just one more LLM startup bro. i promise bro just one more LLM startup and we'll solve AGI bro. it's just a bigger LLM bro. please just one more. one more LLM and we'll figure out reasoning bro. bro cmon just give me 22 billion dollars and we'll solve intelligence i promise bro j

This x1000

It's getting harder and harder for OpenAI to maintain the illusion of rapid progress toward AGI.

Yes, it's done. New approaches are needed. That is why I started Symbolica almost 3 years ago to tackle exactly this.

Just wait until you see what you can do with GPUs when you have AI algorithms that aren't memory or infiniband bound.

if you ask an LLM about say taco recipes every matmul is combining every token with all the knowledge it's compressed from 15T tokens: knowledge about tires, code, chlorophyll, etc in humans, in fMRI this'd like your whole brain lighting up, instantaneously burning every calorie

This is the most impressive thing I've ever seen in the field of robotics.

David I think we’re doing it wrong Whenever value requires exponential scaling it’s time to refactor

The final frontier in AI is structure

Chollet gets it

AGI is an AI system that is capable of generating completely bug free programs which are then in turn used directly to make a novel contribution to science or mathematics without human involvement. Benchmarks don't matter. The only thing that matters is real science.

Intelligence as compression? If you're interested in figuring out how the brain captures the essence of the world, Chomba Bupe's posts on this topic are invaluable. If you're an AI expert with AGI ambitions, his posts reveal why statistical data compression techniques, e.g.,

how can someone very confidently claim that airplanes aren't the pathway to moon travel, when they're clearly the most successful thing to ever happen in aviation? i would get it if they say "it's unlikely" or something, but "airplanes will NEVER fly to the moon" is such a

"Rust powers nothing" he wrote, on the platform hosted on AWS and proxied by Cloudflare, which also directly uses Rust to power critical infrastrucutre. Lmao

It has been occurring, as those of us folks who've started the next wave of frontier AI companies have been saying for ages!

Precisely. And this is the future we're building Symbolica.

Wrapped up a fantastic day at the Oxford University Career Fair! 🎓 Loved chatting with so many brilliant grads about the blend of pure math and AI – and how they can help shape that journey with us. We'll also be at the Cambridge Graduate Jobs Fair on 24 June – come say hi 👋

Bayes' grave is in the cemetery directly behind the Symbolica office.

We've officially entered the next era of AI where big labs are optimizing for getting models to memorize more rather than optimizing them to be more useful for their users.

Ok now that GPT-5 is out I think we can all agree that LLMs aren't AGI and that new reasoning architectures are required. It's time to build.