I'm not convinced that existing complex codebases can be decomposed to the point where these tools can be useful though, sometimes the existing system is just too far gone.

Though it begs the question on whether we can develop an engineering discipline around simplification.

/double_facepalm

I even threw the conversation thread into the internet archive, so that some poor future anthropology student doesn't get their thesis rejected by a GitHub 404 page: archive.is/gN7qe

The longer horizon is even more exciting, though all of the IMEC predictions must be hiding some pretty enormous error bars.

I'm bummed that Intel bailed on 20A, it would have been huge win for them to land RibbonFETs with Arrow Lake. But hopefully 18A will get over the line this year, and TSMC N2 should also be landing GAA FETs in consumer processors around the same time.

emacs, once you get past the first decade.

This would also be the case if you assume OpenAI trained GPT-4 specifically to fit into the DGX H200, it's just neat to come to a similar conclusion from a different starting point.

So assuming ChatGPT is saturating 8x H200s, that gives us 1128gb of memory and if it's running 16-bit weights it means GPT-4 is running 564B parameters.

If we assume an average query takes 5 seconds, and they process inference with a batch size of 32 we get 23k requests per hour at 0.00034kWh each we get 7.8kW. That's mighty close to a single DGX H200 server.