Zhaoyang

Researcher | Agentic AI | Building exciting agent products now... | Exploring agent design, products, and the future of autonomous agents. AI news & insights.

[Zhaoyang @Zhaoyang___]

Highly recommended! 🚀 Stanford's CS329Z: Engineering AI Agents (Fall 2026), now open and led by Diyi Yang (@Diyi_Yang), is a great resource for anyone interested in AI agents.

One of the most interesting takeaways is the history and challenges of AI agents, as well as the emerging topic of proactive agents.

https://t.co/XTPNvsAxov

[Walter Goodwin @goodwin_ml]

We have raised a $720M Series C, led by @ThriveCapital, @lightspeedvp, @Redpoint and @foundersfund, to allow us to run faster in shipping the world's fastest chips for AI.

We've seen over the past year that many of the hardest problems that humanity faces - in science, medicine, physics - may be solved in our lifetimes by AI.

One hallmark of AI applied to the toughest problems is that extreme scaling of inference compute becomes vital. In tackling a Millenium Prize problem, for instance, @OpenAI ran 10,000 parallel agents for 80 hours. Costing over $10M, this level of capability is far from democratised today.

When we study agentic scaling, there are two paths: parallel, and sequential. In just the same way that most hard problems solved by humans don't parallelise efficiently, we see the same with agents: fewer agents, thinking for longer, scales better in every way except for the time that you wait for a response.

Building radically faster inference hardware, which doesn't force a trade-off on throughput, model sizes or context lengths, is a key step in bringing down the cost and time we incur in solving our hardest problems.

We're thrilled to be backed to bring this to the world!

[Google AI @GoogleAI]

Our strongest frontier model, Gemini 4 Carbon, is now available.

Built for deeply complex, long-horizon workflows that push a frontier model’s intelligence, Carbon delivers frontier performance in agentic workflows across complex knowledge work, software engineering and web development, cybersecurity, and scientific reasoning.

While Gemini 4 Argon remains the ideal model for the majority of software engineering and knowledge work, Carbon steps in for your hardest tasks that push a model to its limits — from complex, multi-step debugging to open-ended scientific research.

Available starting today:
— Consumers: use it in the @GeminiApp and AI Mode in Search for Google AI Ultra subscribers
— Developers and Enterprises: build via the API in @GoogleAIStudio and @AndroidStudio, Google @Antigravity for Google AI Ultra subscribers, and Gemini for business
— Scientists and Researchers: apply for a chance to get 6 months free via @Antigravity
— Cybersecurity Defenders: apply to join the Fairwind Program for access to Carbon for cyber

Watch how Carbon brings reasoning and coding together, turning avian research into an interactive 3D simulation of a songbird’s vocal organ that synthesizes the sounds of six bird species live.

[lidang 立党 （劝人卖房/学CS/买SP500/纳100/OpenAI/Anthrop第一人） @lidangzzz]

党哥的RSI Chip的路线图。

stage 1：AI Agent逐步设计出SystemVerilog完备、UVM最小子集完备、完成严格LEC、full stack RTL到生成netlist的EDA，初步达到cadence和synopsys的能力，实现一个功能齐全的Agent-driven EDA；

stage 2：设计一款risc-v的最小指令集RV32I的CPU，验证这套EDA的功能性， 完完全全由AI Agent全自动设计模拟仿真benchmark，确保整个CPU设计正确；

stage 3：AI Agent自己逐步完善AI EDA的整套功能，同时AI Agent设计一款高性能的RVA23指令集的商用级CPU，完成signoff和tapeout，继续验证AI Agent的进化能力、全面完善AI EDA的功能；

stage 4：AI Agent继续逐步实现cuda-compatible tiny GPU，只支持PTX一个非常非常小的subset，实现一个minimum inference-only GPU的实现；

stage 5：AI Agent继续实现一个商用级性能较强的全新架构LPU，实现inference的极致优化，能够让开源模型跑在上面，这样初步实现了RSI loop的闭环；

stage 6：AI Agent开始设计可以规模级组建data center的training GPU和inference-only LPU，Agent自己训练的模型的同时，完成Agent自己逐步改进芯片，让RSI loop闭环进一步跑起来；

stage 7：针对整个计算架构进行大刀阔斧的重新设计，包括CPU（其实现在流行的已经全部SoC化，包括里面的I/O，protocol，communication，其他encoding和decoding），training GPU，各个规格、各个功耗、各个场景下的inference only LPU，把场景和功能进一步细分，重新设计，从半导体层面重新设计整套RSI的硬件体系。

以上每一步都由AI Agent来主导，党哥只是个聋子的耳朵——配的，娶媳妇打幡——凑热闹的。

RSI的最后一步，一定是AI Agent设计EDA，AI Agent逐渐打磨AI EDA，和AI EDA一起进化，最终AI Agent完完全全自己接手芯片设计，AI Agent亲自为自己设计大脑和心脏。

而不是今天其他大厂声称的仅仅某一个环节用AI Agent进行一小步优化。

整个流程会公开在这里：

https://t.co/Hrb3kbt6wy

[Zhaoyang @Zhaoyang___]

These are promising directions for agent research to explore!

OpenAI and other leading AI labs have already made significant progress in mathematics and coding, as these two domains have become natural playgrounds for AI agents. The next frontier will be tackling similarly complex, high-value real-world tasks that can be carried out and verified entirely locally, enabling end-to-end support for compilation, execution, simulation, and rigorous verification through unlimited trial and error.

Proactive agents and goal-driven multi-agent systems will be core technologies in these areas.

[lidang 立党 （劝人卖房/学CS/买SP500/纳100/OpenAI/Anthrop第一人） @lidangzzz]

Agent的力量拉满的超长goal driven应用，我在2026年年初就说过了，一个是爆破数学low hanging fruit，一个是构建超大规模的可验证系统。

数学就是这种最小的系统，我在年初就告诉过你们，我的mac并行跑8个lean4就非常卡了，而OpenAI一定会和一堆数学家来合作，给他们无限的资源来爆破这件事。

接下来的方向是构建其他可验证的超大规模系统。

听好了，作为电子机械自动化全能神的党哥，只跟你说一次，这些价值billion级别的工具，每个后面都是两三个超级无敌巨无霸公司，至少值几十个billion，马上会被逐渐重构，包括但不限于：

CAD，构建整个机械结构建筑的基础，

EDA，整个电子电路设计模拟仿真验证的full stack平台，典型的就是cadence和synopsys，还有无数各个小领域的仿真，

控制信号通信电机仿真，代表方向是matlab simulink和labview，

结构仿真（有限元分析），整个ansys生态，

上面只是几个大方向，ansys下面一大堆工具，cadence和synopsys下面一大堆产品线，simulink打开一大堆组件和工具箱，

这些都可以一个个用goal driven来重构。

[Burny - Effective Curiosity @burny_tech]

Representation-Based Exploration for Language Models: From Test-Time to Post-Training

"Our main finding is that exploration with a simple, principled, representation-based bonus derived from the pre-trained language model's hidden states significantly improves diversity and pass@k rates -- both for post-training, and in a novel inference-time scaling setting we introduce."

[Zhaoyang @Zhaoyang___]

Please, everyone should calm down. You’re not Huang, and honestly, nobody cares about your secrets.

[Jen Zhu @jenzhuscott]

If you don’t have a Chinese honey trap set up for you, you haven’t made it in the tech world.

[elvis @omarsar0]

Must-read paper on scaling multi-agent swarms.

The main focus right now is training agent swarms to coordinate and cooperate better. Here is a promising new work on this.

The problem:

Multi-agent systems usually follow a plan that a developer writes ahead of time.

The plan decides which agent does what and when agents message each other, and it usually stays the same across tasks.

What they propose:

s3 has the agents learn that plan instead. Agents learn through cooperative RL when to work alone, when to share findings, and when to read teammates' work. There are no predefined roles.

With matched training and inference budgets, s3 swarms tested at twice their training size solve over 25x more coordination problems than the strongest baseline. In parallel exploration, communicating agents repeat nearly 50% less work and beat best-of-N by up to 9.4 points.

The agents develop complementary roles on their own, and those roles carry over to bigger teams.

https://t.co/uqGIs6ZACo

Chat with Paper: https://t.co/WtMRctTYSe

[Zhaoyang @Zhaoyang___]

@HolarisSun Congratulations, Hao! 🎉

[Edward Grefenstette @egrefen]

Very excited by the bleeding edge modelling that went into this thanks to new learning methods produced by my team, in collab with Gemini Robotics.

Congrats Michelle Lo, @HolarisSun, @MartinKlissarov, @JonnyCoook, @silviasapora, @Aniket_d98, and Claudia Clopath on the launch!

[Cameron R. Wolfe, Ph.D. @cwolferesearch]

Coming out tomorrow morning. This blog was quite a challenge (but also fun!). My recent blogs have focused on breadth, so I wanted to go deeply into a topic this time. Score centering is interesting and a great topic to focus on given its (alleged) popularity in the labs.

[Zhaoyang @Zhaoyang___]

Hey @X and @XDevelopers , have you recently changed the recommendation algorithm?

I've noticed that liking just 1–2 posts outside my usual academic interests can dramatically shift my feed, sometimes flooding it with content from an entirely different topic.

The problem: The algorithm seems to mistake temporary curiosity for a change in my long-term interests.

What I'd love to see: Keep my core interests as the foundation of my feed, while still allowing room for occasional exploration and discovery.

[Jonny Kaye @k44yej]

@MicrosoftAI will be at NeurIPS in December both Sydney and Paris, I will personally be in Paris walking the floor! Fill my form pinned if interested in meeting us to discuss our models 👇

We are #hiring #ai - follow me for live updates!

https://t.co/8aS7TRc8Ku

[Moshe Vardi @vardi]

In her Sept. 10, 2026 column “Pause AI for Humanity’s Sake,” Peggy Noonan argued, “When something speeds up and gets out of control, it isn’t alarmist to demand prudence and deliberation.”
https://t.co/S0ybnsRnXH

[Rohan Paul @rohanpaul_ai]

New Oxford paper finds that frontier LLM agents fail to coordinate with unfamiliar partners because each one treats its own made-up rules (arbitrary convention) as the obvious choice.

Agents from different companies increasingly must work together with no agreed rules. Past studies used shared scaffolding or complex games that hide why coordination fails.

They tested text and code agents built on models like GPT-5.2 and Gemini 3 Pro in 9 small cooperative games. In Cat & Dog, a safe move guarantees 5 points, while a signal convention pays 10 only if both agents pick the same one.

Paired with other models, text agents scored well below that safe 5, each calling its own arbitrary convention natural. A short shared statement of coordination principles, called a Coordination-friendly definition, led agents to reliably take the safe 5.

If your agent must work with agents you did not build, tell it to avoid unstated conventions and prefer actions any partner reads correctly.

[Zhaoyang @Zhaoyang___]

@taylordsparks @Utah_Football @UUtah Okay

[Ryan Bahlous-Boldi @RyanBoldi]

Happy to announce that Vector Policy Optimization was accepted to NeurIPS 2026!

In case you missed it, we released the code and a shiny new website for the paper: https://t.co/n1LzXvvPpG

paper: https://t.co/XWFuQ2Scjt

[Laura Smith @Feabhsu]

Christmas lights are on! #London