The AI Daily Brief: Artificial Intelligence News and Analysis - Meta is Working on a ChatGPT Killer

Episode Date: September 11, 2023

Today NLW looks at a WSJ report that Meta is pushing for a GPT-4 powered Llama 4 to be released open source; a study that shows that ChatGPT is more creative than MBA students; and an Axios poll that shows that half of Americans believe that AI will negatively impact the US presidential elections. TAKE OUR SURVEY ON EDUCATIONAL AND LEARNING RESOURCE CONTENT: https://bit.ly/aibreakdownsurvey Today's Sponsor: Supermanage - AI for 1-on-1's - https://supermanage.ai/breakdown ABOUT THE AI BREAKDOWN The AI Breakdown helps you understand the most important news and discussions in AI.  Subscribe to The AI Breakdown newsletter: https://theaibreakdown.beehiiv.com/subscribe Subscribe to The AI Breakdown on YouTube: https://www.youtube.com/@TheAIBreakdown Join the community: bit.ly/aibreakdown Learn more: http://breakdown.network/

Today on the AI breakdown, we're looking at a new poll about American attitudes towards AI in the elections. Before that on the brief, the Wall Street Journal confirms that Facebook is working on a chat GPT killer. The AI breakdown is a daily podcast and video about the most important news and discussions in AI. Go to Breakdown.network for more information about our newsletter, our Discord, and our YouTube channel. Welcome back to the AI breakdown brief, all the AI headline news you need in around five minutes. Today we kick off with a story that is something we've been following, but which so far has been largely based exclusively on rumor. Well, now we have that other type of rumor, which is called reportage. And yes, what we're talking about is meta's plans to develop a GPT4 level competitor

that they release in their open-source-ish way. Now, you've probably heard me reference this tweet from Jason at AGI-Cowala from August 25th a couple of times now. Jason tweets, Overheard at a meta gen AI social, we have the compute to train Lama 3 and 4. The plan is for Lama 3 to be as good as GPT4. Jason asks, wow, if Lama 3 is as good as GPT4, will you guys still open source it? Yeah, we will. Sorry, alignment, people. As I said, over the weekend, the Wall Street Journal added their journalistic credibility

to this conversation. The piece was titled, Meta is developing a new, more powerful AI system as technology race escalates. So what information did we get here basically just confirmation of what was in that tweet? According to people familiar with the matter, meta aims for Lama 3 to be several times more powerful than Lama 2, and the focus is on getting it to be at least as good as GPT4. Now, a couple other interesting details. One is around timing.

WSJ writes, META expects to start training the new AI system in early 2024, and also how they're doing it. Remember, Lama 2 was announced at a Microsoft event, but it appears like META is trying to move farther away and more independent from that big company. The WSJ writes, Meta is currently building up the data centers necessary for the job and acquiring more H-100s. While Meta joined with Microsoft to make Lama 2 available on Microsoft's cloud computing platform

Azure, it plans to train the new model on its own infrastructure. Now, in terms of how it would be released, the article says, Zuckerberg is pushing for the new model like Meta's earlier AI offerings to be open-sourced and therefore available free for companies to build AI-powered tools. Now, while that is Zuckerberg's intention, apparently Meta's lawyers are a little bit stressed out about it. They apparently fear potential legal ramifications, including lawsuits around copyright and other

legal issues that may arise if meta-created tools are used to spread disinformation or for other bad purposes. Now, because this is something that most people in the space have been talking about for a few weeks now, there wasn't really much chatter on Twitter slash X about it, but I do think that by and large, what's shaping up is a pretty significant set of releases coming later this fall in early next year. We have, of course, Google Gemini, and then potentially Lama 3, all of which have OpenAI's dominance firmly in their sights. The question, as we asked on shows over this weekend, is whether OpenAI is going to continue to pause at this GPT4 level, or whether the temptation and the economic pressure to advance to GPT4.5 or even GPT5 will be too great. Now, at the current

GPT levels, one of the things that some have contented themselves with is that although these machines are good at doing specific tasks, they don't seem to have the capability for novel or inventive or creative thought. Well, not so fast as a group of Wharton professors who tested MBA students against ChatGPT to see who could come up with the most innovative business ideas. The professors who teach innovation and entrepreneurship gave both GPT4 as well as their MBA students the same prompt. Generate an idea for a new product or service appealing to college students that could be made available for $50 or less. From the human students, they randomly selected 200 ideas, and from GPT4, they generated first 100 ideas just with that prompt,

and then another 100 ideas after providing a handful of successful examples, i.e. a version of training. They measured the results in three ways. First speed or number of ideas per unit of time, obviously chat GPT was going to always win that. The second was market testing using a panel of potential customers. Of those, GPT4 also won. The average purchase probability of a human generated idea was 40%, whereas the vanilla GPD4 was 47% and the trained GPT4 was 49%. Finally, looking for exceptional ideas, they looked at only the top 10%, the subset of the best ideas in the pool. Of those 40 ideas, as only five were generated by students, and 35 were generated by GPT4. Now, ultimately, whether this says more about GPT4's creativity or the lack of creativity

among MBA students, I will leave up to you to decide. Moving on to an interesting area of battleground around AI copyright, there has been quite a discussion in the gaming world around game developers taking advantage of generative AI tools. A post on Reddit from the beginning of the month wrote, the game I've spent three and a half years and my savings on has been rejected and retired by Steam today. The post reads, About three to four months ago, I decided to include an optional chat GPT mod in the playtest build in my game, which would allow players to replace the dialogue of NPCs with responses from the chat GPT API. This mod was entirely optional, not required for gameplay, not even meant to be a part

of it, just a fun experiment. It was just a toggle in the settings and even required the playtester to use their own OpenAI API API key to access it. Fast forward to about a month ago when I submitted my game for early access review. Steam decided that the game required. wired an additional review by their team and asked for details around the AI. I explained exactly how this worked and that there was no AI content directly in the build, and even since then, issued a new build without this mod ability just to be super safe. However, for almost one month, they said basically nothing. They refused to give estimates of how long this review would take, what progress they've made, and didn't ask for any follow-up questions or to try to have a

conversation with me. This time alone was super stressful, as I had no idea what to expect. Then today, I randomly received an email that my app had been retired with a generic Your Game contains AI response. I'm in absolute shit. shock. I've spent years working on this, sacrificing money, time with family and friends, pouring my heart and soul into the game, only to be told through a short email, sorry we're retiring your app. Now, there is a lot that we could discuss about this. It is a brutal reminder of the problems of developing on someone else's platform, but it's also to many just insane. To see this entire field summarily rejected from use in the development of games just seems

pretty fundamentally unrealistic. Why Combinator President Gary Tan write, sad to hear Steam is desal. No AI in games is a idiotic. Now, that wasn't the interesting part, but what was interesting was Tim Sweeney's response. Tim Sweeney is the founder and CEO of Epic Games who wrote, put it on the Epic Game Store. We don't ban games for using new technologies. Now, this happened about a week ago, but I just noticed it, and I think it's a really interesting and telling reflection of where a lot of conversations in AI and industry are going to be. In some places, the rejection of AI is going to be used as a business position, while equally in an opposition, the embrace of AI will, too, be used as a way to divide

the world and Market One's service. It will be another bellwether of where general consumer sentiment is to see which of these positions is most rewarded in the market. Now one more little bit of game lore just to close us out, Emmett Shear, the former CEO and co-founder of Twitch, tweeted over the weekend, Remember Black and White? Published in 2001, Abandoned in 2016, it was an incredible premise. You were a god who raised an avatar creature with commands and Pavlovian feedback. It didn't really work as much as I wanted to love it, but with modern AI dot, dot, dot, dot. Now, someone responded to Emmett and said, Demis Hizizhabis was the lead AI programmer for black and white.

Demis Hizizabas is, of course, now the co-founder of Google Deep Mind. Emmett responds, wait, seriously? That's amazing lore. I had no idea. Hey, Demis, if you're willing to share, what are your thoughts on black and white in retrospect? And do you think someone should take another crack with modern AI again? Demis responds,

it was the most advanced AI we could build at the time. Back in those days, the best AI was found in games, in my opinion. But I dream about what kinds of incredible games could be made with today, modern AI might have to scratch that itch at some point. Anyways, friends, that is going to do it for today's AI breakdown brief. Thanks as always for listening or watching, and I will be back soon with the main AI breakdown. Before we get into the main AI breakdown, I want to tell you about today's sponsor, Supermanage. If you work in a professional setting, you probably have some version of a one-on-one meeting, either with the people that work for you or the people that you work with. Unfortunately,

all too often those one-on-one meetings become glorified catch-up calls. Don't you wish you could jump right to the stuff that really matters? That's where SuperManage comes in. Supermanage AI magically distills your team's public Slack channels into a real-time brief on any employee, any time. Catch up on contributions, work in progress, challenges they're facing, sentiment, everything you need to show up ready for a truly meaningful conversation. And it's completely free. Visit supermanage.aI forward slash breakdown today to start making the most of your one-on-ones. And thanks again to Supermanage for sponsoring the AI breakdown. Welcome back to the AI breakdown. Today we are looking at a new poll from Axios and Morning Consult that's all about Americans' beliefs about how AI may impact

the upcoming elections, and it is not good. So today, what we're going to do is look at these poll results and then look at some of the broader conversation around AI in this election cycle. So first of all, one note about methodology. This poll was conducted between August 10th and 13th, so relatively recently. I point that out because one of the things that I've noticed with AI polls is that you'll often find that they were taken months and months ago, or even last year. And obviously, attitudes and awareness in this space is changing very quickly. But that's not a problem in this case, because these results are about a month old. So here are the two big highlights from this poll. First, half of all Americans expect that misinformation spread by AI will, impact who wins the election. On top of that, one-third say they will be less trusting of the election results because of the presence of AI. Now, obviously, election results and contesting election results is more of a hot-button issue in America than it's been for quite some time. And I do believe that that is important context for this poll. Among those who said that AI will decrease their trust in the 2024 election, among people who voted for Biden last cycle, it was 27%.

whereas among people who voted for Trump last time, it was 47%. In other words, the group that already has more misgivings about election results is more skeptical or more concerned about AI impacting election results going forward. Now, beyond just the election focus, there were some other AI stats that were pretty interesting as well. This survey reinforced the idea that more Americans are pessimistic than optimistic about AI, with 36% saying they're feeling pessimistic versus just 26% saying they're optimistic about the future of AI.

It's also quite clear that this is an issue that does matter to the same. them. When asked, how important is it to you, if at all, that the president understands artificial intelligence, 55% said very important, with 26% saying somewhat important, and just 20% combined saying not too important or not at all important. Those numbers were pretty consistent for members of Congress as well, with 50% saying it was very important that they understand, and 28% saying it was somewhat important. Another interesting result was around the questions of who people thought were best suited to regulate AI. When asked, what do you think is the best entity to regulate artificial intelligence, 9% said Congress, 9% said the White House, 26% said a new department of AI government agency,

14% said the private sector, 7% said global organizations, and the largest portion of respondents said none of the above. AI cannot be regulated effectively. 34% gave that answer. Now, obviously, on the one hand, this shows a general lack of trust in institutions, which is certainly a key part of the political landscape right now across basically every issue, but this is also interesting because senators Richard Blumenthal and Josh Hawley have just started to talk about their comprehensive framework for AI regulation in the U.S., which would indeed create a new agency to oversee artificial intelligence. Last question I'll reference has to do with big questions of AI superintelligence and artificial

general intelligence. Do you believe there is a threshold in the evolution of AI after which humans cannot take back control of the artificial intelligence they've created? 21% said yes, definitely, and 44% said yes, probably. Only 22% said no, probably not. And only 13% said, no, definitely not. Now, among those who said yes, which was 65% of respondents, when asked how soon, the most popular answer was in the next five years with 42% of respondents. 37% said in the next 25 years, 12% said in the next year, with 8% saying in the next 100 years and 2% saying more than 100 years. Another way to put the combined responses of those two questions is that a majority of Americans think there is an AI threshold after which we will no longer be able to control it,

and a majority of those people think that threshold will hit within the next five years, which of course probably goes a long way to explaining why 55% of respondents also want the president to have a good handle on this AI situation. Now, there has been a lot of discourse about AI and misinformation in the elections. And in fact, one of the things to ask when reading these poll results is how much these opinions are manufactured by media accounts. In other words, how much are people scared about election tampering from AI because there are constantly articles about election tampering with AI. Now, of course, mainstream media aren't the only people who are concerned about this. In June, former Google CEO, Eric Schmidt, for example, said the

24 elections are going to be a mess. The reason he said was that social media is not protecting us from false generated AI. Now, in terms of what social media platform should do, Schmidt said, what social media should do is mark all the content, know who the users are, and hold people accountable if they violate the law. It doesn't solve the problem of you and I disagree on facts, but at least it establishes a basis that these are humans who are making these claims. Now, on that front, we learned last week that Google would be requiring the disclosure of all AI-generated content in their political ads starting in November of this year. In a blog post, Google said that political ads that feature synthetic content, that, quote, inauthentically represents

real or realistic-looking people or events, must include a, quote, clear and conspicuous disclosure for viewers who might see the ad. Now, these concerns, unfortunately, aren't just theoretical. In June, Ron DeSantis' campaign made news when they shared AI-generated images of Donald Trump hugging Dr. Anthony Fauci. In February, just a few days before the Chicago mayoral primary, an ad was released with the voice of one of the candidates saying, in my day, no one would bat an eye if a police officer killed 17 or 18 people, which wasn't actually a candidate but an AI-generated voice. And while it wasn't an election-related issue, in May, an image of what looked like a big smoke plume following an explosion at the Pentagon, actually tanked the stock market for several minutes before it was proven to be false.

This has led many to make arguments such as this one written by David Gerwitz at ZDNet, were not ready for the impact of generative AI on elections. Now, one of the things that this piece notes is that given how much AI reduces the cost of creation, that it's very likely that generative AI tools will show up in smaller campaigns who don't have war chests that are as big as national campaigns. Indeed, the piece says, expect to see the biggest impact of generative AI far down the ticket. So what's to be done? Well, some are arguing that we need drastic action and we need it immediately.

Last week, Mustafa Sullyman, the co-founder of Infliction and the former co-founder of Google Deep Mind, penned an op-ed for fortune called ban the use of AI in elections right now. Silliman paints vividly a scenario that many have discussed when he writes, Imagine that three days before an election, a video of a presidential candidate using a racist slur spreads on social media. The campaign press office strenuously denies it, but everyone knows what they've seen. Outrage seized around the country, polls nosedive. Swing states suddenly shift toward the opponent, who, after all expectations, wins. A new administration takes charge. But the video is a deep fake.

One so sophisticated, it evades even the best fake detecting neural networks. A grainy, authentic-looking video or an audio recording of a politician defaming a voting block can be engaging and convincing. However, trusting our eyes and ears is no longer possible. Now, Suleiman's argument is that what should be done is that regardless of any other comprehensive AI legislation, we should make it a policy right now to ban the use of AI in elections. And yet, many are skeptical. Accelerate Harder quote tweeted Suleiman and said,

the belief that banning AI-based speech in elections will somehow stop deepfakes from popping up would be adorably naive if the person arguing for it wasn't also positioning himself to massively profit from AI regulatory capture. The actual answer to this threat is to continue not to restrict free speech and be patient while the electorate adapts to a changing information environment. It might be bumpy for a little bit, but the last two elections were also bumpy and we survived. Now, I wanted to use this quote, one, because a lot of the commentary right now is coming from Mustafa Selleiman and a lot of the skepticism of his content comes from this idea that he's setting up inflection for regulatory capture. Or maybe more broadly, that he's part of a cohort of entrepreneurs, which includes people like

Sam Altman from OpenAI, who are effectively trying to pull the ladder up after themselves, given that they have an early lead, better access to compute, more advanced models, and so don't mind so much if regulation makes it harder for others to compete with them. That's obviously not the subject of today's video, but it's almost like lurking subtext in every conversation we have and makes it much harder to have a good faith conversation because people's assumptions are that anyone from those big companies is just promoting a regulatory capture agenda. But the second part of Accelerate Harder's response that I want to hone in on is this idea that the right answer is to trust the electorate to adapt. At the end of August, the economist actually made a similar point. They published an essay called

How Worried Should You Be About AI Disrupting Elections? The short answer, quote, it matters less than you might think. They write, what could LLMs change in 2024? One thing is the quantity of disinformation If the volume of nonsense were multiplied by a thousand or a hundred thousand, it might persuade people to vote differently. A second concerns quality. Hyper-realistic deepfakes could sway voters before false audio, photos, or videos could be debunked. A third is micro-targeting. With AI, voters may be inundated with highly personalized propaganda at scale. Networks of propaganda bots could be made harder to detect than existing disinformation efforts are. Voters trust in their fellow citizens, which in America has been declining for decades may well suffer as people begin to

doubt everything. And yet they write, this is worrying, but there are reasons to believe AI is not about to wreck humanity's 2,500-year-old experiment with democracy. Many people think that others are more gullible than they themselves are. In fact, voters are hard to persuade, especially on salient political issues such as whom they want to be president. Ask yourself, what deepfake would change your choice between Joe Biden and Mr. Trump? The multi-billion dollar campaign industry in America that uses humans to persuade voters can generate only minute changes in their behavior. Now, the other reasons they say to be not as concerned are that deepfakes still aren't perfect, and that social media platforms are focused on the risks, but I think that the most likely outcome

is what they alluded to before, that people are going to adapt much more quickly than we think to the presence of this sort of AI, and just be inherently much more skeptical of everything that they see. Now, it should be noted that this does not mean that there won't be issues. In fact, we don't really know what it's like to operate in a context where people simply don't believe anything that they see. The fact that the presence of AI forces people to be more skeptical could create ripple effects that are damaging to democracy in their own right. But I am much, much more skeptical of the type of scenario laid out by Mustafa Suleiman and many others like him, where some major crazy deepfake that comes out at the last minute in an election somehow significantly sways it. I just think that especially

within call it 24, 72, 96 hours of the election, everyone is going to be so keyed up and so tuned up looking for anything that even has a whiff of generative AI in it, that it's going to be really, really hard for something like that to actually meaningfully impact things. But who knows? The one thing that is for sure is that this will be a part of this cycle and every cycle going forward. And given that, there's really nothing else to do but dive in and understand it's coming. Thanks for hanging out with the AI breakdown. And until next time, peace.