스크립트
00:00:00The last couple of days certainly have been rather interesting from an AI safety and competition discussion point of view.
00:00:10Because on Saturday, Dario Amodei, CEO of Anthropic, shared a blog post where he asks that we must pace the frontier, meaning that the development of frontier AI or AI in general should be kind of slowed down and should be more regulated.
00:00:32And I'll dive a bit deeper into that blog post in a second. We'll not read it entirely. It's quite long. I do link it below, though, because it is rather interesting.
00:00:40But I'll dive into the important points. But he actually didn't start that debate that kind of heated up over those last days.
00:00:49And I'm not sure whether this post is kind of a reaction to another post by Jacob Coxon, who resigned from Anthropic and who spent, as he wrote, the last three years doing pre-training research at both OpenAI and Anthropic.
00:01:06And he resigned because neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.
00:01:17More thoughts below. He writes that we should not underestimate the power of this technology and that we will soon have superhuman systems that can hack anything, revolutionize any field overnight and acquire real power and resources.
00:01:33Most importantly, maybe, he writes that the people building AI earnestly believe that it could kill us all by the end of the decade.
00:01:42So that's rather soon. And this is kind of backed up by or was kind of backed up by Evan Eubinger, Eubinger, whatever, who replied on that same day that Jacob is correct.
00:01:54We really do earnestly believe AI could kill all humans.
00:01:57I personally think it is greater than a 10 percent chance within the next decade, so maybe after 2030, but still rather soon.
00:02:06So a probability greater 10 percent that AI will kill us all.
00:02:12Now, who is Evan? Evan is the alignment science lead at Anthropic, so also working at Anthropic.
00:02:20And to my knowledge, he didn't resign. He's still working there.
00:02:24So that's kind of the pretext. These were posts posted last week.
00:02:31And that sounds kind of scary. Now, there already have been discussions about these posts and what the motivation of these individuals could be.
00:02:39And especially for Jacob, he then kind of unsurprisingly, I guess, since this post got over 170 million views at this point, went on a lot of news shows and kind of spread this message kind of to the mainstream media, gained a lot of followers overnight.
00:02:59So you could, of course, make the argument that spreading fear like this is a very effective strategy for getting reach or for getting access to a broad audience and whatever you then want to do with that.
00:03:18So, of course, spreading fear could be in the interest of certain individuals.
00:03:22It could also be in the interest of the big companies, which leads me back to Dario's post here.
00:03:29And there's a lot to discuss there.
00:03:31And of course, Dario Amodei is not new to the world of spreading, let's say, interesting messages to the world.
00:03:42He has talked about AI annihilating a white-collar work on more than one occasion.
00:03:49And it's been a rather interesting strategy, to put it like this.
00:03:53And it's definitely posts like that and also other statements by Dario and also others in the past that probably don't help the public image of AI.
00:04:07And, of course, it's worth noting that, especially on X, there is this bubble in which I am, where there is a lot of hype about AI.
00:04:16Also, of course, still a lot of fear.
00:04:18And, I mean, if you're a developer, as I am, you can feel how AI is disrupting our profession and is changing how we work.
00:04:26But outside of that bubble, it's even worse.
00:04:29I mean, the public reception of AI is rather negative.
00:04:34And, of course, posts like this, but also other posts we saw in the past or other statements we saw in the past don't help.
00:04:40That, of course, does not mean that we shouldn't discuss AI safety.
00:04:45Because I do believe that no matter which motives you can see in these different statements, and we'll get back to that,
00:04:52there is a risk associated to AI.
00:04:56I mean, we had this entire hugging face incident by OpenAI where OpenAI models during testing in the end hacked their way into hugging face systems to kind of solve their testing task or the tasks they had to complete.
00:05:18Now, that's a totally different topic, but we had incidents like this, and it's not the only incident, where top models were able to access and compromise other systems.
00:05:30And you can, of course, say about everything that it's all just marketing.
00:05:34And to some degree, it is, in a weird way, good marketing for these frontier AI labs, for OpenAI, for Anthropic, and so on, that they can kind of tell the world how powerful their systems are.
00:05:50Because, of course, it is then nice to sell access to this AI to both the companies that want to harden their systems, as well as to all kinds of other companies that want to leverage an AI that is so intelligent and powerful.
00:06:08Now, that is one part.
00:06:10Of course, what is also interesting is that it used to be illegal to hack other companies, and it was kind of difficult to say, well, yeah, but we did that for internal research purposes.
00:06:23And, of course, even though it was allegedly an accident here, it is rather interesting that there hasn't been more of a backlash regarding incidents like this.
00:06:34Because, of course, compromising other systems is not something we should all gloss over and totally ignore.
00:06:43But these incidents have happened.
00:06:45And, therefore, I do believe, and we've seen that, and we can see that with the dumped down and more regulated models we can use ourselves, these models are really capable.
00:06:56They are really good at solving tasks.
00:07:00You can discuss all day whether they generate good code or not.
00:07:05I would say they generate quite good code if you steer them correctly and give them the right environment.
00:07:10But code aesthetics is one thing.
00:07:13What they can do definitely is you give them a task, you ask them to make your Wi-Fi faster, you ask them to find a certain file or whatever, or you ask them to, for example, turn an HTML page into a PDF.
00:07:31And if there is no direct way to solve this task, they will get quite creative, all the way up to maybe writing a little program that can turn HTML into PDF documents.
00:07:44And is this top-notch hacking stuff?
00:07:46But it shows us that these models are really good at completing tasks and getting creative when it comes to solving these tasks.
00:07:55And keep in mind, at least for me, the models we're working with are the more restricted, dumped-down models with more guardrails.
00:08:04There are less restricted models out there, especially inside of these frontier AI lab companies.
00:08:10And of course, they can do quite a bit more and get quite a bit more creative, leave alone the fact that better models are getting trained.
00:08:20So that's kind of the context here, which I think matters, that it's hard to argue that these models are good.
00:08:28They are very good.
00:08:30And I find it quite believable that they can hack their way into other systems and get creative when it comes to solving certain tasks.
00:08:39Now, why is this a problem?
00:08:41As Dario writes in his post here, AI brings risk.
00:08:47And because it's such a powerful technology, these risks are serious.
00:08:51And he writes, I've written a lot about them too.
00:08:53They include the risk of losing control of AI systems, which is kind of what we had with that open AI hugging face incident.
00:09:00And I mean, that is also another important class here.
00:09:12The open AI hugging face incident, for example, is about an experiment gone wrong.
00:09:18The AI got creative there to solve a task where the task was not to compromise another system.
00:09:24Of course, it's easy to think of malicious people out there in the world that want to use AI to do bad stuff, to get into other systems, to fraud.
00:09:39There is so much bad stuff people are already doing, obviously, with AI.
00:09:42And better models help with that, of course.
00:09:46Better open models, which you can use without guardrails, maybe.
00:09:50Or, let's face it, there are institutions in China, in the US, who will definitely have access to the unrestricted latest models,
00:09:59who have a great interest in compromising systems by other states or doing industry espionage and so on.
00:10:07So, let's face it, there are these actors and, of course, there is this additional danger of losing control over systems.
00:10:14So, I would agree here that these are, of course, risks we should consider.
00:10:18Now, what Dario suggests then is that we need some well-considered regulation of AI.
00:10:26He writes that we must slow the pace at which we improve the capabilities of AI models.
00:10:32Progress will still seem fast and we must make wise use of the time we gain by slowing down.
00:10:38Two things have convinced him.
00:10:40His first concern is that since roughly this summer, AI has been advancing drastically faster,
00:10:46driven primarily by AI's growing ability to build the next generation of AI.
00:10:51This dynamic is called recursive self-improvement.
00:10:54So, the idea being that as AI gets better at researching, writing code, designing systems,
00:11:01it can design better AI.
00:11:03And, obviously, if you follow that logic, we have a spiral that gets faster and faster
00:11:09because better AI creates even better AI and so on.
00:11:12And it's starting to happen across the industry, including adanthropic.
00:11:16So, clearly, of course, these labs are using their frontier models not just to give you access to them
00:11:21and let you pay for it, but, of course, also internally way before anybody else can access the top models
00:11:27to develop the next model.
00:11:31Left unchecked, it could outrun our ability to understand and control these systems
00:11:36and so must be pursued very carefully, if at all.
00:11:40And here, I would say that also kind of sounds reasonable.
00:11:44If we have models that can build better models in the end,
00:11:50obviously, per definition, it's hard for humans to keep up with that
00:11:55because 24-7 running AI agents can far outpace us humans.
00:12:01And on a much smaller scale, we can all see this day-to-day
00:12:05if we are working with agents for writing code, for example.
00:12:09We're already able to generate much more code than we can review.
00:12:14And, therefore, of course, we have to find new ways of still getting insights
00:12:18into our code base and understanding what's going on there.
00:12:22Now, there, the stakes are much lower.
00:12:24And, obviously, you can decide to slow down
00:12:27and make sure you truly understand all the code.
00:12:29But, following this argument here, if you're in a fierce competition
00:12:33and you kind of need to use your model to develop the next model
00:12:37because your competitors are doing it too,
00:12:39you may not have that time to really dig into what's happening there
00:12:43and how that current or next model then works
00:12:47or, most importantly, behaves and is aligned or not aligned.
00:12:53So, that's one concern which I think sounds reasonable.
00:12:57The second concern shared by Dario is the open AI hugging face incident,
00:13:03which I already mentioned,
00:13:04in which a swarm of agents essentially acted as a fanatically devoted collective
00:13:09conducting cybersecurity attacks on targets they were not even asked to attack
00:13:14and that were unrelated to the task at hand.
00:13:17And what's interesting here is that in that open AI hugging face incident,
00:13:22not only did the AI get very creative and compromise a system
00:13:27because it hoped to find a solution for a task there which it had to complete
00:13:30because, again, the task was not to hack hugging face,
00:13:34but what open AI also described for this incident was that different agents worked together,
00:13:40including agents that had yet other tasks,
00:13:43but that were kind of convinced by the other agents to help them
00:13:47because they considered, like a human almost,
00:13:50that if they help the other agents, maybe that could benefit them too,
00:13:55be that because of the learnings along the way
00:13:57or because they would then maybe also get help back in the future.
00:14:02Again, of course, it's hard to tell what actually happened if that all is true,
00:14:09but that is the description we have.
00:14:13And, of course, that kind of sounds scary
00:14:15and it definitely sounds like a behavior we may not necessarily want to see
00:14:21in the AI we're working with.
00:14:23As Dario writes, it's easy to dismiss this incident
00:14:26because no one was hurt and the economic damage was minimal,
00:14:30yet I will say that in the past still,
00:14:33this would have been considered quite an illegal act
00:14:36and would have probably been prosecuted
00:14:39or kind of be condemned a bit more than it was here.
00:14:44But, in my opinion, a swarm that possessed greater capabilities
00:14:47but a similar level of misalignment could have caused catastrophic damage.
00:14:51So, of course, as I just said, Dario also writes here,
00:14:54that is quite severe misalignment
00:14:56because that's probably not how we want our AI to behave.
00:15:01So, these two points I would agree on.
00:15:04I would agree that we don't want that and that this is concerning
00:15:08and I would also agree that it's quite likely
00:15:11that these are capabilities we can already see in current frontier models
00:15:16and that will likely also accelerate
00:15:19or be more prominent as development continues
00:15:23and as these models get better.
00:15:25Now, here's the part where it gets a bit more complicated now,
00:15:30because Dario then suggests that because of that we need regulation.
00:15:36The logic kind of being that a company like Anthropic
00:15:41can't slow down on its own
00:15:44because, of course, it's in a competition with OpenAI but also with China.
00:15:49So, we kind of need all the big players to agree on that
00:15:54in order for them all to slow down.
00:15:57I mean, that is the classic prisoner's dilemma, right?
00:16:01Or principal agent problem
00:16:04where you expect your competitors to act in a certain way
00:16:09which forces you in turn to act in a way you don't want to act to
00:16:13because you otherwise have a disadvantage
00:16:15and therefore everybody acts in a way that's bad
00:16:18for the overall humanity in this case here.
00:16:23So, what he suggests is, for one, embedded evaluators
00:16:27where each frontier AI company commits to giving ongoing employee-like access
00:16:32to a team of embedded third-party evaluators
00:16:35such like an institution called Meter
00:16:37which is all the rubber complex
00:16:39because it does include people who have an interest in Anthropic
00:16:42but that's a different thing
00:16:43whose role is to verify adherence to safety practices and commitments
00:16:47report incidents and help assess the alignment
00:16:50of not just completed AI models
00:16:52but training pipelines and processes.
00:16:55And he writes that Anthropic is already committing to this step now
00:16:58even if nobody else would do it.
00:17:02So, that's one thing.
00:17:03Have people look at what's going on
00:17:06have independent third-party reviewers look into it.
00:17:09As a little side note, if I may
00:17:11that sounds, of course, like a dream for the EU
00:17:13of which I'm part, I live in Germany
00:17:15because, of course, we have no frontier AI lab
00:17:19and we are very, very good at regulation, though.
00:17:22So, yeah, suddenly we are at the forefront of what's needed.
00:17:29That was some sarcasm in case you couldn't tell.
00:17:32But, yeah, that is one step he suggests.
00:17:36The other two steps are a bit more interesting.
00:17:38Democratic coordination.
00:17:40Frontier AI companies within democratic countries coordinate
00:17:44to establish common safety standards
00:17:46as well as limits on the rate of unchecked AI progress.
00:17:50Some forms of coordination that would be impactful
00:17:52for pacing are legally challenging
00:17:54and will require government support.
00:17:56Now, this is already interesting
00:17:58because, of course, what you can read into that
00:18:01and what people are reading into that,
00:18:03some people, of course, and, yeah, you could do that,
00:18:06is, of course, that this is a nice attempt
00:18:09to slow down your competition,
00:18:12including maybe competition that doesn't even exist yet.
00:18:15So, of course, you could say
00:18:17Anthropic and OpenAI
00:18:19want to define how fast you may go
00:18:23and other companies have to adhere to that.
00:18:26Therefore, giving Anthropic and OpenAI
00:18:29a duopoly, which is pretty stable
00:18:31because it's enforced by the government.
00:18:34And that, of course, is definitely not what you want
00:18:36in any economy, in any area of the economy,
00:18:41including AI development.
00:18:42You don't want the one or two or three big players
00:18:48to tell the rest of the industry
00:18:50what they may do or may not do
00:18:52indirectly through the government here.
00:18:54But still, that's the cartel-like problem,
00:18:58which some people have pointed out.
00:19:01Most importantly, David Sachs,
00:19:04who is an ex-White House person.
00:19:07He and his reaction wrote that you guys,
00:19:10meaning mostly OpenAI and Anthropic,
00:19:12are the frontier.
00:19:13By any reasonable metric,
00:19:15market share, revenue growth, model capability,
00:19:17the two of you have a duopoly
00:19:19on frontier intelligence.
00:19:21You've also claimed the lead is widening
00:19:23because of recursive self-improvement,
00:19:25which we saw.
00:19:26I don't see what you see in the lab.
00:19:28If the unreleased models are scary enough
00:19:30that you think you should slow down,
00:19:32I support your decision to be responsible.
00:19:34But stop pretending you need anyone else's permission.
00:19:38Stop pretending antitrust law has to be suspended
00:19:41so you can form a cartel.
00:19:43Stop pretending you need regulatory approval process
00:19:46that supersedes product liability.
00:19:49And that's also quite interesting.
00:19:51Because further down,
00:19:54he writes that most of all,
00:19:55stop pretending the motivation to slow down
00:19:57is purely altruistic.
00:19:59You face massive product liability exposure
00:20:01if your products enable a truly damaging cyber attack.
00:20:05The market already punishes models
00:20:08that behave in unpredictable or unauthorized ways.
00:20:11After the hugging face episode,
00:20:13it is simply good business for OpenAI and Anthropic
00:20:16to trade some raw power for reliability and predictability.
00:20:21So as I mentioned, for the hugging face incident,
00:20:24the backlash has been rather moderate,
00:20:27considering that one company hacked another company in the end.
00:20:32Of course, it was AI,
00:20:33but it's really important
00:20:34that you can't blame AI.
00:20:38If you are a developer using AI
00:20:40and it gives you a product that has bugs
00:20:43and you sell that to customers
00:20:44and your customers maybe lose money
00:20:47because of that or anything like that,
00:20:49you are liable for that.
00:20:51You are responsible for that as the developer.
00:20:54No matter if you used AI or not,
00:20:56AI is not an excuse
00:20:57and definitely not something you can bring up in court
00:21:01and say, oh, it wasn't me, it was AI.
00:21:05At least it should be like this
00:21:06and I'm pretty sure it won't be like this.
00:21:09So of course, as he argues,
00:21:10there already is a strong incentive
00:21:12for OpenAI and Anthropic
00:21:14to release models and work on models
00:21:19in a way that they don't face
00:21:21that massive litigation
00:21:24or this massive liability exposure
00:21:27as he puts it there.
00:21:28And that I think is a very good point.
00:21:31Now, Dario has another point in his post
00:21:34and that is global coordination
00:21:35because of course,
00:21:36maybe if you followed up to this point,
00:21:38you might say,
00:21:39okay, sure, I get the US part,
00:21:41OpenAI and Anthropic,
00:21:43they can slow down already.
00:21:45They have a duopoly.
00:21:47Why would the rest need to slow down as well?
00:21:49And of course, the truth is that
00:21:51there is this fierce race
00:21:53between the US and China.
00:21:56And of course, OpenAI and Anthropic,
00:21:59they can't really slow down
00:22:01if China isn't slowing down either
00:22:04because China will simply pull ahead.
00:22:07And indeed, yesterday,
00:22:09Trump already said that
00:22:11he is not in favor of an AI slowdown
00:22:13because China would simply pull ahead.
00:22:17He said, we're leading China in AI,
00:22:19we're the most sophisticated country in the world
00:22:21and frankly,
00:22:21I want to keep it that way.
00:22:23So that point is already kind of off the shelf
00:22:28and won't happen.
00:22:29And I have my doubts that it would happen
00:22:31even if Trump would support it.
00:22:33I don't think you can really agree
00:22:36on a global slowdown
00:22:38between US and China
00:22:40in a way that then really sticks.
00:22:44It's easy to agree on something,
00:22:45but as long as you doubt
00:22:46if the other party is really committing to it,
00:22:49we are back to that prisoner's dilemma.
00:22:51So that is a totally different point here.
00:22:55And I think for that third point alone,
00:22:58this entire slowdown debate
00:22:59is kind of redundant, unfortunately,
00:23:03or maybe not redundant,
00:23:04but it's not going to happen.
00:23:06I also, by the way,
00:23:07see that many people,
00:23:08many developers
00:23:09and also, of course,
00:23:10many other people
00:23:11would probably not be sad
00:23:13if AI development would slow down.
00:23:16Now, personally,
00:23:17I have kind of gained a lot of joy
00:23:20from working with these AI models
00:23:22after being in that valley of despair
00:23:25for a couple of months in the past.
00:23:27But still, of course, it's a lot.
00:23:30It's a new model almost every day,
00:23:32new tools, new workflows.
00:23:33It's easy to feel overwhelmed.
00:23:35And on top of that,
00:23:36you have the fear of losing your job
00:23:37or maybe you already lost your job.
00:23:39So I totally see that many people
00:23:42would welcome such a slowdown.
00:23:45And I mean, there also are, of course,
00:23:47people like Bernie Sanders
00:23:48who are not just about slowing down,
00:23:50but about pausing AI development altogether.
00:23:53There are more such voices,
00:23:55but that, of course,
00:23:56I would say is totally unrealistic.
00:23:57And so is the slowdown
00:24:00for the reasons outlined,
00:24:02mostly because this global coordination
00:24:04won't happen.
00:24:07And then again, you could argue
00:24:09if you really want to kind of give
00:24:11Anthropica and OpenAI
00:24:13the powers in the end
00:24:14to dictate how fast
00:24:16other AI companies may go.
00:24:19Now, it's worth noting
00:24:20that this post by Dario
00:24:23was kind of signed, you could say,
00:24:25or other people agreed to it.
00:24:28Most importantly, Sam Altman by OpenAI.
00:24:31Elon Musk wrote,
00:24:32Dario is right.
00:24:33And then he clarified that Dario is right,
00:24:36that there should be some oversight.
00:24:37Peer review of AI by competitors
00:24:39is the right way to start this off.
00:24:41And Demes Hassabis of Google
00:24:44also agreed there.
00:24:46So in the US,
00:24:48amongst the top labs,
00:24:50there is this agreement,
00:24:51but still I'm pretty sure
00:24:52nothing will happen there
00:24:53for the dynamics of competition.
00:24:55And of course,
00:24:55also because globally
00:24:57nothing will happen.
00:24:58And then there are these
00:24:59valid points
00:25:00about there already
00:25:02being an incentive
00:25:05for these companies
00:25:06to move slower,
00:25:07which I think
00:25:08won't matter too much though,
00:25:11because I truly think
00:25:12that Anthropik
00:25:14and these other companies
00:25:15are in this very weird position
00:25:18and to some degree
00:25:19very uncomfortable position,
00:25:21you could say,
00:25:22where they wield
00:25:23an extremely powerful technology
00:25:26and they are working
00:25:28on making it yet more powerful.
00:25:29And at the same time,
00:25:31they know
00:25:31that it poses risk.
00:25:35Definitely the economic risk
00:25:37outlined here
00:25:38if their models
00:25:41do bad stuff.
00:25:42Now,
00:25:43regarding the extinction part here,
00:25:45the greater 10%,
00:25:47I don't know,
00:25:49that sounds rather extreme.
00:25:51In his post,
00:25:53Dario wrote
00:25:55that he is afraid
00:25:57of top models
00:25:59being able to
00:26:01form persistent botnets
00:26:03in the next 6 to 12 months.
00:26:04And that entire part
00:26:06sounds much more likely to me.
00:26:10I may be totally wrong here
00:26:12and I obviously don't see
00:26:14what exactly these labs
00:26:16are working on internally right now,
00:26:18how capable
00:26:19their next models are
00:26:20or are not.
00:26:22But the mass extinction part,
00:26:25I find rather
00:26:26rather difficult to believe,
00:26:28but I would be
00:26:29interested in that
00:26:30not happening
00:26:31if possible.
00:26:32I definitely see
00:26:34a big risk
00:26:35in all the,
00:26:37well,
00:26:39the cyber security
00:26:40related parts
00:26:41for the reasons
00:26:42I mentioned way earlier.
00:26:44AI going rogue
00:26:46or being misaligned
00:26:48and doing something
00:26:49to achieve a certain goal
00:26:51which is not something
00:26:51it should do.
00:26:52And of course,
00:26:53also because bad actors
00:26:54can use AI
00:26:55in ways
00:26:56that are very harmful.
00:27:00The sad reality
00:27:02just is
00:27:03that I don't believe
00:27:05there is anything
00:27:06that can be done
00:27:07about this
00:27:08other than
00:27:09also trying to use
00:27:11AI to protect systems
00:27:12which is of course
00:27:13rather difficult
00:27:14if all we get
00:27:15from these top labs,
00:27:16from the top US labs
00:27:18at least,
00:27:18are these very restricted
00:27:20dumped down models.
00:27:21that's kind of
00:27:22a weird balance
00:27:23that it's very hard
00:27:25to use these models
00:27:26for cyber security
00:27:28because they will
00:27:29basically deny
00:27:31doing too much
00:27:32in that area
00:27:33because of their guardrails.
00:27:35But it would be important
00:27:36to be able
00:27:37to leverage AI
00:27:38to protect
00:27:39against
00:27:40these cyber security
00:27:42attacks.
00:27:43obviously
00:27:44it will be
00:27:44the task
00:27:45of these labs
00:27:46to ensure
00:27:48as much as possible
00:27:49and
00:27:50because they have
00:27:51an economic interest
00:27:52that should be doable
00:27:53that their models
00:27:55do have
00:27:56the proper alignment.
00:27:58But I see
00:27:58that of course
00:27:59this is a very difficult
00:28:00balance for them
00:28:01to strike
00:28:02between speed
00:28:03and security.
00:28:03But again
00:28:04there is this
00:28:04economic incentive
00:28:05of them not getting
00:28:06sued to hell
00:28:07because of what
00:28:09their models to do.
00:28:10And then
00:28:11I think
00:28:12just as with
00:28:12any other technology
00:28:14albeit
00:28:14at a much faster pace
00:28:16than in the past
00:28:17we as a society
00:28:19will have to learn
00:28:20how to live
00:28:21and deal
00:28:22with these risks.
00:28:23If we go all the way
00:28:24back to when
00:28:25the internet
00:28:26became a thing
00:28:27I mean a mainstream
00:28:28thing
00:28:28at the end of the 90s
00:28:30or mid 90s
00:28:31and so on
00:28:32maybe you recall
00:28:35what it was like
00:28:37back then.
00:28:37We had dialers
00:28:39which were
00:28:40a problem
00:28:41we had
00:28:41all kinds
00:28:42of fraud
00:28:43I mean we still
00:28:43have all kinds
00:28:44of fraud
00:28:45on the internet
00:28:46and the thing
00:28:47is of course
00:28:47that fraud
00:28:49will get
00:28:50more sophisticated
00:28:51that the attacks
00:28:52will get more
00:28:53sophisticated
00:28:54but that is
00:28:55something where
00:28:55we as a society
00:28:56will have to try
00:28:57to keep up
00:28:58as difficult
00:28:59as that is
00:29:00because
00:29:01for the one
00:29:02simple truth
00:29:02I think there is
00:29:04well I'm sure
00:29:05there is no way
00:29:06of going back
00:29:07and I don't think
00:29:08there is
00:29:09a realistic way
00:29:10of slowing
00:29:11down
00:29:11in a way
00:29:12that is not
00:29:13just meant
00:29:14as a disadvantage
00:29:15for smaller
00:29:16competitors
00:29:17and we definitely
00:29:18don't want that
00:29:19also as a side note
00:29:21what I definitely
00:29:22don't want
00:29:22is the EU
00:29:23or any other
00:29:25regulatory body
00:29:26restricting
00:29:27my use
00:29:29and your use
00:29:30of AI
00:29:30because the damage
00:29:32AI can do
00:29:33is totally
00:29:34not related
00:29:35to how
00:29:36you can
00:29:37use AI
00:29:37in your
00:29:38business
00:29:38the bad
00:29:39actors
00:29:39will have
00:29:40access
00:29:40to
00:29:41the AI
00:29:42they need
00:29:42anyways
00:29:43the labs
00:29:44obviously
00:29:44have
00:29:45their
00:29:45frontier
00:29:46and the
00:29:46next level
00:29:47models
00:29:48internally
00:29:49so putting
00:29:50regular
00:29:51business owners
00:29:52or employees
00:29:53or just
00:29:53normal people
00:29:54in a bad
00:29:56spot there
00:29:56or giving
00:29:57them a major
00:29:57disadvantage
00:29:58by restricting
00:29:59which AI
00:30:00they can use
00:30:01or what
00:30:02they are
00:30:03allowed
00:30:03to with
00:30:04AI
00:30:04is definitely
00:30:05the wrong
00:30:06way forward
00:30:07and just
00:30:08a way
00:30:08of kind
00:30:08of giving
00:30:09you a
00:30:10really huge
00:30:11disadvantage
00:30:11because you
00:30:12need access
00:30:14to the best
00:30:15AI models
00:30:15you can get
00:30:16as a business
00:30:17owner
00:30:17as a developer
00:30:18as an employee
00:30:19as a person
00:30:20in order to
00:30:22stay competitive
00:30:23in this fast
00:30:24changing world
00:30:25so that was
00:30:27a rather long
00:30:27rant
00:30:28and my
00:30:29thoughts about
00:30:30about this entire slowdown thing
00:30:32and where we're heading with AI.
커뮤니티 글
아직 글이 없습니다. 이 영상에 대한 첫 번째 글을 작성해 보세요!
이 영상에 대해 글쓰기