how it works: You say something like: "Fine-tune Qwen3-0.6B on the open-r1/codeforces-cots dataset" And Claude will: ↳ Validate your dataset format ↳ Select appropriate GPU hardware ↳ Submit the job to Hugging Face Jobs ↳ Monitor training progress ↳ Push the finished model to the Hub The model trains on Hugging Face GPUs while you do other things. When it's done, your fine-tuned model appears on the Hub, ready to use. This isn't a toy demo. The skill supports production training methods: SFT, DPO, and GRPO. You can train models from 0.5B to 70B parameters, convert them to GGUF for local deployment, and run multi-stage pipelines. A full training run on a small model costs only about $0.30. Link to the full tutorial in the next tweet!

@KobeissiLetter 25% is insane. Wow

BREAKING: President Trump says he has called China's President Xi and approved sales of Nvidia's H200 chip to China. Trump says 25% of revenue will be paid to the US and the "same approach" will apply to AMD, Intel, and others.

@AmanSharma_554 Daily reminder for new devs to use git ignore and .env

@StockSavvyShay I should have already understood that it would be high due to the ecosystem but I’m surprised by Microsoft’s MAU

$GOOGL Gemini’s curve is the one actually accelerating with It more than doubling MAUs from January to November. $MSFT Copilot trend is different since they have huge enterprise distribution with no corresponding lift in monthly users which tells you Microsoft isn’t converting enterprise access into monthly engagement even as the broader category expands.

@Rainmaker1973 lol does it recycle old news articles for headlines in the browser

New browser tool lets users freeze the internet in 2022 to escape AI-generated content. Say “goodbye” to “AI slop”. A new browser extension called Slop Evader lets you surf the web as if AI never existed. Created by artist and researcher Tega Brain, it automatically filters Google search results to show only pages published before November 30, 2022—the day generative AI went mainstream. The result is a quieter, more human internet: no AI-written listicles, no synthetic stock photos, no deepfake videos or bot-penned product reviews. Just the pre-2023 web, frozen in time. Slop Evader (and similar tools like Kagi’s SlopStop) reflects a growing backlash against the flood of low-quality, machine-generated content that has overwhelmed search engines and social feeds. Brain stresses that the extension isn’t meant to be a forever solution. Instead, it’s a deliberate act of protest—an easy way for everyday users to reject the creeping artificiality of today’s web and demand something better. You lose access to anything new, of course, but for many, the trade-off feels worth it: clarity over noise, authenticity over algorithm.

@thdxr Legit haven’t heard tabs vs spaces in years, wow this is so right

there have been so many religious debates in tech - tabs vs spaces, static vs dynamic typing, fp vs oop all of this is nothing compared to the way people talk about coding agents "codex shall lead me to the promised land beware the deceiver known as opus"

@iruletheworldmo Google will remain since they can always operate their AI team at a loss

it’s not really a bubble. we’re just waiting to see who’s the netflix of ai. they’ll eat all of the value. and there’ll be a lot. seems like google rn.

@thdxr I think most companies use it as free thinking time to improve morale, others target debt areas in hopes for quick win or finding new prioritizations

i still don't get hackathons