-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable
- Thread starter AudubonB
- Start date
But the same 12.4 that was supposed to have 5-10x improvement, for the short period it was released to testers, turned out to be a serious downgrade to the previous version. So what went wrong?
This simple explanation is that 12.4 is not indefinitely delayed. I don't know why you would think that.
I also don't think there is any reason to believe that 12.4 has been performing worse than 12.3. There is no data to support that either.
Knightshade
Well-Known Member
Presumably because it has missed it's original stated wide release date and no new one has been provided? That's literally what indefinitely means so I don't know why you wouldn't think that?
Did you watch AIDrivers video where it shows 12.4 performing worse than 12.3 and requiring more interventions? Also the fact they're already on to 12.4.2 in internal testing hoping to find a version that isn't performing worse than 12.3 so they can wide release it?
OK. I think you may have misunderstood what Altman was saying. He was not disagreeing with Karpathy about the scaling law.
He said, "When the models get smart enough, it shouldn't be about more data, at least for training."
Altman was saying that once the LLM is extremely good, having extra proprietary data won't matter much for business purposes. The improvement you might get won't matter. For FSD, we talk about the march of 9's. But once you get enough 9's you really don't need more. That's all Altman was saying.
The scaling law still stands.
So in the case of FSD, Tesla has vastly more high quality data. Therefore, Tesla's neural net can benefit from the scaling law and improve much faster than any competitor.
It means Tesla will get there first because Tesla has the data advantage.
Agree to disagree @Usain. I'm pretty confident about my calls 
. Also I don't agree with your interpretation of what altman is saying, he also said it's foolish to try to anticipate AI more than 1-2 years into the future.
You haven't addressed a number of the points I made, but you seem to be convinced tesla has the right approach, while I genuinely believe they're manufacturing a growth story to keep the stock price above the required threshold. The timing of the announcements screams that. Occam's razor and all. 10 years in, just as we're consistently seeing negative growth on delivery metrics , fsd is suddenly ready? I don't buy that.
. Also I don't agree with your interpretation of what altman is saying, he also said it's foolish to try to anticipate AI more than 1-2 years into the future.You haven't addressed a number of the points I made, but you seem to be convinced tesla has the right approach, while I genuinely believe they're manufacturing a growth story to keep the stock price above the required threshold. The timing of the announcements screams that. Occam's razor and all. 10 years in, just as we're consistently seeing negative growth on delivery metrics , fsd is suddenly ready? I don't buy that.
Last edited:
I don't think there is a disagreement if you read the words closely and think mathematically. Although I think there is a slight issue with one part of Usain's paraphrase.
The main paraphrase Usain provided from Karpathy is: "performance of an LLM is a smooth, well-behaved, predictable function of N, the number of parameters and D, the amount of training data."
So, first, there are two variables to consider...so without increasing N, there might not be an increase in performance just by increaseing the amount of Data. At some point, you need a more complex model to make use of more data. And, I would expect you would need ever more powerful hardware to make use of a larger and more complex model...so hardware can also be a limitation.
Also, a smooth and predictable increase in performance does NOT necessarily mean that, as data increases, there will always be a significant or even useful improvement in performance. It seems logical that the performance increase would eventually be asymptotic, so there would be diminishing returns with more data. From the LLM perspective...there is probably some amount of data inputs that would provide a "perfect enough" solution so that there would be no perceivable difference jumping from that level of data to training on every known human book, text, conversation, etc.
So, while adding more data above a reasonable limit may provide some "improvement" in the model's ability to provide discussion on, for example, some random and insignificant event in history from the viewpoint of a similarly random person, such things become more and more obscure and might not "matter" relative to your paraphrase of Altman's work. Although, now that I'm thinking of it, there will always be SOME additional data that can improve things, because languages continually evolve...new idioms and cultural references and slang is always developing. So, in that sense, adding additional data which is new and non-duplicative will always add that bit of performance with the model's ability to keep "current."
Usain's next paraphrase was that there was "no known limit to this scaling" but I think that was slightly incorrect compared to Karpathy's statement that there are "no signs of 'topping out.' " Karpathy's video is 7 months old, and his statement would be relative to that time. So, at that point in time, there were still big improvements coming by adding more data and making the models more complicated. Karpathy DID NOT mean that there will never be a 'topping out'...just that, at that time, big improvements were still being made.
With driving, and ignoring hardware limitations, adding more data would NOT mean adding more videos of cars driving in the same boring patterns on well marked roads. After some basic functionality is achieved, the new data needs to capture all the strange stuff -- bad lane lines, confusing signage, unusual merge patterns, random junk in the roads, erratic drivers nearby, blind turns or hill crests, etc. Hopefully the model can ultimately build on the examples it has "seen" and ultimately be able make the best possible maneuvers in any situation -- it should of course interpret and estimate what to do for similar, but different, random junk and/or erratic drivers nearby, etc., just as a human mind does. And, if its range of perception and its response times are better than those of a human, then it might achieve super-human levels of safety. But, naturally, once thre is sufficient training data on driving patterns/laws/signs/etc. in all areas of the world (likely with different models for each region?), there would similarly be diminishing returns with added data...and any added data would need to be carefully curated to provide some new and useful information.
Last edited:
12.4.2 has moved to internal testing. There is no reason to believe that Tesla has no target date for wide release. So "delayed indefinitely" would be an incorrect assumption.
An AIDriver video does not provide data showing that 12.4 actually performs worse than 12.3. It's just a video.
You have completely mischaracterized the situation. And I think you know it.
The 12.4 model has to be tweaked for comfort and tested for safety regressions. Once that process is complete, we will finally be able to judge improvement over 12.3.
Tesla's problem is a model problem. Which is why I find ridiculous that people are trying to anticipate how fast they can solve it. It's obvious the current approach is more of a slider that takes from somewhere to put somewhere else. Otherwise there'd be no regressions, just a straight line of disengagements triggering learning opportunities and thus avoiding the same mistake.
Bingo. Why 12.4.2 and not 12.4? Let's look at the timeline again. In the first week of May, musk was saying that 12.4 has 5-10x fewer disengagements than 12.3. That was 7 weeks ago. Either their internal testing is seriously disconnected from the real world or someone was lying about the internal performance . How did they go from "tweaking for comfort" to continously delaying it and having to train new models?
You can't be serious about moving the goalposts with a straight face. If I go back 6 weeks into this thread people were expecting 12.6 on public release for 8/8
Last edited:
That's fine. Disagreement is what makes for buyers and sellers.Agree to disagree @Usain. I'm pretty confident about my calls
You haven't addressed a number of the points I made, but you seem to be convinced tesla has the right approach, while I genuinely believe they're manufacturing a growth story to keep the stock price above the required threshold. The timing of the announcements screams that. Occam's razor and all. 10 years in, just as we're consistently seeing negative growth on delivery metrics , fsd is suddenly ready? I don't buy that.
I get that there is a temptation to think that Elon is just desperately trying manufacture a growth story when auto sales are lagging. So I do want to push back on your idea that this is all a matter of timing.
First, the kind of end-to-end approach Tesla is now pursuing has always been the final solution. Karpathy told us how Software 2.0 would eat the heuristics eventually. Others, like George Hotz, said much the same thing. The real solution has always been end-to-end. It's just that Tesla has not had the resources to pursue that approach until recently. So that explains the timing. And, it's something Tesla was already working on when auto sales were booming.
Second, if Elon were just pumping the stock, would he really be spending billions on compute and data centers and a new robotaxi platform? It's one thing to say you are going "balls to the wall" for autonomy. But Elon is laying out a huge bet that autonomy will work.
And it is still a bet. Altman is probably right that it's foolish to anticipate AI very far into the future. I agree with all that. Tesla is a risky investment. Big risk. Bug reward.
Knightshade
Well-Known Member
Elon announced an intended date,
That date has passed and no new one has been announced.
delayed indefinitely-- since the length of the delay remains unknown to the same group who previously had been given a specific target date, is factually accurate not any sort of assumption at all.
Perhaps data is also a word you're unclear on the definition of?
12.4 required more interventions, in the same conditions, as 12.3 did. That's data.
It's not ALL the data of course- but it's factually more than the 0 data you claimed existed.
Are you speaking to yourself, or are you repeating the same mistake again mischaracterizing my posts?
If 12.4/.1 wasn't less comfortable and didn't have safety regressions it would've been released. In fact Elon specifically said "if all goes well" it would be--- and it wasn't.
Thus we know, from Elons on words, it did not go well.
(and we even have at least 1 public video showed the inferior performance of 12.4.x so far)
Now they're on 12.4.2 and HOPE that one is finally better than 12.3.x and can get wide released- it's unknown if that's true or if they'll move on to 12.4.3.
So again there's plenty of evidence of the things you keep insisting there's no evidence of.
Yes, this is a new kind of software development. It is indeed hard to anticipate the speed of progress.
But I don't see why a few weeks delay is a big deal. From my decades of experience in software development I can tell you that releases are almost never on time.
I wasn't expecting that at all. I always expect delays.
A few weeks of delay doesn't mean something is fundamentally broken.
You're both right
.Literally, "indefinitely delayed" just means the date is uncertain. That CAN be a very neutral statement.
In common usage, it could mean the delay is anything from a few days to forever (which really means "never"). And, often when a company says there is some "indefnite delay" to a product, it might be a euphemism for "never" with a very negative connotation (ie: we just realized we're going bankrupt, etc.). So, I won't fault anybody for realizing that interpretation as a possibility and suggesting "indefinite delay" is not the *best* and *clearest* description of the current state of 12.4.
I think we are fairly certain that, since Tesla is still testing point releases, some version of 12.4 will eventually be available to regular FSD customers, and that the "indefinite delay" is literally just an "uncertain delay" of somewhere between a few more days and a couple months. Although, there is probably some possibility that some breakthrough on 12.5 will come sooner, and Tesla will skip the 12.4 release entirely...but that would be a GOOD thing, not a bad thing anyway, so the negative connotation often tied to "indefinite delay" still doesn't quite fit.
That all being said, hopefully we can get away from the arguments over semantics and get back to arguments over where our stock is going
.
Last edited:
The stock long term is going up. In the short term the stock is going up and down. I hope that settles it. 
Musskiah
DisGruntled
Kansas Panasonic plant update - production begins 2025 Panasonic applies lessons from Tesla Giga Nevada to new Kansas facility
Musskiah
DisGruntled
I'm not a huge Cathie Wood fanatic. I have never owned their ETFs, so I'm not coming to the defense of ARK as a fanboi. That being said, I would like for you to find me another (college project or otherwise) analysis in 2020 that predicted the 2024 stock price more accurately.
What you suggest is innacurate. In 2020 ARK didn’t forecast a specific number of cars sales at all. Their analysis was based on probabilities (as seen below).
Similar threads
- Locked
- Locked
- Locked
