If you aren't paying attention to the non-diminishing returns of GPT-3, you've missed out on one of the more important things that's happened this year.

Taken together with some other technical advancements, it could be the most important thing. Economic turmoil, global pandemic, and a real shot at reclaiming liberty in the only superpower left to meaningfully oppose authoritarian China – all might be a sideshow compared to the workings of a few engineers.

What is GPT-3, though? It's a text generation engine that can take both prefix prompts as well as do "few-shot" learning, where it sees a couple examples of a pattern and uses the whole general language model to try and continue the pattern it's just seen a small bit of. Gwern Branwen explains is well:

2 years ago, GPT-1 was interestingly useful pretraining and adorable with its “sentiment neuron”. 1 year ago, GPT-2 was impressive with its excellent text generation & finetuning capabilities. This year, GPT-3 is scary because it’s a small & shallow model compared to what’s possible2, with a simple uniform architecture3 trained in the dumbest way possible (unidirectional prediction of next text token) on a single impoverished modality (random Internet HTML text dumps4)  on tiny data (fits on a laptop), and yet, the first version already manifests crazy runtime meta-learning—and the scaling curves still are not bending! The samples are also better than ever, whether it’s GPT-3 inventing new penis jokes5 or writing (mostly working) JavaScript tutorials about rotating arrays.

Modern neural nets are a simplification of what brains do – but not a ludicrous simplification. It seems like neurons fire based on passing a stimulation threshold and being multi-input dependent – there could be other things going on, but these are still full human cells, a lot of the space is taken up by routine parts that don't contribute to 'thinking' any more than your skin does. We don't fully understand neurons, but this is a key subset of what they do, when it comes to building a useful brain.

Humans have about a hundred billion neurons. This is the first neural net trained (without diminishing returns) to have more than a hundred billion 'synapses' (truly, just parameters, but that's the analogy into the computer system). So, not quite there – but at the current rate of increase, it's probably less than 5 years away.

It's producing text I cannot distinguish from human, it can be trained to solve new problems, it has information about the real world and can effectively communicate.

The advances are not just in size, but in approach. MuZero (a different project that's recently been breaking records in terms of computer achievement) doesn't even need us to tell it how to tree search possibility space, imagining what it does and probable reactions to it, back and forth deep into the future – it figured that out on its own, and does it better than we know how, apparently. Learning new problems from a small number of examples is new – and came after their first big successes producing spookily good language models.

In general, it's worth understanding this within the context of technovolatility. That whatever technology does to society, it will increase the magnitude of the changes. If we make something that can contextualize and accumulate knowledge as well as a person (or computer, for that matter), and is as sophisticated as currently existing models, it's possible there will be nothing but the law preventing GPT-3 from being used as a therapist, strategist, inventor, and well, general creator. GPT-4 will likely be able to produce novel video products, as image generation got decently good with GPT-2 and has improved meaningfully. The future is coming fast, which means all other problems might be made irrelevant, depending on how poorly we handle this.

We have almost no idea how to make a safe AI – current understanding would plausibly lead to humanity being trapped in a Matrix. There is a lot of progress to make on questions of safety, all of which will be completely worthless, considering it's probably impossible to have any of the current paths adapt to structures like deep neural nets.

Whatever we want the future to be, we better sharpen our imagination up, because I'm not too excited about what might happen next.