The BIG Problem with AI & Tesla FSD!
ฝัง
- เผยแพร่เมื่อ 16 เม.ย. 2024
- The first 500 people to use this link and code DRKNOW will get 25% off their first subscription with Soylent: bit.ly/3vNiahw.
What happens if/when AI ingests primarily synthetic data? Do we get LLM and Tesla FSD Collapse?! Why could Mode Collapse lead to the the Loss of Creativity?
**If you are looking to purchase a new Tesla Car, Solar roof, Solar tiles or PowerWall, just click this link to get up to $500 off! www.tesla.com/referral/john11286. Thank you!
Join this channel to get access to perks:
/ @drknowitallknows
Thanks to Soylent for sponsoring today's video!
**To become part of our Patreon team, help support the channel, and get awesome perks, check out our Patreon site here: / drknowitallknows . Thanks for your support!
Get The Elon Musk Mission (I've got two chapters in it) here:
Paperback: amzn.to/3TQXV9g
Kindle: amzn.to/3U7f7Hr!
**Want some awesome Dr. Know-it-all merch, including the YEAR OF EMBODIED AI Shirt? Check out our awesome Merch store: drknowitall.itemorder.com/sale
For a limited time, use the code "Knows2021" to get 20% off your entire order!
**Check out Artimatic: www.artimatic.io
**You can help support this channel with one click! We have an Amazon Affiliate link in several countries. If you click the link for your country, anything you buy from Amazon in the next several hours gives us a small commission, and costs you nothing. Thank you!
* USA: amzn.to/39n5mPH
* Germany: amzn.to/2XbdxJi
* United Kingdom: amzn.to/3hGlzTR
* France: amzn.to/2KRAwXh
* Spain: amzn.to/3hJYYFV
**What do we use to shoot our videos?
-Sony alpha a7 III: amzn.to/3czV2XJ
--and lens: amzn.to/3aujOqE
-Feelworld portable field monitor: amzn.to/38yf2ah
-Neewer compact desk tripod: amzn.to/3l8yrUk
-Glidegear teleprompter: amzn.to/3rJeFkP
-Neewer dimmable LED lights: amzn.to/3qAg3oF
-Rode Wireless Go II Lavalier microphones: amzn.to/3eC9jUZ
-Rode NT USB+ Studio Microphone: amzn.to/3U65Q3w
-Focusrite Scarlette 2i2 audio interface: amzn.to/3l8vqDu
-Studio soundproofing tiles: amzn.to/3rFUtQU
-Sony MDR-7506 Professional Headphones: amzn.to/2OoDdBd
-Apple M1 Max Studio: amzn.to/3GfxPYY
-Apple M1 MacBook Pro: amzn.to/3wPYV1D
-Docking Station for MacBook: amzn.to/3yIhc1S
-Philips Brilliance 4K Docking Monitor: amzn.to/3xwSKAb
-Sabrent 8TB SSD drive: amzn.to/3rhSxQM
-DJI Mavic Mini Drone: amzn.to/2OnHCEw
-GoPro Hero 9 Black action camera: amzn.to/3vgVMrH
-GoPro Max 360 camera: amzn.to/3nORGYk
-Tesla phone mount: amzn.to/3U92fl9
-Suction car mount for camera: amzn.to/3tcUfRK
-Extender Rod for car mount camera: amzn.to/3wHQXsw
**Here are a few products we've found really fun and/or useful:
-NeoCharge Dryer/EV charger splitter: amzn.to/39UcKWx
-Lift pucks for your Tesla: amzn.to/3vJF3iB
-Emergency tire fill and repair kit: amzn.to/3vMkL8d
-CO2 Monitor: amzn.to/3PsQRh2
-Camping mattress for your Tesla model S/3/X/Y: amzn.to/3m7ffef
**Music by Zenlee. Check out his amazing music on instagram -@zenlee_music
or TH-cam - / @zenlee_music
Tesla Stock: TSLA
**EVANNEX
Check out the Evannex web site: evannex.com/
If you use my discount code, KnowsEVs, you get $10 off any order over $100!
**For business inquiries, please email me here: DrKnowItAllKnows@gmail.com
Twitter: / drknowitall16
Also on Twitter: @Tesla_UnPR: / tesla_un
Instagram: @drknowitallknows
**Want some outdoorsy videos? Check out Whole Nuts and Donuts: / @wholenutsanddonuts5741
AI Grid Video: • Sam Altman Reveals AGI...
Dave Lee and James Douma video: • FSD v12: Tesla's Auton... - บันเทิง
The first 500 people to use this link and code DRKNOW will get 25% off their first subscription with Soylent: bit.ly/3vNiahw.
I don't understand why anyone would want to eat anything with the name Soylent attached to it what is this‽
“Creative driving” is the perfect phrasesfor FSD
Presumably, the simulated data that Tesla uses is incorporating some randomization term(s) (think, Monte Carlo simulation) that avoid the regression to the mean problem that you are highlighting. My experiences as a science journal editor suggest that this is a common methodology for using simulated data in NN training. And, that randomization could appear both in the synthetic driving worlds Tesla uses, as well as in the possible human driver responses. So, while you are addressing an important problem, there are relatively straighforward fixes IMHO.
Spot on! Looking at some videos of FSD intervention, I often come to the conclusion that the human driver was just being inpatient (or being a control freak) and didn’t let FSD to finish its planned operation.
If FSD is heading for a collision, then an intervention is warranted.
This was a great visualization for what I have been experiencing with the current state of the LLM’s today. Many decades ago I wrote complex estimating models that allowed a major aerospace company to project the impact on labor hours in various disciplines as a result of the changes in the quantity mix of the production aircraft in their extensive portfolio and was extremely aware of this phenomenon in my models. I had to review all of the historical labor data for each aircraft and try to determine what caused the significant outliers and either incorporate logic to account for repeatable occurrences or determine if they should be ignored as one off situations. I do get a chuckle when I find myself arguing with an AI when it produces a result that I know is not complete and then hold it’s hand to guide it to include some pertinent outlier information. Your discussion was on point and needs to be understood better by the general public when they use these impressive tools. They are a good starting point, but there is a large amount of missing information that is needed to fill in the finer details. Thanks for producing the great visualizations to illuminate this phenomenon.
Dr. Know it all. I disagree with the main gist of the video. In the audio presentation the speaker states “Less richer understanding of the world.” My question to him is: Based on what? Who’s understanding? Also what does he mean by "richer?"
The large quantity of data available to AI fosters the discovery of emerging properties because quantity has a quality of its own.
Another point is aren’t "Universal function approximators" also a description of "induction reasoning?" For instance when Johannes Kepler stated the laws of planetary motion it was based on data collected by observations from his predecessors. The “laws” were embedded in the data and Kepler was able to stand back, squint, look, and formulate his laws. This sounds exactly what you are describing with your normal distribution curve.
Also what we seem to be forgetting is that the normal distribution curve changes depending on the prompt. Meaning an exoteric question will reach low probability regions of a distribution shifted from one based on a common question. The speaker said it himself when he stated that he asked AI about the exoteric items and received a low probability reply.
That implies that creativity is now transferred to prompt composition. That translates to knowing what kind of questions to ask. If the question is creative so will be the answer. I have not tried it personally but if one created a prompt for Chat GPT to compose a poem in the style of Wallace Stevens (a poet specifically focused on creativity) using uncommon vocabulary, rhythm, and rhyme. I wonder how creative the poem would be.
As for artistic expression that is now transferred to prompt composition instead of physical brushstrokes, or the glide of the quill over blank pages.
Interesting rebuttal?, but you should get someone to proof-read it for you. It's full of sumple errors.
@@gregbailey45 LOL, my SUMPLE errors are, as the article puts it "beautiful deviations from the normal curve." I started to study English when I was 14 years old and I still have problems with orthography especially prepositional phrases. I just pasted it to google docs and fixed what it found. But thanks for the advice.
Any interaction with the real world will create new real data.
It is matter of recognizing the creative parts
I feel investors should focus on under-the-radar stocks, considering the current rollercoaster nature of the stock market, Because 35% of my $270k portfolio comprises plummeting stocks that were once revered. I don't know where to go here out of devastation.
The safest approach I feel to tackle it is to diversify investments. By spreading investments across different asset classes, like bonds, real estate, and international stocks, they can reduce the impact of a market meltdown.
@@LeandraHarris-db1gv It's often true that people underestimate the importance of financial advisors until they feel the negative effects of emotional decision-making. I remember a few summers ago, after a tough divorce, when I needed a boost for my struggling business. I researched and found a licensed advisor who diligently helped grow my reserves despite inflation. Consequently, my reserves increased from $275k to around $750k.
@@kalenebarinaenwi2417 This is considerable! think you could suggest any professionals/advisors? I'm in dire need of proper portfolio allocation.
@@KatheyWilliams-kh3wr My CFA ’ is “Victoria Carmen Santaella”, a renowned figure in her line of work. I recommend researching her credentials further. She has many years of experience and is a valuable resource for anyone looking to navigate the financial market.
@@kalenebarinaenwi2417 Appreciate this recommendation, hopefully, I can get some insight into where the market is headed and strategies to beat the downtrend when I hear back from Victoria Carmen Santaella.
Re Tesla, they have an advantage in this area which I think you're overlooking. Tesla has perfect information about what data is from humans, what's synthetic and what comes out of their AI. They can choose to train on these three different datasets differently, including not training on AI data at all. They can avoid model collapse this way.
No other company that I can think of has perfect information in this sense. OpenAI gets its text data online. That is going to be increasingly AI generated but they don't have perfect information about what is generated and what is not. That's why they are much more susceptible to model collapse.
Great discussion. I like knowing where knowledge and creativity meet.
great points john, thanks for the new info
Yes, another way of describing it is that LLM follows human herd mentality.
One person in 1,000,000,000 may have the correct answer. Yet given the 999,999,999 incorrect perceptions, then that's what the AI model will present as reality.
We need to build AIs that not only learn what most people think is true, but also apply some logic and testing to determine if it makes sense for that to be true or not, applying the scientific method to do so.
Another example: Einstein was right and the rest of humanity was wrong.
Or, Galileo was right and the Church was wrong. etc.
@@stevenschmidt True.
Unfortunately, some scientists have a better understanding of logic than others. In the climate change debate, and modelling, for instance.
Perhaps going to first principles logic would help AI get it right.
@@stevenschmidt You're describing Artificial General Intelligence and it seems like we're still a ways off from that.
This reminds me of a sci-fi comic book I read when I was a kid (several decades ago)... It was about a man who built a time machine and travelled to the future. There were autonomous cars, but they could only go to places on a menu. Food was all synthetic. I recall the traveler longed for a real steak and finally returned to his own time appalled with how there was no emotion or surprise possible in that future. Hopefully that's not where we're heading!
@DrKnowitallKnows many thanks for introducing the Soylent meal replacement products. I’ve been looking for an alternative to real food and the bottle of scientifically designed food replacement must be much better than old fashioned home made food.
Most of the Tesla TH-camrs seem to educate that AG1 is better than food, but your speech on Soylent convinced me that food is outdated and a plastic bottle of stuff with a scientific formula printed on it has to be better.
Human evolution is so slow !!! so I’m really happy that you have taken the time to research which sponsors have the best product rather than taking the easy route and following the crowd with AG1 !!
Thank you, you really do live up to your channel name. 😅
The theory you present is very intriguing Sir. I suspect you are correct that there will turn out to be a fine line between AI and boring/arrogant AI. Very interesting times indeed.
And what can we expect when AGI becomes sentient?
Perhaps AGI will always be like a savant.
The same correlation can be made with FSD. For FSD to do more complex driving tasks requires more complex decisions and therefore more compute ability. More compute ability will require more energy and ultimately be more expensive. Trade-off is a FSD that gets you where you want to go safely and hopefully in a timely manner. Maybe setting up a dedicated computer in each individual car will improve complex decision making when encountering novel situations. Each individual computer can then be programmed to easily store all necessary local driving information for rapid accessibility. Again, cost and maintenance for such systems may be prohibitive for general users.
Thats an excellent point!
ok, the video was kinda entertaining and interesting, but also made 2-3 assumptions regarding to Tesla's FSD that I don't think are correct.
First one is that they don't filter for FSD driven miles and human driven miles, which I would imagine they do.
Second one is that they don't filter for incorrect interventions where the driver wasn't right. We have pretty little knowledge about the whole FSD development process so we should assume it's quite complex and probably there is a lot of thought put into it. I would also imagine there's an AI labeling for data collected and then probably even human labeling in some kind to put different interventions into categories and stuff. For example, they probably categorized data, so that they can filter out the drivers that do not stop at stop signs in order to be able to train FSD to stop at stop signs.
To make FSD better, knowing future frames is key. Test with 10s of data, while knowing the next 50s. Once FSD can make decisions that take small cues into account, we will think it is magic.
I'm SO relieved to learn that Soylent is PLANTS and NOT PEOPLE!!! If I do try this product, I will definitely be staying away from any green colored bottles.
Why is Dave and James on the cover?
I think the wave form distortion you showed (Nyquist Theorem discussion) does get corrected if you limit the original wave form reconstruction to 1/2 the sample frequency.
I'll have to double-check that, I'm fairly confident it's true. There is an increased noise floor, as a result of decreasing the sample frequency.
I wonder if this is what happens to human minds as we grow up and get older. Reality gets converted into simpler categories that ignore the details. Meditation brings back the uniqueness of each object or situation.
When FSD gets to a point that it flip flops with me I’ll be very impressed 😏
I wonder if Mew is up to the challenge.
I am so interested in learning about what we loose when we advance, and how such loss maybe undesirable. An entire way of life was lost to time when the fens of Lincolnshire was drained and sea became land. An advancements economically, a disaster for those who knew only said lifestyle. They attempted to destroy the pumps, that kept the fens land and not sea, and so became not only displaced but criminals. Thank you for your shared insight and timely contribution. I see connection between the fens then, and the steps we now take to build our Babylon.
"What we loose" is NOT loss! It's a freeing of the spirit, literally.
Progress often causes the loss of something.
This was a wonderfully articulated and illustrated presentation. It sharpened for me the idea of "humans in the loop," which should be a grounding principle for the ongoing development of AI. Thank you!
Hopelessly naive you are.
One way to look at what you are saying is that there is hope for humanity after all, that we sill still have a place in creative endeavors, even as the more plain everyday activities get done by intelligent machines. It may be that only the best people who are truly creative can hit the fine points that intelligent machines miss out on but the teamwork between creative humans and intelligent machines might lead us to higher highs than we could have experienced without the machines we are creating. Let the machines do the drudgery and the mundane, they will probably be far more consistent than people who get bored and stop paying attention anyway. Let people shine in the outliers!
Although I haven’t read/listened to the paper referenced in this video, I do have several questions based on this presentation:
What is different about the way “meat computers” in our heads approach problem solving and creativity vs. the (composition of available materials from our planet energized by electrical energy) computers we use to create A.I.?
Is it a difference in the method of handling directed chaos involved in thinking through a problem or task before a regression is applied to end up with a beautiful painting or song or innovative analysis? After all, the data outside a standard deviation sort may be interesting as a footnote but terrible to be included in a “finished” product.
Perhaps some A.I. based product such as FSD will never be capable of operating independently if it’s expected to perfectly mimic a human. As an example, I’ve wrenched control away from FSD many times because of a difference of opinion. It turns out action proposed by the A.I. (FSD) was more correct than the one decided upon by my personal “meat computer”. FSD was also wrong in a diminishing number of other instances.
I propose for FSD to incorporate winter rallye driving data into the dataset. Maybe get some cybertrucks on the Paris-Dakar, as well. This is where the highly creative magic of driving happens! :-)
But more to the point of the video: what's the AI-training equivalent to sample rate in music digitization? Token amount? And what's the limiting factor for having an LLM being able to keep the noise "in mind"? Total compute capacity? Just time? It feels like some clever math could help out a great deal, here. I mean: AlphaGo was creative, right? How was it able to serve up move 37 and not get stuck in some bland middle of the road position?
Soylent is the sponsor? Are they environmentally conscious? Are they green?
Is there a way to do oversampling in AI? Like in former CD days? Like 192 kHz instead of 41,1 kHz to be nearer at reality.
192kHz isn't closer to reality at all. You don't need more than 2 samples per sine wave. It's a myth that the more samples you have, the closer you approximate the signal. 2 points per sine is all you need. The reason we oversample is because it's very difficult to make a perfect anti-aliasing filter, it's expensive and has distortions. By oversampling, you can relax those filter requirements, you can use a filter with less distortions, and the equipment is cheaper. With expensive components, you only need 2 samples, with cheap components, you get less noise by oversampling, but only because electronics are imperfect. It's impossible to create a perfect filter that passes everything till frequency f, and nothing after that. That's why we always sample at least a little higher.
Reminds me of something Robert Pirsig calls Dynamic Quality in his book "Lila - an inquiry into morals". (Recommended reading.) He also talks about the levels of Static Quality. The top most level is the Intellectual level. The top most level can control and dominate the lower levels. But, Pirsig didn't consider AI, and that is very troubling to me.
Creative driving may have to access another plane to avoid an accident. Consider it to be an "emergency blimp mode".
Can you not track closer to the original curve by adding compute, and therefore higher data resolution? That sounds like the way to get almost all the way there, but yes, I agree - you will always need real data, human generated, as the wild card in the mix.
I recall all the noise around alpha-go was that at some point in the match it made an astonishingly creative move that no one saw coming. Ultimately AI needs to be embodied to be able to experiment with new and creative or random ideas in the real world, but I don't think anyone really wants FSD to be doing that :) for it, boring is probably good 99.999999% of the time.
Don’t, forget that the AI played millions of games, each with different moves that seemed logical at the time. So, the “very creative move” was just a mundane choice to the computer.
It may be possible to push the generative part of machine intelligence to include the outliers. To tune it so that it doesn't just smooth out the curve. Consider the generation of new data similar to what humans do when they brainstorm, just take any idea that comes up and save it no matter how off beat (or how far from the normative curve) it might be. Only after generating possibilities, then we go back and cull the crazy stuff. Or run the simulations and verify the results. Perhaps we could do something similar with intelligent machines when we are looking for unusual ideas or situations.
What a fascinating subject. It immediately made me think of genetic inbreeding. Especially the stuff about 1) losing the outliers, the long tail, the “messy noise”, and 2) the problem of the AI feeding itself its own synthetic data.
That’s functionally equivalent to an isolated population losing genetic diversity and getting worse each generation because they’re mating with close relatives. Your comment towards the end about the need to keep humans in the loop specifically to inject “real data” is exactly how genetic pools remain healthy, by at least occasionally iacquiring breeding individuals from genetically disparate populations.
Maybe that’s the solution to the Fermi paradox: every civilization becomes fat, lazy, and stupid once it creates AI, and loses the ability to venture out into the universe.
…loses interest in venturing into the “unknown”. There’s a big snowball of that coming from “conservative” politicians today even without the presence of an A.I. Just look at all the folks in the USA who wish for a dictatorship rather than accept responsibility as a member of a democracy! Let “George” do it.
Academic funding does a similar clipping of extreme ideas
I would suggest looking into the art of racing in order to keep human creativity in your feedback loop once driving becomes illegal.
I predicted this problem years ago and called it "AI in-breeding".
This is why Star Trek's Data couldn't be funny. Jokes depend on the unexpected. On the other hand, what about the program that developed moves in Go that no one had ever thought of.
As long as AI algorithms are interpolative (decisions are made from data) AI would be less effective. There will always are the edges that are unknown, and thus there would be no solutions.
Soylent looks good, but does it come in green?
You bet it does
Please sell t-shirts that read, "Keep it noisy" with data points off the curve.
Hilarious sponsor name! I prefer Soylent Green
I don't think that I would ever drink or eat anything call solient or solient green. The meaning of solient green from the book isn't something to be discarded or forgotten. Terrible idea... LOL
Does anyone know where they get the raw materials?
@@richardrigling4906some people know 😂
@@PJWey You are assuming there is knowledge after death 😅
@@richardrigling4906In a soybean processing factory. Basically a large chemistry set. You don't really want to know. 🙂
Also as we have more robotaxis we have less human data to mine..
Right this very second I bet DrKnowItAll is geeking the f-out over the new Atlas unveiling. I just rewound it 12 times to pixel peep it. Looks like there's a new robot chassis looking for an AI. Wow... things are getting insane.
Humans sometimes naturally strive for collapse in abilities because too much help taken for granted will weaken and cause more dependency. It's up to each human therefore to have wisdom in focusing to maintain internalized strengths and attributes, while harnessing external tools or implementations at the appropriate time. This was already pointed out way back 1969 when Zager and Evans released their song "In the Year 2525..." 🤔👍
I think it is helpful to remember that we are "training a brain" by establishing neuron connections and the "trained brain" will have emergent behaviors. Also, I don't want my car to try to creatively go up on two wheels to thread through a narrow gap...
fsd is still a speculative ... forget it for now ... like chatgpt it may make things up and/or hallucinate ... energy , cars, chargers etc is where it is at ... fsd supervised is still okay though for general purposes ...
While we are in a geeky mood, what if 8/8 date is a reference to 8x8 as in a matrix?
There are 8x8 POE switches which can switch camera video from eight different cameras (and Tesla cars famously have 8 cameras). Perhaps Elon is enamoured with an 8x8 crossbar POE switch? But why???
Maybe the 8x8 switch is for Optimus which allows it to "focus" its processing power on one of its eight cameras.
Creative driving is most useful when evading tow trucks or police. Maybe Tesla can train on that data?
I’m at about 11:00 - you don’t need to generate synthetic data - I don’t like synthetic data - it is like having a s/w developer write all the tests for his own s/w - he will only test what he has already anticipated in his s/w when it is the unanticipated that is important. You simply have to train another neural net to filter out all the humdrum stuff from your realtime data feed and just keep the interesting edge case stuff to be fed into your FSD training. Where humans are miscorrecting the car there is probably not enough consistency in their behaviour to actually affect the outcome of the training. If you want more edge cases, then put The Stig into a real car and have him drive through the Alps or across Africa or in a Demolition Derby etc…
I like the idea of a Tesla RoboTaxi in a demolition derby ! Would make a great advert if Tesla advertised - would create some interesting training data too
watched til the end - I think we arrived at same conclusion via different routes - You do need to increase the amount of edge-case-related data in your training corpus probably to a disproportionate level, just as you would read poetry not prose if you wanted to write poetry. The question is, how do you do that effectively. Perhaps you do have to go out and actively look for the edge cases. You could try the Stig-in-a-Demolition-Derby approach above but that is a bit scatter-gun. How about looking at all the serious car accidents that have been recorded in recent history, categorising them, reconstructing them physically where possible or virtually where not and then training on this data until your Tesla, put in the same situation, can save (if there is any way humanly possible) its occupants ? Once again - if you could say that in 90% of fatal accidents in the last 10 years, if you had been in a Tesla-FSD you would have survived - this would make great advertising.
Can't you ask an LLM to give you three answers to a request that represent the three areas of the bell curve, common, uncommon, and extreme outliers?
So are we all going to get elevator music all the time if we rely on bots? - AND - we must also consider that if (or when) the true data IS A SINE WAVE, we don't want or need the noise of imperfect measurement or imperfect duplication.
Have done Soylent on and off for years. The liquid bottles are great but expensive af. The powder is relatively cheap on monthly basis but it has to be mixed in blender to get the liquid bottle quality. Shaker bottles just don’t do it. I didn’t realize the powder could even be as good till I recently put it in a blender lol
Many of the powers need time to actually become fully hydrated. I've found that most of them are best if blended and then allowed to sit for some time and then reblended. Some of my drinks I blend a few ahead and so can have one that I can just reblended or shake and get a much better drink from.
@@kstaxman2 this. A blender/shaker bottle will accelerate the process but 30 minutes in the fridge makes a pretty big difference. Powder is good for up to two days (refrigerated) after mixing so mixing your breakfast/lunch the night before is easy to do.
Interesting. I have been using stable diffusion to create art works and some of the results are amazingly creative but I have noticed there is something missing as compared to analog art. What is missing is imperfections and noise. Makes sense in the context of function approximation...
Same for people suddenly realizing that CDs aren't as good as vinyl records at catching the nuance of music.
Interesting problem Doc. If the AI is really just a super expensive and complex parrot mimicking human thoughts and behavior but not actually "understanding" anything, then I can see this being a problem. However if the computer achieves AGI /consciousness then it should be able to notice all data and experiences and incorporate those tokens of reality creatively similar to humans. Is it possible that this ability becomes the definition we have been searching for, of "consciousness"?
Ironically this TH-cam platform does the same thing, it shows me stuff it thinks I would like. Giving me a narrow choice based on my previous picks. Choosing from these narrow choices, narrows it even further. Netflix does the same. This is why I like thrift stores because the choice of goods or clothes is so varied. Look into any other men's store and your choice of clothes is black, navy, or grey. It is this dumbed down, same old rotating hamster tread wheel we are all led towards. No spice or spark of imagination. It would be a shame if Ai did the same thing.
Smart content ,Thank you DR.
Throughout history, humans have explored. That’s how we grow our knowledge, and capabilities. There are many specialized field of information, but there have always been those who explore new things. Now that AI seems to be honing in on the average, we need to make an AI that will explore.
"while you're down there ..." is a bit too intimate, to my taste
The training data curators bias the AI
It sounds like AI tends to converge on conventional wisdom, which is often wrong.
AI verses RI….I think I would rather have abstract RI. I agree we should careful. A copy of a copy of a copy. End result: background radiation…Static…lol
Badge the AI with what level of education or real-world it can perform! Like this AI can only give you a college-level education, but not graduate school
I use AI all day long in my cryptology research and TH-cam postings.
(I just used AI to clean up the above sentence)
My son an IE/MBA uses AI now constantly in his design of production/manufacturing facilities
from investigating a client's needs
to design of the assembly lines
to quote the project
to implement project management
to run the facility
He figures that the above are easily worth 3 Industrial Engineers and 2 assistants
and he pays a total of about $50/mo for the AI software
All the above are available to him and senior management of the client in the form of a LLM
where they can ask all kinds of questions about their new facility of AI, and he doesn't have to
spend a second supporting 90% of their questions and concerns.
And he has only been using AI for a year now.
He probably represents just 1% of how AI will eventually change every business model.
I think purely synthetic data is a bad idea past some point of the FSD being good enough and having sufficient market penetration.
If you took all of the data from FSD driving issues and edge cases and then had the AI brute force trials of that data to find the optimal /alternative solutions to real world data; then I have to believe the eventual outcome will be better than throwing artificial spaghetti at the wall. Synthetic data is not real world data and the rubber meets the road in the real world., the real world is where you want your FSD to be at it's best in. At some point you also need to say good enough before you start washing out the goodness.
But can’t you keep adding the bizarre situations to the training dataset? Or is that too simplistic? I mean in the context of FSD.
AI will need to graduate from the training wheels of imitation learning to RL with real-world feedback. This is what everyone is working on right now, including Tesla. They wouldn't say FSD will become superhuman if they didn't see a path to removing humans from the loop.
Great paper I have to read it.
Humans are great at handling the daily chaos, and that's also the reason why I love psychology. I appreciate the unpredictability of human behavior; yes, of course, there are also normally distributed behaviors, but you cannot completely predict behavior; there will always be extreme samples in every study. Math makes the world artificially predictable, which is necessary for technology and AI because AI is still not a complex living being that knows more states than 0 and 1. However, with the current architecture, AI will only possess a normally distributed intelligence, which cannot produce something truly new. Therefore, it is not creative. Its more like a illusion of creativity . Additionally, any art produced by AI would be a normally distributed new copy of art that humans have created.
FSD 12 will become more and more predictable as it improves, which is a good thing because the extreme behaviors of us humans are what cause accidents.
So, how do we synthesize creativity? How do we invent AI humor? Without creativity, will AI ever be self aware?
I don't think AI will make people less useful or intelligent. The definition of work will be more inclusive. Work should benefit humanity wether you knit caps or go to Mars. Everyone lives well because robots do the work we don't want to. Jeannine
Let's be real here !
FSD still is terrible in snow and Ice and allthough it's pretty good in rain, it still has troubles in that kind of weather !
Teslas allthough having a bunch of cameras, still have blind spots, so FSD can't 100% reliably reverse without human supervision .
The lack of radar sensors don't help either !
It still can't read all the road signs!
There is more to road sign recognition than beeing able to read speed limit signs !
Robo Taxis for sure will not be real by 8th of August.
Well to be fair, Elon never mentioned the year.....😂
Whatever you do, don't get the new Soylent Green. Watch the movie instead.
JPEG compression
The solution is bloody obvious, multimodal. Use data OTHER than text. Why is this even a question?
Instead of 'keeping humans in the loop', wouldn't it be equally valuable to keep AI's out of the loop? Scraping AI-generated pablum is similar to modeling FSD after average drivers... without the stochastic failures like distracted driving, drunk driving, falling asleep, etc. End product would be a world of Sunday morning, 'little old lady/man' traffic which would certainly be maddening. Internet content is already mind-numbingly stupid -- on average -- so using it as training data is fraught. This is 'garbage in, garbage out' to the nth degree.
Isn't continuous reprocessing all current data fundamentally incestuous?
It seems the next step will have to be providing AI with an independant sour e for new data.
Keep in mind that humans only "learn" when confronted with new information that contradicts our current knowledge base and world model
Quantum behavior and reletivity being two examples
This is a long ways away. The reason Ai will advance faster than humans is because of it compiling our collective knowledge. Once this is done AND fully exploited by humans, which will birth new creativity in the process, then we will have this problem.
Is there no meaningful 2D information on a plane of the sine wave data? Say you’re a little 2D creature on a sheet of paper walking the path of this sine wave and your crazy 2D buddy takes off to your left and right along the raw data. Your friend can take off in any forward direction (wrt the sine wave) but the only rule is that you must cross paths at the same location/time along the sine wave. So with this, there are so many parameters to play with: constant speed along sine wave (forcing variable speed on raw path), constant speed on raw path (forcing variable speed on sine wave), directionality of chosen raw path, distance from sine wave, distance travelled, etc. Is there a math term for this? Can’t we learn the “personality” of this staggered behaviour? Is there not a finite number of raw paths that can create the given sine wave exactly? This might be a key to developing a controlled noisy behaviour that pushes past human mimicry.
All that just to say AI science can’t do what we spiritual beings do creating art and outliers.
So AI is normalizing humans. I think that will bridge gaps where there is a lack of consensus but at the same time increasing the value or impact of uniqueness and individuality.
You are correct and honest. I appreciate that.
Seems just like what Answers with Joe in reusing Adobe fill over and over quickly becomes wildly boring
This problem is likely a consequence of Godel's Incompleteness theorem. Humans will always need to be involved in extreme edge cases.
Disagreed. There's more and more real world data from the fleet that is fed to the system, there is no inevitable need to heavily rely on fake/synthetic data.
Genius individuals (say IQ over 140) have the same problem - they are usually trained by teachers in the 100 to 120 IQ range... thw "garbage in - garbage out" problem demands filters to focus training to bias toward the genius side of the curve (and yet recgonize that most encounters will be in the salmon colored area)...
I undertand the scalability is a problem but you re only using humans as ground truth with shadow mode. Not sure how the disengagements work, but i m sure they d have to check, either manually or with some mindblowing pipeline, to know if the problem is to be considered or not.
Point is, you don t feed video of humans driving (of course you do but), you simply feed as perfect driving as you can, who cares who s driving? This means the data collection inside the system is not memeking at all humans it s memeking a Tesla created perfect driver that doesn t actually exist.
How long would you guess each clip that s fed would be? I bet the avg one is shorter then 10 seconds.
Maybe knowledge collapse for AI but not for humans, who in my world still have their own brains and will use them.
I am rooting for humans and humanity, I do not give a shit about AI. Humans are the problem and we humans are also the solution.
This not a difficult problem or solution to understand.
I think of our current way of training eg Fsd is that the computer looks at the vision and then goes does what statists predicts a driver would do.
The ultimate would be for us to have a model of reality that we could call ai and the computer can check with it's understanding of reality to decide what to do.
So when Fsd is trained in future we are explicitly updating fsds understanding of reality.
Would be nice if we could actually read the reality that Fsd has for its understanding for driving.
Guess this does sound like writing the code in C which we have just abandoned😅
So you're saying there will be a place for humans in ten years.Well, that's good news.👍😄
AI may become statistically dumb and dumber
The drink is based on soy? I hope not...
LLMs have plateaued. There will be nothing new in the next 2 years and we will also not have fsd.
I totally agree! FSD will always need human input instead of trying to improve itself with a dumbed-down synthetic feedback loop. After three years of using FSD, it is improving in average driving situations, especially with version 12 but not in odd ball edge cases. A big area where FSD could rapidly improve with human input is allowing a driver to program the car's path for unmapped areas. How come we don't have an on-screen function to correct navigation errors?
Just add noise back in when processing the tokens.
If only it were that simple... LOL most noise is unwanted and not worth using. So adding all noise back in would only weaken what is usually the best selection. The issue is how do you get the useful outlying information separated from the noise that's useless. If you add in all the noise you get a mess.
AI Idiocracy
I have noticed this phenomenon in writing. I try to avoid sequential reuse of the same verb, noun, or modifier in my writing specifically to keep it lively and interesting for my reader. That means I frequently make use of terms that are uncommon in AI-generated text. I try not to go overboard because doggedly applying the Thesaurus to one's prose can be pretty annoying. However, I think I retain a slight edge over Chat GPT in terms of creative expression.
The creativity in driving comes exactly from edge cases. There is no need for it otherwise.
So you need creativity for solving edge cases, and you need edge cases to create the creativity!
Why the weird eyeballs all the time. Makes me pass you by...