For all the quibbles, you still have to admire Moravec's insight, and the predictive power of cold hard extrapolation.

Of course you can quibble: Moravec thought the AI would be embodied, he thought it would be directly programmed rather than evolved, he thought we would be uploading our minds, and his prediction assumed we would have human-like AI in a supercomputer by 2010 already.

Arthur C. Clarke described it as "the most awesome work of controlled imagination I have ever encountered". Today Moravec sees that we're "on track, more or less".

In 1988 he put the date of human-like intelligence at 40 years in the future, and then went on to explore the consequences. His later work imagined "mind fire" spreading into the universe, turning matter into computation.

Moravec looked at the trend of how the computing power you could get for a thousand dollars was doubling every two years, and then extrapolated out to the 10 trillion operations per second (teraops) you would need for a human-like mind. He assumed scaling is enough.

And just from a personal perspective, if I wanted to start a project in the past I would have had to block several days for the programming, but now I can do it in between all the meetings and small admin stuff, without all the endless hours lost on bug fixing and pandas wrangling.

I used to think the main advantage of open source was the flexibility - the ability to go into the code and do deep customisations, then throw thousands of parallel jobs on a cluster, but perhaps the familiarity of the agents with open source will be the bigger selling point in the future.

This has surely dawned on many people already, and folks from my team are way ahead of me on this, but it was only this summer that I had time to play with the agents properly.

With relatively little prompting they can go and do sophisticated things that would have taken days in the past. With future agent tooling, this will only improve.

And when they do get stuck, they can go and look it up, search the issue tracker and read or even edit the actual source code. Closed source gives them much less to work with.

Just the sheer amount of online code and examples with PyPSA that the agents are trained on means they’re super-familiar with the tooling, concepts and conventions of open source tools.

Thanks to @maxparzen.bsky.social for helping me track down all the references!