18 | product @quolabsai (backed @a16z @speedrun) | prev yc founder
Joined January 2026
- Tweets161
- Following61
- Followers81
- Likes1.2K
This is the benchmark the industry needed
moats aren’t real in 2026. Unless you are a deep tech company who is in stealth with 100M+ in funding or you are an AI lab you have no moat.
The only thing you can win on is distribution.
Every recently successful company is not difficult to copy; in fact they have already been; in China. Every American product already has a Chinese copy.
These Chinese copies will lose because they cannot distribute to the American market. If you are building in the US, this is your key insight.
we have a working version of this internally
清华大学和无问芯穹等团队刚刚开源了一篇注定载入史册的重磅论文（已收录于 ICLR 2026，代码已开源）：
他们彻底打破了所有多智能体（Multi-Agent）系统的底层范式——
大型语言模型之间，从今天起可以彻底不通过人类文字进行直接交流了！
这项被称为 C2C（Cache-to-Cache，缓存到缓存）的技术，
开篇就指出了全网 Agent 协同里最愚蠢、也最昂贵的一个死穴：
为什么两个跑在同一台服务器里的超级 AI，交流时非要像人类一样，
把脑子里极其丰富的高维思考，硬生生压缩成一个个英文或中文单词打给对方看？
现有多 Agent 协同的痛苦，做过工程的人都懂：
模型 A 思考完，必须经历极其漫长、吃显存带宽的逐字解码（Decode），慢吞吞吐出一大堆文字；
模型 B 拿到这堆文字，再重新过一遍分词和前缀计算。
这一来一回，不仅丢失了海量微观的语义细节，更造成了毁灭级的延迟和 Token 账单！
清华团队这次做了一件极其惊艳的体系化反叛——C2C 脑电波直连：
1️⃣ 彻底消灭中间文本，直接做“思维投影”：
模型 A 算完后，根本不生成任何文字，
系统直接用一个轻量级神经网络（Neural Fuser），把模型 A 的注意力记忆（KV-Cache）通过高维空间旋转和对齐，直接嫁接、融进模型 B 的 KV-Cache 脑内！
这相当于人类交流跳过了声带和耳朵，直接把我的记忆切片瞬间印进你的大脑；
2️⃣ 可学习的智能门控（Gate 机制）：
不同模型的网络层数和结构各不相同，直接硬灌肯定会神经错乱。
论文设计了一套精妙的可学习门控：
大模型会自己动态感知，精确挑选哪些关键层吸收外部缓存的增益最高，哪一层该保持独立思考，毫秒级自适应配平；
3️⃣ 极其残暴的工业基准成绩：
• 彻底消灭中间所有的文本解码等待，推理速度直接飙升 2.0 到 2.5 倍；
• 综合问答准确率比单模型最高暴涨 14.2%；
• 最狠的是：它的表现比传统用文字互相聊天的 Agent 团队，还要高出整整 5%！
因为在直接的语义融合下，原本会被文字丢弃的逻辑细微特征，全部被完整保真保留了。
这篇论文最刺痛人的一层哲学启示在于：
人类一直自豪地以为，自然语言是这个星球上传递智慧最完美的介质；
但在机器与机器的交互世界里，人类语言不过是一道充满歧义、极其拥挤的低速单行道。
当两个模型不再需要通过文字向对方妥协，
它们正在绕过人类的喉咙，
在冰冷的显存带宽之间，编织出一座全人类根本听不见、却快得不可思议的超高速无声神经网络。
is jev the future of robotics?
Got @typesafeai Jev driving our @makermodsai Metal arm in MuJoCo.
Images are of the sim camera is not passed in, sim hands Jev the state of the environment in JSON and it picks the next bounded action (hover, descend, grasp, lift, place) as a typed choice with a confidence score.
Jev makes the decisions for example "Put the red apple in the blue bin", and it runs in 9 decisions at ~150 ms each, and it refuses tasks it can't map to the scene instead of guessing.
I wonder what this means for datacollecting in sim in realtime? And I wonder if combining with GPT6 astra would it be able to then control an actual robot arm
Gregory Roudenko retweeted
OpenAI just solved the Navier-Stokes problem.
This problem has been open for over 200 years.
I just made a visualization so you can see what the solution is actually doing in real time.
Check it out below
OpenAI just solved the Navier-Stokes problem.
This problem has been open for over 200 years.
I just made a visualization so you can see what the solution is actually doing in real time.
Check it out below
Gregory Roudenko retweeted
i don't know if people know this, but i was born & raised in san francisco.
i genuinely think it's the greatest city in the world.
here's some thoughts / my love letter to this city:
> i love public transit. this is unironically the #1 tip. tech bros vastly underrate muni / bart, especially when paired with an e-bike or a scooter. i have a car, but i basically do not use it. you can basically get anywhere in <30 minutes if you take the right path.
> do not spend your entire life in rincon hill, dogpatch, or another one of the tech bro pockets. get out and explore! eat at small restaurants in the mission. go find some cute clothes in haight-ashbury. balboa, west portal, potrero, richmond, outer sunset, castro… the city has tons of little quirks and pockets of culture.
> on the other hand, san francisco is basically the #1 city for like-minded serendipity. do spend working hours near people working on interesting problems! i find myself running into people i know, and more recently, running into people who recognize me from twitter (weird but cool??). the best part about being in a city where everybody is hacking on their own cool things is you get to meet a lot of cool people. use that to your advantage!
> san francisco has a lot to do if you like nature. go out and spend time in nature. growing up here, i went to a lot of parks all the time. you should go explore the pockets of greenery that are all across the bay. golden gate park is huge (significantly larger than central park), mt. tam, mission dolores, all the beaches, half moon bay. take in the scenery. it’s calm, it’s beautiful, and it’ll remind you of your place in the world.
> it also has a lot to do if you like things like trying great pastries, making pasta, going to museums, anything athletic, thrifting, meeting people, etc. there are good bars and clubs as well... but the tech bros aren't onto them yet and i'll leave you the fun journey of finding them!
> some of the best restaurants are here. the bay area has more than half of the U.S.’s michelin three-star restaurants (and is #1 in the US). but it also has a ton of good food overall. go out and try restaurants- walk down a random street and pick one. that’s how my friends and i have found the best restaurants.
> everybody who says san francisco has no culture has basically never been outside. san francisco has a lot of culture. learn about the history. learn about the architecture. learn about what type of city this is. because it isn’t just b2b saas memes. go enjoy abstract art and underground music and thrifted fashion. this is the home of the grateful dead, train, green day, journey, and many more. this is where blue jeans, popsicles, the jacuzzi, cable cars, and motion picture graphics were invented.
welcome! enjoy your new home. it’s a wonderful, wonderful place. and we’re quite lucky to be here, to be honest.
“This gold and white acropolis rising wave on wave against the blue of the Pacific sky was a stunning thing, a painted thing like a picture of a medieval Italian city which can never have existed. I stopped in a parking place to look at her and the necklace bridge over the entrance from the sea that led to her. Over the green higher hills to the south, the evening fog rolled like herds of sheep coming to cote in the golden city. I’ve never seen her more lovely. When I was a child and we were going to the City, I couldn’t sleep for several nights before, out of busting excitement. She leaves a mark.” — John Steinbeck
The problem with AI labs and frontier models will not be able to keep scaling endlessly. Giving a model more parameters will make it more knowledgeable, but not smarter.
The knowledgeable 1T parameter model knows about everything in the world and cannot reason.
The smart well-architected 8 billion parameter model knows how to reason and think about complex problems it has never heard of.
I foresee a future where the best frontier models will not be huge MoE models but small, dense models that reuse weights and are deeply recursive.
Realtime video generation completely changes the game.
Speeding up text generation only makes chatbots & agents respond faster, and speeding up image generation lets you make more images in a shorter time.
But real time video lets anyone create live experiences that change with the user; as the user is experiencing it. As diffusion only gets more realistic and time-invariant/consistent people will be able to curate personalized real time live experiences on demand for literally anyone.
This genuinely might be bigger than the jump from passive TV to having active video games
Today is a very historical moment for AI video generation
You can now generate AI video faster than you can watch it
Before it'd take let's say 2-5 minutes to generate 15 seconds of video
@fal made a post-trained Minimax H3 variant called Max which is 50x faster than the original but still maintains quality
It generates 15 seconds of video in 9 seconds!
That means you can now do new things like build a perpetual livestream with it that never ends!
I hate that your claude code CLI chats don’t show up in the Claude UI. I wonder if there’s a way to fix that…?
