Coincidentally, taking advantage of Lao Ma's stunning flying car today, I filled in the pit I buried a few days ago. Regarding Jev, Strata, Atlas, these three things that I think are fun in the recent AI circle I use a movie scene to string the three things together and say that Lao Tie who is interested in asking AI or Research himself without talking about technical details What is this scene? - Terminator! Imagine the T-800 being sent back to 1984 with the mission of finding and killing Sarah Connor! At this point, he needs to possess the following three basic functions: understanding space, making instant judgments, and thinking without networking (there is no Skynet for you to connect in 1984) Coincidentally, these three correspond exactly to Atlas, Jev, and Strata 1. First, let's talk about Professor Li Feifei's Atlas - Understanding Space Schwarzenegger (T800) walked into a room and glanced at what was probably behind the door and under the table What Atlas is currently doing is this: providing a few photos and predicting what it would look like to look at a different location LLM predicts the next word Video models (such as Seedance) generate images down the timeline Atlas predicts a new perspective Nowadays, many video models refer to themselves as world models. A few months ago, Li Feifei divided them into categories: renderer, simulator, and planner Simulators can become renderers, while planners must be based on a powerful simulator T800 takes a quick glance and knows what's probably inside the house. This is a simulator. Calculating what will happen if you jump and where the bullet will hit is also considered a simulator. Choosing between chasing, dodging, or driving into a collision most likely requires a planner, while a renderer is basically not needed The goal of Atlas should be to create the strongest simulator and upgrade to a planner in the future The roadmap looks like this- Static 3D space (current) → Dynamic (4D) → Editable → Interactive → Simulator as planner (endgame) The first step of the Long March has just begun, but you really can't achieve the Terminator without this. LLM alone is not enough 2. Then there's Jev - Instant Judgment Do you still remember that scene? The T-800 was knocked on by the administrator at a motel, and a row of candidate replies popped up in its field of vision. Choose one and go back out without any language organization. (I remember choosing the most rude sentence...) This is Jev: there is no need for LLM to spit out a paragraph, nor for COT to think through the chain of thought, just select from the given options and give how certain it is. TypeSafe's Jev is like this: it does not generate text, only selects from the given options, and provides a calibrated probability. Reply 0.97 → Go straight back The highest is only 0.60 → Don't go back for now, let the slower but smarter big model think about it In the TypeSafe demonstration video, it is used to determine whether an invoice is fake. Traditional LLM takes 8 seconds, while Jev only takes 0.1 seconds Fast, because it only does multiple-choice questions and does not generate text; Accurate, relying on black technology RLCD - Reinforcement Learning for Calibrated Decisions - Reinforcement Learning for Calibration Decisions The architecture has not been made public yet, and I feel that it is most likely still the Transformer base, just cutting off the generation scenario directly. It can be seen as an ASIC for LLM in a specific context - not chatting, not writing code, only making high-frequency decisions, compressing and simplifying the various steps of traditional LLM judgment into a true or multiple choice question, with a speed and cost difference of one to two orders of magnitude It's a bit like System One in Daniel Kahneman's famous book "Thinking Fast and Slow", the traditional LLM calculation system two I feel like a new AI path has opened up beyond generation - decision-making/judgment Anyway, if the T-800 is judged by LLM's current speed, it will definitely not be able to complete any tasks in the movie. It will be like a snail (why does it suddenly remind me of lightning in Zootopia...). At this time, Jev emerged, and the second puzzle of Terminator function was completed 3. Last piece of puzzle - Strata - Local running of large model In 1984, there was the Internet, but the first was slow to burst, and the second was only in the military and schools. Besides, there was no Skynet on the cloud to let you connect! You still need to install a local model in the T-800's mind to travel back Strata enables a consumer grade graphics card with 12GB of video memory and 64GB of RAM to run the 125 billion parameter Qwen3.8-Flash Next model Black technology first compresses the parameters from 16 bits to 2-3 bits quantitatively, reducing the size by several times. Then, the commonly used 20% expert (copy) is thrown into the video memory, all experts are thrown into 64GB of memory, and finally a 30GB large lookup table (quick reference manual during training) is thrown into the SSD Completely offline, data privacy. Of course, the ability has been discounted due to compression, and reading long documents is also slow. However, high-end game consoles can run billions of models, which is a problem from scratch We don't know the specific parameters of the graphics card, memory, and CPU in the T-800. Anyway, Strata allows you to run larger parameters in a limited hardware environment, based on the MOE large model (currently only supporting the Qwen model, but the idea is universal to the MoE model) At first glance, it seems that we are getting closer and closer to the era of Skynet and Terminator! @drfeifei @theworldlabs @typesafeai P. S - Which rich guy ordered the Roadster? Can you take me to experience it in 2028? Please