Andrej Karpathy on Tesla AI data engine

Lex Fridman Podcast full episode: • Andrej Karpathy: Tesla...
Please support this podcast by checking out our sponsors:
- Eight Sleep: www.eightsleep.com/lex to get special savings
- BetterHelp: betterhelp.com/lex to get 10% off
- Fundrise: fundrise.com/lex
- Athletic Greens: athleticgreens.com/lex to get 1 month of fish oil
GUEST BIO:
Andrej Karpathy is a legendary AI researcher, engineer, and educator. He's the former director of AI at Tesla, a founding member of OpenAI, and an educator at Stanford.
PODCAST INFO:
Podcast website: lexfridman.com/podcast
Apple Podcasts: apple.co/2lwqZIr
Spotify: spoti.fi/2nEwCF8
RSS: lexfridman.com/feed/podcast/
Full episodes playlist: • Lex Fridman Podcast
Clips playlist: • Lex Fridman Podcast Clips
SOCIAL:
- Twitter: / lexfridman
- LinkedIn: / lexfridman
- Facebook: / lexfridman
- Instagram: / lexfridman
- Medium: / lexfridman
- Reddit: / lexfridman
- Support on Patreon: / lexfridman

[@LexClips]

Full podcast episode: kzbin.info/www/bejne/mZXMdWBvgrKjmJI Lex Fridman podcast channel: kzbin.info Guest bio: Andrej Karpathy is a legendary AI researcher, engineer, and educator. He's the former director of AI at Tesla, a founding member of OpenAI, and an educator at Stanford.

[@rb8049]

My interpretation is the intel crisps are abstracting results in plots. My experience is what matters.

[@nicholasmarquez4076]

This is simple , necessary and very badass !!!

[@hagestad]

And that is why all Teslas brake a lot since this data is mined from customers cars and saved locally on SSD before being send to the mothership. Those SSD disks tend not to last very long.

[@ReflectionOcean]

By YouSum Live 00:00:14 Data engine: Iterative improvement of training sets. 00:00:24 Enhancing neural networks through diverse data sets. 00:00:40 Importance of catching rare scenarios for network improvement. 00:01:50 Human role: Strategizing data collection and execution excellence. 00:02:36 Task prioritization based on system feedback and goals. 00:03:16 Individual experience as a source of truth and intuition. 00:03:23 Insight from personal interaction surpasses statistical analysis. 00:03:58 Elon Musk's hands-on approach for system understanding. By YouSum Live

[@rb8049]

I wish it was in the error 0.000001% now. My impression is error is more like 10%. Still crossing my fingers that the training will use my car’s feedback soon.

[@distasteee]

The thing is - at what point is the AI safer than a human driver that's texting or been drinking or dealing with a screaming kid in the backseat etc. -- have we not already crossed that threshold?

[@Lascarnn]

Musk will be able to postpone and postpone for years... And a lot of amateur investors are buying this, which raise stock price to absurd levels, etc. If they won't finish it in 10, 20 years, Musk will still get his huge payment. Meanwhile, Mercedes is currently better than Tesla when it comes to autopilot.

[@martincday007]

Deploy, Data Source, Notice a Problem, Boost, Label, this seems like a relatively crude way of learning as it relies on a problem to have occurred, before it can be addressed by human intervention and intuition. That would be fine in a safe environment but self driving vehicles are not operating in a safe environment nor are they operating in an environment where it is only signed-up Tesla aficionados that will be affected. AI needs to be able to complete the Boost and Label aspects of the problem solving process without human intervention. Until it can do that, the process is effectively building in an "acceptable" number of potential harm and deaths to Tesla drivers and more significant those who are non-Tesla customers. A responsible company would take the first step and rename "Full Self Driving" to something like "Advanced Driving Assist". Only an irresponsible company's would market a product as one thing and then in the small print describe it as something completely different. Level 5 autonomy is probably still a decade away, why not celebrate being among the most advanced level 2, although it is interesting that Mercedes' Drive Pilot has become the world's first fully certified Level 3 autonomous driving system and is now available in Germany.

[@KingKwaZulu]

Did anyone else check the speed because this dude speaks so fast

[@v1kt0u5]

No, but sometimes I play Ben Shapiro at 0.75x

[@blueredbrick]

Not really. But yeah a bit of staccato cadence.

[@curtaustin8119]

When a human is confronted with a previously unseen situation, he/she can quickly figure out what to do; they *think* of something. They generally require far less training - a much smaller data set. A few photos versus dozens. I've not seen a specific effort by Tesla to pursue whatever process humans use that accomplishes this ordinary magic. But the AI community must be working on it. Does it go by some inscrutable name? Or simply "artificial thinking"?

[@bobo-cc1xw]

How long did you study for your driving test? Or feel good driving? Have you tried driving in a foreign country with different road layouts and rules

[@curtaustin8119]

@@bobo-cc1xw Huh. Here I am, confronted by this unforeseen comment, but I can't think of a response. Wait! I got one: What good are "Deer Crossing" signs, since deer don't know how to read?

[@bobo-cc1xw]

@@curtaustin8119 As a African driving around america first time i do recall driving around noticing these weird signs for Xing which i thought was tourist signs for a mythical beast or something and paid them no attention. It took me a long while to work out what a Xing was. I recall the scene in i think Knightrider 2000, surely the future of our AI overlords. Where evil new kit super AI sees a dear ahead and steps on the gas as it will do no damage and to clear the obstruction. Does my car need to go to school for years to learn what a Deer is? or even better a kiboko?

[@davidradtke160]

I think your confusing adults learning new tasks related to previous skills vs the time to learn from scratch. My 5 year old still has a lot of minor issues with language/speech. 5 plus years is a lot of time to train on English. I learn juijitsu techniques faster then most people, because I wrestled for years in high school. Am I really learning something new in a few examples and practices? Or am I adding and adjusting techniques I learned previously over years of wrestling practice for my new setting. With the latter I already had 100s or thousands of very close examples to the new behavior.

[@Joshua-rp2nq]

They're working on it :) it's called Zero-shot learning (or few-shot learning)

[@juhanamantymaa9407]

More like Andrej Carpathy

[@chethanslthippeswamy]

the question which was missed from the interview was "How did they got rid of ultrasonic?"

[@jaselee117]

He answered in a different video. How is easy, they just didn’t put it in lol. Why is because humans only drive with visions so roads are designed to be driven visually, and splitting the team to focus on different sensors is inefficient then focusing on just one aspect

[@westsiyed]

am I the only one know noticed he wearing his Apple Watch wrong? Isn't the crown suppose to be outward facing?

[@user-yh1px5wl4q]

Let's talk

[@user-yh1px5wl4q]

Chat me up

[@bonmec101]

You can wear it inwards facing as well. That’s how I prefer it. That’d enable you to press the crown with thumb which is more ergonomic than forefinger imo.

[@blueredbrick]

No.

[@landocycle]

You choose which way you want it during setup.

[@eleson00]

So a Tesla needs to be exposed to scenario, and fail at it, to get data to learn from?

[@lylestavast7652]

I ran a testing group for several years. Testers can be pretty creative at imaging what might be unusual and crafting scenarios to expose those concerns. So some scenarios will come from such efforts and they'll keep mining existing imaging for different aspects of it they were filtering around as time passes. Shadows, reflections, random things like plastic bag flying at you in some random flight pattern... kid kicking soccer ball out of an alley... bird crap, things falling off the back of a truck... (I've been hit twice by such items in my years driving)... it'll be interesting to see how creative they get to augment handling the expected, vs the unexpected. A guy I knew in Texas had a small airplane fly over and land in front of him in an emergency on a country road... so many odd things happen :)

[@CoalOskold]

Data sets that look at the world threw all of the animals eyes

[@user-yh1px5wl4q]

Let's talk

[@user-yh1px5wl4q]

Chat me up

[@darkmatter6467]

The content is great, but audio quality is crap. Everything is extremely compressed, dark, muddy.

[@DaCarnival]

"Crap" is subjective, but compared to most podcasts this is fine, and compared to a lot of youtube it's great. I also think this works great on phone speakers (though I'm listening on over ear headphones myself).

[@vivekacharya3652]

me first

[@user-yh1px5wl4q]

Let's talk

[@user-yh1px5wl4q]

Chat me up

[@intrametaarchi1015]

I think that Tesla represents the Military industrial complex, not everyone can see that.

[@attilarza2488]

The scientific name is: Muskius Scammius Insidius Tradius.

[@FredericJacquemin]

is it me, or LF is really, but REALLY bad with articulation? Hey Lexi, can you please pay attention to that? that would be great for everyone, you're like shooting off the content just by not articulating.

[@Tolo-rx6tx]

!

[@user-yh1px5wl4q]

Let's talk

[@user-yh1px5wl4q]

Chat me up