We need to make AI an inflection point for semi's. Push aside the old guys (like me.) Create a fine-tuned model trained on everyone's data (which can remain secret), and build open SW, prompts, scripts, tools, meta languages and build a powerful ecosystem that draws in newcomers!

Yes, that's the idea here (at Efabless) where people like me have been spending decades fighting for open-source hardware. We have been promoting and encouraging open-source repositories on github for hardware designs, and that is increasing the amount of material that the next generation of LLMs will be trained on. We have been doing it long enough that ChatGPT can be told to create a verilog wrapper module for the Efabless Caravel user project---and it will do it.

Brian, We are working it every day. Some big names are helping. Can't say more (Because traditional companies like to be quiet and keep things to themselves.) but it is absolutely our goal to prevent the downside scenario you outlined above and use this wonderful new capability to help along with open source EDA , IP, PDKs and rapid fabrication to massively increase semiconductor innovation.

To have any hope of success, the examples that are being used for training need to be of high quality... unfortunately, the stuff that is out there is typically not the code that someone would ship in a product.

@Borivoje Nikolic: I concur with that opinion. We need a lot of quality RTL code to train an AI. It is not clear to me, though, what is better for training an LLM: Large projects of well-written code, or small snippets of code examples and discussion like one finds on sites like Stack Exchange. For the latter, there is a lot of accompanying material about why something should be coded a particular way, why one way is right and another way is wrong, and examples of multiple ways to achieve the same result. All of those are valuable training that you can't get just by feeding the AI a bunch of well-written projects.