Conversation

, and Jim Keller

238

Jim Keller basically defined x86-64 whilst at AMD and comments on the complexity of variety of computer chips and technology involved in Tesla tech.

notes that he's the software to Jim's hardware, that one of his proudest achievements was in getting CS231n started at Stanford. Majority of audience have read his work or heard his lectures.

asks what's missing in current AI tech and whether Tesla use open source existing stacks or custom. Jim notes the efficiencies gained by specializing and removing overheads between components. Elon/

@karpathy

note this can give 10x the power at a tenth the cost.

Specifically that bottlenecks between the hardware and software can destroy much of the gains in a pipeline / integration - essentially Amdahl's law but across a pipeline.

Zaremba et al. (as I know him 🤣) asks the timeline for self driving cars.

makes a bold statement of two years for full self driving, three years for an order of magnitude better than humans. These gains are not binary, they can still be incremental even in driving.

Question regarding slaughter bots - "any competent drone company could do it".

@shivon

notes that biological and chemical attacks are right now quite possible with minimal work but that society has guarded against it.

argues regulation is two years for study, ...

two years for discussing how to do oversight, then first rule six years later.

argues the backlash from the first large scale AI disaster scenario where we had no preventative regulation will hamper progress far more than having had AI regulation in the first place.

Q: re how invasive neural link tech would need to be.

4:49 AM · Dec 8, 2017Twitter for Android

thinks it will be invasive but similar to lasik. Without it equivalent to your ears against a factory's wall - no good information. Sees this as necessary to maintain competitive advantage vs rapid progress of AI.

Replying to

Even if half of all new produced vehicles are autonomous, it's 2%/year that would be pushed to unemployed.

thinks AGI is going to be an issue before self driving unemployment - thinks five to ten years before AGI beats humans generally. Insanely optimistic estimate O_o

thinks there's no limit in energy for humans - we have a useful fusion reactor in the sky - our limit is purely on how many photovoltaics we have.

Ask to Jim: what are the exciting architectures? On chip memory to ensure data exactly next to compute and partitioning problems to have high bandwidth interconnects where you need it.

Q to Elon: How do we convince society as a whole that AI will be better, convince them to give trust, when statistics and intuition can be at odds.

thinks over time statistics will win out - even anecdotally / occasional video evidence to overwhelm negative anecdotes.

says that of the tragic failure cases in the past, they had extensive investigations (internal + govt) to analyze + show it wasn't an obvious fault of Tesla. He notes there will be many continued battles going forward however.

Q re what AGI will look like to arrive -

@karpathy

and

both claim it will be the jump that catches us. Extreme sub human behaviour to super human with no long duration "near" human level. It won't work until it really does.

Continued Q re AGI - Exponentials seem to continue until they don't want hardware hitting limits (Moore's) far before requirements of human brain. Jim disagrees, that he has seen this fear in the past, but that the sheer force of humanity's progress keeps pushing hw forward.

Technology has continually been predicted to stop moving forward but that Jim plans to ignore that the rest of his life as it hasn't been true so far.

Q re adversarial protection on Tesla vehicles. Stop signs on a short.

says stop signs are fine - they'll be geocoded - and that the real issue will be "bullying" of AI cars. You'll always be able to cut off a self driving car as it won't fight a human on that.

notes simpler adversarial issues exist - highly reflective back of trucks, roads painted on walls (Wile E Coyote!) - that much of life will provide inadvertent adversarial problems to deal with.

Q re ethics: Military AI will be a fearful fight when nation states decide that AI is a competitive capability for warfare.

asks whether self driving cars will have unintended societal consequences - urban sprawl, ... Also questions whether progress in public transportation would be better.

states that no one wants better public transportation if they can get personal cars.

Q re who will buy these self driving cars and who will gain benefits of automation if it - will it be Airbnb of your car? Cars will be more accessible as car costs will be shared (minimal overhead for not always in use). 3D tunnels will allow less congestion.

notes that progress is slow - Tesla released their car in 2006 showing the future but auto execs were straight up blind. Flat earthers but in tech.

End of conversation. My missed question (

@shivon

kept noting neural link will allow direct domain knowledge into brain (ala Matrix) but isn't that a terrifying vulnerability? We saw in the last year how vulnerable to coercion humans are, ...