Edward Welly @Ed_Welly

新消息：一名 Reddit 网友称，他告诉 Fable 5，如果不能把 80 美元变成 5000 美元，GPT-5.6 就会“干掉 Anthropic”。 结果 Fable 5 直接开满杠杆，做了超过 1 万笔加密货币交易。

Nature research paper: Universal cell embedding provides a foundation model for cell biology go.nature.com/4eSkxmZ

I hear a lot of people talking about how LLMs will revolutionize research. But I don't see many quantitative measures to support this. Until I came across this paper which compares LLM-generated ideas against published research: arxiv.org/abs/2607.01233 Key findings: LLMs produce reasonable ideas, but they cluster around synthesis over novel problem framing. It is worth reading it!

Can LLMs predict the next World Cup champion? Goodfire partnered with @EternisAI to improve how LLM forecasters use available evidence and manage uncertainty. We found models were overconfident in their predictions – but probes significantly improved calibration. (1/6)

NIH’s Unified Funding Strategy will allow us to take a more comprehensive approach to our funding decisions. While peer review remains essential, we’ll no longer rely on a single score. Our new model using multiple inputs will help us make more informed decisions on the research we fund and be better stewards of scientific opportunity. Learn more: bit.ly/4oEnxpL

I was clearly wrong about Anthropic. They are obviously currently the leader in AI. No company has released a model as good as Mythos/Fable and they will undoubtedly have Mythos 2 ready soon. And I would never cut them off in a way that hurt them badly, even as a competitor. That’s not my style. Tesla open sourced its patents and we made the Supercharger network available to all competitors, even though we could have made it a walled garden. SpaceX launches competing satellite systems with no increase in price or use of unfair terms. Even my worst enemies can attack me on this platform. …

AI 2040 has been released! What a day. AI 2027 was probably the most forward-looking blog about the future, and much of it has come to pass with remarkably accurate timing. Alongside Aschenbrenner’s “Situational Awareness,” it is probably the most important blog for theoretically preparing ourselves for the future. That makes me all the more excited about AI 2040! I’m going to read it immediately, and I can only recommend that you do the same.

An AI summary of Biomni capabilities, by @NotebookLM

What if AI could learn the shared structure of human health across diseases, modalities, and time? Excited to share our new preprint introducing RisQ, an AI model that learns a shared representation of health across diseases, modalities, and time. 🚀 Some highlights: 🧬 A single model integrates genetics, biomarkers, lifestyle, medications, environmental factors, and longitudinal clinical history. 🌍 Trained in ~500,000 UK Biobank participants, RisQ generalizes to the independent All of Us cohort without retraining, outperforming strong disease specific and multi disease baselines! 🔬 The learned representation reveals reproducible cross disease risk programs and connects them to rare genetic variation, recovering established biology while also highlighting new biological hypotheses. Building on exciting advances such as Delphi, RisQ extends this paradigm by integrating multimodal information and learning a shared representation across diseases, modalities, and time. This representation can then be queried to estimate the risk of hundreds of diseases over different prediction horizons. For me, the most exciting aspect is not that RisQ predicts disease better. It is that the results suggest human health has a learnable shared structure that can be leveraged for prediction, patient stratification, and biological discovery. We believe this is another step toward foundation models of human health, helping us move from modeling individual diseases to understanding the patient as a whole. Congratulations to our outstanding co-first authors Paul Hager, Benedikt Roth, Niklas Bühler, and Diyuan Lu for driving this ambitious project. 👏 And a huge thank you to fantastic partner labs, true collaborative (and fun!) effort with Rückert and Schnabel labs pulled together by Paolo Casale! Preprint: medrxiv.org/content/10.648… #ComputationalBiology #DigitalHealth #FoundationModels

keyboard_arrow_left Previous keyboard_arrow_right Next

High-throughput profiling of chemical-induced gene expression across 93,644 perturbations nature.com/articles/s4159…

Researchers in Science report the development of a general-purpose biomedical AI agent that can help automate biomedical research workflows. The authors say their results point “toward a future in which AI agents work alongside human researchers to accelerate biomedical discovery from basic research to translation.” Learn more: scim.ag/3QYElfh

Today, we're excited to share that Biomni is published in @ScienceMagazine. Biomedical research is still fragmented, manual, and difficult to scale. In this work, we introduce Biomni - the first general-purpose biomedical AI agent with an integrated biology environment that can reason, plan, and execute end-to-end scientific workflows. We show that, with the right environment and harness, AI can automate large-scale omics analyses, orchestrate laboratory robotics, optimize molecular properties, and even train new AI models for biology. We also introduce a reinforcement learning recipe for continually improving biomedical AI agents, enabling open-source models to achieve frontier-level performance. It's surreal to look back. We started the Biomni project in early 2024, when agentic AI was still nascent. It is exciting to see tens of thousands of biologists collaborating with agents every day to accelerate science. Try Biomni: biomni.phylo.bio Read more: science.org/doi/10.1126/sc… This work is not possible without this truly inter-disciplinary team: @serena2z @hcwww_ @YuanhaoQ Minta Lu, Ryan Li, @yusufroohani Lin Qiu @shiyi_c98 Gavin Junze Di @rickwierenga @kavi_deniz Sherry @TianweiShe Shruti Jennefer Xin Zhou @MWheelerMD Jon Bernstein @MengdiWang10 @PengHeAtlas @zhou_jingtian @SnyderShot @lecong Aviv Regev @jure @StanfordAILab @genentech @phylo_bio @arcinstitute @UW @berkeley_ai @RetroBio_ @tamarindbio @Princeton @UCSF

GitHub 上又炸出一个离谱项目，这玩意直接把市面上所有的 AI 痕迹探测器给干废了。 项目叫 harshaneel/humanize，作者死磕了 50 多篇 2024 到 2026 年的前沿学术论文，提炼出了一套“终极去 AI 味”指令集。它不是什么需要复杂配置的软件，而是纯开源的底层 Skill 文件，直接喂给你的大模型就能产生质变。 降维打击：9大仿生杠杆，碾压主流 AI 探测器。 全模型通用：Claude、ChatGPT、Cursor 随便挂载。 法医级质检：自带扫雷，逐句给你标红机器味证据。 零成本白嫖：不用买代写服务，开源规则直接拷。 🔗 GitHub：github.com/harshaneel/hum…

"LLM-as-a-Verifier: A General-Purpose Verification Framework" The key idea of this paper is that it does not ask for one rough score, it reads the model’s full uncertainty over scores, which helps to make the judgment much more fine-grained. This approach lets agents pick better solutions, track progress, and learn from denser feedback.

现在很多人聊 Research Agent，默认期待是： 读论文、找 gap、想 idea、做实验、写 paper。 但耶鲁大学的这篇论文问了一个更深入的问题： LLM 生成的研究想法，和人类研究者真正做出来的 paper idea，差距到底在哪里？ 论文的核心概念叫 research taste，可以理解成「研究品味」：你通常会发现什么样的问题，用什么方式把它变成一个 contribution。 它的实验设计挺巧。 作者从真实论文里抽取 human idea，再反推 4 到 8 篇最可能启发这个 idea 的 prior works。然后把这些 prior works 的标题和摘要给 LLM，让模型在同样的文献上下文里生成一个新的 research idea。 这样比较的重点，就从「单个 idea 看起来新不新」推进到「一批 idea 的分布像不像人类研究者」。 论文一共用了 11,683 个 human ideas，覆盖 ML 会议论文和 Nature Communications 里的自然科学论文；模型包括 Claude、Gemini、GPT、Qwen、DeepSeek 等 9 个设置。 结果很有意思：人类 idea 的分布明显更宽。 在人类论文里，只有 12.1% 的 idea 属于 bridge opportunity，也就是把不同文献、方法或证据流连接起来。 但在 LLM 生成的 idea 里，这个比例变成了 47.1% 到 64.2%。 方法层面也类似。 人类论文里，synthesis / unification 只占 5.1%；LLM 这里是 22.5% 到 38.7%。 简单说，模型很喜欢把研究问题理解成： 这里有两个东西，可以把它们整合一下。 论文还发现，给模型更多上下文也没有明显解决这个问题。full-paper context 版本没有让分布更接近人类；开启 thinking mode 甚至可能让模型更强化这种 bridge-and-synthesis 倾向。 这篇最有意思的地方在于，它没有简单说 LLM idea 好或不好，而是指出了一个更隐蔽的问题： LLM 可以生成很多看起来合理的研究想法，但这些想法的类型可能高度集中。 这对 AI for Science 和 Research Agent 很关键。 对于真正强的 Research Agent，难点在于能不能形成更宽的 problem-finding 能力：发现 failure、识别机制、提出测量工具、构造系统、修改局部假设，而不总是回到「整合两个已有东西」这条安全路径。 自动科研如果缺少这种多样性，最后可能会变成一种很流畅的同质化机器。 📎 arxiv: arxiv.org/abs/2607.01233

Andrew Ng just dropped a 3-hour course on how to become an AI Engineer in 2026: • 00:00 - How to build agentic AI systems • 04:25 - Future of AI engineering • 23:38 - AI Prompting full course • 2:52:17 - Creating an app with AI in 30 minutes This 3-hour watch could replace 10 AI engineering courses on the internet. Watch it today, then read how to run a self-improving system in the article below.

Codez @0xCodez

4 weeks ago

x.com/i/article/2065…

57 362 2K 5.3M 7K

离谱了兄弟们，现在 AI 圈发论文的速度，简直是把人当牲口用。每天被新模型轰炸根本看不完？ GitHub 上有个老哥实在受不了了，直接搓了个“全自动读论文机”，这玩意简直就是打工人的究极外挂。 项目叫 auto-paper-digest，一条极度硬核的开源自动化流水线。它不仅帮你盯盘，还能把晦涩的 PDF 嚼碎了喂到你嘴里。直接打通了 Hugging Face 和 NotebookLM，主打一个用 AI 魔法打败 AI 论文。 🔥 核心操作有点狠： ▪️ 无情盯盘：全自动追踪 Hugging Face 每周爆款论文，一个不漏。 ▪️ 甩手掌柜：自动下载 PDF 直接塞进 NotebookLM，不用自己啃。 ▪️ 降维打击：一键把干巴巴的论文生成视频解说，通勤直接当播客听。 ▪️ 私人智库：所有摘要自动打包归档，变成随时可搜索的周报数据库。 白嫖党和科研党狂喜，以后再也不用贩卖信息过载的焦虑了。这波操作确实是懂的都懂。 🔗 GitHub 传送门：github.com/brianxiadong/a…

MIT published a paper arguing that every AI model on earth is secretly converging on the same "brain." The paper is called "The Platonic Representation Hypothesis." The claim inside it is one of the strangest ideas in modern machine learning, and once you see it you cannot unsee it. For years, everyone assumed that a model trained on images and a model trained on text were building fundamentally different things inside themselves. Different data. Different architecture. Different world. A vision model learns what a cat looks like. A language model learns what the word "cat" sits next to. Two separate universes with no reason to line up. The researchers checked whether that was actually true. They took 78 vision models and a stack of large language models, and measured how each one organized concepts internally, not what they output, but the shape of the relationships between ideas in their heads. Which things they treat as close together. Which things they treat as far apart. Then they compared the shapes across models that had never seen each other's data. The shapes were lining up. And here is the part that should stop you cold. The bigger and more capable the models got, the more their internal maps agreed with each other. A better vision model and a better language model don't drift apart. They converge. As if they were both climbing toward the same summit from opposite sides of a mountain. The authors put it in a line that sounds almost like a joke, borrowed from Tolstoy: all strong models are alike, each weak model is weak in its own way. Then they took it one step further, and this is where it stops being a curiosity and starts being unsettling. They found that how closely a language model's internal map lined up with a vision model's internal map actually predicted how good that language model was at reasoning and at math. The models that saw the world more like the other modality did better at problems that had nothing to do with images at all. So the question the paper asks is the obvious one. If a model that only reads text, and a model that only sees pixels, and a model trained on a completely different objective are all drifting toward the same internal representation as they get smarter, what is that representation a representation of? Their answer is the thing that gives the paper its name. They argue the models are all converging on a single shared statistical model of the reality that generated the data in the first place. Text is a shadow of the world. Images are a shadow of the world. Sound, touch, everything, different shadows cast by the same underlying thing. And a big enough model, trained on enough of any one type of shadow, starts reconstructing the object casting it. Plato said this in 375 BC. The allegory of the cave. Prisoners chained facing a wall, watching shadows, mistaking the shadows for reality, while the real forms exist outside the cave, casting everything they see. The MIT team took his allegory literally and pointed it at neural networks. The training data is the shadows on the wall. The model, they argue, is slowly turning around toward the fire. They even proved a version of it mathematically. Under certain conditions, a whole family of learning algorithms is provably pulled toward representing the same underlying statistical structure, the co-occurrence relationships baked into reality itself, regardless of whether they're fed words or pixels. Different sensors, same answer. The implications the paper draws are the part that should matter to anyone building this stuff. If it's true, then to build a better language model you should train it on images, because pictures carry information about the same reality that words are trying to describe. They cite evidence this already works. It means translation between any two modalities gets easier the smarter models get, because they're all speaking dialects of the same underlying language. And it means, their words, that hallucination might decrease with scale, because a model converging on an accurate model of reality has less room to invent things that reality doesn't contain. Now the honest part, because the authors are honest about it and a viral thread that skips this is lying to you. This is a hypothesis, not a verdict. On their own measurement, the alignment between vision and language models climbs clearly with scale but only reaches about 0.16 on a scale where 1.0 is perfect. They flat-out ask in their own paper whether that number means strong convergence with noise on top, or weak convergence with a mountain still left to explain. The clean math only holds in an idealized world where nothing is lost between reality and observation, which is not the world we live in. Some things a picture can show that a sentence never will, and vice versa. And in domains like robotics, they see no convergence yet at all. So it might not go all the way. The cave might be deeper than one paper can measure. But sit with the shape of what they found anyway. Systems built by different labs, on different continents, trained on different senses, for different reasons, with no coordination, are independently drifting toward the same internal picture of the world. The smarter each one gets, the more they agree. And the thing they seem to be agreeing on is reality itself.

吴恩达： Loop 有三层，大多数人只看到第一层 “Loop Engineering”最近很火，但很多人只关注第一层。 吴恩达在最新指出：Loop 其实有三层，代码循环只是最基础的一层。 第一层：Agentic Coding Loop AI 自己写代码、自己测试、自己修复迭代（分钟级循环）。 第二层：Developer Feedback Loop 开发者从 QA 变成产品经理：审视产品、调整方向、提供高质量上下文和判断（小时级循环）。 这里的核心是人类的上下文优势。 第三层：External Feedback Loop 真实用户反馈、A/B 测试、生产数据等（天/周级循环）。 这一层最慢，却决定产品最终生死。 总结： 真正厉害的不是让 AI 自己跑代码，而是设计好这三层循环，让人类和 AI 各司其职、相互强化。 未来工程师的角色正在扩大，越来越需要兼顾产品战略。

Andrew Ng @AndrewYNg

2 weeks ago

“Loop engineering” is a hot buzzphrase after mentions of it by Boris Cherny (Claude Code’s creator) and Peter Steinberger (OpenClaw's creator) went viral on social media. Loops are now a key part of how we get AI agents to iterate at length to build software. In this letter, I’d

380 2K 8K 628K 10K

看完 Karpathy 这段 State of GPT，终于能把 ChatGPT 的底层逻辑讲清楚了。 1. 预训练 先吃掉海量互联网文本，学会预测下一个 token。 这一阶段最贵，吃掉 99% 训练算力。 2. 监督微调 把模型从“会续写文字”，调成“会按人类指令回答”。 3. 奖励模型 让人类去比较多个回答，告诉模型哪个更好。 4. RLHF 用强化学习，让模型更接近人类喜欢的回答方式。 所以你平时用的 ChatGPT，不是原始 base model。 它是被一步步训练成“助手”的模型。