Why did Elon tweet that he needed "hardcore coders" rather than AI experts?

My guess is that Elon wants some out-of-the-box thinkers: people not trained in orthodox AI methods which may be insufficient for what the company wants to accomplish.

He also might want just the opposite. Elon is the "Chief Architect", he probably has his vision, and will surround himself with the tools to create said vision. No need for AI experts or machine learning PhDs; just hardcore, "yes" man, coders.

calisnow said:

Just got me thinking because the tweet was really non-specific - it didn't call for machine learning experts, machine vision experts or anything else - just "hardcore" coders.

Click to expand...

While I cannot know what Elon was thinking, I can tell you in my experience, there is a big difference between those who research stuff, and those who code stuff. I've met very few researchers who are really good at coding (there are a few...very few). It's a different skill set.

Give me a coder and I can explain what needs to be done, and even though they aren't experts in the particulars if they can write good clean reliable code, with all the edge cases thought of and tests written, I'm much better off than if I had the "expert" write the code herself. Often I find that by working side by side with each other to express an algorithm as code, the final solution is much better than the original proof of concept code. This is due to the coder's being much more aware of memory usage, code speed, cheats and tricks, and always on the lookout for the corner cases.

Just as brilliant building architects don't necessarily make good brick layers, AI experts don't necessarily write good code. Tesla probably has the expertise, now they need to express that as fast reliable code.

woof said:

While I cannot know what Elon was thinking, I can tell you in my experience, there is a big difference between those who research stuff, and those who code stuff. I've met very few researchers who are really good at coding (there are a few...very few). It's a different skill set.

Give me a coder and I can explain what needs to be done, and even though they aren't experts in the particulars if they can write good clean reliable code, with all the edge cases thought of and tests written, I'm much better off than if I had the "expert" write the code herself. Often I find that by working side by side with each other to express an algorithm as code, the final solution is much better than the original proof of concept code. This is due to the coder's being much more aware of memory usage, code speed, cheats and tricks, and always on the lookout for the corner cases.

Just as brilliant building architects don't necessarily make good brick layers, AI experts don't necessarily write good code. Tesla probably has the expertise, now they need to express that as fast reliable code.

Click to expand...

I think woof nailed it.

I'd just tag on that my assumption is he wants people who are great at "figuring things out" -- on the fly, in areas they aren't familiar with, as the story develops, etc. This is what the best coders I've known are fundamentally excellent at, and it seems to be something Elon hugely values. (as anticitizen13.7 basically said as well)

woof said:

While I cannot know what Elon was thinking, I can tell you in my experience, there is a big difference between those who research stuff, and those who code stuff. I've met very few researchers who are really good at coding (there are a few...very few). It's a different skill set.

Give me a coder and I can explain what needs to be done, and even though they aren't experts in the particulars if they can write good clean reliable code, with all the edge cases thought of and tests written, I'm much better off than if I had the "expert" write the code herself. Often I find that by working side by side with each other to express an algorithm as code, the final solution is much better than the original proof of concept code. This is due to the coder's being much more aware of memory usage, code speed, cheats and tricks, and always on the lookout for the corner cases.

Just as brilliant building architects don't necessarily make good brick layers, AI experts don't necessarily write good code. Tesla probably has the expertise, now they need to express that as fast reliable code.

Click to expand...

There's that, but more important is that Tesla demands that its developers work insane hours, and you have to be pretty hardcore to want to work 70 hour weeks.

Would add that according to Elon AI could be the end of civilisation (ok, an exaggeration) but more importantly - AI can be unpredictable which is not what you want from safety critical systems. Imagine every car learnt to respond to the road differently based on the type of roads it'd been driven along - unless you keep track of exactly what inputs it receives and how it adjusts itself it is a nightmare from a problem diagnosis point of view.

AI is great, it has its uses and will be a big part of the development of autonomous vehicles (for example, machine learning can be used to identify the best absolute algorithm for detecting lane boundaries), but its use in the end product will be limited if present at all (as in, the car you drive won't be doing much learning 'on-the-job'.

If you watch the NVIDIA CEO talk about what is needed for Autonomous driving, it would appear human habits in real life driving as that currently being collected with Release 7.0. is part of what needs to converted to code. I like JHH's "how do you teach a baby to play ping-pong" comparative: GTC 2015: NVIDIA DRIVE PX Self-Driving Car Computer and Deep Learning (part 8) - YouTube

I can imagine Elon is looking for the best coders to convert what is learned by Tesla and NVIDIA self-learning to software code that can be compiled to Tesla Auto Driving releases to follow. Therefore, coders need no knowlege of AI, since as you watch the video, AI is not invented but observed. It took coders at IBM 3 years to code the "human" dessison process to make "WATSON" become a JEOPARDY Champion -- Despite its massive data basess.

Autopilot does not require any AI.

I disagree that researchers are generally not good coders. I'm a computer science professor and I've worked with a lot of PhDs (and produced a few as well). Places like Google love hiring PhDs and they are very very good at coding.

And for the assertion that "autopilot does not require any AI." Hmm... depends on how you define AI. Computer vision is usually distinct from AI these days. But imagine a truly autonomous vehicle that is suddenly confronted with having to either sideswipe an adult or collide full-speed into a bassinet (which may or may not be occupied). It's actually a moral question which action to take.

I have seen the future... and the future is...

... lawsuits.

ozweepay said:

I disagree that researchers are generally not good coders. I'm a computer science professor and I've worked with a lot of PhDs (and produced a few as well). Places like Google love hiring PhDs and they are very very good at coding.

And for the assertion that "autopilot does not require any AI." Hmm... depends on how you define AI. Computer vision is usually distinct from AI these days. But imagine a truly autonomous vehicle that is suddenly confronted with having to either sideswipe an adult or collide full-speed into a bassinet (which may or may not be occupied). It's actually a moral question which action to take.

I have seen the future... and the future is...

... lawsuits.

Click to expand...

Most code PhD's write is garbage left behind while they're patting themselves on the back that has to be completely re-written to function in the real world.

I think you guys are taking "coder" too seriously, I don't think he wants mindless keyboard monkeys, just people that really understand how to build systems.

Keepin' it real...

Systems vs. Coders

AWDtsla said:

Most code PhD's write is garbage left behind while they're patting themselves on the back that has to be completely re-written to function in the real world.

I think you guys are taking "coder" too seriously, I don't think he wants mindless keyboard monkeys, just people that really understand how to build systems.

Keepin' it real...

Click to expand...

Implementing space and time efficient code and data structures, along with the User Interface for an AI system someone else develops is what I think of when I hear the term hard core coders. As you say, people that understand building systems. Computer Science PhDs are less interested in real-world coding (unless that is their field of study, e.g. software testing) and more interested in creating new efficient data structures for their own ideas, as opposed to, say, implementing a GUI for someone else. Most PhDs are paid to be creative, not to be coders. The "garbage left behind" by a PhD may be the kernel of a great idea that gets trashed during an implementation that is rushed to meet a schedule.

Its an extremely unorthodox way to recruit developers, I can tell you that. One assumes that the signal to noise ratio resulting from that tweet is not very good.

ozweepay said:

I disagree that researchers are generally not good coders. I'm a computer science professor and I've worked with a lot of PhDs (and produced a few as well). Places like Google love hiring PhDs and they are very very good at coding.

And for the assertion that "autopilot does not require any AI." Hmm... depends on how you define AI. Computer vision is usually distinct from AI these days. But imagine a truly autonomous vehicle that is suddenly confronted with having to either sideswipe an adult or collide full-speed into a bassinet (which may or may not be occupied). It's actually a moral question which action to take.

I have seen the future... and the future is...

... lawsuits.

Click to expand...

Hm...well he's hiring for the "autopilot software team", but then says "to achieve generalized full autonomy". Maybe he should clarify what he's asking for.

A fully autonomous car would probably need an AI to handle unexpected real world situations and make moral decisions while driving. It is for this reason that I am very skeptical of ever having quality autonomous cars on existing roadways (without updating the roadways), and think it will ultimately be easier to make life multi-planetary than perfect autonomous cars. However, I don't this Musk is asking for this. Maybe he should be, but I doubt it's what he is envisioning. More like a giant software system with a giant database, but it will still happily plow over both the adult and the bassinet if that's where the software decision tree happened to take it. And it still won't be truly autonomous; there will still be situations where it doesn't know what to do, at which point, it either has to turn control over to you, pull over and wait for a tow, or make a "best guess". The first isn't full autonomy, the last isn't going to fly (accept maybe over a cliff), which just leaves the middle option, which they could claim is full autonomy by arguing technicalities, but we would all know it really isn't.

If I was 20 years younger, this would be a fun challenge.