Drop the most iconic photo of you.

Silicon is reflective in it’s raw, elemental state, but when it’s refined into wafers for use in chips it’s iridescence is striking. Our branding is a harkening to this process of turning metal to machine, refining raw materials into a world-changing device.

Silicon empowers investors to own enterprise-grade GPUs through tokenization while enabling datacenter operators to scale through crowdfunded infrastructure.

We're excited to announce our rebrand to Silicon. This marks a major milestone in our mission to transform GPU compute into a new financial asset class.

Scary part: one hot slutty gurl could derail the entire thing...

It actually makes perfect sense to have a team of cracked zoomers on DOGE. You want people who look at a system and say “this sucks, let’s start from scratch” - not “how can I work in this bureaucracy to further my career?” Can you imagine if it were Google employees instead?

The only SF government service run with rigor and ruthless efficiency

True

There is a ton of misinformation about DeepSeek r1 model. Tl;dr it really is as impressive and powerful as the hype suggests, but: a.) You basically NEED to use chain of thought to use it properly, it's generally the norm that you can use suboptimal technique for a model and still get pretty good results, with this one you need to play by its rules. b.) There's a lot of chatter about how "you can run R1 with print(hilariously_low_amount_of_VRAM)" but what people are referring to is running one of the distills, and this isn't really helped by packages like ollama calling it "deepseek r1 7b" and then you drill down into the docs and see it's qwen 7b with r1's reasoning embedded. You aren't running R1 itself without 300+ GB of VRAM at the lowest available quants. Distills are a downstream benefit, but running a distill isn't running r1. c.) I don't think anyone is going to beat deepseek's API price - not us, not our competitors. It's just not possible. It's orders of magnitude below what you'd pay on OpenRouter on a token cost per model parameter basis. You can get 1m tokens on openrouter on a 7b model for $0.10, or you can get 1m tokens on deepseek's API for this 671b model for $0.14. I think this API is just being run as a huge loss leader to pump hype about the model, but regardless, it exists and it's very hard to compete with. It's more feasible for us to show that you can run the distills on RunPod (which are perfect to run in serverless) or perhaps a lower bit quant of R1 in a pod. Long story short: incredible model, deserving of the hype, but also has the most complicated meta-environment in an open source LLM we've seen to date. Brendan from our team wrote a great piece about this - posting link below

New counterculture just dropped; Pro crypto, anti memecoin. Will be interesting to see how it develops.