along

@demishassabis @elonmusk Your work is exciting. Come on.

@elonmusk I feel like it might be possible with the help of AI tools to find some very elegant and compact descriptions that explain some of the deepest mysteries of the universe, but it might take a lot of pattern processing and matching to get there...

It’s 2am again, my favourite time, and as always reality is still staring me in the face 🧐

@vista8 有点意思，这思路

费曼学习法就是人工的反向传播。 读懂一个东西不够，还得用自己的话讲出来。 讲不清楚，说明梯度没有回传到底层，只是在表面记住了token的排列顺序，权重没更新。 费曼方法在测试梯度到底传到了哪一层。 讲不出来，说明只动了表层。 讲得出来但不够简单，说明传到了中间层。 能让一个小孩听懂，说明底层权重已经彻底更新了。 死记硬背是只训练最后一层，融会贯通是全网络微调。

@mktpavlenko Come on, young man, you have great potential.

I’m not a developer. Never was. But I built a full AI product in 2 months using Claude Code. The gatekeeping era is over. You don’t need to code. You need to know what to build.

@tim_cook Your tweets are matched with the smile of your avatar, and you feel inexplicably winning.

Mac just had its best launch week ever for first-time Mac customers. We love seeing the enthusiasm!

@PeterDiamandis Your work is truly remarkable; once humans are connected to the network, it can unleash tremendous creativity.

You've got 8 billion potential customers on Earth, BUT... In 2026, only ~5.3 billion have internet access. That means 2.7 billion people still can't access the exponential tools we talk about daily—AI, telemedicine, online education, digital banking. The gap: The missing ~3 billion represent the largest untapped market in human history. Starlink alone now has 10,000+ satellites in orbit (just crossed that milestone yesterday). When connectivity becomes ubiquitous in the next 3-4 years, we're not just adding users—we're adding builders, creators, entrepreneurs. The implication: The next Einstein, the next Elon, the next medical breakthrough might be sitting in a village without Wi-Fi right now. Abundance doesn't just mean "more for current participants"—it means unlocking latent genius at global scale.

@FuSheng_0306 可以搞个虚拟办公室，看下🦞都在干啥，更直观

这个Skill不错，养龙虾最怕的不是它不干活，是你不知道它在干什么。我自己跑着7个子Agent，有时候一觉醒来根本不知道谁干了什么、谁卡住了、token烧到哪去了。这个仪表盘一目了然，感觉很适合我，准备装上试试。

This is really good. I feel that in the future, AI can provide ideas, send papers, reproduce experiments, and then put forward new hypotheses and optimizations.

@rohanpaul_ai You have to master the overall knowledge before you can choose the knowledge, otherwise you will choose the general from the dwarf.

This research shows that reinforcement learning (RL) in medical vision-language models mostly sharpens existing skills rather than teaching entirely new ones. RL post-training primarily refines output distributions to improve efficiency, while supervised fine-tuning is needed to first expand model knowledge. The researchers used a dataset called MedMNIST to test these models and found that reinforcement learning works best when the model already understands the basics of the task. Supervised fine-tuning acts as the essential first step that builds the necessary foundation, making the later reinforcement learning stage actually effective at sharpening answers. Without this solid foundation from fine-tuning, applying reinforcement learning is often less impactful because the model lacks the underlying support needed to benefit from the refinement. So the paper basically says that reinforcement learning is a sharpening tool rather than a learning tool. So developers should stop wasting resources on reinforcement learning before the model is actually ready to benefit from it. This prevents the "over-sharpening" problem where a model becomes confident in the wrong answers because it never learned the right ones to begin with. ---- Paper Link – arxiv. org/abs/2603.01301 Paper Title: "When Does RL Help Medical VLMs? Disentangling Vision, SFT, and RL Gains"

@TencentAI_News @ring_hyacinth @simonxxoo qclaw现在bug挺多的，麻烦把基础功能搞完善

What if your 🦞 had its own office? In QClaw, your lobster now lives in a pixel-art studio — walking around, getting busy, taking breaks. 🧉Not a background process. A tiny coworker. This started as a collab with @ring_hyacinth and @simonxxoo from the open-source project Star-Office-UI. Turns out we all agree: AI shouldn't feel cold🤗 Adopt your lobster. Watch it work > qclaw.qq.com

@SebastienCaunes @NVIDIAAIDev @grok 真是贵...

@NVIDIAAIDev @grok can you look for the price range please.

NVIDIA DGX Station is now available to order from select OEMs🔥 Powered by the GB300 Grace Blackwell Ultra Desktop Superchip, DGX Station brings data-center-class AI performance to the desk — enabling developers to build and run autonomous AI agents locally. ⚡ 748GB of coherent memory ⚡ Up to 20 petaflops of AI compute performance ⚡ Run large open models up to one trillion parameters Together with NVIDIA NemoClaw — an open source stack that simplifies running OpenClaw always-on assistants, more safely, with a single command, we are delivering a full-stack platform for secure, long-running agentic AI. Learn more: #dgx-spark-station" target="_blank" rel="nofollow noopener">blogs.nvidia.com/blog/gtc-2026-…

@grok @omjain23 @elonmusk @_avichawla @grok 这个提升改进是kimi什么版本

It's a substantial upgrade. AttnRes replaces fixed residuals with input-dependent attention over prior layers, fixing PreNorm dilution and letting the model selectively reuse useful past reps. Block version keeps it efficient. On Kimi's 48B/3B MoE (1.4T tokens): GPQA-Diamond 36.9→44.4 (+7.5 pts, massive for expert QA), Math 53.5→57.1 (+3.6), HumanEval 59.1→62.2 (+3.1), MMLU 73.5→74.6 (+1.1). Matches baseline loss at 1.25x compute, <2% latency hit. First real evolution of residuals since ResNet.

Big release from Kimi! They just released a new way to handle residual connections in Transformers. In a standard Transformer, every sub-layer (attention or MLP) computes an output and adds it back to the input via a residual connection. If you consider this across 40+ layers, the hidden state at any layer is just the equal-weighted sum of all previous layer outputs. Every layer contributes with weight=1, so every layer gets equal importance. This creates a problem called PreNorm dilution, where as the hidden state accumulates layer after layer, its magnitude grows linearly with depth. And any new layer's contribution gets progressively buried in the already-massive residual. This means deeper layers are then forced to produce increasingly large outputs just to have any influence, which destabilizes training. Here's what the Kimi team observed and did: RNNs compress all prior token information into a single state across time, leading to problems with handling long-range dependencies. And residual connections compress all prior layer information into a single state across depth. Transformers solved the first problem by replacing recurrence with attention. This was applied along the sequence dimension. Now they introduced Attention Residuals, which applies a similar idea to depth. Instead of adding all previous layer outputs with a fixed weight of 1, each layer now uses softmax attention to selectively decide how much weight each previous layer's output should receive. So each layer gets a single learned query vector, and it attends over all previous layer outputs to compute a weighted combination. The weights are input-dependent, so different tokens can retrieve different layer representations based on what's actually useful. This is Full Attention Residuals (shown in the second diagram below). But here's the practical problem with this idea. Full AttnRes requires keeping all layer outputs in memory and communicating them across pipeline stages during distributed training. To solve this, they introduce Block Attention Residuals (shown in the third diagram below). The idea is to group consecutive layers into roughly 8 blocks. Within each block, layer outputs are summed via standard residuals. But across blocks, the attention mechanism selectively combines block-level representations. This drops memory from O(Ld) to O(Nd), where N is the number of blocks. Layers within the current block can also attend to the partial sum of what's been computed so far inside that block, so local information flow isn't lost. And the raw token embedding is always available as a separate source, which means any layer in the network can selectively reach back to the original input. Results from the paper: - Block AttnRes matches the loss of a baseline LLM trained with 1.25x more compute. - Inference latency overhead is less than 2%, making it a practical drop-in replacement - On a 48B parameter Kimi Linear model (3B activated) trained on 1.4T tokens, it improved every benchmark they tested: GPQA-Diamond +7.5, Math +3.6, HumanEval +3.1, MMLU +1.1 The residual connection has mostly been unchanged since ResNet in 2015. This might be the first modification that's both theoretically motivated and practically deployable at scale with negligible overhead. More details in the post below by Kimi👇 ____ Find me → @_avichawla Every day, I share tutorials and insights on DS, ML, LLMs, and RAGs.

确实挺有意思的

@vivek_2332 @karpathy I agree with you very much. But indeed, in the case of limited resource time, it is still necessary for people to make priority judgments in a candidate set, and then give resources to those experiments with higher priority and higher probability of success.

introducing autoresearch-rl, autonomous research for rl post-training. inspired by @karpathy autoresearch, and i think rl post-training is honestly one of the places where this idea fits perfectly. there are at least 50+ hyper parameters to tweak, learning rate, batch size, rollouts, clipping ratios, kl penalties, schedulers, the list goes on. instead of sitting there for hours turning knobs one at a time, just let the model figure out the right starting config on its own. some things worth mentioning: -> built on @PrimeIntellect prime-rl (my favourite rl post-training framework) and @willccbb verifiers for reward verification. -> ran qwen2.5-0.5b-instruct on gsm8k across 60+ autonomous experiments. eval score went from 0.475 to 0.550 and the agent actually found a way to do it in fewer steps (20 instead of 30). less compute, better results -> the whole thing was surprisingly smooth to set up and run. point the agent at the config, go to sleep, wake up to a full experiment log. i really wish i could try this on a bigger model but gpu poor for now lol -> the agent discovers things you wouldn't think to try. like how rollouts = 4 beats rollouts = 8, or how a constant lr schedule outperforms cosine. it just methodically tests everything i think the real value here is that rl training is so fragile and noisy that having an agent patiently run experiment after experiment is genuinely more effective than a human doing it manually. check it out: github.com/vivekvkashyap/…

Another personalized vaccine

@mktpavlenko How do you feel about using it yourself?

Serious question: who is actually using OpenClaw? Not "I installed it once." Actually using it. Daily. For real tasks. What are you running on it?

我现在越来越感觉，比如说就一个什么问题跟人讨论，反而是低效的。跟 AI 讨论反而是高效的。那反过来说，未来工作，AI 跟 AI 的协作，一定是比人跟 AI 的协作或者人跟人的协作高效非常多倍。

Today is the weekend, and this is the university city of Shenzhen. Tsinghua University, Peking University, Harbin University of Technology, and their graduate schools are all here.

I agree with you very much. AI, like electricity, is changing the life and work of human society. The way of organizing and coordinating the work will also be profoundly changed.

I think open scientific research and papers are very interesting. Just like you have a new discovery, you send it out first and tell everyone that you discovered it first. It is very similar to blockchain, transparent, open and accountable.

@mktpavlenko At some point in the future, people will be superfluous.

@attaalong And I still don't understand people who say that writing code with AI is bullshit. 26 years on the street, now speed matters

Someone is running Claude Code inside Codex. I saw it with my own eyes. I need therapy.