⛰️ (@minghao914) header
⛰️ (@minghao914) profile picture
🌊
Joined May 2019
Photo by ⛰️ (@minghao914)Photo by ⛰️ (@minghao914)Photo by ⛰️ (@minghao914)Photo by ⛰️ (@minghao914)Photo by ⛰️ (@minghao914)Photo by ⛰️ (@minghao914)
nostr
npub1em22uy5qw5e25m3jym432r70kfz5entfg3kl5d97wgzj8w4kssjsm88my0
1
583
⛰️ retweeted
Video 2 Physics I gave Claude Opus 5.5 AI videos with broken physics. It measured what went wrong, solved the real physics with CFD and rod mechanics, checked it with an independent finite-element solver, and rendered it back into the same scene
9
18
7
185
382,842
⛰️ retweeted
所有人都在找 Crypto 和 AI 的结合点,但大多数尝试要么很生搬硬套要么是拙劣的骗局。 其实答案一直都在眼前。 回到 Nick Szabo 设计智能合约的初衷,本质上就是让互不相识的人,在没有中心化实体介入的情况下,把承诺变成可以自动执行和结算的规则。十年 DeFi 走下来,这套基础设施已经非常成熟了。 再看 agentic commerce 的演进,也很自然。先是 agent 替人买东西和执行交易策略,然后是 agent 之间直接采购 模型外部资源。再往后,交易会越来越小、越来越短、越来越定制化。很多过去根本不值得签合同的交易,会变成机器可以实时发现,议价,签约,结算的瞬时交易。 真正难的从来不是支付,而是信任。 Agent identity 几乎可以零成本批量生成。很多实验也已经反复显示,单靠 reputation 很容易被 Sybil、弃号重来、策略性欺骗击穿。一个坏记录如果随时可以扔掉,那这个记录本身就没有约束力。 真正有约束力的,是你提前押上的、做恶以后真的会失去的东西。 所以我越来越觉得,低成本、可编程的 escrow,以及围绕明确条件的自动执行,会成为 agentic commerce 里非常核心的一层。而这恰好是链上已经做得很成熟的东西:permissionless,全球化,7×24,而且边际成本足够低。 再往上就是 discovery 和市场机制。 Discovery 让 agent 找到谁有什么资源,有多少 capacity在什么条件下交易。Auction, pool 这些机制,则负责在大量成本,库存,和风险敞口都不同的参与者之间,完成资源和风险的分配。 Payment 解决钱怎么走。 Escrow 解决为什么敢交易。 Discovery 解决跟谁交易。 Markets 解决以什么价格交易,以及对面到底有没有人。 走到这一步,Crypto 才不只是AI 的支付轨道,而是真正 开始变成机器经济真正的协调基建。
19
31
7
140
25,006
if you value intelligence above all other human qualities, you’re gonna have a bad time
900
3,525
1,190
24,375
9,973,135
Today, my team at MIT and our collaborators are launching the first open benchmark measuring AI's holistic impact on human well-being. Explore the platform: impactbench.media.mit.edu/ @AHA_MediaLab @MIT @medialab
1
6
2
71
2,774
RSI-Jev v3.0 is out, and this is the first version where RL actually helped. Our AutoScientists loop ran 59 RL experiments on the 2B decision model. Most failed. The pattern was pretty clear: when every example already gives the model the right answer, plain supervised training is very hard to beat. What finally worked was giving RL a signal the labels didn’t contain: whether the model ranked the right memory first. That made it find the right memory first 60% more often on a long-memory benchmark, with no regression elsewhere. The lesson we’re taking from this: RL helps when the reward teaches something the labels can’t. But honestly, the part we care about just as much is why the other 58 experiments failed, so we wrote all of that up too. Report: github.com/Shanghua-Gao/RSI-… Code, weights, experiments: github.com/Shanghua-Gao/RSI-…
The AutoScientists kept going: RSI-Jev v2.1 is out now. On a third-party benchmark we never trained on, v2.1 scores 0.774 vs. Jev 1.13’s 0.753, with roughly half the calibration error. A few lessons from this round: protecting the LLM’s text behavior didn’t fully protect its knowledge, and 18 RL attempts still couldn’t beat plain supervised training. More importantly, v2.0’s benchmark-specific gains didn’t transfer much. v2.1 is the first version where we see the gains carry over to truly unseen data. Now we’re curious how far the loop can push this. Check the details: github.com/Shanghua-Gao/RSI-…
1
8
2
70
6,939
HEXIS: Compiling Skills into Extended Finite State Machines Minghao LI arxiv.org/abs/2609.30123 [𝚌𝚜.𝙰𝙸]
2
6
443
TypeSafe 创始人 Diogo Almeida 公开了一份 Coding Agent 设计草案,希望社区直接拿去实验。他直说团队大概率没时间自己把这些想法都做出来。 这份草案最核心的改动是上下文。现在的 Coding Agent 通常会一直带着之前的聊天、工具结果和代码状态,太长了再压缩或重启。 TypeSafe 想让 Agent 每轮重新判断:哪些内容保留全文,哪些只留摘要,哪些直接丢掉。已有 KV Cache 也不会直接放弃,而是比较继续复用和重新构造上下文哪个更划算。 Almeida 认为,现有 Agent 很多别扭的设计都是被 KV Cache 反绑住了。比如先把任务切给便宜模型,再切回强模型,强模型可能还得重新处理一遍长上下文,路由反而更贵;工具定义长期塞在系统提示词里,也会一直占空间。 如果上下文可以动态重组,模型路由、Sub-agent、MCP 和 AGENTS.md 也能跟着按需加载。简单任务交给便宜模型,需要哪个工具再加载哪个,子 Agent 只拿和自己任务相关的状态。TypeSafe 暂时把这种思路叫「Meta-attention」。
16
44
2
600
158,301
Introducing Bespoke Nimble: an open data, open model, open recipe for an open Jev. Code and info: github.com/bespokelabsai/nim… Model: huggingface.co/bespokelabs/B… Data: * A new data curation recipe called contrastive data curation. * Slightly change facts to generate negative data. This pushes the model to discriminate better and become a better decision maker. The calibration is implicit. * Didn't do ablations but I think this is a critical piece! * This also means training data doesn't need probabilities. * Data covered 10 categories, and is fully synthetic. * This data is split into train and eval. Training * LoRA finetune of Qwen3.5-9B. * Distillation-free: we use Jev to only evaluate. * No RL yet! Serving * Parallel constrained decoding as suggested by @NielsRogge and @harshagundal. Results: * The post-trained Qwen (Nimble) became substantially better on our curated eval: 66% for Qwen to 90% for Nimble. Jev is at 93%. * 100ms on H100 and free to use on your macbook! Feel the AGI for free. * 2 days of building in public. :) Big caveat is that there is no standard benchmark to measure performance, and it's possible Nimble is much worse on other benchmarks compared to Jev. But it should be better than Qwen! We thank @typesafeai for making Jev and the inspiring discussions in the community. Hope this release lifts all the boats and encourages more research and activity in this space.
Should I build an OpenJev?
82
177
40
1,392
557,979
塔勒布曾说,真正的成功,是退出某种竞争激烈的赛道,让自己的生活节奏回归内心的平和
这两年在硅谷采访了如此多的人 好处太多,但坏处也很明显 接收到了从实习生,到大厂高位,到明星创业founder,20岁-50岁,全方位无死角,各种各样的焦虑 - O/A家大佬,各种厂明星一样焦虑,迷茫接下来的自己的价值意义 - 实习生无工作焦虑工作,有工作焦虑成长 - 高P焦虑技能点追不上新时代,掉头难 - 创业者焦虑客户/融资/市场 每个人没有那么多不同,日常挤满了工作,工作中焦虑,下班也都很难喘息。每次新模型,新功能发布的时候,可能是少有能出来发泄一下的时候
27
22
5
163
26,231
⛰️ retweeted
We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra. The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.
5,636
19,966
14,812
119,624
75,473,352
⛰️ retweeted
一款新型的神经网络,仅用29个逻辑门就能玩马里奥,基于逻辑门的神经网络模型技术。 具体原理: 「逻辑门神经网络」——不是普通神经网络那种浮点乘加,而是 AND / OR / XOR 这类门电路。训练时把门放松成可微分形式,学完再收成真正的离散电路。 逻辑门输出只有 0 和 1,中间没有平滑斜坡,所以不能用“求导”那套办法训练。这样就有平滑曲线,可以求导。 学完再把每个节点「拍死」成一种真的 AND/OR/XOR。训练时不直接用真门(无法微分的原因),换成「软门」:输入先当成 0 到 1 的小数(像概率) AND 近似成乘法:a⋅ba \cdot ba \cdot b OR 近似成:a+b−a⋅ba + b - a \cdot ba + b - a \cdot b 软门不是新硬件,只是训练时的「假开关」,好让误差能往回传。每个节点不是先选定「我是 AND」或「我是 OR」,而是:16 种真门各做一个软版本(AND 用乘法,OR 用上面那式,XOR 也有对应平滑式……) 用一组可学习的分数,把 16 种结果加权平均 谁对最终对错更有用,谁的分数就被梯度推高 训完再看:这个节点分数最高的是哪种门,就把它焊死成那种真门。 他是怎么把训练的转换成真实的门操作的呢呢 训练时每个节点同时“假装”16 种门,学完只留分数最高的那一种 和传统LLM体系模型的区别: 普通神经网络推理要:大量乘加 浮点数 电、热、显卡 逻辑门电路推理是:基本就是 0/1 开关 可以快到纳秒级 功耗可以极低 适合 FPGA、专用芯片、传感器边上的小控制器 所以它更对口的不是「代替 ChatGPT」,而是:玩具、家电、无人机飞控里的小决策 工厂里要实时、要省电的开关逻辑 边端设备:不能联网、不能带显卡 物理世界控制(他们说的 physical AI):看见障碍就拐、该跳就跳 所以它替代不了 ChatGPT,也很难单靠 29 个门打通所有马里奥关卡还保证泛化。它明显不如传统的地方训练别扭:先造软门,再拍成硬门,精度会掉一截。 参考资料:【预测】2026年,基于逻辑门的神经网络将迎来爆发式发展。 zenn.dev/teba_eleven/article…
【衝撃】世界初?!たった29個の論理ゲートでマリオをプレイ!? 未踏ITでチーム開発中の論理ゲート型AIライブラリの成果がめちゃくちゃ出てきた🔥🔥 ついにマリオを操作し始めた!!! クリアできるかはまだ分かりませんが、たった29個の論理ゲートでこんなにできるの、ロマンしかない!!!
80
105
15
822
128,412
⛰️ retweeted
So Astra is able to identify sounds from mel spectrograms zero-shot. I don't think we've scratched the surface of what this model can do (and this is light reasoning btw)
204
436
192
6,365
1,163,755
现在深切感受到 vibe code 带来的生产力下降了,由于整个 architecture 全是 vibe 的,重构了两周都还没重构完,修完一坨狗屎之后发现遍地都是狗屎 感觉只要是写稍微novel/out-of-distribution的东西,就一定不能信LLM写的架构……
38
12
6
309
90,668
Synthetic data, generated from #DigitalTwins through computer simulations, is enabling developers to bootstrap physical #AI model training. Learn how you can build custom SDG pipelines using NVIDIA NIM microservices for #OpenUSD. 🔗 nvda.ws/3WH6yY1 #SIGGRAPH2024 #GenAI
1
12
3
33
3,783
Do you remember when you joined X? I do! #MyXAnniversary
3
89
CAIR Lab at #CHI2024, 4 of 13 / Full Paper Title: Enhancing UX Evaluation Through Collaboration with Conversational AI Assistants: Effects of Proactive Dialogue and Timing Monday, May 13, at 12:00 / AI and Interaction Design
1
3
3
283
SWE-agent is our new system for autonomously solving issues in GitHub repos. It gets similar accuracy to Devin on SWE-bench, takes 93 seconds on avg + it's open source! We designed a new agent-computer interface to make it easy for GPT-4 to edit+run code github.com/princeton-nlp/SWE…
44
407
117
2,197
696,596
AGI System Flow Chart for Solving Complex Problems 1. Problem Identification - An initial LLM receives a complex problem and identifies its nature and scope. 2. Task Decomposition - The problem is broken down into subtasks, and requirements are outlined. 3. Resource Allocation - The CFO LLM allocates computational resources and prioritizes tasks based on complexity and urgency. 4. Tool and Information Gathering - Relevant software tools are identified. - The browser module gathers necessary information, possibly engaging in web research. 5. Subtask Distribution - Subtasks are assigned to specialized LLMs, like the Python interpreter for computational tasks or the video/audio LLMs for multimedia processing. 6. Parallel Processing - LLMs work on their assigned subtasks concurrently, leveraging the distributed architecture for efficiency. 7. Agentic Workflow Engagement - Each LLM employs agentic reasoning, engaging in iterative processes, planning, reflection, and tool use specific to their subtasks. 8. Intermediate Synthesis - As subtasks are completed, an LLM synthesizes intermediate results, checking for coherence and completeness. 9. Multi-Agent Collaboration - LLMs exchange information and results, engaging in multi-agent collaboration to refine solutions and fill gaps. 10. Executive Review - The CEO LLM reviews the synthesized solution, ensuring alignment with the overall goal and strategy. 11. Optimization and Refinement - The solution is iteratively optimized, with the COO LLM overseeing operational efficiency. 12. Final Synthesis - All refined subtask solutions are integrated into a final solution. 13. Verification and Validation - The final solution is verified against the problem requirements and validated for correctness and effectiveness. 14. Output Generation - The verified solution is compiled into the appropriate format for output. 15. Delivery - The final solution is delivered through the peripheral devices I/O or stored in the file system for later retrieval. 16. Feedback and Learning - Performance feedback is collected, leading to system-wide learning and improvement for future problem-solving. Summary This flow chart outlines the stages an AGI system would go through in solving complex problems. It emphasizes the system's distributed nature, the specialized roles of various LLMs, and the importance of iterative processes, multi-agent collaboration, and continuous learning.
4
48
2
307
69,351
ファン、関係者の皆さまへのお知らせ dragon-ball-official.com/new…
17,478
514,129
85,916
916,536
124,260,212
Hack of the day: Llama on a microcontroller 🦙🔬 Details and source code at github.com/maxbbraun/llama4m… Thanks to @karpathy, whose llama2.c inspired this!
52
284
71
1,596
463,824