Taking advantage of today’s wild ride by Old Ma, let’s fill in the gaps I left a couple days ago about three intriguing AI projects I’ve been following: Jev, Strata, and Atlas. I’ll tie them together using a movie scene—no technical details, just a narrative. If you’re curious, ask your AI or dive into the research yourself. The scene? Terminator. Imagine the T-800 sent back to 1984 with a mission: find and eliminate Sarah Connor. To accomplish this, it needs three core capabilities: understanding spatial environments, making instant decisions, and thinking offline (no SkyNet in 1984). Coincidentally, these three projects map perfectly to Atlas, Jev, and Strata. 1. First, Li Fei-Fei’s Atlas — Understanding Space When Arnold (the T-800) walks into a room, he glances around and instantly knows what’s behind the door, under the table. That’s exactly what Atlas does: given a few images, it predicts what the scene would look like from a new viewpoint. LLMs predict the next word. Video models (like Seedance) generate frames along a timeline. Atlas predicts new perspectives. Many video models now call themselves “world models.” A few months ago, Li Fei-Fei categorized them into three types: renderers, simulators, and planners. A simulator can become a renderer. But a planner must be built on top of a powerful simulator. The T-800 glancing around and knowing what’s in the room? That’s a simulator. Calculating what happens if he jumps or where a bullet will hit? Also a simulator. Deciding whether to chase, hide, or ram a car? That likely requires a planner. Renderers? Barely used. Atlas aims to be the strongest simulator—and eventually evolve into a planner. Its roadmap looks like this: Static 3D space (current) → Dynamic (4D) → Editable → Interactive → Simulator as Planner (endgame) We’re just at the first step of a long journey—but this is essential to building a Terminator. Pure LLMs won’t cut it. 2. Next, Jev — Instant Judgment Remember that scene? The T-800 is in a motel room when the clerk knocks. Instantly, a list of response options pops up in its vision. It picks one—no sentence construction, no thinking. (I think it chose the most rude one…) That’s Jev: no need for an LLM to generate text or follow chain-of-thought reasoning. Just pick from predefined options and output a calibrated confidence score. TypeSafe’s Jev works like this: no text generation. Only selection from fixed options, with calibrated probabilities. Reply with 0.97 confidence → send it immediately. Highest confidence only 0.60 → hold off; defer to a slower, smarter model. In TypeSafe’s demo video, Jev determines whether an invoice is fake. Traditional LLMs take 8 seconds. Jev takes 0.1 seconds. Why so fast? Because it only does multiple-choice—no text generation. Why so accurate? Thanks to RLCD: Reinforcement Learning for Calibrated Decisions. The architecture hasn’t been disclosed yet, but I suspect it’s still built on Transformers—with the generation component entirely removed. Think of it as an ASIC version of an LLM: can’t chat, can’t code—only makes high-frequency decisions. It compresses and simplifies all the traditional LLM decision steps into a single binary or multiple-choice judgment—speed and cost improved by one to two orders of magnitude. It’s like Daniel Kahneman’s “Thinking, Fast and Slow”—Jev is System 1; traditional LLMs are System 2. This opens up an entirely new path for AI beyond generation: decision-making and judgment. If the T-800 relied on today’s LLM speed, it’d be useless in the movie—moving slower than a sloth (suddenly I’m thinking of Zootopia’s Flash…). Jev is the second missing piece of the Terminator puzzle. 3. The final piece — Strata — Running Large Models Locally Yes, the internet existed in 1984—but it was painfully slow and limited to military and academic use. And there was no cloud with SkyNet to connect to. The T-800 had to carry its own model inside its head. Strata enables a consumer-grade GPU with 12GB VRAM and 64GB RAM to run Qwen3.8-Flash-Next, a 125-billion-parameter model. The magic? First, quantize parameters from 16-bit down to 2–3 bits—shrinking size several-fold. Then, keep the most frequently used 20% of experts in VRAM; store all experts in 64GB RAM; finally, store a 30GB lookup table (a training-era cheat sheet) on SSD. Fully offline. Data privacy guaranteed. Sure, performance is reduced due to compression, and reading long documents is slower—but now a high-end gaming PC can run a trillion-parameter model. That’s the leap from nothing to something. We don’t know the exact specs inside the T-800’s brain—but Strata lets you run larger MoE models on constrained hardware (currently only supports Qwen, but the approach is generalizable to all MoE models). Put it all together: we’re getting closer than ever to SkyNet—and the Terminator era. @drfeifei @theworldlabs @typesafeai P.S. If any rich folks have ordered a Roadster—can I ride along in 2028? Please?
