back to writing

We're All Just Pattern Matching Systems That Got Good at Using Tools

the jamestown colony was the first permanent english settlement in north america. they suffered numerous hardships including mass starvation, resorting to eating vermin and even cannibalism - while the 'less civilised' native peoples thrived in the same area.

there are many stories like this one.

how?

it turns out that homosapiens, individually, aren't that smart.

when it comes to raw intelligence at least - many other species have/had bigger brains. yet, here we are. billions of us, spread across the earth like a great plague - planning our expansion to the stars.

there's no single consensus explanation for how this happened, it's likely a complex intermingling of things - joseph henrich's 'collective brain' theory is almost certainly part of the picture.

it goes something like this -

the "secret of our success" is that, as a species, we are hyper-social learners. Individually, we're not that smart. but our collective brain is vastly more powerful thanks to social learning and cultural transfer.

and a big part of the collective brain story, is tools and tool use.

we're all just pattern matching systems that got good at using tools

lots of animals pass down knowledge. birds pass down migration routes, whales share hunting techniques and chimps literally teach tool use to younger chimps.

but humans really cranked the ratchet.

from stone tools to smartphones, thanks to increasingly sophisticated tools and the transfer of techniques and knowledge to use those tools.

when humans solve problems, we're often doing two things:

1 pattern matching from our training data - life experience plus accumulated and inherited knowledge from the 'collective brain'

2 selecting which tools to use based on that experience

think of a single technological breakthrough that would have been possible without the collective brain or the use of tools?

in some sense, we might be overestimating what humans do, and underestimating what these new systems will be able to do - given the right token budget, context and tools.

pre-training data is essentially a compressed representation of the 'collective brain'

two things we shouldn't take for granted.

1 the 'collective brain' is cached somehow, in the training data. enough to allow current systems to engage in coherent human conversation. this is wild.

2 reinforcement learning actually works. a model can be further trained/tuned to think and plan and use tools in a human-like way. this is also wild.

explores how RL + LLMs might be enough for AGI.

further, we know that new methods of training and rl are being explored, allowing models to interact and learn from their environment via verifiable rewards, beyond human data and preferences.

nathan lambert's work on RLVR (reinforcement learning from verifiable rewards) shows how models can learn from binary feedback rather than human preferences.

and david silver's work on moving beyond human data suggests we're entering an "era of experience" where AI learns independently.

it's not crazy to imagine that - by the end of the year, we'll have systems capable of autonomously invoking a 'planning' tool and efficiently allocating tokens to solve more and more complex problems, with little to no human intervention.

weighing the limitations

no doubt, there will be limitations - creativity, intuition, serendipity, continual learning and cultural context will be irreplaceable, in some ways. maybe we don't see truly 'novel' innovations for some time.

kenneth o stanley's work on "the myth of the objective" suggests that breakthrough innovations often come from pursuing novelty rather than directly targeting objectives.

but in other ways, we should expect that machines tapped into the collective brain, trained to explore their environment, design plans and use tools to solve problems will lead to some crazy progress we can't even begin to imagine.

especially considering the reverse limitations - human and cultural evolution are bottlenecked by biological birth and death; these systems are virtually immortal and omnipresent, limited by compute.

the cognitive core

View on X
- efficient models that sacrifice encyclopedic knowledge, for raw capability.

this kinda reflects how humans climbed the food chain, and how some humans climb the human food chain.

reasoning and system 2 thinking, delegation and heavy tool-use.

References