Leo

Most investors still value SpaceX as a rocket company. Silicon Valley’s top venture capitalists see something much bigger. Their thesis: rockets are simply the transportation layer for the next economy. The real value lies in AI infrastructure, orbital computing, satellite networks, robotics, and eventually space-based energy. If AI’s biggest bottleneck becomes power rather than chips, orbital solar-powered data centers could transform computing economics. Starship lowers launch costs, Starlink connects the network, and AI becomes the operating system. It’s an ambitious vision and much of it remains speculative. But the greatest investments often look impossible before they become inevitable. The question for investors isn’t how many rockets SpaceX launches next year. It’s whether SpaceX is building the foundational infrastructure for a multi-trillion-dollar space and AI economy. That’s the bull case the market is being asked to price

How can you tell if you're on the latest chatgpt image generator? It still cannot draw molecules for me (but everything else is quite good).

We've been defining the actual frontier of AI for science: - created the most powerful agents and generative models of our biomolecular world - coupled with the most accurate models of state - proved again and again ability to break novel ground in drug design >>> now I'm looking for those who want to push frontier LLMs even further, in an environment where you see the impact on real science the next day

The bottleneck for physical autonomous agents was never cognitive. It was state reconciliation. When you pipe a frontier model through standard cloud IoT APIs, the latency bleed ruins the execution loop. The agent thinks faster than the physical environment can react, creating a desynchronization between digital logic and physical reality.

Here is the kind of performance you can expect if you use @Cloudflare, talking to mix of services (r2, kv, d1) with the database in another region, even when batching to keep RPCs to a bare minimum. It's typical to see 500ms d1 queries. Tuya cloud is the alternative way to go

It's insane how quickly you can build throw-away prototypes with Claude now. I made a timeline recording debugger in about ten minutes. Warning: this is partly fake data! It would surely be many weeks to make this production ready.

GM ☕️

something fun is coming!

Only on china 💯

thanks for having us!

would be great to see other applications besides classification too. tried CTMs on pathology images few weeks back, excited to see what comes next

debugging Lang graph

its not possible to use multiDrawIndirect/Instances when rendering fat lines, but if you use one big buffer is possible to tweak vertexAttribPointer offset in a tight loop and issue 5k draw calls in ~1ms

wow, the thing that "never happens" where you forgot that you put in a sleep somewhere actually happened to me. I was wondering why my CAN bus was so slow! free 50x speed-up. (I think I did this initially to space things out on my oscope.)

my child has almost all of its limbs again.

hope you guys had an amazing time at the event, and pleasure to meet everyone in the expo! see you guys in the next tuya summit event

i was up until 3AM fixing demo bugs. for some reason it doesn’t work on mobile 🥲 how late did you stay working last night?

my inner hoarder wishes there was a reason to keep old computers and add them to a bigger and bigger cluster. sadly any real use case I can think of is just a better fit for cloud

Sliced up the image into the patches that would be seen by the ViT (clip-vit-large-patch14-336). Does this explain the mistake it makes on ESPEON?

Old school LLaVA is really good, 13B model fits nice and comfy in ~12GB VRAM at 4bit quant and it even does OCR.