Charlz Guybon posted:

Things seem to be going...poorly

https://twitter.com/ChrisMurphyCT/status/1640186536825061376

BrainDance posted:

The "synthesis" was my own guess in my quick one off example that I typed literally as an example of potential capabilities of AI, not as me stating this is the thing that AI will definitely do in exactly this way. But as an example to show the kinds of things emergent abilities in AIs lead to which you took incredibly literally and as a complete thesis for some reason.

The articles are from my bookmarks, and absolutely do show the potential for AI in drug development among many, many other things. A quote from one itself, "AI can be used effectively in different parts of drug discovery, including drug design, chemical synthesis, drug screening, polypharmacology, and drug repurposing."

So, you can go ahead and argue that AIs wont do a thing I literally never said they would do. We may have plenty more drugs than we can currently test, but that doesn't mean help in identifying new pharmaceutical pathways to targeting different diseases is unhelpful or unwanted, (otherwise, what are they even paying my friend for?)

"nor does the description from your friend" I didn't even tell you what his description was.

gurragadon posted:

I just want to address this point before the thread gets too far along. When I asked to remake an AI thread because the Chat-GPT thread was gassed I was told to keep it vague. The reason given for gassing the Chat-GPT thread was it was too specific to Chat-GPT in the title and the title was misleading. In the thread I hope people will refer to the specific AI programs they are referring to, but unfortunately or not this thread was directed to stay vague.

BrainDance posted:

ChatGPT itself likely won't, but it's hard to say what even larger language models will be capable of because, like I was saying, we've seen a bunch of unexpected emergent abilities appear from them as they get larger.

What I was getting at from the start is this, it's not a surprise it's learned to do chemistry better than gpt 3.5, even if it wasn't really intentionally trained to https://www.jasonwei.net/blog/emergence

Just making the models larger makes them more capable of doing things, sometimes some things that weren't even explicitly trained to know or even expected to know. Even when it comes to being factual, GPT4 still hallucinates a lot, but you look at very small models and they're not coherent enough to be accurate, you look at larger models and they lie a lot and are wildly off base, larger and they're more accurate, larger still, more accurate, etc.

A thing Ive been saying since I sorta stumbled into finetuning an incredibly small model (1.3B) into being roughly as good as GPT3 on one specific task (but only that specific task) is that I think transformers can potentially be very useful for a huge variety of things we haven't really tried them on yet if they're purpose trained for that specific thing. GPT is very general, so it's going to be a limited jack of all trades. But a model that follows the same principles and is only trained to do one specific task but of the same overall complexity might be very, very capable and very useful at that one task.

BrainDance posted:

It's like a couple of you guys are trying to take things extremely literally and completely miss the point. Yes, I am aware it can't really do chemistry. You need arms to do chemistry. ChatGPT is in some sense immaterial and has no real corporeal form which you also need to do chemistry.

The whole point was that emergency abilities appear in AIs as they get larger, sometimes unexpectedly. Language models at a certain complexity are sometimes able to create a kind of new information from the things they were trained on. That was literally it. AI in a broader sense has applications in many areas. I believe LLMs as they get more complex, non-general ones, may be able to do some of this (not literally gpt4 chatgpt.) What this will actually be in the future? It's not entirely predictable.

Edit:

That is not at all what happened. And I can't see a way you could read the conversation and come to that conclusion. A person said AI wouldn't be used that way because we don't need assistance discovering drugs, we have too many drugs we can't test already. And those papers were cited to show, we already use AI for this, there is already a drug discover niche for AI.

esquilax posted:

The GPT-4 paper includes some discussion of its capability to use outside chemistry tools in the context of potentially risky emergent behaviors - specifically the capability to propose modifications to a chemical compound to get a purchasable analog to an unavailable compound. See section 2.10 (starting pdf page 55) in the context of section 2.6. I don't understand the chemistry but I'm guessing this is what the tweet was about, layered through 3-4 layers of the grapevine.

https://cdn.openai.com/papers/gpt-4.pdf

StratGoatCom posted:

I would avoid those people, because generative AI is poison for IP.

StratGoatCom posted:

Nope, it is very long standing doctrine that machine output cannot be copyrighted, as attempts at brute force of the copyright system would be anticipatable since Orwellian book kaleoscopes. Machine or animal generated, versus stuff touched up with is is not going to be allowed.

quote:

The Office also agrees that the selection and arrangement of the images and text in the Work are protectable as a compilation. Copyright protects “the collection and assembling of preexisting materials or of data that are selected, coordinated, or arranged” in a sufficiently creative way. 17 U.S.C. § 101 (definition of “compilation”); see also COMPENDIUM (THIRD) § 312.1 (providing examples of copyrightable compilations). Ms. Kashtanova states that she “selected, refined, cropped, positioned, framed, and arranged” the images in the Work to create the story told within its pages. Kashtanova Letter at 13; see also id. at 4 (arguing that “Kashtanova’s selection, coordination, and arrangement of those images to reflect the story of Zarya should, at a minimum, support the copyrightability of the Work as a whole.”). Based on the representation that the selection and arrangement of the images in the Work was done entirely by Ms. Kashtanova, the Office concludes that it is the product of human authorship. Further, the Office finds that the compilation of these images and text throughout the Work contains sufficient creativity under Feist to be protected by copyright. Specifically, the Office finds the Work is the product of creative choices with respect to the selection of the images that make up the Work and the placement and arrangement of the images and text on each of the Work’s pages. Copyright therefore protects Ms. Kashtanova’s authorship of the overall selection, coordination, and arrangement of the text and visual elements that make up the Work.

quote:

The Office will register works that contain otherwise unprotectable material that has been edited, modified, or otherwise revised by a human author, but only if the new work contains a “sufficient amount of original authorship” to itself qualify for copyright protection. COMPENDIUM (THIRD) § 313.6(D). Ms. Kashtanova’s changes to this image fall short of this standard. Contra Eden Toys, Inc. v. Florelee Undergarment Co., 697 F.2d 27, 34–35 (2d Cir. 1982) (revised drawing of Paddington Bear qualified as a derivative work based on the changed proportions of the character’s hat, the elimination of individualized fingers and toes, and the overall smoothing of lines that gave the
drawing a “different, cleaner ‘look’”)

quote:

To the extent that Ms. Kashtanova made substantive edits to an intermediate image generated by Midjourney, those edits could provide human authorship and would not be excluded from the new registration certificate.

Count Roland posted:

A side effect of Dall-e and similar programs is that there's a lot of AI art being generated, which shows up on the internet, which is trawled for data, which is then presumably fed back into AI models. I wonder if AI generated content is somehow filtered out to prevent feedback loops.

gurragadon posted:

I'm a little confused as to why OpenAI can't train ChatGPT on copyrighted works as long as they aren't just replicating the work wholesale. Anybody who is training to be a writer will train themselves on copyrighted works, and every work is derivative of some experience the writer had. No writer emerges from the ether to release fully new ideas into the world, they would need to have there own language to do it.

Obviously, when ChatGPT reproduces a Mickey Mouse script or Midjourney copies a picture of Mickey Mouse you have a claim. But to claim that something couldn't train on copyrighted materials would basically change how art and writing come to be. I mean if I couldn't train to write on anything not copyrighted, that would be like open source stuff and anything before 1924 I guess?

Edit: I know ChatGPT doesn't have personhood so it's not a direct 1 to 1 comparison. But I don't how else you could really train AI to be useful without copyrighted material. It would be so far behind, like an AI that only had information about a time before computers somehow.

gurragadon posted:

Personally, I think copyright law is written in a way that tramples over people's rights and using information that is available in the world isn't trampling on peoples rights. Not from a tech start only point of view either, excessive copyright laws just stifle creativity and innovation in my opinion, which is where my stance is coming from. Everything is built on something else; you can't build something entirely independent.

Your point about people paying or not paying for the information we consume is a good one. Is there a way to quantify the amount of money an average person pays out for media of all kinds through there life? I get a ton of copyrighted material for free just by virtue of existing in society. I see images that are copyrighted, read copyrighted things and hear copyrighted music. But I DO pay for some it, that is extremely true.

Would that be a fair amount to pay for a "copy" of ChatGPT? ChatGPT is a written format specifically so I think it would be cheaper than other forms. I guess OpenAI could get a library card and that would open up a ton of writing. A lot of stuff that isnt available would be left out though unless OpenAI did specific agreements with universities and other institutions.

SCheeseman posted:

You're continuing to miss the point that the biggest problems generative AI creates won't actually be solved by obstructing only a subset of the businesses wanting to develop and exploit this stuff. I'm all for legislation that actually helps people who had their lives ripped apart by automation (the actual problem that capitalism has been causing for a long time!), but all you're advocating for is maintaining an already hosed status quo.

gurragadon posted:

For this application I think it is inappropriate to apply copyright law at all really. ChatGPT was trained on 45 terabytes of text, which is just an insanely large amount of data. I don't think anyone copyright owner can claim any kind of influence on the program itself. An individual text is just a tiny bit of data that doesn't exert any influence by itself, the program needs a huge amount of text to make patterns.

Maybe if your copyrighted material was tokenized in ChatGPT like some of these glitch tokens were that were posted in the previous ChatGPT thread. There were certain usernames from Reddit that were from a counting subreddit that ChatGPT made a token out of because they were seen so often and other things like that.

But if you made OpenAI get permission from every content creator to create the ChatGPT it would be prohibitively expensive in time and money. I mean you're right, that does read as me complaining it would be too inconvenient to pay people for their work. ChatGPT needs so much text, and it needs to be diverse if it going to be useful, that I think it is valid. It just seems to be fundamentally unworkable with the way AI technology works currently for one company to be able to pay copyright fees on every piece of writing ChatGPT saw.

Which if you don't think we should use AI programs or have them its fine and could be correct, but I don't know if copyright law is the way to go about it because it's kind of an indirect way of banning AI. It's definitely a big problem with ChatGPT and the likes devaluing art and writing.

Do you not have a problem with ChatGPT because it is free? It probably won't be forever, but would a theoretical free AI program trained this way be ok to use copyrighted data?

I think these models should be treated differently because they are different in the way they are constructed. The model needs the writing to become useful. I'm not really saying it should be treated like a human exactly, but the way it works is just so similar to how humans learn that I don't think we can have AI without giving it a lot of data.

The "If a person thing" is really valid and it is very off base for this current generation of AI programs. But the problem I have with it is that we don't know when it won't be so off base and the improvements are pretty quick, which makes it hard to avoid veering into for me.

SCheeseman posted:

Impede, obstruct, whatever. In any case what you want isn't going to make AI art generators uneconomical, it'll make the 'legal' ones economical only for the entrenched IP hoarders. What you want changes nothing about how people will in actuality be exploited and may even serve to make it worse!

Gentleman Baller posted:

One thing I don't really understand here, is what the legal difference is between text interpreted computer generated works and mouse click interpreted generated works?

My zero legal education impression is that I could open Paint, tell it I want a 400x600 picture with a star in each corner and the text, "Barry's Mowing Service" written in Comic sans in the middle and I could secure a copyright on that image. If I was to ask a home-ran image generator to do the same thing, using incredibly precise language rather than mouse clicks to tell it what to draw and what font to use, would that be okay under current law? Is it based on the precision of the human input? The randomness of the AI input? Or something else?

cat botherer posted:

It's not, though. In Anglo systems, this stuff falls under fair use. IDK much about other systems, but the answer is not to increase the power of IP holders. That just empowers rent-seeking behavior and creates unintended consequences. I think people really need to take a deep breath here.

XboxPants posted:

Yeah, that doesn't even seem to be the big issue to me. Let's say I'm the small artist who draws commissions for people for their original DnD or comic book or anime characters, and I'm worried how AI gen art is gonna hurt my business. Worried my customers will use an AI model instead of patronizing me. So, we can decide that we're going to treat it as a copyright infringement if a model uses art that they don't have the copyright for. That will at least keep my work from being used as part of the model.

But, that won't protect me.

Disney, as has been noted, has an incredible library of art including every Marvel comic going back 70+ years. They may not have my works to use, but they'll just train a model based on the art they have access to.

So, as that artist who was doing commissions for people, I'm just as hosed now. My old customers just get a license to use Disney's AI now, instead of using Midjourney. StratGoatCom's suggestion hasn't helped me at all.

StratGoatCom posted:

Do you know, have links outside of the AI sphere?

KwegiboHB posted:

I refuse to link directly to the gofundme, I don't want them funded. You can easily find that with a simple web search if you want to.
I'm saying, scroll down on this page https://copyrightalliance.org/about/who-we-represent/

No one in that group should be begging regular people for money for lobbyists. It's a classic bait-and-switch. "Hooray everyone we did get the laws changed! In exactly the ways that help us and not you! By the way, you can't draw even fanart of our stuff anymore. We're off to play with our new AI models that you can't use now! Oh, and firing another 7,000 people while we're at it."

IShallRiseAgain posted:

That uh doesn't actually change anything. Like sure there a lot of unions, trade associations, and artists' rights groups that are part of it, but that doesn't change the fact that companies like Disney are on it too. It makes it a bit more unclear about the actual motives, but it doesn't change the fact that companies which hold a lot of rights to art see some advantage to advocating for it. Also, these unions and other artist advocacy groups are pretty much forced to fight for this because its members very much want this even if they don't understand the full implications of what would actually happen. I don't think their members would care or believe it if they tried to explain the actual probable consequences of this going through.

Owling Howl posted:

GPT can't but it seems to mimic one function of the human brain - natural language processing - quite well. Perhaps the methodology can be used to mimic other functions and that map of words and relationships can be used to map the objects and rules of the physical world. If we put a model in a robot body and tasked it with exploring the world like an infant - look, listen, touch, smell, taste everything - and build a map of the world in the same way - what would happen when eventually we put that robot in front of a mirror? Probably nothing. But it would be interesting.

KillHour posted:

I should stress to both sides of this conversation that the reason AI researchers are talking about safety is it's really important that we need to solve a lot of very difficult problems before we create AGI. So there is both an acknowledgement that what we have now isn't that and also that we have no idea if or when it will happen. Nobody thinks a LLM as the architecture exists today is going to become an AGI on its own. The problem is we don't know what is missing or when that thing or things will no longer be missing.

"How sentient is ChatGPT" isn't really an interesting question, IMO. "What could a true AGI look like architecturally" is more interesting, but not super pressing. "How can we ensure that a hypothetical future AGI is safe" isn't very flashy, but is incredibly important for the same reason that NASA spent a ton of money smashing a rocket into an asteroid to see if they could move it.

Nobody sees this poo poo coming until after it happens.

KillHour posted:

Neither of those problems require a deep understanding of AI research and I think most AI researchers would object to being told to work on this as clearly outside of their field of expertise. AI researchers are, and should be, concerned with the issues of the technology itself, not the issues of society at large. The latter is what sociologists do (and there is a large crossover between sociology and technology, but it's a separate thing). Your framing of AI ethicists is more like "Techno ethicist" because the actual technology in use there isn't particularly relevant. I also object to your framing of these issues as more "real" than misalignment. Both are real issues.

Lastly - killer post/av combo there.

Leon Trotsky 2012 posted:

His argument kind of seems like a situation where he is saying: "Don't make AI evil" and is not really a useful assessment on a practical level.

His argument is essentially: AI will be able to do many great things that help society, but bad people exist and could use AI for bad purposes. Therefore, we should not pursue advanced AI.

The same thing could be said of computers and the internet, but we didn't shut those down because they enabled trillions of dollars in fraud and scams over the course of their existence.

His other concern is the "Terminator Scenario" that seems like a crazy reason to shut down research decades in advance of this situation. It also doesn't really prevent someone like China or the U.S. government in secret, from pursuing these kinds of research if you ban it from the public.

The concerns of "if AI is available to everyone, then bad people can use AI" and "it is possible for the worst outcome to happen" are both obviously true, but not useful in determining what to do about it.

quote:

Until last year, he said, Google acted as a “proper steward” for the technology, careful not to release something that might cause harm. But now that Microsoft has augmented its Bing search engine with a chatbot — challenging Google’s core business — Google is racing to deploy the same kind of technology. The tech giants are locked in a competition that might be impossible to stop, Dr. Hinton said.

quote:

Down the road, he is worried that future versions of the technology pose a threat to humanity because they often learn unexpected behavior from the vast amounts of data they analyze. This becomes an issue, he said, as individuals and companies allow A.I. systems not only to generate their own computer code but actually run that code on their own.

Leon Trotsky 2012 posted:

He says he didn't sign on to one of the letters calling for a moratorium on AI research because he was still working at Google, but he agreed with it.

the other hand posted:

For those who have the time to listen to (very) lengthy audio, this guy’s YouTube channel has a lot of interviews with leading people in AI research and industry. Some of his recent guests were the Boston Dynamics CEO, a computational biology professor from MIT, and the CEO of OpenAI, which makes the GPT software (video below).

I’m not qualified to provide any commentary on these, but they seem intended for the general public and I’m finding them super interesting and educational. The OpenAI CEO interview includes topics like political bias, AI safety, and AGI.

https://www.youtube.com/watch?v=L_Guz73e6fw

quote:

Peace, love, and Hitler: How Lex Fridman's podcast became a safe space for the anti-woke tech elite

In early October, Ye culminated a dayslong spree of inflammatory comments about Black and Jewish people with a tweet declaring, "I'm a bit sleepy tonight but when I wake up I'm going death con 3 On JEWISH PEOPLE." The outrage was immediate. Companies like Balenciaga and JPMorgan announced they were cutting ties with the rapper formerly known as Kanye West. Twitter blocked him.

The scientist turned podcast host Lex Fridman, meanwhile, invited Ye to appear on his podcast.

During the 2 ½-hour interview, Ye doubled down, comparing Planned Parenthood to the Holocaust and blaming his personal and financial problems on Jewish doctors and record-label executives. Fridman, who's Jewish, pushed back at times, telling him, "When you say 'Jewish media,' there's an echo of a pain that people feel."

But Fridman still aired the interview in its entirety.

"I believe in the power of tough, honest, empathetic conversation to increase the amount of love in the world," Fridman tweeted alongside the link.

Over the past few years, Fridman, 39, has gone from an unknown academic researcher to a social-media celebrity, professional agitator, and member of Elon Musk's inner circle. In 2019, while working at MIT, he coauthored a controversial study of Tesla's Autopilot concluding that human drivers remained focused while using the semiautonomous system. Musk, Tesla's CEO, was so enamored that he flew Fridman to the company's headquarters to tape an interview. Seemingly overnight, episodes of Fridman's podcast began racking up millions of views.

In his podcast, Fridman asks world-renowned scientists, historians, artists, and engineers a series of wide-eyed questions ("Who is God? "What is the meaning of life?"). It all seems innocent enough. But recently, "The Lex Fridman Podcast" has become a haven for a growing — and powerful — sector looking to dismantle years of "wokeness" and cancel culture. Episodes include Joe Rogan talking about being spied on by intelligence agencies (8.7 million views); Jordan Peterson questioning the validity of climate-change models (9.5 million views); and, of course, Ye ranting about Jews running the media (4.6 million views). And as some of tech's most vocal leaders have taken a sharp right turn — the bitcoin mogul Balaji Srinivasan's slamming the Food and Drug Administration's regulations, Musk's blaming "communism" for turning his transgender daughter against him — Fridman has become their mouthpiece.

His podcast provides a reputable forum (he's a scientist, after all) to discuss vaccine safety, race science, and the importance of traditional gender roles. Fridman has said his ultimate goal is "to allow each other, all of us, to make mistakes in conversation." But Lior Pachter, a computational biologist at Caltech who's become one of Fridman's most outspoken critics, said some scientists and academics fear Fridman is contributing to the "cacophony of misinformation."

As Fridman's career has blossomed, his ambitions have grown. He's pitched a startup to "put robots in every home" and talked about launching a social-media company (though there doesn't appear to be movement on either front). In December, Fridman asked Musk if he could run Twitter. "No salary. All in. Focus on great engineering and increasing the amount of love in the world. Just offering my help in the unlikely case it's useful," Fridman tweeted at the CEO, to which Musk replied: "You must like pain a lot. One catch: you have to invest your life savings in Twitter and it has been in the fast lane to bankruptcy since May. Still want the job?"

Though Fridman has touted affiliations with MIT and Google, AI and machine-learning experts who spoke with Insider said Fridman lacks the publications, citations, and conference appearances required to be taken seriously in the hypercompetitive world of academia. When AI professionals on Twitter challenged Fridman's research methods for the Tesla study, he blocked them en masse. Soon after the study was published, Fridman moved from his prestigious MIT lab to an unpaid research role.

But to the hordes of young men fed up with the so-called mainstream media, Fridman is a truth teller. He's a genius scientist who doesn't talk down to them and who's challenging the status quo, one interview at a time.

"If you're into flat Earth and you feel very good about it, that you believe that Earth is flat, the idea that you should censor that is ridiculous," Fridman said on the neuroscientist Andrew Huberman's podcast. "If it makes you feel good and you're becoming the best version of yourself, I think you should be getting as much flat Earth as possible."

---

At the beginning of each episode, Fridman addresses his listeners from his home studio, wearing a boxy black suit and sitting in front of a black drape. (He's joked that he looks like a "Russian hitman.") He welcomes fans in an ASMR voice before launching into hourslong conversations about everything from extraterrestrial life to sex robots. Since launching the show in 2018, he's taped more than 350 episodes.

When it comes to Fridman's interviews, he's all peace and love. "I'm a silly little kid trying to do a bit of good in this world," he's told his listeners. He presents himself as a hopeless romantic striving for human connection. His only foes are the critics and journalists intent on spreading cynicism and negativity — and even to those people Fridman responds with heart-emoji tweets.

The Hallmark-card persona Fridman has cultivated makes it all the more jarring when he muses about whether Hitler would make a good guest. "If you talk to Hitler in 1941, do you empathize with him, or do you push back?" he asked on a recent episode. "Because most journalists would push, because they're trying to signal to a fellow journalist and to people back home that this, me, the journalist, is on the right side. But if you actually want to understand the person, you should empathize. If you want to be the kind of person that actually understands in the full arc of history, you need to empathize."

He's also expressed interest in interviewing Andrew Tate — the misogynist influencer Fridman's described as "an interesting person who raises questions of what it means to be a man in the 21st century" — even after Tate was detained in Romania on human-trafficking and rape charges.

Fridman's podcast wasn't always so incendiary. It launched in 2018 as "The Artificial Intelligence Podcast," drawing a modest audience of autodidacts with conversations about self-driving cars, robotics, and deep learning, a form of AI that aims to use neural networks to replicate the processes of the human brain. As the podcast gained momentum, bigger names like Mark Zuckerberg and Jack Dorsey came on to talk about topics such as the metaverse and life in a simulation.

"The Lex Fridman Podcast" offered a rare opportunity to listen to four-hour conversations with luminaries of tech and science. Fridman's open-ended, nontechnical questions made complex topics feel accessible to layman listeners. To this day, many credit Fridman for inviting guests from across the political spectrum, even when he disagrees with their views. Bhaskar Sunkara, the founder and publisher of the socialist magazine Jacobin who appeared on Fridman's podcast in December, praised Fridman's interviewing style.

"I really do think there's a deep earnestness," Sunkara said. "That's, I think, part of his strength, the fact that he can just have these 'aw, shucks' conversations with people."

But in the past year or so, there's been an undeniable shift as a rage has built among tech magnates who want to say it like it is, with no repercussions. They've decried identity politics and political correctness in the mainstream media. They want unfettered capitalism and deregulation. They don't want to hear the whining about how their products affect people's health or happiness or well-being. They don't want to be criticized for hoarding wealth. Scientific facts, they argue, don't care about your feelings.

Fridman has latched onto the momentum. His guest list these days has turned into a who's who of the "intellectual dark web," a movement of controversial thinkers claiming to offer an alternative to woke culture.

Fridman makes the perfect spokesman for these ideas because he seems so harmless. He's not vitriolic like Ben Shapiro, creepy like Jordan Peterson, or crude like Joe Rogan. He's just a sweet-seeming, self-deprecating guy who wants to ask questions and hear everyone out. Even Hitler.

But Pachter warns that Fridman is pulling the wool over people's eyes.

"He says, 'I'm just asking questions,'" Pachter said, "but actually he's sort of two-siding issues where they don't really have two sides."

---

Fridman was raised in Moscow in the 1980s, where he developed a deep suspicion of socialism and communism. His grandfather was a decorated World War II gunner, his grandmother a survivor of the 1930s Stalinist famine that killed millions of Ukrainians.

When Fridman was about 11, his family relocated to the Chicago suburbs. He's said the adjustment to American culture was difficult. Years later, when Nan Xie, Fridman's fellow doctoral candidate at Drexel University, asked Fridman if he still spoke any of his native tongue, he replied, "I speak 120% Russian," Xie recalled.

By the time Fridman entered academia, his family name was well established. His brother, Gregory, and his father, Alexander, who was one of the Soviet Union's most accomplished plasma physicists, were both star professors at Drexel, where Fridman earned his undergraduate and doctoral degrees.

Fridman majored in computer science, focusing his studies on active authentication, a way to identify a device's user based on their habits and gestures. Fridman's Ph.D. advisor, Moshe Kam, joked that he often had to pull Fridman's head out of the clouds, as his lab desk was stacked with books on philosophy, beat poetry, and Russian expressionism.

In 2014, around the time he was finishing his doctorate, Fridman debuted a YouTube channel: a dry mix of AI lectures and interviews with martial-arts experts. He also tried his hand at poetry. The works, published on a since deleted blog, mostly focused on love and longing, which would become omnipresent themes in Fridman's podcast, with titles including "Oversexed Wolf Among Undersexed Sheep" and "You Can't Quote Nietzsche While Exposing Your Breasts."

One haiku, titled "A Blonde College Girl," reads: "a blonde college girl / climbs a tree in a tight shirt / and black yoga pants."

While he pursued his side projects, Fridman joined Google's secretive Advanced Technology and Projects team. He worked for Project Abacus, a task force studying how to use active authentication to replace traditional passwords. Fridman left the company after six months. (Google did not respond to requests for comment about Fridman's tenure.)



At his next stop, at MIT's prestigious AgeLab in 2015, Fridman really began to cultivate his public persona. His team aimed to use psychology and big-data analytics to understand driver behavior. An AgeLab colleague of Fridman's told Insider that Fridman increasingly focused on attention-grabbing studies with results that were harder to quantify. Many of these studies were published only on ArXiv, an open-access service that accepts scholarly articles without the kind of quality control required by other publications.

The former AgeLab staffer recalled Fridman telling him that he wanted "to amass a giant amount of followers."

Fridman started to do just that. He opened up his lectures to the wider Cambridge community, packing people in so tightly that they'd have to sit on the stairs. He uploaded the lectures on his personal YouTube channel with MIT's logo prominently displayed, and he updated his Twitter profile photo to one of him looking professorial, standing in front of a blackboard full of equations. (Pachter noted that the equations were from a course entirely unrelated to Fridman's own field of study.)

Filip Piękniewski, a computer-vision and AI researcher who's become one of Fridman's most vocal critics, said it was "clear from the very beginning that he's positioning himself to become one of the celebrities in this space."

---

In 2019, while working at MIT's AgeLab, Fridman posted his controversial Tesla study online. It found that "patterns of decreased vigilance, while common in human-machine interaction paradigms, are not inherent to AI-assisted driving" — in other words, that drivers using semiautonomous vehicles remain focused. The findings were a shock to the industry, contradicting decades of research suggesting that humans generally become distracted when partially automated systems kick in.

The MIT seal of approval was likely enormously valuable to Tesla. Its Autopilot feature had come under intense scrutiny over several widely publicized fatal crashes involving Tesla vehicles. The company was hit with a class-action lawsuit that described Autopilot as "essentially unusable and demonstrably dangerous." But Musk insisted his technology did not require human intervention, going so far as to brand the feature "Autopilot" instead of "Copilot."

Academics in AI began to pick apart the study's methodology. They criticized the small sample size and suggested the participants likely performed differently because they knew they were being observed, a phenomenon known as the Hawthorne effect. Missy Cummings, a former MIT professor who's served as the senior advisor for safety at the National Highway Traffic Safety Administration, has called the report "deeply flawed."

"What happens is people like Elon Musk take this and then try to make a big deal about it and say, 'I've got science on my side,'" Cummings said.

"I personally do not consider Lex a serious researcher," Cummings recently told Insider.

When Anima Anandkumar, a well-known AI expert, tweeted in 2019 that Fridman ought to submit his work for peer review before seeking press coverage, Fridman blocked her and many of her colleagues, some of whom had never even engaged in the discussion. (Fridman did not respond to several interview requests and requests for comment. He told the historian Dan Carlin during one episode: "Journalists annoy the hell out of me ... So I understand from Putin's perspective that journalism, journalists can be seen as the enemy of the state.")

Many of Fridman's peers had another reason to be suspicious of the study. Fridman's admiration for Tesla's CEO was well documented: He was an active participant in Tesla fan forums, he'd been photographed with Musk's Boring Company flamethrower, and in a 2018 tweet that's since been deleted he asked Musk to collaborate on a fully autonomous cross-country drive. Musk had even tweeted about Fridman's Tesla-friendly research in the past.

One former MIT colleague of Fridman's said many people in their field believed that being closely associated with Musk could be a career boon. "Lex was relatively excited to get in touch with Elon Musk and get into his good graces," said the former colleague, who asked to remain anonymous to avoid professional repercussions.

A week after Fridman posted about the study on Twitter and Tesla message boards, Musk invited Fridman to Tesla's offices. There, Musk sat for a 32-minute interview for Fridman's podcast in which he argued that within the next year Tesla's semiautonomous systems would be so reliable that if a human intervened while driving, that could actually decrease safety.

Soon, coverage of Fridman's study appeared in tech and business publications, including this one. Fridman's show became a sensation. Before the Musk interview, the podcast's catalog had a total of 1 million downloads. Suddenly, it wasn't rare for episodes to garner millions of views, with guests including Bridgewater's Ray Dalio and Facebook's chief AI scientist, Yann LeCun.

But as well-known billionaires flew out to chat with Fridman, the study — along with a second Tesla-centric study Fridman published — was removed from MIT's website without explanation. That year, Fridman quietly switched from AgeLab to an unpaid role in the department of aeronautics and astronautics. In 2020, Fridman rebranded his show as "The Lex Fridman Podcast."

An autonomous-vehicle investor and strategist described the removal of the study as disastrous. "Is there anything more damning than being disappeared from the MIT website? Like, post-Epstein, that's basically the next worst thing to actually being arrested," the investor said, referring to the school's scandalous history with Jeffrey Epstein. (The investor asked to remain anonymous, saying: "I've already been victimized, threatened, and harassed by the army of Musk people. He's Musk's guy and I just don't need it.")

None of Fridman's current MIT colleagues that Insider contacted agreed to be interviewed for this story. Only Sertac Karaman, the director of the Laboratory for Information and Decision Systems, where Fridman has since relocated, provided a brief statement: "Dr. Fridman has been a research scientist at MIT LIDS and in our research group. I have known him for many years, and been very impressed by his ideas and his research accomplishments."

The automotive journalist E.W. Niedermeyer told Insider that the 2019 study helped create the enduring perception that an automated system is ultimately safer.

An AI expert who knows Fridman thought that perhaps the podcast host abandoned academic rigor in pursuit of fame. (The expert was also granted anonymity so as to avoid harassment from Fridman's fans.)

"It's almost as if he sold his academic soul to get an interview with Musk," the expert said.

---

Today, Fridman lives in a spartan four-story townhouse in Austin, Texas. He shares the house with only a fleet of Roombas he says he once trained to scream in pain as a test of human empathy toward robots. (He joked on Huberman's podcast that he decided not to upload the experiment on his YouTube channel because it risked affirming some people's perception of him as a "dark person.")

Fridman often refers to the "Russian soul," an identity that helps explain his reverence for both science and romanticism. In 2019, Fridman incorporated a startup rooted in the idea that everyone should have a personalized AI system that would not only know them inside and out but provide tenderness and emotional support.

The company, called Lexhal, would encompass a more advanced version of the algorithms that already pervade our lives — like Netflix's recommendations for what to watch or Amazon's suggestions for what to buy. Fridman told Huberman that the goal was "exploring your own loneliness by interacting with machines."

Fridman offered the example of a smart fridge that witnesses its owner's late-night ice-cream binges. "You shared that moment — that darkness or that beautiful moment — where you were just, like, heartbroken," he said. "And that refrigerator was there for you. And the fact that it missed the opportunity to remember that is tragic. And once it does remember that, I think you're going to be very attached to the refrigerator."

While it doesn't appear Fridman has made much progress with Lexhal beyond ideations, Fridman seems obsessed with the idea of what it means to be a person in this age of technological advancement — and more specifically, what it means to be a man. He's embraced the kind of hypermasculine grind culture that has swept Silicon Valley in the past few years in which primarily men are looking to extend their lives and increase their sperm count. "I love fighting," Fridman told the philosopher William MacAskill on an episode of his podcast that's since been deleted. "I've been in street fights my whole life. I'm as alpha in everything I do as it gets."

Fridman has said he maintains a rigorous schedule of long, intensive work sessions followed by 10-minute social-media breaks. He eats a carnivore-keto diet complemented by supplements from brands like Athletic Greens that sponsor his show. His punishing exercise routine includes judo, Brazilian jiujitsu, and long runs where his only soundtrack is brown noise that allows him to meditate.

As he wrote in a 2014 poem called "Instructions for Being a Man": "a man should be alone and quiet. / exploring but certain: a dog without a chain. / a man should work hard, even when tired, / and never, never, never complain."

For Fridman, Musk appears to be a guiding light. Musk has appeared on the podcast (three times, in fact). So has Grimes, Musk's former partner, as well as her close friends Liv Boeree, the poker champion, and Aella, the sex researcher. Even conversations with other guests that ostensibly have nothing to do with the Tesla CEO often turn to praise of his various projects. Meanwhile, Fridman's profile has risen with every Musk retweet. (Musk's account is Fridman's most-interacted-with account, while Fridman's is in Musk's top 15.)

The two men are at the helm of a new wave of techno-optimism positing that the quicker we bring along bold technological advances, the better. Piękniewski described it as a "semireligious movement" where anyone who calls for guardrails is seen as an enemy of progress.

In recent months, as the general population has interacted with artificial intelligence in unprecedented ways, the popular narrative has taken on an increasingly fearful tone. Chatbots trained on large language models like ChatGPT have gone on deranged rants and threatened users. Pundits have predicted that automation will replace millions of jobs and produce a flood of misinformation. Doomsayers say AI could spell the end of humanity. In the past two weeks, Fridman has hosted both Sam Altman, the CEO behind ChatGPT, and Eliezer Yudkowsky, a researcher convinced that AI will wipe out the human race.

Yet Fridman remains bullish. His body of work seems to center on the idea that individuals can be trusted to use technology to become better versions of themselves. Just as he argued that Autopilot drivers wouldn't fall asleep at the wheel, he maintains that his fans can separate fact from fiction in an algorithmically driven mess of information.

As Fridman's influence grows, so does the criticism of him. The autonomous-vehicle investor and strategist called him a "useful idiot." Reddit and Hacker News, Silicon Valley's favorite message board, went wild after Fridman posted a 2023 reading list including "The Little Prince," "1984," and Fridman's personal favorite, "Animal Farm." Both platforms are full of frustrated fans accusing Fridman of relying too much on the MIT name and lacking expertise on complex topics.

Nassim Nicholas Taleb, the mathematical statistician known for writing the book "The Black Swan," blasted Fridman on Twitter in January. "Sending mail with heading 'Quick Note From MIT' when podcast has nothing to do with MIT is shoddy as hell," Taleb tweeted.

Taleb said he'd turned down 10 invitations from Fridman to appear on his podcast. "I have been convinced that there is something that is not solid about the person," Taleb told Insider.

Whether Fridman has the credentials or smarts to be this tech figurehead remains to be seen. But one thing's for certain: He has the support of some very powerful people.

"How do you become a priest of the religion?" Piękniewski said. "Well, the best way is to get blessed by an even bigger priest."

SCheeseman posted:

What interests me about this is where the line for parody and satire is drawn now that AI tools allow for uncannily realistic impressions. Dudesy reportedly implicitly talks about the artifice of their "AI" on the show with their fanbase being in on it, the word kayfabe comes up often on the show and and in the fan community, so it's not like this could be framed as some kind of conniving scam. They tipped their hand to anyone paying attention (I wasn't).

Are voice transformers over the line? Deepfakes, even if it's for the purposes of comedy? Is parody infringing if it's unfunny?

quote:

"Defendants always presented the Dudesy Special as an AI-generated George Carlin comedy special, where George Carlin was 'resurrected' with the use of modern technology," the lawsuit argues. "In short, Defendants sought to capitalize on the name, reputation, and likeness of George Carlin in creating, promoting, and distributing the Dudesy Special and using generated images of Carlin, Carlin’s voice, and images designed to evoke Carlin’s presence on a stage."

SCheeseman posted:

Claiming it's from Virtual Carlin in the same way that pro wrestling is claiming that it's a sport.

Since when was parody something you accuse someone of?

The act explicitly makes light of the death of George Carlin in the act itself, it's not an imitation of George Carlin but an imitation of a now dead George Carlin who has been resurrected with a (fictional) AI, something that is part of the greater meta narrative of the Dudesy podcast (or less charitably, deception, but that's harder to argue this now that it comes out that the fanbase was in on it). Is that enough "comic effect", taking from the typical dictionary definition, to define as parody? Maybe not, but I don't think it helps the case of the Carlin estate.

SCheeseman posted:

This reads like commentary to me and it takes up about a third of the act, using the premise of an AI of George Carlin as part of it's commentary, going on to touch on issues related to using AI generated likenesses of real people and other politics associated with it. They're making clear points about the implications of what they're doing (or appearing to do).

quote:

The purpose of the media's use of a person's identity is central. If the purpose is "informative or cultural" the use is immune; "if it serves no such function but merely exploits the individual portrayed, immunity will not be granted."

Staluigi posted:

quote:

The internet loves a fiasco. Whether it be 2017’s infamous Fyre Festival, 2014’s sad ball pit at Dashcon, or last year’s muddy hike for freedom at Burning Man, we love to marvel at events that make big promises but flop spectacularly. It’s the online equivalent of slowing down in your car to look at a giant wreck.

Enter February 24’s disastrous Willy Wonka chocolate factory experience in Glasgow, Scotland, which instantly became a viral sensation when pictures emerged of the sad spectacle. Organized by House of Illuminati, whose AI-heavy website promotes the company as a place where “fantasy and reality converge,” the event was advertised as an immersive experience for families themed around the classic Roald Dahl novel and movie franchise (and with tickets that cost up to $44). But the end result was anything but magical.

The event took place in a largely empty warehouse/venue that had been decorated, so to speak, with a few scattered candy-themed props, hanging plastic backgrounds, and a small jumping castle. As bewildered families wandered the space, they were greeted by actors dressed in shoddy costumes who read from scripts they later said they suspected were “AI-generated gibberish.” Rather than feasting on an array of sweets, children were given a few jellybeans and a quarter cup of lemonade. Soon, angry parents called the police as they demanded refunds from organizers, who canceled the event before the day was out.

While many photos have emerged of the Wonka fiasco, one picture in particular has grabbed the internet’s attention. It shows an actor wearing the Shein equivalent of an Oompa Loompa costume and looking slightly dead in the eyes as she stands in a smoky room behind a table covered in so much scientific equipment that countless people online compared it to a “meth lab.”

So who was she? Meet Kirsty Paterson.

While a subsequent photo shows there was at least one other Oompa Loompa at the table that day, it’s Paterson who has emerged as the viral star and defining image of the Wonka fiasco. X posts have proclaimed her as one of the standout memes of 2024 so far. People are comparing the image to works of art or making art of their own showing Paterson as the Mona Lisa. Others are wondering if she is really the missing Kate Middleton in disguise. Many are predicting, with good reason, that she’ll be among the Halloween costumes you see this year.

Paterson, who first spoke with the Daily Mail about her experience with the event, opened up in an interview on Wednesday about what it’s been like to go viral for one of the most mortifying days of her life, and for a picture, she stresses, does not actually resemble her in real life.

Tell me about yourself. What do you do when you’re not dressed as an Oompa Loompa?
I’m 29. I want to act more in children’s entertainment because I enjoy it. I’ve got a lot of energy, and I like being around kids. But it’s more of a side hustle. I’m also a fire dancer, and I’m just trying to build up my yoga teaching at the moment.

How did you get roped into this event? Was it listed somewhere?
It was listed on Indeed. I don’t normally get my acting jobs through Indeed, but I just thought, All right. To be honest, I was a wee bit skeptical, because it was not through an agency. They were offering £500 for two days of work, so I decided to go.

You went to the warehouse on Friday for a dress rehearsal. What happened at that meeting?
I was shocked, to be honest with you. I wasn’t expecting it to be like that. It did seem like there was a production going, but to me it wasn’t a finished production — just the start. It was the first time me and the other actors met. I’ve never had a script the night before ever in my life, so when I got the script the night before I was like, Oh, this is not …

They kept going on about how you could just improvise. I was skeptical of them saying that, too, because if you’d written a script, then you’d probably pride yourself on what you’d written, right? You wouldn’t want people to improvise! I kind of thought it was AI-generated, but by this point I’d signed the contract.

Were you shown the costumes that day? The set?
This is the mental thing about the costumes! Given the amount people were paying to go and the amount they were paying us, I thought they’d have sufficient costumes. So at this point, I was like, “What is it you want us to wear? What’s the makeup like?” Because it’s a Willy Wonka experience, you need to have good costumes, because it’s all meant to be a bit imaginative. But we didn’t know what we were wearing at that point.

They said they were going to be working through the night or whatever, so I just assumed it was going to be a lot better.

What happened when you turned up on Saturday?
It was the exact same! I was like, I don’t know if I actually want to do this. But I’d signed the contract, and part of me didn’t want to disappoint the kids going to this. Honestly, it was bad enough. I knew it was shocking, but I know I’m good at what I do, so I was like, If I can bring a wee bit of something good to this …

At this point, they gave us the costumes. They were so inappropriate for what it was. It was just strange, and they were really cheaply made. It was almost like they were secondhand.

Were you wearing orange body paint in that photo? I couldn’t tell.
No! They didn’t have any makeup or anything. It’s something I’ve never experienced ever in my life.

By this point, I’m judging myself for letting this go on as long as it did. But we got the costumes, and I started seeing the kids coming in, and they were all dressed up, and I was just feeling so guilty.

Tell me about that moment as the kids started coming in. Did you see the innocence leave their eyes?
They were quite upset. I think they were confused. But it was more the parents. I just looked at them, and I think they must have known when they looked at me. An older couple said to me, “I really, really hope you get paid well for this.” And I was honest and I said, “I’m this close to walking out. This is not what I signed up for.” But I didn’t want to let the people around me down. The actors I was working with are amazing people, and this has got nothing to do with them. So I just thought, I’m going to make the best of this.

When the first round of kids came in … Don’t get me wrong, it’s still an incredibly, shockingly bad set, but we did our lines and everything well. We just had a bit of fun with it. I don’t know how else you can put sprinkles on poo poo, but we were trying to be the sprinkles on poo poo.

I was going around and feeling really embarrassed. After we did it the first time, the organizers were like, “Just abandon the script and let the guests walk through.”

When you say “walk through” — how long are we talking here to walk through this event?
I’m telling you this would have taken about two minutes. I’ve never experienced anything like this in my life.

Let’s talk about this photo that’s gone viral. What are you doing at this moment here?
It doesn’t even look like me! I was thinking when it came out initially, Oh, people aren’t going to know it looks like me. Maybe I can get away with this. But no, it went completely global.

In the exact moment of the photo, they’d told us to abandon the script. They had this “Jellybean Room,” but they eventually ran out of jelly beans. I was already rationing the jelly beans to three per kid, and that was me being generous. I wanted to give the kids all the candy. So we had no jelly beans, and people were coming up to me. It was just humiliating. I was starting to get angry. The other Oompa Loompa came over at this point and I went, “Where is everyone?! Why am I left here on my own?! Where is everyone else?!” You know how they talk about “me contemplating my life”? This is me contemplating my life.

I’m laughing about it now, but I was so angry for the kids and the parents. I know people spent a lot of money coming here. It’s a disservice to what I do. Eventually, I just walked off. I was like, “I’m done.” But someone got a picture of me.

Are you aware just how viral this picture is?
I’m not going to lie or sugarcoat this: This has been quite a lot for me. I find it funny and I can make a humorous joke about it, but the flip side of this is that this is embarrassing for what I do, and I hope this doesn’t tarnish that. It’s as if it’s been edited. Obviously, because it’s such an ugly photo, people were commenting on it and saying I look ugly or like a meth head. I found the negative comments really hard, but I do see the funny side of it. I know I’m all-right looking.

I suppose that’s good it doesn’t resemble you.
Me and all my close friends recognized me! And my friends know I’m the only one who would do something as daft as this.

To be honest, I don’t know how viral this has gone. I turned off my phone for a day because I was like, Right, this is too much. My phone was going off constantly, and I didn’t know how to navigate it.

So how big has this gone? Because this is the first interview I’ve done where I’ve really spoken about it.

It is very viral. I’m not going to lie. But a lot of pictures from the event have gone viral.
I don’t blame it for going viral. I just hope everyone gets their money back, because I’ve not been paid for this. None of the actors have. So it’s like we’re kind of going through all this for nothing.

I just need to see the funny side of it. Anyone who knows me knows I can take a joke. I just didn’t expect it to go this wild.

Have you seen the people saying they’re going to dress up like you for Halloween?
No, but if you want to, you can!

Boris Galerkin posted:

How good are the LLMs at being multilingual, like how a person raised in a multilingual household would be?

Been thinking about this because the other day I was watching TV and there was a random scene where the language switched to [non-English language]. I understand and speak this language, but there was a word I didn't know.

I said "Siri, what does [foreign word] mean in English?" and Siri couldn't understand that I was switching language for that word. I would have accepted an explanation in English.

I tried "Siri, [in foreign language: what does this word mean in English]?" and Siri transcribed my words into the closest approximate English syllables I guess, which was gibberish. I would have accepted an explanation in the foreign language.

I asked about this in the iPhone thread since it was a Siri question and I know Siri isn't a LLM (right?) but it's spurred some additional discussion about how Siri sometimes can't distinguish perfectly between numbers like "15" and "50" for some people.

This is just an example. But in real life real world, when I talk to my family or some friends we do switch language like that and it's completely natural/normal.

Should clarify that this is in the context of voice assistants.

SaTaMaS posted:

It's pretty hard to define a sense of "language" other than as a collection of associations between words, which is exactly what LLMs are.

GABA ghoul posted:

You are making a lot of assumptions about how language processing works in human brains.

Natively bilingual people can effortlessly jump between speaking language A, language B or a random mixture of them in just a single sentence without any feeling that they are speaking different languages. It feels more like "language" is just another classification in the brain that each word and grammar rule has. Like an apple is associated with the category of "fruits" and broccoli with "vegetables", this word/grammar rule is associated with "English" and that with "German". Like I can effortlessly speak only about fruits, I can effortlessly use only English words.

I think it would be really strange if humans had some specific natural mechanism in the brain to separate languages because AFAIK it wasn't really a common scenario in our evolutionary past that you needed to use more than one language.

SaTaMaS posted:

People such as Ferdinand de Saussure have already worked to define language. He argues that words acquire meaning through their relational positions in a language system rather than through direct links to the external world (aka structuralism). Tokens aren't the same as words, but words are made up of tokens. LLMs rely on the statistical relationships between words in large datasets to predict and generate text, and the context in which a word appears is crucial for its interpretation.

SaTaMaS posted:

It's not at all true that an LLM can't learn arithmetic problem solving, for example https://machine-learning-made-simple.medium.com/how-google-built-the-perfect-llm-system-alphageometry-ed65a9604eaf

GABA ghoul posted:

This is what you wrote:

Humans don't have a "sense of language" either when speaking habitually, as any natively bilingual person will tell you. They freely interchange words and grammar of language A, language B or the pidgin mixture of both, without even being consciously aware of doing it. At no point is there any reasoning about the abstract concept of language A, B or the pidgin involved(unless you make a conscious decision to edit the generated speech in your head to strictly conform to one of them before you say it out aloud).

You assume that the way LLMs generate speech is different from the way humans do, which we just don't know. It all looks suspiciously similar at a first glance though.

SaTaMaS posted:

I think you're really underestimating the technology here. There's no reason it can't be trained on arithmetic problems and I assume another LLM already has been. "All" that AlphaZero did in order to beat all humans at Go was combine a Transformer (aka the same fundamental technology as an LLM) that suggested some next likely moves with a Monte Carlo Tree Search to figure out which of the suggested moves would be the best. There's not a lot that looks like human reasoning going on in either case, and the LLM/Transformer was only 50% of the solution.

Lucid Dream posted:

It's really not very complicated. The LLM itself is not very good at math, but it can be trained to utilize external tools such as a calculator. Whether you think that counts as the LLM "learning to do math" or whether it counts as the LLM "solving the problem" is obviously subjective. I think it counts, but we can agree to disagree.

quote:

Well, they could attempt to calculate the answer and they'd probably do a hell of a lot better than me if you gave me the same problems