@VictorTaelin Maybe you can make a licence saying something like "free if you have less than $100M revenue, otherwise make a deal with us". I think they did something like that for llama 2 back then
@VictorTaelin I think the hardware approach they describe and also the one you took is not the best for hardware ICs. I think it is actually vastly suboptimal. For instance, the approach described by tendrils is actually von neumann. In yours, I am surprised by the presence of "pointers"
@VictorTaelin If I get what you describe, you would need to send pointers/label data across the chip in all directions, all the time, right? This sounds *not optimal*
@ID_AA_Carmack I think a safe bet would be to guess that we would try maximizing yield. From a design pov it would mean smaller dies. Forget about H100-like monstrosities and start pushing haaard on transfer speed per pin
@ID_AA_Carmack If you can delegate the email writing to ai, is it worth making sentences in the first place, anyway? (Would you use those features when writing an email to a loved one?)
@ThePrimeagen New business model: a company which does actual exams and tests for AIs. No automation, humans evaluating AIs. Would it be more reliable?
@Fumeon1 Fais gaffe les oraux c'est la roulette (jle sais mΓͺme moi j'ai tout eu avec mes 25 de qi), mais saint louis je me demande qui sont les profs de nos jours
@teej_dv Non-empty list type or list type with asserted length. You assert length, and in one of the branches you know at the type level that your list is not empty, or of length 2, or 3 or whatever compile-time value. So it would be list-into-tuple type guard
@tmaiaroto@sudoingX Just in case: I did the thermal pads mod on my 3090 fe (recommended by crypto miners at the time), and it went from crazy loud to almost silent (the fans never hit 50%), this is a great QoL improvement...