1. ML models are off the leash
Want to read a camera frame and identify a person, there’s a model for that.
Want to transcribe user’s speech to text, there’s a model for that.
Want to respond to a question with a sensible answer, there’s a model for that.
These things weren’t so good. You had to tack on endless heuristics and ended up with extremely brittle stuff that broke on the first unseen case. That’s changed.
2. Hardware is becoming very accessible
As of today, you can buy a soldering kit on Blinkit! Yes, I bought one.
Add to that the fact that can now order a dozen sensors, from gyroscope to ultrasonic to humidity, on Amazon and get it delivered next goddamn day. If you’ve ever scoured the narrow lanes in electronics markets in midday heat trying to source a random-ass component you know what I’m talking about.
3. There’s a teacher available. 24/7/365.
The trouble with hardware is that the cost of mistakes is higher. You can break things, burn things or even worse blow things. A dedicated lab with a supervisor would be very lovely. Somebody who can correct you when you are doing stupid shit. But you don’t have the luxury outside colleges.
That’s now possible. The LLM is my prof. I ask the LLM everything. If it could speak its mind I’m pretty sure it says “dude do you anything besides connecting wires and bolting nuts?”
4. Writing software is going to be fundamentally trivialized.
With each improvement and iteration in the next LLM, the model gets much much better at writing code. There is no real need write it yourself as long you can clearly and unambiguously spell out what to do.
And hardware coding is actually more boilerplate than you can imagine. Sure there’s logic. But there’s a ton of code just describing your breadboard connections so that the code knows which bits should receive signals. The lift you get from LLMs to get this done is crazy.