I don't think Tesla is going to drop FSD R&D anytime soon. They've also built out amazing software infrastructure under Karpathy's leadership to make use of their fleet of hundreds of thousands of cars to run shadow experiments with newer models in the background and continuously collect more difficult and interesting training data so they can improve their models and their internal test-sets for tracking metrics. They are also a relatively lean team from what I've heard, so it's not like they are bankrolling hundreds of very expensive ML PhDs at the moment. It is debatable if they (or anyone else) will "solve" FSD with current approaches, but I think the technology is already pretty fascinating and useful.
Yup. “The best minds of my generation are thinking about how to make people click ads.”
If you want to do fundamental research, the jobs are very scarce and, in many cases, not particularly good in terms of salary, stability, or location. A postdoc in the life sciences makes $50-65k/year, often either in a) a high cost-of-living area (Boston, SF) or b) a land-grant institution in the middle of nowhere, which is tough with a partner. These are usually short contracts too--mine is renewed annually. Faculty and Pharma jobs pay better, obviously, but are also pretty thin on the ground.
I like doing research that makes the world better, but doing so is an incredible luxury, even coming from a decently middle-class background. If my family were even slightly poorer, there's no way this would be possible and if one of them were to get sick or hurt, I can't imagine how I'd be able to stick it out.
It doesn't have to be this way, obviously. We could fund more stable positions--and I think it'd probably work out to more/better science per dollar spent. But right now...we definitely don't.
And it's not only to click ads, there is a whole lot of engineers (even more in poorer countries such as Brazil) who end up working in banks and finance, mostly to deal with complex spreadsheets.
It's a complete waste of educational investment.
There used to be a big incentive for the smart scientific and engineering minded types to go into academia, but now the majority go into either financial services or tech work because that is where the high salaries and actual problem solving are.
Though I don't know if it was an intentional reference or just a coincidence.
That's better than when the best minds are thinking about how to kill people.
Anyways, it is a good thing that even if the end goal is far from noble, most of the times, invest enough in R&D and positive things come out of it.
> numerous lives it would save
If you want to save lives, work on malnutrition, reducing coal energy, water access, basic vaccines, etc. You'll have a much greater impact.
This eliminates the most deadly accidents (head on collisions) and makes overtaking completely stress free.
It's a simple intervention that you can do almost anywhere, and it saves lives immediately. You only need to invest a bit in concrete barriers and paint.
I wonder why almost no other country does that.
Edit: You think our covid restrictions was successful?
They're talking about road layouts.
https://www.statista.com/statistics/525530/sweden-death-rate...
The common point that you won't see huge structural improvements until human driven cars are gone is true and it's going to be a WHILE.
Even if you had perfect self driving technology the cultural and political capital you'd have to expend to get rid of recreational cars is just astronomical. At that point you may as well ban all of the fucking things, the electric ones aren't better they just shovel the suffering around a bit.
The externalities of driving are huge (both positive and negative) but the negative externality of driving accidents is nearly a trillion dollars per year in the US.
On the other hand, nobody has any idea how to get Americans to change their minds on guns, cars, or red meat, and there is no foreseeable course of action that would work. That's not to say it's impossible, but I don't see how to get from here to there. This isn't a case of people just not knowing that if they drove more carefully they'd be at less risk of dying. Public education is not even remotely sufficient to the task of accomplishing the behavior changes we're talking about. Hell, I bet you could spend a hundred billion dollars and not even get to the point of a solid majority agreeing there's a problem to be solved, much less make any progress on solving it.
It isn't difficult or expensive in its own terms. You don't do it with "public education", you do by consistently dramatising the results you want in popular media, and demonising the results you don't.
Give it ten years of consistent messaging from multiple seemingly independent sources and it's done. Give it twenty five and it's so done the alternatives are no longer thinkable.
The difficult and expensive part comes from the enclosed nature of political power in the US, which has a choke point on the kinds of messages that are allowed to appear in popular media.
Interesting that your list doesn't include alcohol.
I'm always astonished at how those most skilled at software, turn around and explain how software is better than humans at task $x. Yet, I've never used a piece of software without bugs. I've never seen a piece of software, not real world capable, which even remotely close to dealing with things, as well as humans.
And the more complex the software, the more bugs. Bugs abound in complex software. Full self driving is as complex as the linux kernel, easily.
How many bugs are in the Linux kernel? Right now?
To even hope, even begin to hope that things will be mostly bug free, bug free beyond human "bugs" when driving, we'd:
- Need to 100% freeze the platform, so the car has zero changes. Ever. Like a Volvo, 10+, 20+ years with only fixes for flaws.
No change in drive train. No change in how the car handles, etc.
- A 100% feature freeze fix for years and years, only bug fixes, no feature add.
It's just not happening overnight. And when it does, cars are going to require an addition of massive sensors, it's not going to be camera or radar, it's going to be camera, radar, lidar, motion, wind, and 50 other environmental sensors.
It's going to be mainine computing power in triplicate. Connectivity and sensors in triplicate. 100%, no way is the AI CPU even slightly network connected. No over-the-air updates.
Bugs cannot happen. Bugs kill. Bad hardware kills.
I do get that having a car self drive, in the desert, in constantly warm temp is an easy thing. Try it at -40C, in a snow storm, with a snow filled road, in the country with no ditches, no lines to see (due to snow), in the dark, when windy, periodic white-outs, and more.
Heck, I'll say there's a chance when AI can drive on a frozen lake. And yes, humans do and can.
So far, all AI driving is monoculture.
Not to mention, how about a 20 year old car? Or even 10? "Buy new?" Now you hate the environment! And 20 year old cars are driven daily by humans, no they aren't the primary cause of accidents, because, when driving you notice issues in the feel of how the car drives.
Can AI "feel" that? Sure! But that means loads and loads of sensors! And cost! ANd repair bills!
To be honest, my 2050 thing is likely mostly due to this. Cost.
Speaking of Sweden's road safety rate...
https://en.wikipedia.org/wiki/List_of_countries_by_traffic-r...
Denmark (like most countries) are definitely better at not getting killed on the road, but closer to a factor of two by miles driven and four per capita.
They don’t matter per se. It is just a stat I had in my head that I could provide without googling anything to show how different both countries are. There are so many factors to consider with regards to driving infrastructure and habits, even density is not enough.
Waiting for L5, even if it actually is achieved in the current AI cycle, will introduce a whole new class of problems once self driving cars become a meaningful share of traffic, e.g. coordination/cooperation between cars and so on.
Indeed, a worthwhile intervention. But when do you think Tesla will be auto-updating the FSD rickshaws in Liberia?
https://en.wikipedia.org/wiki/List_of_countries_by_traffic-r...
Traffic backups and delays aren't just created from people crashing, it's because when you take a a dozen tiny humans and you put each of them in mostly > 15foot long cars, you now have 12 people managing to take up over 180 feet of space.
Speeding up emergency services is a solved problem. You take a lane that is filled with single-occupant vehicles, and you dedicate that to efficient forms of transit, whether it be a high-speed bus lane or a bike lane. Emergency services will not only opt to use that the majority of the time, they will also be less likely to crash with another vehicle due to not needing to weave lane to lane.
Hope this helps.
Or else it exists in some intersection of corporate benevolence and the work not actively harming the ad company. Which is just not an arrangement I trust sorry.
Plus I mean that's two people. How many people work at google and fb and how many who do get to decide they're going to do something more important than ads?
We are talking about "best minds of our generation". That means dozens.
But even a regular good engineer/researcher can probably find a place where they can work on something they're interested in. I did.
I understand that HackerNews is tech centered, but some times when I read "the best minds of my generation are doing X" I wonder if the commenter can see the flashes of brilliance shining beyond their field or what they know as "tech."
It’s not really about interesting for most people, it’s about is there a market for this and ultimately can I pay my mortage with this? I’m all for noble pursuits but don’t put down the trillion dollar smartphone, social media, PC markets that have long established themselves for the product we don’t even know is possible yet
This is an absolutely ludicrous statement. A very large number of people absolutely do prioritize what is interesting to them when it comes to their career. No one is going to be dirt poor, busking on the streets to make ends meet because they decided to work on self-driving tech. Nor is Tesla in any dire situation that they can't afford to fund their well-paid team of 20-25 ML engineers.
This is so asinine. Just because some R&D project is the most profitable thing FAANG companies can invest in doesn't mean it's actually the best allocation of resources for the people working on it.
Market fundamentalism is a brain disease.
I am sure Tesla employees can pay their mortgages.
So sure, he could leave, but to do what? To start building something similar from scratch somewhere else? He could go to Waymo to find that he went from being the project leader to an underling, and that Waymo is not using the approach he wants to use (he designed the Tesla project, and it's close to pure deep learning). IMO, unless something goes really wrong at Tesla, he will stay, because they're so close to success, surely he wants to see what happens, because the payout will be huge.
Tesla's R&D spending in their earnings report was ~1.5 billion. That's a small number for unlimited. To contrast, their competitor at GM, with a significantly smaller market cap, is spending 6.2 billion.
> They've already deployed said hardware in the wild, and they're probably about as close to Waymo to true full self driving
Can you cite a source for this? Tesla is "probably about as close to Waymo to true full self driving" is an absurd claim. Waymo publicly states they're at level 4 autonomy (https://twitter.com/Waymo/status/1347286935535017986). Tesla states they're at level 2 autonomy (https://www.autoblog.com/2021/03/09/tesla-full-self-driving-...).
Absolutely no one actually working on the autonomous side of the automotive industry, myself included, believes this. Tesla is around the middle-bottom of the pack in terms of actual driverless operating capabilities. They're at the top of the pack only in marketing and reckless deployment before the tech is ready.
The problem here is the "level" system everyone has been told is the standard, which heavily implies that if you're at "3" while someone else is at "2", then you're closer to "5". This isn't really true though because solutions at Level X don't necessarily scale/work at Level Y. I dislike using the level system for this reason.
I think Tesla is closer because they are certainly getting there[1], and I believe their model (relying entirely on CV) is the one that has the greatest likelihood of stretching to FSD (Level 5 autonomy). Mostly because there is only one system we're aware of that can achieve FSD (humans), and the sensory package there is entirely visual.
Waymo is not trying to get to L5 at all. They are a strictly L4 technology and their focus is clearly on servicing where the market is, metropolitan cities.
L5 is a pipe dream for anyone; it's a moving target no one can ever reach. It essentially means the car can drive anywhere, in all conditions and without the driver needing to take over at all. So it really means Tesla can't declare FSD as "done" until it works for everyone, everywhere and in every condition. How close do you think they are for that to happen? Does that also mean if it encounters a unique situation and gives up, their L5 claims fall flat? Given their current poor performance [1], they have a long way to go to even let drivers take their hands off the wheel, so I'm not holding my breath here (vision stack or not).
The real end game for self driving is L4 because that's what you can reasonably promise to your users. That it works in a defined, tested area and when required conditions are met. This is why literally every SDC company is L4-only, they never promise L5. Either Tesla is focusing on the wrong problem instead of providing value to its paying customers or they are blatantly misleading customers by promising "L5 by the end of the year" every year.
And yes, I would say my bar is a little lower than yours seems to be, as hinted at in my first comment. As long as the system can respond with a solution that is equivalent or better than what a human driver would do the vast majority of the time, I see that as FSD.
It seems like your bar is lower for Tesla (you're okay if it doesn't work always, yet you consider it true self driving), but extremely high for Waymo - who are actually fully driverless - just because it's geofenced. Sounds like you're not applying the same standards to both.
Maybe it looks like my standard is tailored to fit Tesla, but it's actually not, it is merely the behavior I would want if I ever purchased a self-driving car. Specifically, I want a self-driving RV where I can program in anywhere I want to go and it will take me there.
I think Tesla's approach can easily get us there in my lifetime. I'm skeptical of Waymo's.
It's still pretty vague
Not saying they will succeed here, but just saying that to give up on FSD would be corporate suicide at this point.
[0] https://en.wikipedia.org/wiki/Tesla,_Inc.#Founding_(2003–200...
A borderline non-sensical accusation w.r.t. PayPal, since he definitely didn't takeover PayPal, he founded X.com and got a payout through that after the merger and subsequent IPO + acquisition by eBay.
For Tesla, he was the first investor and took an active role in the company, starting in 2004 – less than a year after its inception and four years before they produced their first car (the Roadster). By the time the Roadster actually hit production, he was the CEO. Was the Roadster (a Li-ion sports car) Eberhard's idea? Absolutely. But it seems very unclear that it actually would have existed if not for Musk. And the rest of Tesla's vision/direction (that has made it the world's most valuable tech company) seems to have come almost entirely from Musk.
He wholly founded SpaceX and they're doing pretty well.
You may be thinking of a leveraged buyout or LBO.
A hostile takeoever is one where the management of the company are excluded from acquisition talks. The potential acquirer goes over their heads and proposes a deal with the shareholders directly.
In any case, I certainly don't think Musk's relationship with Tesla qualifies.
Edit
> I never heard that he didn’t pull his weight from Peter thiel or the others
I don't know. A letter of no-confidence to the board in order to replace Musk with Thiel doesn't exactly shine a lot of good light on Musk. From https://www.cnbc.com/2019/11/02/why-elon-musk-says-taking-va...
> At the time, he was CEO of X.com, and company executives were not pleased with his leadership. While Musk was on the plane with Justine, executives delivered a letter of no-confidence to the company’s board, pushing Musk out as CEO, and replacing him with Peter Thiel.
Though in regard to tech stacks, I also thought X.com might've been using .NET and (if that had been the case) then to my mind it was the better option, compared to what many people were using for scrappy Linux web backends in that era – PHP (which I say as someone who wrote his first web backend in PHP). But after looking up the timelines, I realize that this wasn't the case. Musk was replaced as CEO in 2000 and .NET was released in 2002.
Maybe you could have timed the bubble and exited with a pile of cash, who knows. If you want to make those kinds of bets you could make them right now.
Some people made out of course, but in general ~2001-2004 was not a great time to be in tech.
Per Wikipedia: "Within the merged company [PayPal], Musk returned as CEO. Musk's preference for Microsoft software over Linux created a rift in the company and caused Thiel to resign. Due to resulting technological issues and lack of a cohesive business model, the board ousted Musk and replaced him with Thiel in September 2000."
But, in general, salaries were mostly unexceptional in tech even in Silicon Valley, a lot of companies went bust, and even most of the surviving firms--including the big systems/storage/chip/networking/etc. players--went through some very lean years with a bunch of layoffs.
We’re in the same place with crypto right now. People will look back and say “oh jeez if only”. The early bazillionaires are already minted but there are a lot of crypto millionaires made every day. It’s just a lot of people dumping their life savings into BTC/ETH and enjoying the ride.
For comparison to Musk's other companies , it's analogous to Bill Cantrell or Mike Griffin or that guy from sea launch at SpaceX.
Some in this group were very briefly there like Tim Hanson.
Some I think are misreported to not be there , unless anyone has any extra information. For example , Paul Merolla was there at the last conference , even though he's reported as not being with the company.
Some I think are also advisors like Sabes , unless something has changed.
One of them is just a neurosurgeon , but they already have a whole surgery department with a lead now.
Less snarky response, perhaps if we saw more worker coops we'd care less about being a founder and more about delivering a result.
So even though the expected value is (correctly, in the medium term, IMO) super high, everyone would get nothing based on current or near term revenues, which is exactly the situation VC exists to rectify. What you're seeing is the free market working correctly as more founders jump in to take VC money and make bets on their ability to execute.
Interestingly one of the things Marx and Proudhon violently agreed on was that cheap credit is one of the most radical things you can create. The reasoning is simple: The cheaper credit is, the less power employers have over employees. To take a hypothetical extreme: if credit is free and guaranteed, an employee can walk out at any point and take whatever time they need to figure out what to do, at which point employer power of employees is nullified.
As such, at some point along a gradient of cost and access - and it may well be an impossible point to reach - capitalism becomes indistinguishable from socialism, in that capital ownership becomes a matter of choice.
The research isn't there yet, which is why Elon should be trying to retain people who have deeply studied the hippocampus and brain. What the heck is he doing? Find the best people in this domain. The potential is there with the electrode technology he's built. It's an AI problem now, to make sense of the data. Supervised learning mostly. You can measure what the brain does when certain things happen. He needs more researchers in the fundamentals. I think if they've built the robot that inserts the electrodes non-invasively, they can continue to add additional electrodes and get an even more accurate measurement of what's happening elsewhere in the brain. This tech will takes years and years to master. Neuralink a science company. Or at least, it should be at this point if I'm taking it seriously. God knows there are hundreds of tech companies selling SAAS software for domains that have negligible research. Neuralink shouldn't be like that. Don't overpromise and underdeliver just yet, causing best people to leave. This company has tons of potential.
This is an unproven hypothesis. Since we do not know how computation actually happens in the brain, we can't say for sure that electrical signals can be matched to the actual computations happening, even if they are correlated. It is still possible that the electrical signals we are measuring are just byproducts, and that the real compatation happens at a chemical level or inside each individual neuron or even in microtubules as Penrose believed.
The nice part is that this means that neuralink's attempts could be extremely useful to our understanding of the brain even if they fail - it would actually be an amazing discovery if they could show that you CAN'T deduce the brain's intent from observing electrical activities.
We do know for sure though that you can use electrical signals to interface with motor neurons, so at least improvements in prosthetics should be something that neuralink can realistically deliver.
The problem comes when you want to send a signal back into the brain - but even then there's a chance that the brain could rewire itself to take care of redirecting and reprocessing the fuzzy and inaccurate input.
Reasonably accurate output has the potential to revolutionise prosthetics (feedback aside).
I'm not thinking of something like echos, but more of auxiliary signals - e.g. Perhaps they are signals to increase energy supply to a particular are of the brain, or for controlling 'peripherals', while the main computation happens in other channels.
It's also an interesting question to see whether the analysis techniques could work on a simpler and clearer machine, one where we know for sure that is controlled only by electrical signals - could we infer what a CPU is doing by just measuring electric currents in various areas, without any knowledge of the code running there? This is an experiment that is eminently doable, especially if we design a core specifically for this purpose (making a large and slow core in order to focus purely on the problem of analysis, not difficulties in probing and data collection).
Not a neuroscientist, but it seems like the problem is more involved than just training a DL model on a data. We need detailed maps of the brain before we put any electrodes in it. Meanwhile, full mouse brain map is likely a decade away [0]. Humans have a bigger brain, imagine how long it’d take. It’s just a lot of manual labor which takes incredibly long time. Keep in mind that you also need to map thousands, if not millions of brains to model population heterogeneity.
The interesting aspect of your comment is in fact: what is the capital outlay and time needed for the science to actually get there.
Elon has a track record of severely under-estimating those, but he is also smart enough (via PR / govt grants / leveraging past successes, etc ...) to stretch both beyond anything reasonable, to the point of actually reaching success (e.g. TESLA, SpaceX).
I own a Tesla, I love SpaceX, I think Elon Musk is insufferable.
This is both a strength and weakness. How many times have people wished CEOs actually knew what they were talking about? As long as Elon knows enough to build a decent team and lead it well, he doesn't need to know everything.
From the outside it looks like he sees himself as 'one of the engineers' which is something I deeply respect, even if there's major issues as well (FSD and Hyperloop being the big ones).
> major issues as well (FSD and Hyperloop being the big ones).
That Vegas loop too... That's going to be a truly awful tunnel fire someday, and the safety systems there seem seriously lacking. (Also, the whole project feels like a grift.)
Risk taking is amazing. I’m not sure if I want to live in a risk averse world where ideas are nipped in the bud by CEOs of GE and Honeywell.
More information https://www.youtube.com/watch?v=-RPMt_FS-s8
No single personality can be 100% perfect. Let’s not ignore pretty much the entire world changing revolution of EVs and reusable rocketry, while focusing on the guy’s public persona and how he behaves that’s apparent to us and project what he must be like in general. Seems deeply and profoundly unfair.
Why can’t someone have flaws which is ok to point out, but then ignore their accomplishments. Opposite also stands! Fanboys of Elon think he is some kind of a messiah to save the world from global collapse but turn a blind eye to his rather rude behavior?
It’s like populism has a reverse backside of the coin.
History respects Edison, Jobs, Tesla, Carnegie, Rockefeller, Hughes for what the accomplished while acknowledging their individual flaws. I think Musk will be remember similarly. Depending how successful his ventures are over the coming century he might even be remembered more than any of them.
That doesn’t read as “haters gonna hate” to me, but rather as a narrow criticism of just the one aspect of Musk’s businesses that was brought up.
Elon sure did. He decided, apparently out of pure greed, to invest big into Bitcoin, driving the price up and with it increasing the carbon emissions of it, enough to quickly wipe out basically all carbon savings Tesla has worked to achieve for its entire company lifetime.
Our current financial system cannot.
The environmental cost of Bitcoin is not that of the energy it uses, it is the cost of the dirtiest energy it forces others to use.
There is always pollution. Some forms of energy have more and some have less. It is never good to just waste energy, that will always cause excess pollution, no matter the type of energy used.
It doesn't matter how energy is produced, bitcoin will always be harmful.
At least in the case of SpaceX, they do not seem to be as reckless as you paint them. After 20 years in space industry, they have 0 fatalities on record, which is uncommon for a corporation of this size and history.
And they scrub launches whenever they are concerned about weather or anything off-nominal that their sensors indicate. Which led to an excellent safety record for Falcon 9s once the technology matured.
She hasn’t gotten a ton of press but it’s clear she is a key reason for the successes you described above.
Which Musk can take some credit for—-hiring well and not getting in her way.
Reasonable estimates like what ? If your company works on something that has never been done before there is not much data at hand to make good estimates is it ?
I also have colleagues who promise small concrete incremental deliverables in predictable timelines. They rarely talk concretely about the big end goal pie in the sky’s timeline, but generally deliver the big thing sooner anyways.
The answer isn’t making reasonable estimates of an inestimable thing, it’s providing reasonable estimates for the reasonably estimable pieces. Instead of “we’ll have ultra mega science in three to five years!” it’s, “we’ll have this specific piece of the stack built by 2022 and we hope it’ll enable these other slightly less concrete things towards ultra mega science.”
There is a reason why legally Autopilot/FSD is a driver-assist system, but for marketing purposes robotaxis are just around the corner.
There is a fine line between 'overly optimistic' and misleading, and more often than not, Elon crosses the line.
See https://elonmusk.today/ for an overview (if a bit biased).
My point isn't that these companies don't get there, it's that they take much, much, much longer than estimated.
Note that they have done zero work on building a anything other than the rocket body to get there. Making the interior volume habitable for 2 years of interplanetary travel is going to be non-trivial.
There are enough senior software engineers in this world that will have a crazy hard time making a good estimate. There was this HN comment once that mentioned that learning how to estimate work is it's entire own discipline.
It can be done and trained, according to that comment, but the comment also seemed to imply that it seems very unusual to do so.
I can't find the comment unfortunately, I read it years ago here.
Elon Musk has been pretty successful at getting top talent to push the limits of engineering, but with neuroscience, it seems the research isn't there yet. What I'm really interested to see is how long Karpathy will stay at Tesla as other self driving projects drop like flies (ATG, Zoox, and most recently L5)