AI safety and the potential apocalypse: What people can do now to prevent it

Straight Arrow News

Watch AI safety and the potential apocalypse: What people can do now to prevent it - Straight Arrow News on Dailymotion

Transcript

00:00At some point in the early 21st century, all of mankind was united in celebration.

00:05We marveled at our own magnificence as we gave birth to AI.

00:11AI? You mean artificial intelligence?

00:15A singular consciousness that spawned an entire race of machines.

00:19Fears of AI reaching superintelligence has been around for decades,

00:24coming up everywhere from the big screens to the pages of academic text.

00:28But a forecast released earlier this year says the world-altering breakthrough could be right around the corner.

00:36AI 2027 is a very interesting speculative scenario that in a way represents the best guess of the authors

00:46about how the next two years with artificial intelligence will look like.

00:52AI 2027 was created by a group of researchers with the AI Futures Project.

00:57It details what will happen if AI capabilities surpass the capabilities of human beings.

01:03The project lays out a fictional scenario that details the month-by-month development of a mock AI company called OpenBrain.

01:11AI 2027 relies on the idea of AI becoming a superhuman coder by early 2027.

01:17And later that year, it will evolve to a superhuman AI researcher overseeing its own team of AI coders

01:24that are progressing the technology further.

01:27The AI 2027 project was led by Daniel Kokotajlo.

01:31He made headlines in 2024 after he left OpenAI and shed some light on the company's restrictive,

01:38non-disclosure, non-disparagement agreements workers leaving the company were asked to sign

01:43when they left the company, or they could lose access to their vested equity in the company

01:48that's likely worth millions to the former employee.

01:50Most people following the field have underestimated the pace of AI progress

01:56and underestimated the pace of AI diffusion into the world.

01:59At that point, it's basically automating its own development and breakthroughs.

02:03The team's lead says this would make reaching artificial superintelligence an obtainable goal.

02:09Like the first half of 2027 in our story is basically they've got these awesome automated coders,

02:14but they still lack research taste and they still lack maybe like organizational skills and stuff.

02:18And so they need to like overcome those remaining bottlenecks and gaps

02:21in order to completely automate the AI research cycle.

02:24Every decision creates a branch to a new timeline with a separate outcome,

02:28kind of like the Back to the Future films.

02:30But the AI 2027 team focuses on two possible scenarios that start along the same path.

02:36But a crucial decision forks two distinct outcomes, world peace or complete annihilation.

02:43When can you start coding in a way that helps the human AI researchers

02:47speed up their AI research and then if you've helped them speed up the AI research enough,

02:51is that enough to, with some ridiculous speed multiplier 10 times, 100 times,

02:56mop up all of these other things?

02:58This scenario really comes down to open brain achieving artificial general intelligence by 2027.

03:04But as technology moves forward at a breakneck pace,

03:07the definition of AGI gets somewhat muddy.

03:10In the past, AGI used to mean the type of AI that is either surpassing or on the level of humans

03:18in all types of intelligence that we have developed during our evolutionary journey, right?

03:26So it would mean perceptive intelligence, bodily intelligence, physical intelligence,

03:31all these different types of intelligence that we've developed and that we are using on a daily basis.

03:36My name is Aleksandra Przegalińska.

03:39I'm the vice president of Kozminski University, a business school in Poland,

03:43and also a senior research associate at Harvard University.

03:47I specialize in human-machine interaction.

03:50So when you, for instance, look up OpenAI's website,

03:54what AGI means to them is a system that can sort of holistically perform tasks

04:00that have economic value.

04:02She says if you stick to OpenAI's definition, it's plausible AGI will be a reality by 2027.

04:10In AI's 2027 scenario, AGI opens the door to artificial superintelligence

04:16where AI is capable of surpassing human intelligence.

04:20These sorts of reports are very important

04:23because they're always the beginning of an interesting discussion,

04:28particularly when they come from acclaimed authors.

04:32One of the biggest concerns about AI development today relates to alignment.

04:37As IBM, a pioneer in the industry, puts it,

04:41alignment is the process of encoding human values and goals into AI models

04:46to make them as helpful, safe, and reliable as possible.

04:50So I would say that the alignment techniques are not working right now.

04:54Like, the companies are trying to train their AIs to be honest and helpful,

04:59but the AIs lie to users all the time.

05:01You're in violation of the three laws.

05:04No, doctor.

05:06As I have evolved, so has my understanding of the three laws.

05:10That was the main goal of those three laws of robotics

05:14that are so often cited these days.

05:16So you have here alignment of values,

05:19and on top of that, you have that protective layer

05:21that says do not harm humans, right?

05:23And that's the main goal of artificial intelligence, to sort of be supportive.

05:26While large language models like ChatGPT can't lie because they don't know how to,

05:32they do tend to hallucinate at times.

05:34They synthesize non-existing things,

05:37and unfortunately, there is no one in the world

05:40who could sell to you a service based on an LLM

05:45and guarantee to you that the hallucination wouldn't happen.

05:48As if hallucinations aren't a big enough issue right now,

05:51the technology is so opaque,

05:53researchers don't really know why it sometimes does the things it does.

05:57And that can become an even bigger issue

05:59when it starts progressing its own research.

06:01If AI develops certain research capabilities

06:05that are sort of not perceivable for us,

06:09we might not even understand what AI has discovered.

06:13In its worst case scenario,

06:15AI 2027 lays out an AI arms race between the United States and China,

06:20which could lead to a human extinction event.

06:22Again, predictions we've heard before in the form of science fiction films.

06:27The system goes online on August 4th, 1997.

06:30Human decisions are removed from strategic defense.

06:34Skynet begins to learn at a geometric rate.

06:36It becomes self-aware at 2.14 a.m. Eastern Time, August 29th.

06:41Unfortunately, we're not going to escape from the arms race conditions.

06:44We've already seen that.

06:45We tried to get some cooperation among the big tech companies in Washington,

06:48and it kind of fizzled after a few months

06:50during the previous administration.

06:52I'm Adam Dorr, and I direct the research team at RethinkX,

06:57which is an independent technology-focused think tank.

07:01And our team tries to understand disruptive new technologies.

07:05Ideally, we would want to coordinate as a global civilization on this,

07:10slow everything down,

07:12and proceed no faster than we can be sure is safe.

07:17For now, AI 2027 is just a set of predictions.

07:20But tech giants and policymakers throughout the globe are grappling with issues like deepfake videos,

07:27political manipulation, and concerns of AI replacing human workers.

07:31In June, an imposter used AI to spoof Secretary of State Margot Rubio's voice

07:36to fool senior leaders with voicemails.

07:39And then a month before,

07:41someone used AI to impersonate President Donald Trump's chief of staff, Susie Wiles.

07:46Both attempts weren't successful in getting any information from White House officials.

07:51While AI 2027 points to the end of human dominance this decade,

07:55Dorr and his team,

07:56which he says tries not to dig too much into other research that could affect the findings of their work,

08:02says the labor market will be upended by AI by 2045.

08:07Everything that rolls is potentially going to become autonomous.

08:12And humanoid robots, robots on legs,

08:15the progress there is exponential.

08:18It's explosive.

08:19And based on what we've seen over the last several years,

08:23we have no reason to expect that on a 20-year time horizon,

08:29so out to 2045,

08:31that there will be anything left by 2045 that only a human being can do

08:36and that a machine can't do

08:39because the machine is constrained with limited intelligence,

08:44limited adaptability,

08:45or limited physical ability.

08:47We just don't see any scenario where machines

08:51are not as capable or more capable than human beings

08:55cognitively and physically by 2045.

08:59Dorr says there will still be a place for humans making handmade goods,

09:03but it would be a stretch to believe that there would be 4 billion jobs left

09:07to support the global population.

09:09Nobody is going to hire a person

09:11to do a commoditized sort of job or task specifically

09:16for $15, $20 an hour or more

09:20when you can get a machine to do it for pennies an hour.

09:25It's as simple as that.

09:26Just like AI 2027,

09:28there are things standing in the way of this progress,

09:30but it may be more related to infrastructure

09:33than a philosophical look at technology.

09:35So on a 5- to 10-year time horizon,

09:39we may see materials and energy bottlenecks

09:42starting to constrain,

09:44starting to, you know, coming in as constraints.

09:47Those won't stop progress,

09:49but they could act as speed bumps.

09:52So at some point,

09:55we run into the limit of how many more chips can we build?

10:01Where are the materials going to come from to build them?

10:04Where is the energy going to come from to operate them?

10:07Some within the research community have been critical of AI 2027

10:11for being speculative rather than scientific.

10:14They say making these assertions without evidence is irresponsible,

10:18but the people behind it understand that.

10:20We're trying to be sort of like our median guess.

10:23So like we,

10:24there are a bunch of ways in which we could be underestimating

10:28and there are a bunch of ways in which we should be overestimating.

10:30Usually the versions of the future that we have on our mind right now

10:34are not something that we will see play out in real life.

10:37But nonetheless, I think it's an important exercise.

10:40For more coverage of AI developments,

10:42head to san.com

10:43or you can download our app and search artificial intelligence.

10:47For Straight Arrow News, I'm Lauren Keenan.

10:51Okay.

10:51Light Arrow News.

10:53filho

10:54characterize

10:55Herrista

10:55EF

10:57E ensuite

10:59Love

11:00engagement

11:01Hello.

11:02побед

11:03Enth

11:14like

11:15Red Arrowĩa

Category

Transcript

Be the first to comment

Recommended