Tesla wants to take machine learning silicon to the Dojo To quench the thirst for ever larger AI and machine learning models, Tesla has revealed a wealth of details at Hot Chips 34 on their fully custom supercomputing architecture called Dojo. The system is essentially a massive composable supercomputer, although unlike what we see on the Top 500, it's built from an entirely custom …
AI based on, for instnce, all Dickens' books, which is his lexical clone, Dickens himself, contains his bias (as his BIOS), all his knowledge and preferences, everything, and which can easily be learned-unknown-to-Dickens/ trained for everything — about 5-7 MB, contains about 1.000,000 structured and weighted phrases. At the same time, such a clone is quite reasonable, you can talk to him and it really thinks and learns new. I know, I've tried: see my articles in NIST TREC QA. You can make one by yourself and try, if you doubt my words.
This is despite the fact that models like BERT are completely stupid, cannot be used in any reasonable way except for writing meaningless texts and insane Translating. Besides that these enormous models are incredibly expensive, complex and very difficult to train.
You'd think that, at some point, someone somewhere in Tesla is actually go to stand up and say, "This has gone too far, we must stop and just concentrate on building cars, not computers". That'd probably take a change at the top.
It's a very strange situation. The people working on this must realise that it's all very unlikely to succeed (i.e. it will not result in a fully self driving vehicle), even with such silicon largess, and that effectively they're taking their company for a ride. Ok, so the boss is encouraging them to do this, but you'd think that at some point the sheer scale of it all will make even the most ardent of powerpoint engineers wonder about the wisdom of keeping saying "yes", taking the money and smiling. I'd not want to be there when the brown sticky stuff does eventually hit the fast spinning fan of reality, as it must at some point surely do.
If they're having to build this kind of thing for "training", what does that mean for all the "fully self driving capable" packages that Tesla have sold? Or, has that one already been quietly dropped and forgotten about?
Oh well. Perhaps when the boss wakes up having been forced to spend $44 billion he didn't want to, some sanity will prevail. Or is this thing destined to help Twitter seek out and destroy non-monetisable, non-profitable bot accounts?
Alternatively, Tesla have seen the writing on the wall as all the "traditional" car makers are now on the EV gravy train and they have way more experience at building and marketing car and this is a way of either creating a USP or something to spin out when Tesla finds it's no longer a disruptor or innovator. This does seem like a pretty serious and big operation to run inside of Tesla with likely applications to general AI/ML R&D, not just or specifically to self-driving cars.
this thing, whatever it is, exists. Look, we are talking about “paedo guy” here. If it was real, he would have done umpteen pre announcements about forthcoming pre announcements and then pre announced slipped delivery dates e.t.c. When is that semi being delivered again Mr Musk? Oh, and that thing with unbreakable windows that got broken at the pre launch. Still, we can all look forward to his robot being delivered next month was it?
The main description sounds like a quantum leap in SoCs with the ultimate GPIO interface :-)
It might need a slightly larger case than a Pi to put it in though :-)
I'd imagine quite a few chip designers have thought about doing this, but have been unable to convince TPTB to run with it, needing a market to sell it to. Tesla have decided to Just Do It for their own uses. Assuming it lives up to expectations, they'll either go to market with it or others will copy the idea.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
