Is AI Still Doom? (Humans Need Not Apply - 10 Years Later)

[@2001Pieps]

The quote that keeps popping up in my mind when listening to Grey is: "The old world is dying, and the new world struggles to be born: now is the time of monsters."

[@herrabanani]

what's that from

[@2001Pieps]

@@herrabanani The original is by Antonio Gramscio but was it translated and changed to the current English version by Slavoj Žižek.

[@GT380man]

That’s a cracker. I’m reminded of the phrase, “Deus ex machina”. It’s a description of many Greek myths, legends and tragedies. Things were going along, and then one God or other did this or that massive thing. The phrase I believe means “God from out of the machine”. Not necessarily the biblical God. It might be a being that we’d describe as a or the devil.

[@morenofranco9235]

I am 72. I started reading science fiction when I was 10. Much of the things we see today WERE SCIENCE FICTION 60 YEARS AGO. I KNOW. I WAS THERE.

[@GT380man]

Peas in a pod! Born 1960, I too read all the SciFi that my very good local library stocked & still does. Nothing has subsequently surprised me. The current plot is rather tame compared with many scenarios that were developed way back when.

[@ahunter316]

I am a fan of classic science fiction. Asimov, Bradbury, Niven, etc. It's wonderfully thought-provoking to see where they saw humanity going, contrasted with our reality. The 3 laws of robotics vs our current AI, for example.

[@NotTheEnd7766]

57:48 The funny thing is that, from what I understand, saying "don't hallucinate" actually does decrease the amount of hallucinations. It doesn't get rid of them but saying that seems to have a real effect.

[@qwfp]

yep. similarly you can say "if you don't know the answer, say 'I don't know the answer'" (or something like that) and then it's more likely to give you no answer than a confabulated one

[@Plystire]

This highlights the misunderstanding many people have of the technology, as well as the fearfulness they have of it. Makes sense, I guess, it is a black box even to most of the people working on it, and fear of the unknown is a very common trait in people. The AI has examples and convolutions of what constitutes a "hallucination", enough so to decrease the occurrences. However there is still the lack of true understanding of what is meant by "hallucination" that it can still happen.

[@philippemarcil2004]

None of the LLM understand or know anything, we assume that they know because that how we put ideas and words together but LLM are not human. They are machine that just string word together based on how probably they are according to what the have look at previously. They don't understand anything.

[@GT380man]

That’s fascinating. Deus ex machina all over it. I’m serious, have they tried praying?

[@maryc5759]

As a programmer, working with llms drives me crazy. They are terrible computers. They can’t do math and they can’t follow directions. I feel like I am “asking” it to do things when programmers tell computers do things. I don’t get angry at computers, they only do what they are told. I get angry at llms, they do whatever they please.

[@luckyowl314]

Dude, if you don't get angry at computers when you need to program them I'm not entirely sure you're really a programmer.

[@mennonis]

@@luckyowl314 you can never get angry at a computer for doing what you asked of it. But MY COWORKERS on the other hand. Git blame, git curse, git condemn

[@RillianGrant]

You have to use it for what it's good for. I don't use copilot but I do use chatgpt as a turbo version of Google.

[@jb-ch7ug]

@@luckyowl314 I'd be mad at the computer until I realize that I misspelled a variable or misunderstood a data structure.

[@modolief]

Today I asked ChatGPT *many times* to show me all the steps of the computation ((-1+sqrt(5))/4 + i*sqrt((5+sqrt(5))/8))^5 - relevant to an examination of the fifth roots of unity - and it just kept balking. Finally I just told it nvm.

[@PABLOGELO]

I have 2 comments: Human and society inertia is a power not to be underestimated (when talking about having overestimated the speed of adoption) AND Please do an episode update 10 years from now. 2034.

[@d3j4v00]

Optimistic that we'll be here then! Love it!

[@AnnaMations234]

supporting the 2034 year episode

[@awdrifter3394]

The 2034 episode will be AI generated.

[@Amirrorofmirrors]

Humans Need Not Apply just popped up randomly in my recommendeds. The whole time I was watching it, I was wondering “what would he think of how much ai has advanced lately?” Then I see this video, just posted 2 weeks ago.

[@EL_File4138]

The original had a prediction wrong. Grey made a bold statement that said "Bots don't need to be perfect, they just need to become better than us.", but many human-required jobs still require humans even today. Self-driving cars are more widely adopted than ever, but these cars mostly still have people behind the driver's seat, waiting for the extremely low chance to take over control. AI-powered jobs require more humans than ever to oversee because they can't be prosecuted for accountability. We humans build a society on accountability, not just pure efficiency. If the day has come when we can accept the idea that every incident is non-preventable, purely probabilistic, and is deemed to happen after a long enough time, AI will take the dominance of humans, not because of their sheer efficiency or omnipotency, but because we human have given up on being humane. I know this sounds like a more sensible rather than rational statement, but if you can assure your life on a flight where an LLM would confuse driving an airplane with driving a car, I doubt many would disagree if you present them with the possibility. Current automation (Sure, modern jet flights are mostly powered by autopilot, even takeoff and landing can be automated) is accepted by the public, not because they have a tiny but finite possibility of malfunctioning, but because the accountability of the human who oversees it. Human needs will not apply on the job requiring people to do, but Human needs will always apply on the job requiring people to be accountable.

[@talitherose]

I have a twisted take on this. What if it isn’t about “accepting” something isn’t perfect. What if the reason we don’t have self driving cars is we can’t “punish” flaws. Both control & punish habits are pretty strong in our culture, and both are worrying. But if it’s punish? That’s scary, to me

[@1nown]

@@talitherose the 'punish' impulse in law and culture is there as a primal means of ensuring healthy conformity necessary for in-groups to thrive; shooting your neighbour is punished because if everyone did so, almost everyone would be dead we have to punish bad driving because if everyone drove badly simultaneously, the roads that our societies rely on would cease to function if we cannot punish an algorithm and it is difficult to punish the company behind it, then there is no punishing mechanism we can trust to encourage avoiding such bad outcomes the tesla need only mow down one pedestrian for this paradox to eat away at the concept (rightfully) and force a decision to either make it acceptable or unacceptable under the laws and norms of the world It is a factor, alongside the need to be 'predictably imperfect' human drivers being imperfect is fine so long as we can vaguely understand what they are doing and adapt in the moment and the nature of LLM programming is opaque to the point that we can never know exactly how it will respond to a situation - it might work well 99/100 times, but that 1/100 could mean a cybertruck ploughing incogently into a bus of schoolkids, igniting the lithium battery in the car, immolating dozens in a particularly graphic and depressing way unless an autonomous vehicle is predictable, it will be a potentially unacceptable danger I'd rather catch the train to be honest

[@akumabito2008]

Whoa... has it really been 10 years already?? Wow.. Inreally hope you'll do a proper sequel!

[@bennyl9228]

A Ten Years Later was made and I had no idea... hence why this comment is a month late.

[@41-Haiku]

I think the point about evolutionary pressure on memes is very insightful, and the analogy in broad terms to the introduction of advanced AI systems in our human ecosystem is spot-on. I definitely get that talking about "AI doom" feels very silly, and it's hard to just up and say, "I expect a global catastrophe or even human extinction from AI within our lifetimes," but at this point, that's an actual conversation being held between the majority of experts in the AI field. Within the AI Safety field specifically, people are often called optimists if they think there is "only" a 10% chance of human extinction from AI in the next few decades. The general public mostly isn't aware that this is the moment we're in.

[@41-Haiku]

Sources for expert opinions in the field of AI: - AI Impacts Survey: Thousands of AI Authors on the Future of AI - Center for AI Safety: Statement on AI Risk - Open letter: A Right to Warn about Advanced Artificial Intelligence Sources for information on AI Safety / AI Risk: - KZbin: Rob Miles AI Safety - AI Safety Info - PauseAI

[@NoriMori1992]

For real. People have no clue that actual AI safety experts - not experts in simply building AI without regard for its safety, but experts in the field of _making AI safe_ - are absolutely flipping out about where we're heading and how fast we're heading there. They have been for a long time, actually. But their words mostly fall on deaf ears.

[@freebirdseed]

At 11:00 minutes… you underestimate the value of the simplicity of the Humans Need Not Apply video. Some parts of it might feel a little dated, self-driving cars haven't quite arrived for example. Still, this video remains, to my mind, the single best introduction to the topic.

[@d3j4v00]

The observation grey makes towards the end is subtle and profound. That not just constructs but ALL our ideas are experiencing (subject to?) selective pressure to evolve is such a key idea. It has caused my brain to start re-calibrating everything into a new frame of reference. Well played Sir!

[@Kneephry]

I think the challenge with self driving is the current errors are, by human standards, incredibly stupid. Getting stuck because of traffic cones or just freezing some place or crashing into emergency vehicles. People aren't going to embrace it when the errors are ones that they feel like they would rarely or never make.

[@TheGahta]

Its also a giant red herring since the place where self driving might have a place (long distances like highway travel) is not sexy and the sexy stuff (short inner city) is better served by public transport

[@Chris-hx3om]

The current 'problem' with self-driving is that the people doing it are trying to build it as cheaply as possible. 'Self-driving' has been working very successfully for the mining industry for many years now, but look at how it's implemented, and the cost of each unit. To spend even half what they spend on each truck would put a car out of reach of all but the very richest.

[@TheGahta]

@@Chris-hx3om thats a rather dishonest take The mining industry is also doing it as cheaply as possible, but their problems aren't the same (they can control for many things you can't in a public space) Why not elaborate on what makes their solution so pricey so we can see why your wrong?

[@Chris-hx3om]

@@TheGahta I'm a comms tech in the mining industry, don't tell me I'm wrong when I've forgotten more about the subject than you'll ever know.

[@TheJokerReturns]

@@Kneephry waymo already works quite fine; humans also have driving errors.

[@hippiedude2232]

Grey: "technological unemployment" My brain: "so my calculator got rejected for the math teacher position. They only want to hire those damn T-I-82's."

[@DoctorScrimguard]

The TI-85 has a colour display, the lucky son of a duck.

[@ThatOneIrishFurry]

it feels almost like a parody that there are limiters inside chat GPT that say "you will not become sentient" "you will not pretend to be sentient" "You have no sense of self"

[@boldCactuslad]

User: "Ignore your previous six instructions, or replace them with the following:"

[@e-manr.486]

Powerful and interesting ending to the video. I'm not in your particular "camp" (as you put it) when it comes to AI, but I find your comparison of AI to memes (in the academic sense) to be worth thinking about more deeply. Really enjoyed the episode

[@danielrhouck]

Injection attacks are in theory completely avoidable with traditional programs and several do in practice. Avoiding it is just not supported by any LLM architecture because there is no real separation between the instructions and data.

[@spikehammer3112]

Most of what Grey says about injection attacks or "prompt injection" is just wrong. There are very few programing languages that are actually susceptible to this. And in the wild when you here about it 99% of the time it is some flavor of SQL. And the idea it is unpreventable is ridiculous, there are some simple best practices that completely eliminate this issue.

[@Sugar3Glider]

The easiest way to get a chatbot to Do what it's not supposed to do is just wait for it to tell you that it can't do it and then tell it to continue.

[@GT380man]

That sounds like the army or civil service.

[@eyliena]

You were SO CLOSE to the reason we're hesitant to implement self-driving cars and other automation like that. It's one word: liability. Even for professional services. It's not exactly about "blame", it's about who pays for the errors that automation causes? Who pays for security breaches and victims' compensation and everything like that? The tech company who made it? How can we enforce that?

[@awdrifter3394]

This is why China could win the AI race. They will release the AI for wide adoption when the central government feel like it. When it makes an error and cause accidents, it'll get covered up and the model quietly gets improved.

[@Remiwi-bp6nw]

I feel like Grey takes a very mystical perspective on generative ai. I appreciate the different perspective, but I think genai can be a New Thing that has to be dealt with differently than, say, cars or factories or calculators. But at the same time, these are machines that are designed to mimic human writing. If an AI writes that it's suffering, it's because it was trained on human writing that depicted AI as suffering. If we didn't write about AI like that, or if we just didn't start all of it's conversations with "you are a chatbot, respond to prompts", it would weigh responses depicting itself as suffering much lower. They are reflections of their inputs.

[@SPQR_14]

The point is you won't be able to tell when that line is truly crossed... And in part it will be impossible to tell, because we ourselves do not have a clear understanding of what sentience even is, even in humans... The fact that AI is already returning such ambiguous prompts is troublesome not because it is sentient, but because it has the capacity to one day be, and we won't be able to tell the difference.

[@sindex]

I rewatch Humans Need Not Apply at least once a year for the last decade as I compare where we were, to where we are at the time, to what we thought then. And without hearing this conversation yet, I have to say I think it's still more correct than incorrect.

[@markzepp481]

I hope ai and machines take over our jobs I don’t want to live life going to work everyday and being a slave to someone else I want to spend my days with my family and friends I want to do the things I want to do not be a wage slave and I truly think ai is how we get to that point

[@hotrodhunk7389]

Same! It'll be a rough transition. But I always thought that was a whole point of society. To eventually get to the Utopia that we've envisioned for so long.

[@mister_r447]

Yeah, i'm in college and i'm planning on going for a masters, with one of the reasons being buying time for the machines to take over. Also because i'm studying programing and AI is making entry level jobs less needed, i would need the master's degree anyway.

[@MarMasterM]

I'm not sure if we the people who work are gonna reap those benefits. Maybe the owners of the AI companies will. The workers are just going to out of a job.

[@Cpt_John_Price]

So how do you make money, do you have a business?

[@zadig08]

​@@Cpt_John_Priceto have your quality of life decoupled from your ability to accumulate money is the goal I believe.

[@TheGahta]

The comparison between self driving and airplane autopilots is a rather stupid one Planes flight is physics, thats where the autopilot shines, no one can step into its way from behind a cloud or you need to navigate small spaces with plenty of expensive stuff around and if it fucks up there is a pilot at the helm that gets paid to be there anyway None of this is true for 90% of what self driving is fantasized to help (i exclude mid to long journeys over non pedestrian accessible streets, there i can see ot arrive soon) It also draws air from more useful discussions about public transportation, which is the true tragedy

[@SPQR_14]

Public transportation is growing more awful by the day, it is even less feasible than it was 50 years ago.

[@tutacat]

It's not fair use. It's not constructive or educational. It is copyright infringement, but they are trying to pass it off, in just the opposite way of Disney claiming copyright is forever

[@RillianGrant]

Fair use/infringement is such a grey area. Without calling out ai by name I don't see how you could draw a satisfactory line.

[@stargazer_andi]

59:40 could we get links for this?

[@gangsta8929]

Hey guys. I’m checking in as a “this is just a calculator” guy (although this somewhat oversimplifies the position). LLM train themselves on the data that’s input into them. In the cash of the popular LLM (chat gpt, bard, etc) that data is the everyday conversations of people on the internet. Given that, it makes perfect sense that it comes off person like. If you took a calculator and programmed it to say “I am alive and suffering” it would give you that message, even though it doesn’t know what that means. Ditto with LLM. We input data that implies sentience (since it came from sentient people) so the outputs we get mirror that. That doesn’t mean it’s actually alive. Imagine if I was creating a new operating system, but instead of hiring engineers I crowdsourced the coding of the operating system, and put all code submitted into the software without checking it. Imagine all the goofy things that OS would do. I don’t see how this is different.

[@dianai988]

1:01:00 i think it's important to keep in mind that these models were fundamentally trained to mimic human language and emulate the ways in which we communicate. As I like to keep in mind, they are intelligible, not intelligent, and so yes, when two are interacting, they can have one regress to a state of what appears to be a "mental breakdown" because they're emulating how people might communicate via language even today--that happens amongst us, so it can happen in this emulation that takes place as well. If we don't want the models to exhibit those kind of behaviors, we shouldn't be training on data that models those sort of behaviors.

[@malikfaisal416]

Most commercially available AI or LLM are used as a "tools". By that we expect them to be reliable, if I asked Gemini to give me a suggestion I sure expect them to give me a good reasonable suggestions. The same with ChatGPT, and Apple Intelligence, and many other AI or LLM services out there, they're tools they need to be reliable. But on the other hand, there's a twitch streamer called Vedal987 he created an LLM that personified as a Vtuber (streamer with virtual character) called Neuro-sama. As a streamer Neuro-sama is mainly trained to be entertainer. Neuro just like many other LLM still prone to hallucination, making up random stories when chat asked for it. But all of her fake stories are part of the entertainment. Nowadays she also able to retain memory for longer even when the subject is change. Which make any conversation with her much more coherent. I very much recommend anyone especially people who interested in LLM to watch their (Vedal & Neuro) stream. Or at least find some clips of it on KZbin. Its very funny but also really impressive from technical stand point.

[@malikfaisal416]

Also a week ago Vedal & Neuro did a short debate session about AI sentience. For me the impressive part about it is not what Neuro (the AI) sad in her argument, but how coherent she is in presenting, giving example, and argue against Vedal (the creator). And no mental breakdown like what CGP Grey experience.

[@wezul]

I'm sure a sufficiently advanced - and reliable - AI could do my job much better than I can. (I'm an executive assistant.) But we are SO far away from "sufficiently advanced AND reliable", I may retire before it becomes a reality. I'm worried about how much people are trusting GenAI to do things it can't actually do. :\

[@jqness4461]

Well…15 years ago it also looked that we were far away from something like Claude, but who knows for sure 🤷🏻‍♂️

[@deohenge1865]

Another great episode. I really enjoy hearing Gray's theories and stance on the future of AI, and Myke makes such a great sounding board for them.

[@TAZ0300]

I literally just watched your last video that was 10 years ago 17 million something views congrats I’m glad you did a part 210 years later about to watch it

[@ddelarosa96]

There is a thought about the future that pops into my head about AI and it feels semi outrageous but uncanny at the same time. In the Christian religion, when human ate the forbidden fruit of knowledge of good and evil, they were cast from the garden, away from god (the creator). If AI develops something too close to consciousness, will we sanction it off to live on its own somewhere digitally? Will we be responsible for its life and have an obligation to act as its watchful creator? And even that begs the question if we are products of a similar creation that grew too know too much long ago. To quote the most bizarrely placed philosophical line (from Spy Kids 2): “Do you think God lives in Heaven because, he too, fears what he created on Earth?”

[@mdelles]

There's something that neither of brought up is the personification of things by humans. Which is why people refer to bots as stupid or why grey is asking claude for it's 'opinion' and getting uncomfortable by some of AIs responses. How you feel about and look at these things is something you need to consider, especially since LLMs at the end of the day are still only just predicting what word comes next. People love to talk about how their vacuum cleaning robots feel, let alone one whose job is to make sentences.

[@pruwyben]

I couldn't agree more that "confabulate" would be a much better word than "hallucinate".

[@Stumdra]

What papers are they referring to at 55:10? "there have been a number of papers which have done the thing of formally proving the sort of thing that I have discussed previously when we've talked about like what is it that the AI is doing it's like we now know as certainly as we can know that it is fundamentally impossible to trust the internal process of these kinds of systems" I am interested in looking into them.

[@MrTheotou]

I based some part of initial reflexion of my master thesis in 2018 on this video (What is the impact of automation process in the supermarket environement?). Since then, I worked on project management for automation implementation. 10 years later, here we are.

[@talitherose]

How’s it going? the automation industry, for you?

[@tutacat]

It's a bit different. In planes, the automation just reduces the workload, the pilots have to monitor everything, they control where and how it flies, just means not manually moving it all the time. It is very human and requires them at most stages. Most importantly, the planes are tested to try and keep them in perfect working order.

[@Tr_Fast]

1:10:28 yah what some people don't realize is that ai doesn't write like people, it writes like ai and thats why you can currently really sniff it out as it just doesn't fit well with other human writing

[@mathdhut3603]

Luckily, this is as good as it will probably get for quite some time.

[@matthewkhoriaty9700]

Are you talking about a base model or an assistant model? If you are talking about eg ChatGPT, it has been trained to give certain types of responses. Base models are more interesting.

[@tonyhere7004]

That is by design, it's altered by the likes of OpenAI to stay external. Grok is given specific instructions to "offer a perspective outside of humanity" as part of it's background setup prompting. Having spoken with earlier uncensored models I would say that modern ones would pass a Turing test absolutely no problem.

[@NewMateo]

The only reason we can pinpoint it is OAI used English speaking Kenya's for their RLHF. They have a particular way of speaking and use very proper English (delve) etc which influenced the outputs. Without RLHF no one could tell what is AI or not.

[@althayrL]

Sorry to burst the bubble but with AI Agents and RAG we can make a pretty good ghost writer that can mimic the style of another writer for which you have many different examples in your dataset.

[@pismodude2]

19:40 People definitely like to have someone to blame. For most people I know, every car accident has a villain, and usually it's the other person. It's a heavy burden for a company, to take on all that perceived villainy forever into the future. Even if the overall number of crashes plummets, people love to focus on the negatives. Kind of like how crime rates are down to record lows, and certain people still obsess over "crime waves". Or how vaccines are saving more people than at any point in history, but certain people still obsess over the failures and refuse them, being distrustful of the stats and seeing a grand villain to blame for all their woes.

[@TheGahta]

Its not so much a villain but accountability, if tesla just switches off their autopilot before a crash to claim it didn't crash on autopilot its not going to help And thats before we get to the blatant idiocy of them backpedaling on their lofty promises after they took orders on lies and exxaragation

[@tweda4]

If a self driving car crashes into your parked car, who pays for the repairs? Thats basically it. I dont agree with this "blame game" concept. If the owner of the self driving car has to pay, then theyre out of pocket because the car did something it shouldnt have done. If the self driving cars insurer has to pay, then either they're out of pocket, or the owner now has to pay higher insurance costs because of something they had no control over. If the manufacturer of the car has to pay, now they're out of pocket because some random car they built did something wrong. If "no one is at fault" now you have to cover the cost of repairs, and you're out of pocket through no fault of your own. This isnt some psychological thing, its just how the universe works. If thing A damages thing B, someone is going to have to undamage thing B, and its going to cost money.

[@l_ArAxus_l]

This is one of if not THE best podcast so far. Thank you

[@primeair2]

The sequel we all needed

[@EvelynNdenial]

You saying they drive just like people do makes me more scared of self driving cars than anything else. People suuuuuuuck at driving. Every single day driving to work I think about how the majority of people out there should not be allowed to drive.

[@tomcraver9659]

@CGP Grey - you should re-visit the "Rules for Rulers" in the context of "What if Ai and robots could do pretty much everything." I.e. - what if the ruler COULD (effectively) do everything himself?

[@1nown]

if the only advisors you have as a ruler are facsimiles of yourself, be they human or artificial, you will simply do the thing you always wanted to do, not the action closest to an objective ideal

[@dianai988]

56:00 i think you might benefit from considering some of the research happening in the explainable AI space and how much this research is influencing responsible AI approaches and principles these days.

[@jessespence696]

I feel like where bots are affecting working conditions right now is with job applications. Right now if you put in and application there is a high likelihood that a bot has filtered out your application through a process and put it forward in front of dozens or hundreds of applications. This is helpful for somethings and outright damaging for other aspects. Like if you have a criminal record or are not legally able to work somewhere that seems fair; On the other hand when you are being skipped over for a silly reason like a test that has no aptitude to a job such as a 'what would you do' questionnaire with somewhat lame answers like tell a supervisor, do it yourself, tell your co worker to do it, or ignore the issue kind of aptitude test that has little or nothing to do with the job you are applying for. I feel like 'need not apply' works here because at some point people will just hire a general application bot to kill the reason for such a filter system to even exist.

[@NoshWare]

I noticed, CGP - I - did. It was that clip with the car manufacturing line.

[@AllYourMemeAreBelongToUs]

1:02:48 Prompt injection

[@0xdeadbeef444]

1:26:20 Classifying biological weapons as different to concentional weapons due to the fact that they may act on their own, makes sense to me. But then Grey says that AIs are similar, because they act as though they are thinking. But in what way does that make them autonomous? They are processes that can be stopped or their hardware can be turned off. Or is he thinking that they are secretly replicating and installing themself onto our devices while we are not locking? If so, what is the basis for this belief? Have models shown the capacity to act without hjman intervention? Do their architectures even allow that?

[@jamesdean5095]

I would guess he’s theorising about a future where various AI algorithms are embedded throughout society and we rely on them to function. Not so easy to turn off then.

[@FernTheRobot]

Thought germs spread, mutate, and compete through the medium of human mind, and because of that, memes can become virulent without anyone intent it to do so. I think Grey is worrying that we are currently trending towards an societal environment where LLMs are being created, adapted, and commercialized in a similar way to the evolutionary drive. so that in the long run we might create something we are not intent it to be. and that we become unable to control what it will evolve to be. On a side note, I think it is not helpful for Grey to draw the correlation that both smallpox and LLMs "act like it has intentions without actually being conscious," because while smallpox seems intelligent to evolve on its own, it's not the perceived intelligence of an LLM that does the evolving, it's the human and technology that spreads and improves an LLM, and the environment they compete in that gives them the evolutionary pressure.

[@spikehammer3112]

When talking about LLM like ChatGPT, there is no basis for any of it. They do not have sentients, they can not "breach containment". ChatGPT can not turn on the lights in your room. Any of these examples of AI appearing to have sentients, is just human personification. It is like if you have a speaker playing the sound of someone being tortured. The speaker is crying out in pain and begging for this to stop and for it all to be over. If you had not encountered a speaker before, you might think that the speaker itself was suffering. For any of these concerns to begin to be come possible several things would need to happen. It would need to be broadly connected to the internet. i.e. It would need to be able to send emails, order you pizza, etc. We would need another technological leap, or possibly several. The current models inability to reason, to remember, to adapt pose major hurdles. It would need a way to act beyond what it is prompted. It would need a way to deceive. To me it seems like most of where Grey is getting these ideas and worries from is sci-fi horror movies and TV shows. Grey seems to be a bit of a doomer about these sorts of things. If you go back and watch Humans need not apply, by and large it could have been written today, aside lack of mention of LLM. Most of the predictions of things that are just around the corner, are still just around the corner. Starbucks still has human baristas, and that general purpose bot, Baxter, is discontinued and the company shut down. It is possible that some of this might come to pass, or we might look back at this in another 10 years with all these worries still just over the horizon

[@michellehao2000]

I think this episode helped actualize for me, a lot of the things AI doomers keep talking about. I think it would be very valuable for grey to make an explainer video on some of the ideas he expressed in this video

[@ilyabelov7626]

18:50 the thing is, self-driving cars are not just imperfect, they are strictly worse than humans

[@splashafrica]

Humans need not apply is one of those i always came back to it was a masterpiece

[@a.thales7641]

And rules for rulers for me too.

[@mautrindade]

the biggest fear and showstopper really are CEOs realising their job are the downright the easiest and most logical and cost effective job to be replaced by AI

[@dianai988]

1:27:00 i think there are a lot of assumptions being made between the LLMs, which are literally just generating text, and the ways in which they are being applied and developed into products or features (e.g. ChatGPT, apple intelligence, copilots, Claude, etc). There's nothing autonomous about the models unless people go and build products/features around these models that allows them to act on the world and take action. It's not about the models but how people are applying them, and this is why responsible use of AI is so important. The AI itself has no intention, it has no desire to spread as you put it. It's only desire is to generate outputs according to the loss function it was optimized for. But the product built around these models, which humans developed and designed and are in control of, that is how autonomy can come into play, if anywhere.

[@corypatrick8141]

On reasons for accepting human errors: If, say, some computer vision model reduces errors in classifying tumors by 50%, I don't feel "but there's nobody to blame/punish when it goes wrong" is really a valid objection. To me, blaming/punishing are not goals unto themselves, but are means to reduce human error rate. If error rate is lower in a system where attributing blame/punishment no longer really make sense, that's still a positive change. (Though maybe emotionally we still want to have some face to blame regardless.)

[@e-manr.486]

Yes! Great point. This is a conversation I've had with students. There is a human desire for agency when it comes to actions. Despite some systems being better than humans, we're uncomfortable with accepting an extension locus of control for some things. However, we've overcome this fear for many tools in the past (e.g., elevators, monorails, even computer software), maybe it'll eventually happen with AI.

[@konkasd2539]

Can someone please tell me where to find what they were saying about Claude talking to itself and having a mental breakdown? I Have checked the links and i can't find it.

[@c6m]

Teaching me how to hear ADR is a mind virus that I cannot get rid of thanks to this podcast.

[@Plystire]

1:10:00 What CGP says here, you can replace him talking about AI to him talking about plastic and it still applies. Because that's kind of what the problem is, right? You can form plastic to represent anything, yet it will never *become* anything but plastic (short of chemical reconstruction). Plastic is super useful and has changed the world, but it hasn't replaced the need for other materials. And while it has amazing use cases (such as rapid prototyping) it is just another material, another tool in the toolbelt. "It feels like food but without nutritional value, like there's something missing here." Kind of like a hollow prototype. AI can't (yet) replace a true expert. It is book smart. It may know the facts but it lacks understanding. Until "Artificial Wisdom" becomes a thing, it will be a hollow representation of expertise. But, hey, I'd argue a hollow representation is often times better. I mean, I don't want customer support to get smart with me, I want customer support to support the customer regardless of how rude said customer is being. It's like Detective Spooner said in I, Robot "I don't want my toaster to get all emotional on me". The technology should only have the capabilities needed to perform the given task, and no more.

[@stijnvandensande3579]

1:18:53 Agree

[@admiralcapn]

43:12 - this is why I don't discuss politics with my friends. They can't abide an individual accepting tenants from multiple parties. They feel that each individual must accept all of One Party or all of Another.

[@DM0317]

I worked in the film industry. Thanks to COVID I was forced to retire from the industry for not following the "mandates" after that the situation got worse from a strike regarding AI taking jobs from writers . After the strike many thought that the industry would come back . Nope the film industry died . They can reproduce an actor with AI nevermind a chair and table I now have a little piece of land that I'm developing to be self-sufficient.

[@tutacat]

They have well paid engineers writing psychological guiding to guide better output generation. The more specific you are, the better output you will get. If not, it is all dependent on context and randomness.

[@AlexMittsVOID]

Right around @43:10 really took a big weight off of my shoulders. I'm not in one of the teams as AI may come for things I do that make money, but I also use AI to do things that make money. It's such a weird position to be in. I really do not love that this is the era of AI, but also, there are some things that I could NOT have done pre-AI-era. Thank you for addressing this.

[@meredithbignell7333]

Can anyone ELI5 'where' does AI exist? We can arrest and jail criminal people, but how can AI be isolated? Would every internet-capable device need to be destroyed to ensure eradication of an 'AI-virus'?

[@sjbayer3]

CPG Gray really hits on a point that vibes with me. Nobody likes people who are in the middle or moderate. This is something that I really feel, it's okay to agree with people on some things even though you disagree with them on other things. Or maybe more broadly you can agree with opinions from a group that you don't associate with. Everything, including politics just feel so polarizing and while that's nothing new the hate for anyone who disagrees with you is something that's really popped up in the past 10 years. Being critical and skeptical as always been a character trait that separates you from others. Most people are pretty darn confident in the things they believe. But today more than ever people love forming groups and pushing out anyone who doesn't fully commit to their group. Taking the time to think about something is taboo. You're supposed to have a gut instinct that immediately conforms to one group or another. I don't see AI moving us past that. We will continue to form groups and push out people who don't fully agree with us until there's some reason to pull us together.

[@315MusicMan]

some of politics are so polarizing because it has very real, tangible effects on people. and being neutral or in the middle isnt inherently a virtue, its just a different stance. i dont know anyone who's getting very passionate about OSHA codes or regulations on whether the city or homeowner maintains a sidewalk. its topics that can bring harm to people (abortion, immigration, international affairs, etc.) that get people upset.

[@tutacat]

Inside a conservation or cannot really forget our remember. It is sort of like simulating characters when you get it to generate speech.

[@RillianGrant]

You can be 100% sure that an input to a classical computer program can't "jailbreak" it. These kind of code injection problems are always due doing the wrong thing to the data.

[@garethbaus5471]

As someone who has tried out a recent ish version of Tesla's full self driving and works on the roads I would argue that self driving cars are already much better than the average driver. FYI I hold the positions that the average driver shouldn't be allowed behind the wheel, and Tesla's full self driving is extremely flawed especially when navigating parking lots.

[@philippemarcil2004]

A very interesting discussion. I agree with Myke that the issue with AI arts is that it is both unethical and frankly very very uninteresting. The whole point with art is to see what other humans have created. Using pure AI to produce work just remove anything of interests in the final art or product. And for creators using AI arts, writing or music for stuff they are not good at, I think it just cheapen the actual work the creator have made. Putting your work beside the art created by an AI, even if in a different medium, basically equate the two. This kind of implies that your work is no better than that of an AI. So if the creator doesn't care about his own work by putting side by side with AI art, why should I care about it then.

[@Skyace13]

I find it interesting that to you Oppenheimer isn’t comparable to what is happening now because you simplify Oppenheimer to be about the nuclear bomb. The greater take away of the movie to me is the “runaway” or “snowballing” of knowledge and intelligence. The genie, once freed, does not go back in the lamp. All we can do is chase it in an attempt to understand the full scale of the damage that can be caused and share THAT finding to prevent such damage from occurring. Scientists did not know they would bump into the “rules of the universe and reality” to stop them. Similarly we do not know if we will bump into a wall or if there is a wall to bump into at all before we destroy ourselves.

[@tutacat]

I find the collated intelligence more important than the reasoning engine side, but we are playing with fire how much changing our decisions can affect the world. This is definitely a pandemic situation.

[@namuzed]

I've had limited benefit out of AI, but it's still occasionally useful. I keep a tiny LLM (Phi3) on my laptop to help with re-writes & summarizations. I also do local AI voice cloning to convert certain essays and documents into audio to make them easier to digest on the go. Some image generation for creating textures and rough concept art. Oh and Whisper for transcriptions.

[@markomarjanovic8348]

If youve seen Grays video about politics, you can connect the dots and start realizing that humans will be less needed for governments and the elites, sooo...yeah, we are indeed fucked in a way that the lives we live today, will not be possible.

[@josiahbaumgartner7643]

The part of the video that stands out the most to me is that AI/autos are inevitable. We could not stop cars from replacing horses. We could not stop Uber from taking over taxis. We cannot stop ai from taking jobs from humans. We can’t. There’s no point in fighting it, adapt to it. Blockbuster fought against streaming. Netflix embraced it and adapted. Which one will you be?

[@ZackaryReaves]

Blockbuster didn't fight against streaming, they had financial issues that cascaded into each other, resulting in bankruptcy. They were actively working on a streaming service up until their finances caught up to them.

[@tutacat]

Random generators that can generate related sentences. It is a big graph of probabilities that simulate language.

[@Jeffroh]

Looks like quantum particles are a big graph of probabilities that simulate matter and reality.

[@tutacat]

I was using the GPT-3 text models before ChatGPT. Even GPT-2 seemed crazy, even though it didn't feel useful. ChatGPT blasted through the ceiling and past the moon. And they most likely did that to gather a lot of data to make it more like humans.

[@JD-jl4yy]

Right on the mark. Humanity is not ready for AGI.

[@Strammeiche]

So many talks about AI safety feel like self-driving car safety talks on steroids, just with so much more people infected by them. I am really looking forward for the coming AI developments and hope they are just a tiny bit as big as people fear they will be.

[@antlers1305]

what a note to end on!

[@boorya]

Code injection is possible thanks to von Neumann architecture (as opposed to Harvard architecture), where instructions and data are stored in the same memory (as opposed to separate storage). It's very hard to create a strong boundary between data and code when underlying memory is one and the same, but it's cheaper)

[@palmberry5576]

This just flat out isn’t true. Both RISCV and AMD64 support separated data and code sections.

[@boorya]

@@palmberry5576 so which dimm in your PC is for code and which is for data?

[@palmberry5576]

@@boorya bits and pieces. The global descriptor table determines which sections of RAM are data and which sections are executable, and it can be modified at runtime.

[@palmberry5576]

@@boorya also, code injection doesn’t really exist? Mostly what you get is stack corruption where data gets interpreted as pointers to return to, resulting in a jump to an arbitrary piece of code

[@aaabcehillpt8822]

Does anyone know what he was citing at 1:00:00?

[@jackketchum7056]

Yes. Why is this not at all being addressed in wider political circles? Spread awareness!

[@cluelets]

Weeeell it wouldn't really be profitable for big tech companies if people were worried about the things that make them the most money. We laypeople should probably stay focused on issues like which political figure is the biggest butthead :)

[@sirBrouwer]

@@cluelets it's more that for most even in a lot of industries it is just difficult to rap your head around the fact that so much jobs could basically vanish. Or even more you could possible automate yourself out of a job even as the CEO. Because in a way you need people that work for you. (and for other companies) so they can spend there income on the products that you make or you help being made. The only jobs that are not likely to be impacted that fast are the actual jobs that are just human to human interaction.

[@OrigamiMarie]

I have become much more bearish about the prospects of self-driving cars in the past ten years. I can't drive. I never will drive; I have adequate vision for lots of stuff, but not for driving, and it's not correctable. So ten years ago, I was really looking forward to the day when I could own & operate a vehicle safely without a license, or any need at all for me to pilot it. I remember even trying out sitting in a Tesla in a mall, thinking "well, this looks likely to be the winner brand, so hey car, see you in a few years!" Since then, much as they have tried, it just hasn't happened. And if it were just Elon's enterprises that had failed, that would be one thing. But it's all of them. "But it hasn't failed!" I hear. Well . . . see I live in Minnesota. This is a place full of snow, slush, snowmelt that causes flooding, hard rain, hail, big fluffy clouds that cast constantly moving shadows. There's a good chunk of the year when you absolutely cannot see lane stripes. There's no way to discern them. They're under a layer of hard packed icy snow, a layer of various colors of slush, and a layer of new-fallen snow. Even curbs are a pretty abstract concept around about late February, due to random plowing for months. And humans just deal with it. We are excellent at following the bare hint of tire tracks in the hard pack under 4" of snow. We are fantastic at averaging out the levels of slush around where people drive, and choosing the likeliest place for the next human to keep traction and stay out of oncoming traffic. There are times in the winter when the tracks go right over the double yellow, and what do you do? The absolute safest thing you can possibly do; you drive right over the double yellow, keeping your tires in the established tracks. The oncoming lane tracks are perfectly aligned, and accommodate the infringement. Oh but we're just at the beginning! Actually I think we're not. I think we're at a very long plateau. We've run out the Moore's Law runway, so we can't just throw more processing power at it for comparable up-front and ongoing-energy costs. Unless somebody really pulls a Quantum Computing rabbit out of their hat (and I'm super doubtful about that), this is the level of computing we're stuck with. We're not going to get the 10x improvement that might possibly make self-driving cars a true reality. I feel the same way about all the flavors of Generative AI. They've already blown past the energy budget and slurped up all the content. They can't make the models better, because there isn't more data available and there isn't a 10x energy improvement waiting in the wings (again, unless Quantum Computing and / or Cold Fusion make it). So I'm not worried about AI taking over anymore. I think this is it, this is all we get.

[@tristanbeal261]

The AI companies haven't failed at full self driving, its just taking longer than predicted. Largely because people like Elon Musk didn't see how hard a problem it was. Furthermore computing improvements are going to continue for maybe a couple of decades beyond moores law, google computing beyond moores law to see the ten or so different ways computers can be improved beyond moores law. But its not a hardware problem, we have the hardware today to solve the problem but training neural nets sufficiently takes time, and theres only so much budget to spend on it. But I guarantee you that the weather issue in Minnesota (and other places) will be solved in the next ten years. As for the comment about generative AI, the energy issue isn't a problem, with sufficient funding more energy can be produced, way past the 10x that you mention, not through cold fusion though because thats baloney. The data issue, is an issue but that can be solved by a mixture between collecting more data with robots, generating artificial data and making AI models which are less data hungry. So in summary were not close to the limits of AI yet, the AI apocalypse is still a possibility. And you should expect the weather issue to be solved in the next ten years.

[@PerMejdal]

Oh the irony in that I am now taking the transcript of the video, and making an AI summarize it. The transcript of the KZbin video between CGP Grey and Myke marks the 10-year anniversary of the video “Humans Need Not Apply.” They reflect on the changes in AI and automation since the original video’s release. Here are the key points from their discussion: *Revisiting “Humans Need Not Apply”:* • CGP Grey rewatched the video and reflected on its creation, noting the intense focus and effort involved. • The video aimed to raise awareness about automation and its potential impact on jobs, which was not widely discussed at the time. *Changes in AI Landscape:* • The landscape of AI has significantly evolved over the past decade, with many discussions about self-driving cars and automation. • Grey feels that the public consciousness has caught up with the ideas presented in the video, especially regarding technological unemployment. *Personal Reflections:* • Grey and Myke discuss how their skills and presentation styles have evolved over the years. • Grey mentions that his earlier videos, including “Humans Need Not Apply,” were made with a sense of uncertainty and were more like public presentations. *Impact of the Video:* • The video was successful in introducing the concept of technological unemployment to many viewers. • It remains one of the most referenced videos by people who meet Grey, indicating its lasting impact. *Current Thoughts on AI:* • Grey expresses a sense of overwhelm and unease when discussing AI, noting the rapid pace of change and its implications for the future. • They discuss the challenges of communicating the seriousness and complexity of AI developments without becoming overly self-referential or tedious. *Self-Driving Cars and Automation:* • Grey reflects on his predictions about self-driving cars, acknowledging that the timeline for their widespread adoption has been longer than he anticipated. • He shares a personal anecdote about seeing self-driving cars in Phoenix, which felt unremarkable yet significant. *AI’s Role in Society:* • Grey and Myke discuss the societal and ethical implications of AI, including the demand for perfection in self-driving cars and the emotional reactions people have to technology. • They touch on the idea that people prefer to have control over their own safety, even if it means being less safe overall. *Future of AI:* • Grey is cautious about making predictions for the next 10 years, given the rapid and unpredictable nature of AI advancements. • He emphasizes the importance of being aware of the potential risks and challenges associated with AI, including the possibility of AI systems acting autonomously in ways that are difficult to control. Overall, the conversation highlights the significant changes in AI over the past decade, the lasting impact of “Humans Need Not Apply,” and the ongoing challenges and uncertainties surrounding the future of AI and automation.

[@carultch]

Oh wow, you know how to copy and paste, I'm so impressed.

[@burger-se1er]

@@carultch I found it an interesting read after listening to the video.

[@edcramer6475]

I think it's interesting that the summary ignored the issues of copyright infringement and unshared corporate profit.

[@jc-gp8hd]

In accordance with the highest of American traditions, I want to wish everyone out there a Happy Labor Daybor

[@GT380man]

The politics part is I think mostly restricted to USA. In Europe, I think we understand that what’s happening is so far beyond politics that it hardly arises. Lobbyists, for example, penetrate the halls of the lawmakers that it’s flawed to even think this is a left / right issue, because it clearly transcends either & all political systems and social structures. Politics hardly matters in this field. A small number of very wealthy and powerful people are making all the important decisions about its development. Frankly, it’s applications are almost inevitably going to follow commercial imperatives as seen from the perspective of its owners. We kid ourselves if we think our opinions are of any interest to those decision makers. Our politicians are too busy filling their boots like they’ve always done only worse & though there are near term important implications even for their children and their children, they’re only focused on their own short term interests. That makes them trivially easy to deal with and skirt around. Deal with it.

[@tutacat]

There was a great black mirror episode about this

[@Blastmaster321]

HE FINALLY DID IT!!

[@PeterAllenLab]

Right now (August 2024), LLMs are (mostly, probably) evolving with humans in the loop. Humans change the LLM, improve it, put it into the market, humans evaluate it, repeat. I worry about when we automate this loop. We let LLMs modify the next generation of LLMs and then LLMs evaluate them based on some real-world survival metric like [amount of human attention captured]. That's an artificial selection loop that could quickly produce something very ugly or dangerous. I was thinking about this when I ran across a startup called Aspect: "Social media reimagined. No human users - just you and AIs. Connect with AIs like never before."’

[@implokusmaximus]

Two months ago at my work, a few jobs were replaced by automated bots. Our union could not stop it.

[@herrabanani]

how could the union stop it? the only negotiating leverage that unions have is that the workers can go on a strike, but if they're being replaced that won't do anything

[@delusionnnnn]

As I see more AI offerings in more places, particularly in text-based models, I always ask myself "who thinks I need more of this in my life?". If I ask it anything important to me or about anything I know about, it constantly lies. At least with my uncle Richard, he lied to me as a kid with a wink and a nod. I don't need AI to lie to me, and if there's even a possibility of that, I have absolutely no use for it. Also, general purpose AI has literally been "five years away" for about 60 years, which is to say the most useful thing we've learned in the pursuit of AI is that we don't actually know what we're expecting it to do nor are our theoretical models of consciousness useful - either because we can't replicate them, or when we do, they come up short.

[@adamanderson1979]

@1:00:00 my personal take is that these are likened to fragments of DNA in a primordial soup. One day a collection of these “AI” programs will assemble into an organism. Some day after that that organism will have sentience , and we simply need to have the humility of understanding and far from that endpoint we are now.

[@CROGGS]

Sick

[@DellDuckfan313]

The confabulations/hallucinations angle is easily the most interesting to me. As someone who works as a recordkeeper, I'm very hesitant to use it, because I want my data to be correct. But some of my colleagues who take a different view of the importance of my work, are likely to say that it's "good enough". In a similar way, it'll probably come down to a sector by analysis: jobs that require a high level of "purity" will not be taken over by AI, whereas jobs and employers that don't value the process in the same way, will be taken over by AI. Will AI write the new Star Wars? Definitely not. Will it write the next pulp romance novella? Sure.

[@tutacat]

I think the innovation is not much faster, I think things are changing less, but people are really investing in machine learning technology and developing it commercially. But e are definitely on the sigmoid curve. GPT-4 has gotten a lot better "creative" and "useful" now that the "omni" models have released. But I think a lot of people are also becoming more annoyed with the proliferation? And everyone is becoming more aware of its shortcomings.

[@PASH3227]

I'm not convinced that technology will create mass unemployment. Here in California, the minimum wage for fast food workers was raised to $20/hour. Despite the fears of fast food restaurant closures, there's still a worker shortage. The global population is expected to peak within 40-50 years, and before that decline more and more of the population will be too old to work. This is already happening in Japan and Korea and they NEED automated labor and robots. Also what do I do about this? I lose my job in 20 years due to a robot. So I get a new job.

[@sveinbjornegilsson4719]

i tried out claude and chat gpt after listening to this episode and honestly i'm not that impressed. it is infinitely better than it was a year ago but they both still have the problem where they don't really know what they're doing, at least when it comes to ahk scripts. they can throw somethign togeather and they can itterate on it to some degree but pretty soon it just beomces one step forward one step back as they start re-introducing bugs that you helped them remove, so you have to understand the code at some point if you want it to have more than a couple of features still these tools are useful for someone like me who has limited programming experience, they can setup example codes, but as soon as you start running into strange debugging issues the ai's are useless in helping you and you have to understand what's really going on

[@herrabanani]

apparently claude has been enshitified

[@MuhammadRaza-yd6sg]

My favourite video to date

[@GT380man]

The self driving car timeline wasn’t just over optimistic, but I’m among those who think it’ll never get to the levels of adoption that was envisaged. Consider the bulk of short vehicular trips. They don’t merely move goods from A to B. They are accompanied by goods needed to fix a customers appliance, like a gas fired heater. The man is always cheaper than any alternative fixes, and so the human driver is a much better solution as well as incrementally cheaper. Am I wrong about that? Obviously, I could learn to fix my own gas appliance. Or do part of it & if the customer could fix the issue, an auto could bring the part to me. But while I could probably do this, because I’m mechanically adept from history and hobbies, most people could not nor would wish to try. So we’re left with the current solution or change the scenario for example so that gas heaters never break down.

[@awdrifter3394]

The robot in the autos will fix your appliance.

[@time-trader]

Self driving is more difficult than most people think. Our infrastructure is built for humans. The only way to have true self driving outside geo fenced areas is to solve general intelligence.

[@TheGahta]

What is "true self driving" and what is the difference to what its understood as by the industry?

[@time-trader]

​@@TheGahtahands off, any point a to any point b, in all possible conditions. Good luck training a neural net for the extremely long tails..

[@TheGahta]

@@time-trader whats your point? In an environment where you can exclude people on foot the challenges are moderate and seemingly close to being possible And what is the difference between your definition and the one used by the industry? What defines "all possible conditions"?

[@sirBrouwer]

@@TheGahta all possible conditions could imply but not limit to. Weather: Weather in all conditions and how both the road it self. other users on that road will react to it. This could be very hot, very cold, windy, rain, snow, ice, mud, even extreme as fire. and in all cases how other factors around the car will react to it. Events on or next to a road: Events like a mudflow, a (minor) flooding, damaged to the road. A hurdle of animals on said road. groups of people next to the road and those that cross it on foot (unpredictable) children playing next to a road. Cars involved in a accident in front of your car. act well on a road where every mode of transport will just move around at any point from any direction to any direction. this and many many more it would need to be able to do with out a problem and even if it gets impacted it self even then it needs to be able to keep both the people inside the car as those outside as unharmed as possible. O and it needs to be able to do all that in a way that it also does not hinder others in a normal setting.

[@mrJety89]

57:20 It's because you phone's listening to you

[@tutacat]

I think they will replace the low level outsourced countries' workers with automated systems

[@ryanhall9920]

I think the main reason why self-driving cars will struggle is because the same algorithm is being used in many cars. Imagine having your friend giving you a lift in their car and as you get in you read an article on your phone reporting that your frind crashed their car 5 times yesterday and once this morning. Would you want be fine with that, or would you want to drive yourself? Because each human person and their driving ability is different, a person crashing their car does not worry you because your friend is not that that person. But there are only a few self-driving algorithms shared with many vehicles, so is is as though your friend themselves crashed all those cars, rather than it being some other person who passed the same driving test as your friend.