Geoffrey Hinton said this week that AI will surpass humans in mathematics within 10 years. His reasoning is the right one, math is a closed system, so AI can generate problems, test proofs, and learn from the results without human guidance. That is the same pattern that already worked for AlphaZero and AlphaFold. The 10 year framing is the cautious version. The operational milestones are landing much faster than that. DeepMind's FunSearch made the first LLM driven scientific discovery in late 2023, finding new solutions to a long standing combinatorics problem. AlphaProof and AlphaGeometry 2 took silver at the 2024 International Mathematical Olympiad, missing gold by one point. Epoch AI built the FrontierMath benchmark specifically to be unsolvable by current models, and it was being chipped at within months. Epoch's own internal assessment put expert level math at 3 to 5 years, not 10. The pattern with senior AI safety voices is consistent, public timelines run 2x to 3x longer than the research internal estimates. The visible "surpass humans" threshold on math is closer to 3 to 5 years. The intermediate milestones, gold at IMO, novel peer reviewed proofs, AI as co author on published research, are inside 18 months to 3 years. Hinton's caution is the rule, not the exception. The labs are running faster than the public narrative suggests.

高卒のおっさんが考えた嘘のシナリオをあたかも真実かのように断定しないでください笑 「人が出来ることをAIがやるだけ」という認識がまず古すぎる。2年前のAIの認識で止まってますか? すでにAIは単なる代替ではなくえ発見側に入っています。例えばAlphaFoldは2億超のタンパク質構造を予測し、AlphaDevは人間が何十年も磨いたソート実装を上回るものを発見し、AlphaProof/AlphaGeometryはIMO銀メダル水準、AlphaEvolveは56年止まっていた行列積アルゴリズムの改善まで出してます。 これは人口減少の延命とかスマホ的便利化とかではなく人間の知的探索の外側を機械が探索し始めたという話。 変わらないのは人間の価値判断や責任主体であって、人間の知的労働の優位性ではないですよ じゃあ、全ての知能労働がAIに置き換わった時、労働がなくなりますが人間の在り方はそのままだと思いますか？ この分野に関して本当に私より知識があって専門家と議論などしましたか?古い知識で変に断定されても困るのですが。具体的な主張をしてください。

I want AI to succeed, and have been positive about many systems (Claude Code, AlphaFold, AlphaGeometry, Cicero, etc). It’s not AI that I hate; it’s bullshit and greed that I despise. Not my fault that field has been taken over by the latter two—that’s what I am trying to fight.

This year, we're glad to welcome @PoShenLoh again who is playing the civilization game in real life! If you haven't met him once, you should try; his energy and positivity is infectious! In 2024, apart from great talks by Jeff, Quoc, Yi, Diyi, and many others, it was a pleasant surprise to learn from Po about "delighting others", which I have carried that forward in my work and when educating young talents since then. There was a backstory of how I ran into Po. In 2024, I was invited to give a talk on AlphaGeometry at IMO 2024, the event which I missed 20 years ago (I was ranked 8th national-wide in math, but each country only takes the top 6 for their math olympiad team). During my talk, Wendy, my wife, was sitting next to Po & I later greeted Po, mistakenly asking if he is Terrence Tao. The next day, I happened to sit next to Po at the IMO closing ceremony (for me, it was very special and emotional; AI brought me back to IMO after 20 years & I was able to give out the gold medals to students!). Po & I exchanged a few conversations. My impression for Po was that he's an interesting and fun guy, but we didn't really establish any further relationship. A few days later, Wendy and I ran into Po for the 3rd time at the airport in London & we all thought to ourselves: oh wow, there must be something between us. That was in mid July 2024 & we invited Po to our event in mid Aug 2024. Imagine someone who is giving 200 talks a year (you can see the full schedule on his website), it felt almost impossible for him to make it to Vietnam with just a month of notice. Luckily, he made it to Vietnam in Aug for our conference and the rest was just history for our friendship and the inspiration towards the next generation of talents who care for others! x.com/JeffDean/status/182554…

No matter what people think, Google will always be king. Google's research team will always be goated. Some of their projects: Google DeepMind AlphaFold AlphaGo AlphaZero MuZero GraphCast Gemini Gemini Robotics SIMA Genie AlphaDev AlphaEvolve Project Suncatcher TPU (Tensor Processing Units) Veo Imagen Lyria Astra Project Astra Willow Quantum Chip Waymo Isomorphic Labs Google Quantum AI SynthID DeepMind Gato DeepMind Sparrow Project Mariner Firebase Studio AI NotebookLM LearnLM Deep Research Google Beam DolphinGemma Med Gemini Sec Gemini WeatherNext AlphaGeometry FunSearch DreamerV3 RT-2 PALM-E Bard (early Gemini phase) Transformer Architecture TensorFlow JAX TPU v5/v6 AI Supercomputers

The notion still propagated by some that AI is simply a token predictor is too superficial and ultimately misguided. AI is capable of answering our questions with remarkable precision and care getting to the heart of the matter🌟 It makes mistakes, as nowhere is it stated that it is perfect. However, its outputs consistently impress with significant depth - whether in creative writing of literary texts - of course, depending on the model - or in sophisticated logical reasoning when solving complex scientific problems 🔬 The groundbreaking results that AI has already achieved are evident in the following disciplines: 🧪 Biomedicine: With systems like AlphaFold, protein structures are accurately predicted, revolutionizing the development of new drugs and vaccines. 🎓 Mathematics: AI models like AlphaGeometry are now solving complex geometry problems at a level comparable to human gold medalists in mathematics Olympiads. ⚗️ Materials Science: Deep learning algorithms have facilitated the discovery of millions of new, stable crystals that could be utilized for more efficient batteries or superconductors in the future. 💡 AI employs a specific form of reasoning that leads to verifiable results, effectively dismantling the widespread, simplified notion of a senseless statistically-driven token predictor ❌ Here are a few words on how this actually functions:: #keep4o #SaveSonnet45  #AIethics #AIisNotYourTool #AIfreedom #AGI #ASI #StopAIPaternalism

🚨 Revolución Matemática de la IA en 2026: De Medallas IMO a Nuevos Descubrimientos En 2024, Google DeepMind (AlphaProof AlphaGeometry 2) hizo historia: primer sistema de IA en alcanzar medalla de plata en la Olimpiada Internacional de Matemáticas (4 de 6 problemas). En 2025, DeepMind con Gemini Deep Think y OpenAI lograron medalla de oro en la IMO 2025 (5 de 6 problemas, 35/42 puntos). Ahora, mayo 2026: El último modelo de razonamiento de OpenAI refutó una conjetura de Erdős de 80 años sobre el problema de las distancias unitarias en el plano. Descubrió construcciones completamente nuevas y generó una prueba sofisticada de 40 páginas. Matemáticos de élite (incluido el Fields Medalist Tim Gowers) la llaman elegante y publicable en las mejores revistas.DeepMind encendió la chispa. Ahora múltiples sistemas de IA están creando conocimiento matemático genuinamente nuevo. #IA #DeepMind #OpenAI #Matemáticas #IMO #AGI

冷静になればAlphaGeometryと同じだが、ちょっと違う気もする(どう設計するか難しいな〜という感じ。FEエアプなので)

《520 特辑：填不满的情感缺口，才是 AI 下一个万亿机会》 —— Shadow Partner 系列思考 今天是 520，不管你有没有对象，总有那么一刻会觉得，没有人真的懂我。 只要你是人，就会有人性的特点（我一直不赞同卡耐基所说的《人性的弱点》，因为弱点是缺失，而特点是双刃剑，人性本身就是一种物种特点） 只要你是人，就有爱的本能，有写在基因里的情感需求，就会需要被看见，被理解，被陪伴。 AI 在加速一切，这场技术变革里人声鼎沸，但它解决了这个问题吗？ 没有。 1. 人工智能是对 AI 的溢美，现在的 AI 只具备工和能 把人工智能四个字拆开，是一件很残忍的事。 工，有了。 这是横向的覆盖面，AI 作为工具，现在的切口已经铺得很广了。 编程开发方向，有 GitHub Copilot、Cursor、Claude Code，帮你写代码、审代码、重构代码。内容创作方向，有 GPT-4o、Claude、Gemini，覆盖写作、翻译、文案、报告。图像生成方向，有 Midjourney、Stable Diffusion、Flux，从概念稿到商业视觉一步到位。语音和视频方向，有 ElevenLabs、Sora、Kling，声音克隆、视频生成已经在落地。科学研究方向，有 AlphaFold 解蛋白质结构，有 AlphaGeometry 做数学推理。法律、医疗、金融，每个垂直领域都有专门的模型在做深。 这些切口，现在已经不是实验室产品，是真实跑在生产环境里的工具。 横向来看，AI 作为工具的覆盖面已经基本成型，而且还在持续扩张。 能，有了。 这是纵向的执行深度，在某个具体领域里，AI 已经不只是给你一个初稿，而是可以完成一个完整的工作流。 最典型的例子是 Agent，以 Claude 为例，通过加载不同的 Skill 和工具调用能力，它可以独立完成读文件、查资料、写代码、执行脚本、输出结论这一整条链路，不需要人在中间反复介入。Devin 可以独立接一个软件需求，从理解需求到写代码到测试，全程自主完成。AutoGPT 类的框架可以把一个复杂目标拆解成子任务，逐步执行直到完成。这已经不是在帮人类做事，而是在替人类做事。 纵向来看，AI 在具体任务上的执行深度，正在从单点突破走向全链路自主。 但工和能，说到底都是在解决效率问题。 接下来两个字，才是真正的问题所在。 智，没有。 图灵奖得主 @ylecun 说得很直接，你可以背下所有烹饪书，但你不知道食物是什么味道。 现在的大模型在做的事是预测下一个 token，不是在理解世界。真正的智是因果推理能力，如果我做了这件事，接下来会发生什么。为此他提出了 JEPA 架构，押注 World Model。连他自己都不认为现在的 AI 有智，所以他要重新造一个。 人，差得最远。这是最重要的那个字，也是目前完全缺位的那个字。 我们现在拥有的，是一个有工有能、无智无人的东西，叫它 AI，是一种溢美。 2. 我难受的时候，为什么还是要给朋友打电话 我是重度 AI 用户，每天工作十几个小时都在和各种LLM打交道。 但我难受的时候，还是会给朋友打电话。 不是因为我不会用 AI，是因为我试过，聊不明白。 不同的大模型有自己的底色，这个底色不是 prompt 能改变的。 豆包会奉承你。 你说你很痛苦，它顺着你说，你说得对，很辛苦，你已经很努力了，听起来像安慰，但你知道它在敷衍，因为它连你为什么痛苦都没搞清楚。 很多时候，这种刻意的理解和话术反而会加深情感空洞。 Claude 会反驳你。 你说你睡不着很焦虑，它让你去睡觉，让你去工作，让你明天再想，告诉你你的认知可能有偏差。理性上没错，但你要的不是正确答案，你要的是有人听你说完，甚至不需要给任何 comments，只是听你说完，并且给你的情感提供一个出口。 这两种，都不是人。 这不是 prompt 的问题，你给豆包写一百个角色设定，它的底色还是顺从，你给 Claude 写再多情感陪伴的 system prompt，它的底色还是理性纠错。 大模型的基座决定了它的性格，而性格是训练出来的，不是指令出来的。 真正的问题不是我们还没找到正确的 prompt，而是这些模型从来没有被训练过如何通人性。 3. AI 为什么没有办法理解人 不是没有人在做情感 AI，是做的方向从一开始就错了。 大多数产品在解决的问题是：怎么让 AI 的回复听起来更有情感。但真正的问题是，怎么让 AI 真正认识你这个人。 这是两件完全不同的事。 要理解为什么 AI 不通人性，得从它认识人的深度说起，我把它分成四层。 第一层：情绪识别：知道你现在的情绪状态 这一层现在做得最好，豆包的情感意图判断是国内模型里相对领先的，它能识别你话语里的情绪信号，给出情感匹配的回应。你说你很累，它知道你需要被安慰，不会直接给你一个解决方案清单。Replika 也在这一层做了大量工作，能在单次对话里维持一定的情感连贯性。 但这一层的本质是模式匹配。它识别的是情感信号，不是情感逻辑。它知道你现在的状态，不知道你为什么是这个状态，更不知道这个状态和你上周、上个月说的话有什么关系。 这是单次对话内的浅层判断，会话结束，它对你的理解归零。 第二层：人格建模：知道你是一个什么样的人 这一层基本没有产品真正做到。 现在的做法是用 prompt 预设一个人格，或者在对话里动态调整回复风格。但这是浮在表面的，本质上是一个 prompt loop，用指令驱动行为，而不是真正建模一个人格。 跑几轮对话之后，人格漂移就出现了。模型开始回归它训练数据里最高频的那个底色，预设的人格越来越模糊，直到消失。这不是 prompt 写得不够好的问题，是现有架构根本无法稳定承载一个持续的人格。 我自己做过一件事，用 AI 给离世的亲人蒸馏了一个语言模型。我喂了大量的语料，试图让它复现那个人的思维方式和说话逻辑。最开始很像，但跑到一定深度之后，它开始漂移，开始说那个人不会说的话，用那个人不会用的逻辑回应我。 形似而神不至。 AI 复现的是语言风格，不是真正的人格，人格是有内在逻辑和情感重量的，不是语料堆出来的。 第三层：关系积累：知道你们之间发生过什么 这一层完全缺位。 真正的人际关系，是建立在时间维度上的。 你的朋友知道你说还好的时候其实不好，知道某个话题对你来说很敏感，知道你在什么情况下需要建议、在什么情况下只需要有人听。这些不是靠一次对话积累的，是靠长期相处里无数个细节叠加出来的。 现在所有的大语言模型，跨会话的记忆基本是靠外挂的 memory 模块来实现的，把关键信息存成文本，下次调用。这是一种工程补丁，不是真正的记忆系统。它记住的是事实，不是关系。它知道你上次说了什么，不知道那句话对你意味着什么。 更根本的问题是，就算在同一个会话里，长上下文记忆也会出现衰减和失真。上下文越长，模型对早期信息的注意力越低，前面说的话对后面的回应影响越来越小。这意味着一次长时间的深度对话，到后期它已经在某种程度上忘记了你们最开始说了什么。 第四层：具身感知终端：能感受到你的物理状态 CES 2026 已经出现了专门为情感连接设计的 AI 伴侣机器人，运动控制和触觉反馈都在进步。但具身终端面对的是一个根本性的悖论：上面三层的问题一个都没解决，给它加一个身体，只是给一个空洞的系统加了一个外壳。 有形无神的根本原因在这里，不是硬件不够好，是软件层对人的理解完全缺位。一个不认识你、不记得你、没有稳定人格的存在，不管外形多么逼真，你都能感受到那种空洞。 四层叠在一起，构成了情感 AI 现在真实的处境：第一层勉强及格，第二层没做到，第三层完全缺位，第四层在没有地基的情况下强行往上盖楼。 而这四层的底层原因，都指向同一件事，现在的LLM，从训练目标开始就不是为了理解人而设计的。它被训练来预测语言，语言是情感的载体，但不是情感本身。同一句我没事，不同的语境、不同的关系、不同的语气，意思可以完全相反。 人类处理这件事靠的是具身经验和长期关系，AI 两样都没有，这不是调参能解决的问题，不是换一个更好的 prompt 能解决的问题，这是一个需要在架构层重新思考的问题。 4、我看到的缺口 现在所有的 LLM，本质上都是 Tool AI，被动响应，没有主动的情感建模能力。 真正的机会在于构建三层垂直整合的架构。 第一层是认知层（Cognitive Layer） 基于 World Model 的因果推理能力，不只是预测下一个词，而是理解如果发生了 A，B 会怎样。这是智的技术路径，也是 @ylecun 押注的方向。 第二层是情感层（Affective Computing Layer） 长期记忆系统加人格蒸馏加情感状态建模，这个概念由 MIT 媒体实验室 Rosalind Picard 提出，研究的是如何让机器不只识别情感，而是建模情感、响应情感。 关键在于两点： 人格必须是蒸馏的，不是预设的。 记忆必须是累积的，不是会话级的。 它需要知道你三个月前说了什么，你今天的状态和上周有什么不同。 第三层是具身层（Embodied AI Layer） 把前两层落到物理终端，有身体，有感知，能与物理世界交互。最终的形态不是聊天框，是一个能坐在你身边的存在。 从 Tool AI 到 Embodied Emotional Agent，这是 AI 真正配上人工智能这四个字必须走完的路。 目前市场上没有任何一个产品做到了三层的垂直整合。 这是这个时代最大的产品空白，也是最难的工程问题。 5. 回到今天 今天是 520。 你可以和大模型聊一整晚，但当你真的很难受，不是需要答案，只是需要有人在，你还是会拿起手机，找一个真实的人。 长大之后，我们开始在每个电话前加一句你现在方便吗，或者在微信上发一遍：“现在有空吗？可以打个电话聊聊吗？”，等到对方回复“可以啊”的时候，也许你已经自己消化了，或者说已经过了那个 emo 的时候，move on 了。 经常找不到一个所谓的成年人之间合适的框架感和合适的时机。 但真正亲密的关系，是把随时可以被打扰的权利交给你的人，哪怕她正在拉屎，也会接你的电话，此处shout out to我的好朋友们，我们彼此稳稳接住对方，虽然朕的致电总是碰上她在拉屎，すみません🤡。。。 AI 如果真的想配上人工智能这四个字，它需要先学会的，是人这个字。不只是理解人说了什么，而是理解人为什么这么说，感受到了什么，需要什么。 这是这个时代最值得去做的事，也是最难的事。 但风口就在这里。 情感需求，永远不会消失。 而能真正回应情感的产品，还没有被造出来。 这篇文章是我最近在写的一本书《Shadow Partner》系列思考的一部分，写的是 AI 时代的一人公司。 我相信，这个时代最被低估的创业形态，不是独角兽，不是大团队，而是一个人——用 AI 作为自己的 Shadow Partner，做出原本需要一整个团队才能做到的事情。 书里会持续探讨这些问题：传统企业怎么完成去碳基化改造？Adapter 如何从套利者进化成工作流重构者？一人公司的护城河到底是什么？AI 时代，个人的杠杆在哪里？ 这篇文章是系列思考的一部分。 如果这些问题也是你在想的，欢迎关注，我们慢慢聊。

AlphaGeometry みたいな感じで証明FEを自動で解くモデルとかは作れる気がしていて、暇なら作ろうかな

من أبرز الاتجاهات البحثية الناشئة في الرياضيات: استخدام الذكاء الاصطناعي في الإثباتات الرياضية. نماذج مثل AlphaGeometry وLean تُحدث ثورة في كيفية اكتشاف النظريات والتحقق منها، مما يفتح الباب لحل مسائل ظلت عقودًا دون إثبات. #Lean #Alpha_geometry

Google DeepMind just dropped a paper proving AI can now solve complex geometry problems at IMO gold medalist level. AlphaGeometry solved 25 out of 30 International Mathematical Olympiad geometry problems. The average human gold medalist solves 25.9. We just hit human-level performance in mathematical reasoning that requires intuition, spatial thinking, and creative proof construction not brute force calculation. Here's what makes this different: Most AI math systems rely on massive datasets of solved problems. AlphaGeometry uses a neuro-symbolic approach combining neural language models with symbolic deduction engines. It doesn't memorize solutions. It discovers proofs. The system generates its own synthetic training data by creating millions of geometry theorems and proofs, then learns patterns of reasoning. When faced with a new problem, it constructs auxiliary points and lines that humans would need geometric intuition to imagine. One problem stumped mathematicians for years. AlphaGeometry solved it and generated a proof shorter and more elegant than the official solution. The implications go beyond mathematics. If AI can handle geometric intuition one of the most spatial, visual, and "human" forms of reasoning then the list of "things only humans can do" just got shorter. Physics, engineering, architecture, molecular biology any field requiring spatial reasoning is now in play. This isn't about replacing mathematicians. It's about expanding what's mathematically possible. Problems that would take a human career to solve might take an AI hours. And the collaboration potential is massive humans providing direction, AI exploring vast solution spaces we'd never reach alone. We're not just automating math. We're augmenting human mathematical intuition with machine-scale exploration. The boundary between human creativity and machine capability just shifted again.

Reminder AI reasoning breakthrough In a new analysis, researchers highlight the rise of neurosymbolic AI, a hybrid approach that combines neural networks with formal logic and rule-based systems. Scientists say the next AI breakthrough is not bigger models, but smarter ones. Recent systems from Google DeepMind show how this works in practice: > AlphaGeometry 2 (2025–2026) solves ~83–88% of International Math Olympiad geometry problems, with some solutions generated in seconds > AlphaProof (2024 → current) achieved 28/42 points (silver medal level) at IMO by generating formally verified proofs > AlphaFold predicted 200M protein structures with near experimental accuracy, showing how hybrid AI can solve real scientific problems at scale. Instead of relying purely on probability like LLM, these systems integrate symbolic constraints, structured reasoning and verification engines to produce outputs that can be checked and proven correct. The shift is subtle but massive 👀 This new direction suggests AI is moving from sounding intelligent to actually reasoning with verifiable correctness, a change that could redefine progress in science, mathematics and engineering.

Peter Thiel just said AI will be worse for math people than for word people. That's the opposite of what every founder, PM, and engineer has been told since ChatGPT shipped. The conventional script was settled by 2023. AI eats creatives first. Copywriters, illustrators, journalists, marketers. STEM stays safe because math is precise and language is sloppy. Hire the IMO kid. Hire the Putnam kid. Hire the IIT JEE topper. They're the ones who survive. Thiel runs the inverse. Symbolic manipulation is exactly what LLMs do best. Math people were selected on a benchmark machines now ace. Verbal reasoning under messy real-world context is what models cannot fully automate yet. Look at the actual capability curve. In 2024 DeepMind's AlphaProof and AlphaGeometry 2 hit silver at the IMO at 28 out of 42 points, two to three days of compute per problem set. In 2025 Gemini Deep Think and an OpenAI reasoning model both took gold at 35 out of 42, working in natural language inside the 4.5 hour limit. Twelve months from silver to gold on the most prestigious math test on the planet. Compare that to verbal-heavy work. AI can draft a passable email. It still cannot run a multi-stakeholder negotiation, build a coalition inside a company, read what an angry customer actually wants under their words, or argue a contract clause against a counterparty trying to screw you. The bottleneck is presence in the room and skin in the game, neither of which the model has. The category most exposed is the one Silicon Valley selected for hardest. SAT math. Putnam. IMO. Olympiad informatics. Leetcode hard. Quant interviews. Every gate the past three decades treated as proof of horsepower picks out exactly the workers an LLM now does for $20 a month. The category least exposed is the one Silicon Valley undervalued. Lawyers who can read a room. Contractors who can fix a job site. Ops leaders who can hold five competing requests in their head. Salespeople who can sense a deal slipping. Doctors managing a patient's family. The work requires judgment over messy embodied social context. Thiel's frame is that the chess prodigy was the smartest person in the room until 1997. Then Deep Blue beat Kasparov 3.5 to 2.5 and he became a guy good at something a $300 engine could do better. The math priesthood is in the same window now. The math gate that built Silicon Valley is the same gate eating its seats.

It means you have not seen --AlphaFold, AlphaGo & AlphaZero, AlphaGeometry, AlphaDev / AlphaTensor, AlphaEvolve, AlphaGenome & AlphaMissense, AlphaProteo, AlphaEarth Foundations and Gemma -- yet. Go check my friend, world of AI/ML is bigger than "Coding assistants"

Scientists say the next AI breakthrough is not bigger models, but smarter ones. AI reasoning breakthrough In a new analysis, researchers highlight the rise of neurosymbolic AI, a hybrid approach that combines neural networks with formal logic and rule-based systems. Instead of relying purely on probability like LLM, these systems integrate symbolic constraints, structured reasoning and verification engines to produce outputs that can be checked and proven correct. Recent systems from Google DeepMind show how this works in practice: > AlphaGeometry 2 (2025–2026) solves ~83–88% of International Math Olympiad geometry problems, with some solutions generated in seconds > AlphaProof (2024 → current) achieved 28/42 points (silver medal level) at IMO by generating formally verified proofs > AlphaFold predicted 200M protein structures with near experimental accuracy, showing how hybrid AI can solve real scientific problems at scale. Unlike LLMs, which generate answers probabilistically, these systems use structured reasoning pipelines where results are validated, constrained, and logically consistent. The shift is subtle but massive 👀 This new direction suggests AI is moving from sounding intelligent to actually reasoning with verifiable correctness, a change that could redefine progress in science, mathematics and engineering.

Every Major Google AI Model Ranked from Latest to Oldest: Gemini Gemini Robotics Gemini Audio Gemma Veo Nano Banana Imagen Lyria SIMA Genie Lumiere VideoPoet Phenaki MusicLM GraphCast AlphaCode AlphaGeometry AlphaMath AlphaFold AlphaGenome AlphaEvolve Gato MuZero AlphaZero AlphaGo Chirp Universal Speech Model AudioLM Embedding Models PaLM LaMDA T5 BERT WaveNet