AI getting cheaper may make compute more expensive. Dylan Patel says Anthropic has reached $50M revenue per MW against $10–15M of compute cost. Reinvesting the spread into training lets labs outbid others on power. Better models → more revenue → higher compute prices.

I counted 1.05 million lines in a simulation I built with AI. About 242,000 are tests. At 100 finished lines per hour, that is 10,500 hours. I spent roughly 20–30 active hours steering it. The comparison is crude. A 350–500× difference is still hard to ignore.

Jalapeño makes NVIDIA’s OpenAI investment look smart. If recursive self-improvement happens and most economic activity shifts to frontier labs, OpenAI could be worth more than companies selling hardware. NVIDIA’s OpenAI stake could eventually be worth more than NVIDIA itself.

Sol working on a design for a worm robot that can be 3D printed. Told it to color parts differently and use lines so that it can find defects (of which there are still many).

How long until we don't need GPUs anymore and we have compute on new substrates?

Nothing you did before the singularity matters

software moats are dying faster than physical moats if everyone gets the same reasoning models, the edge moves to things the model cannot instantly copy: owned experiments messy data robots lab throughput weird supplier relationships actual products in the world

i believe the biology restrictions in frontier models are honest signal, not just pr when labs start gating capabilities by domain, the useful thing is no longer “can it answer bio questions?” it is “can this system help run discovery loops that touch the physical world?”

the next biotech company might look less like a traditional lab and more like a recursive machine ai chooses targets ai designs candidates ai designs the lab and robots robots run experiments data comes back ai iterates humans mostly choose taste, constraints, and ambition

Fable 5 blocks everything that could be related to bio work. In this case the word "lab" seems to be the issue.

openai + ginkgo running 36,000+ experiments is the part people should stare at not because 40% cheaper protein synthesis is the final product because the unit is becoming: ai searches the experiment space robots hit reality new data trains the next move

biology is probably where ai stops being just software because the data you need does not already exist on the internet you have to manufacture it in cells, assays, robots, failed experiments, weird measurements that makes the lab part of the model

ai biology is crossing a weird threshold not “a model reads papers and suggests ideas” more like: model proposes robot tests data comes back model updates company learns at that point the biotech company is less an org chart and more a closed loop

Fable 5 just one-shotted a robot design that can pick and place plates in a lab. Something Codex didn't get right even after a week of running.

the important thing about the mythos release is not just "ai got better at biology". it is that biology/chemistry is now a frontier capability category. general model gets safety rails. trusted version gets the lab-useful thing. probably the shape of the next few years

models can help you realize almost any idea now which makes it more obvious how surprisingly uncreative many people are once they get unlimited options

if you actually believe agi is coming, it is weird to keep building strategies for the world where it does not that is the trap. intellectually price in the fast world, then emotionally act like the slow world is safer

i like the idea of ugly robots more than perfect humanoids sometimes robots built from whatever parts are available, for dark factories nobody has to look at, redesigned by ai every week not elegant. useful

humans are just not built to feel exponentials things look slow, then normal, then obvious, then suddenly everyone says it was inevitable ai progress is annoying because you have to reason about it before it feels true