Hardware Is Software + The Use of AI/ML in Hardware
AI could cut manufacturing costs by 74% as labor savings cascade through entire supply chains, not just individual factories.
AI could cut manufacturing costs by 74% as labor savings cascade through entire supply chains, not just individual factories.
One thoughtful comment unlocks the ability to create threads.
real... I also think we'll hopefully see these wonderful tools get weaponize to tackle the true core: increasingly complex semiconductors are a bottleneck!
AMD recently put out a thing about accelerating chip development at the software level which is great, but what I'd throw millions of dollars (or my entire time of life) at is dramatically increasing IC and more complex chip volume/export at the same or even higher QA/reliability standards. why do worry about RAM bottleneck when we can just take risks to do it with all this excess cashflow instead of outsourcing it? tough nut to crack between hard verification and manufacturing + labor bottlenecks and strategic partnerships/relations, but that needs to be solved to keep it going.
then we solve the next bottlenecks, the physical & political ones, get more materials, go to the Moon for mining, etc etc... byproducts of what the ultimate goal was for the unit-economical value it yields
This is such a great read.
There's one thing I feel ... throttled in my, admittedly hobby, work in this space; assembly. Designing a PCB or picking parts is easy. Running the hardware experiments, assembly (excluding PCBs) is hard. Or at least labor intensive.
Another thing that resonated with me; finding hardware company who DON'T automate the hell out of everything, buying them and automate the hell out of it. I'll be shopping for just such a company.
This is in TLDR, but I wanted to raise it specifically because I'm delving deeper into my satellite product... and I feel I strongly agree!
I do think there's a lot more to "pushing the envelope" of what is possible with hardware even with GenAI assistance, because it requires a ton of knowledge (capture, collection, or regurgitation) about the space (ha). This goes for other primary and secondary sector industries, too!
For the mean angry mostly-vacuum above that pokes fun (radiation) and drags down sats, and for how cluttered + how much debris there is now in LEO/lower MEO, it becomes the question of "how can I engineer this thing to last far beyond the payload or the mission", not just passively / steucturally but also actively: modular upgrades, durable software, algorithms to detect and avoid other sats while station-keeping, etc.
An interesting take is you no longer need dozens of engineers to do the tedious work, if their knowledge is captured and now automateable so they can instead focus on iterating faster (unless I'm a defense prime program director and I or the customer wants to employ a lot of engineers, or any number of other scenarios... that's just a people & incentive problem!).
I'd be curious to hear from my fellow hardware integrators or builders what this is looking like for you! (I'm also a huge compute design nut, e.g. nand2tetris but tensor-scale)
I'm generally sympathetic to the view that progress in everything (including atoms) is downstream of progress in intelligence.
Like this piece argues, even "hardware" firms like Nvidia are really just design firms that outsource the actual manufacturing. I think the only companies that will be safe from AI disruption are the true scale manufacturing businesses that operate on thin margins and make it up in volume.