正在加载视频...
视频加载失败
Introducing Real World Engineering Bench (RWE-Bench). AI isn’t great at designing rockets, engines, or chips yet. With better evals and RL environments, it will be. Software was just the beginning. The physical world is next.
167,771 次观看 • 20 天前 •via X (Twitter)
40 条评论

The RWE-Bench leaderboard is live now. We scored 11 frontier models on 20 real physical engineering tasks, everything from designing engine components to routing circuit boards to simulating the flow through turbine vanes. Physics solvers check every design against the strict requirements set by our engineers. A run only passes if it meets every single one of them. GPT-6 Astra leads at 37%.

Agents are given actual engineering tasks and then work in sandboxes using the same CAD and CAE tools that human engineers would have access to. One mechanical engineering example: design the crankshaft for a 3-cylinder diesel engine. The agent has to ensure the counterweights balance the system without exceeding the mass limit, and its fillets survive peak firing loads and millions of fatigue cycles.

One electrical engineering example: route the DDR3 interface between an FPGA and 2 memory chips. That’s 77 nets on an 8-layer board with 295 parts already placed. Every trace has to stay skew-matched to its strobe and hold its impedance. The board fails if one net is off.

One aerospace engineering example: simulate flow through high-pressure turbine vanes. The agent needs to mesh the passage, set up and run CFD sims, and solve for surface heat transfer and pressure loss.

We’re building harder tasks in more engineering domains. To define the frontier of real-world engineering, we need RL environments designed by industry experts, hard enough to seem impossible for today’s models. If you want to build them with us, we’re hiring engineers and researchers.

If you want your model to master real-world engineering, or if you just want to learn more about how your model performs today, send us a DM. Read more here:

Bofff que calidad!

Lets make it!! Insane team and opportunity

10/10. Extremely excited about as a developer... and industrial engineer!!! From @AISaturdays_ES to the moon 🔥🚀

lfg!

Congratulations on the launch!

👏

Congrats to the team!

let's goo 🚀

came for the benchmark, stayed for bach (nice music choice)

This is amazing, congrats!

Insane product, Bort! Such a big opportunity ahead

:)

@maketaster you may like this. I wonder if test data could be tied back to train the models, in particular for a given business.

@paigeinsf

🚀🚀🚀 amazing work!!!!

Giving SI rocket-design homework makes “show your work” a much more interesting instruction

🔥

🚀🚀🚀

epic engineering eval :)))

I love Humanity.

Increibleee

Fantastic to see this come together :)

let’s go!

Great job… and that’s only the beginning!

🔥

Very interesting

Real world engineering is the next frontier. Excited for you guys to take the world there!

🚀 🤖

Insane team! Congrats

Way to go! Physical AI is getting 🔥🔥

🚀🚀🚀

We’ve been doing this internally for electronics! Will definitely give it a go

This is game changing 🔝🚀🚀🚀

This is absolutely phenomenal. Would we be able to contribute to the next set?

