The 10 Biggest AI Events of 2023
00:00Today on the AI breakdown, we are looking at one list of the 10 biggest events in AI of2023 and arguing about how correct it is and what might have been left off.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.Hello, friends, we are in end-of-year coverage time.That means it's list season.And I actually love this time.I think that lists are really fun.
00:34I think they're great ways to reflect on what has been and what will come.And so today we are looking at an R's technical list of what they're arguing are the 10 biggestAI stories of 2023.They call it a song of hype and fire, which is a great headline.So what I'm going to do is I am going to read excerpts from their list and then talk a littlebit editorialize, if you will, about whether I agree how big this event is or whether I havesome disagreement.Now, one thing that I will start with right away is that I think that if I were curating
01:04list, I would make ChatGPT's November 30th, 2022 launch an honorary 2023 event. Of course,in point of fact, it was 31 days before 2023, but there's no denying that it was the eventthat kicked off the long year of 2023 and this entire phase of generative AI. So to the extent thatyou have creative and editorial control, perhaps a worthy inclusion. So these are roughly inchronological order, sort of. And so we'll just go through them as they are presented.First on their list, Bing Chat loses its mind. They write, in February, Microsoft unveiled BingChat,a chatbot built into its languishing Bing search engine website. Microsoft created the chatbotusing a more raw form of OpenAI's GPT4 language model, but didn't tell everyone it was GPT4 at first.
01:51Since Microsoft used a less conditioned version of GPT4 than the one that would be released in March,the launch was rough. The chatbot assumed a temperamental personality that could easily turn on usersand attack them, tell people it was in love with them, seemingly worry about its fate,and lose its cool when confronted with an article we wrote about revealing its system prompt.Some people thought BingChat was sentient, despite AI experts' assurances to the contrary.It was a disaster in the press, but Microsoft didn't flinch.And it ultimately reigned in some of BingChat's wild proclivities and opened the bot widely to the public.Today, Bing Chat is known as Microsoft co-pilot and it's baked into Windows.
02:23Now, I think, to the extent that we are looking for singular or seminal events,The specific one to mention when it comes to Bing Chat is the conversation had with New York Times reporter Kevin Ruse.That's the one where at various points it said that it loved him. It gave itself a name. It started to reveal parts of its quote-unquote personality.It basically felt like there was a whole set of things going on behind the scenes that were not just an LLM predicting the next word in a non-sentient way.That article, I think, had a huge impact on starting the nervousness feeling that many have continued to have about AI throughout the year.Next up, U.S. Copyright Office says no to AI copyright authors. In February, the U.S. CopyrightOffice issued a key ruling on AI generated art, revoking the copyright previously granted to theAI-assisted comic book Zaria of the Dawn in September 2022. The decision, influenced by the
03:11revelation that the images were created using the AI-powered mid-jurney image generator,stated that only the text and arrangement of images and texts were eligible for copyright protection.It was the first hint that AI generated imagery without human-authored elements could not becopyrighted in the U.S. The stance was forfurther cemented in August when a U.S. federal judge ruled that art created solely by AI cannot becopyrighted. Now, this is a significant issue. Copyright issues are dominating the legal conversationaround AI, but it's not just this question of whether AI-created works can be copyrighted,but the question of copyright and AI training. In fact, that, I think, is the bigger issue,
03:46although there's not one crystal clear event, that a list like this might be able to point to.Still, what these current suits find in terms of whether models broke copyright in training oncopyrighted works, will have very significant implications for the future of AI model training.So I think that if we're going to identify copyright, we should identify the whole batch ofissues around it.Next up on the list, the rise of Meta's Lama and its open weights direction.On February 24th, meta released Lama, a family of large language models available in differentsizes that kick started in Openweight's large language model movement.
04:18People soon took things into their own hands when they leaked Lama's weights, crucial neuralnetwork files that had previously only been provided to academics onto BitTorrent.Soon, researchers began fine-tuning Lama and building off of it, competing over who could buildthe most capable model that could run locally on non-data-center computers.In tandem, Meda's Jan Lacoon quickly became a vocal proponent of open-AI models.In July, meta-launched Lama 2 and even more capable LLM, and this time they let everyonehave the weights.And in early December, mixtral 8x7B reportedly matched GPD 3.5 in capability, which was a landmark
04:47achievement for a relatively small and fast AI language model.Clearly, companies with closed approaches such as OpenAI, ironically, Google and Anthropic,are going to have a run for their money in the coming year.Now, this I definitely agree with,and I think it's completely reasonable to point to Lamaas a way to sum up the whole insurgency of open source approaches.I think the best exemplar of this realization came back in Maywhen a leaked internal Google document claimed, speaking of Google,
05:13quote, we have no moat and neither does open AI.The note basically argued that the breakout forcein generative AI in LLMs that year was not one of the big labs,but was the incredible amount of innovation in the open-eastern,source space. They also noted somewhat ruefully that meta was getting a lot of the value of thatbecause it was happening in their ecosystem. Now, recently, Mistral has definitely been stealing some ofmeta's thunder, but I agree with Ars Technica when they say, the 2024 is poised for a big fightbetween increasingly capable and increasingly smaller open source models versus the large models
05:45of the closed labs. Next on the list, GPT4 launches and scares the world for months. On March 14th,open AI released its GPT4 large language model with claims that it exists.limited human-level performance on various professional and academic benchmarks,and a specific document that describes attempts by researchers to get a raw version of GPT4 to play outAI takeover scenarios. That set the doom ball rolling. On March 29th, the Future of Life Institutepublished an open letter signed by Elon Musk, calling for a six-month pause in the developmentof AI models more powerful than GPT4. The same day, time published an editorial by Les Wrong founderEliezer Yudkowski, advocating that countries should be willing to, quote,
06:21destroy a rogue data center by Airstrike if they are seen building up a GPU cluster thatcould train a dangerous AI model because otherwise, quote, literally everyone on earth will dieat the hands of a superhuman AI entity. This section then goes on to talk about a number of otherpivotal AI safety moments, including Biden giving remarks, Jeffrey Hinton resigning from Google,Biden meeting with tech CEOs at the White House, AI executive signing a statement warning thatAI could end humanity. And concludes, eventually the fear and hype began to settle down, butthere's still a contingent of people who are convinced that a theoretical superhuman AI is anexistential threat to all of humanity, bringing a bubbling undercurrent of anxiety to every
06:54AI advancement. Now, this very much needs to be split into two, maybe three, maybe more differentbullets on our top events of the year. First of all, GPT4 wasn't just significant because it scaredpeople. It was significant because it was incredibly powerful and has remained state of theart throughout the entire year. Indeed, it's only in the last two weeks that we've had somethingthat purports to beat GBT4 in, of course, Google's Gemini Ultra, but since no one can actually usethat model until next year, we don't actually know if it does indeed outperform GBT4,and the way that they are measuring it is imprecise relative to the way that GBT4 was measuredagainst the same benchmarks. Point being, GPD4 is significant because of how much it pushed
07:35the field, and the very surprising situation that for nine months now, no one has been able tooutdo it. It is leading to questions at the moment of whether we're actually running up againstsome technical barriers, and is worthy of a spot basically all on its own. Now, many of the thingsdiscussed in this bullet, however, I think are also deserving of spots on this list. The six-monthpause letter, while obviously ineffectual in actually getting a six-month pause, absolutely did itsjob in jump-starting the AI safety conversation in the public mind. I think one could make an argumentas well that Jeffrey Hinton's resignation from Google and his subsequent media tour were the secondpart in a one-two punch that really put AI safety on the mainstream public's map. I'm not as sure
08:16that I agree with the idea that the fear has settled down. And indeed, I think that these conversationsare actually going to come to a head around specific policies heading into next year.Hello, friends. One quick note before we get back to the rest of the episode, registration forJanuary's AI Education Beta is now officially open. It's open until just Friday at 1159 p.m. Easterntime. You can find the link to learn more and register at bit.ly slash AI beta. Now, this is an experimentthat I've been running all throughout December, in which every day I drop a new video tutorialor a case study, and usually partner it with a challenge, the idea of which is to get youlearning about all of these new different AI tools, as well as specific strategies for the
08:58most frequently used, like ChatGPT or Dali, and then gets you actually testing them out in thereal world with real use cases, and hopefully applying them back to your personal or professionalpursuits as well. The first month has gone incredibly well. People seem to be really likingthe video content as well as the incredible community that's forming. And part of that is thatit's a group of really serious people. This is a paid experience. It's $20 a month. Part of thereason for that is that I want you guys to judge this content on the basis of whether it's actuallyworth that much to you. And second, I wanted it to be full of really serious people who are intenton applying AI to their lives in some real and significant way. Anyways, I would love to have more AI
09:36breakdown listeners participate in January. Content will start on January 3rd after the end of theholiday season. And again, the link to find out more and to register is bit.ly slash AI beta.That's BIT.L.L.Y. slash AI beta. And now back to the show.Next on ours's list, AI art generators remain controversial but continue to grow in capability.2023 was a big year for leaps and capability from image synthesis models.In March, Mid Journey achieved a notable jump in the photorealism of its AI generatedimages with version five of its AI image synthesis model, rendering convincing people with five-fingeredhands. The pace of change didn't stop, with V5.1 coming in May and V5.2 launching in June.
10:16Also in March, we saw the launch of Adobe Firefly, an AI image generator that Adobe saysis trained solely on public domain works and images found in its Adobe Stock archive.And OpenAI's Dolly 3 took prompt fidelity to a new level in September, raising interestingimplications for artists in the near future. On this one, I don't really have that many notes.I agree entirely that 2023 was the breakout year for image generators. They are not theoreticaland future-oriented like video generators.They are here now.They are being used every day.
10:45They're being used multiple times a day by people like me.20203 was absolutely the year of the AI art generator.Next, bullet, AI deepfakes have a deeper impact.Throughout 2023, the wider implications of image, audio, and video generatorsbegan to take hold.Several controversies emerged, including fairly convincing AI-generated images of Donald Trumpgetting arrested and the Pope in a puppy jacket in March.Also that month, news broke about a scam where people were mimicking the voices of people's
11:09loved ones using AI and routing it through telephone calls to ask for money. Now, this one is interesting.I would actually argue that AI deepfakes have had a dramatically smaller impact this year than manypeople would have thought they would have. For example, those Trump arrest photos didn't actuallycause any real controversy. Neither really did the Pope in a puffer jacket photo. Yes, they showed howgood the technology was, but they didn't influence public opinion in a pivotal moment. And indeed,even when it comes to these scams, while they are very serious in getting more so,it's just been more quiet than I would have thought.Now that said, I think we are heading for a much more chaotic 2024 when it comes to all of this.
11:46The election cycle, for example, means that there is a context where deepfakes could havea significant impact, and there is definitely a growing crisis around AI-generated nude picturesthat is not going to be easy to solve.Next event on Arge Technica's list is the one that I most definitely disagree with being on thislist. It's AI writing detectors promised results but don't work. They write, the emergence ofchat GPT led to an existential crisis for educators that rolled over into 2023, with teachers andprofessors worrying about synthetic text replacing human thought and class assignments.Companies quickly emerged to capitalize on these fears, promising tools that would be
12:21able to detect AI written text. We soon began hearing stories of people being falsely accused of editingchat GPT to write their work when, in fact, everything had been human written. Now, they're notwrong that AI writing detectors didn't work. I just think it's sort ofvery quickly became the reality. It was a very passing fancy, in other words, that AI detectorswould be able to solve the problems of an AI impacted education field that quickly went out the window.Education is just going to have to change based on this, and detectors aren't going to do a thingabout it. Next, AI generated hallucinations go mainstream. In 2023, the concept of AI hallucinations,the propensity for some AI models to convincingly make stuff up, went mainstream thanks to
12:58large language models dominating the AI news this year. Hallucinations resulted in legal trouble.In April, Brian Hood sued OpenAI for defamation when Chad GPT falsely claimed that Hood had been convictedfor a foreign bribery scandal later settled. And in May, a lawyer who cited fake cases confabulated byChad GPT got caught and later fined by a judge. Now, once again, it's not that this isn't an issue.I'm just not sure it's a top 10 event of the year. Once again, in a similar way to deepfakes,I think that there were probably far fewer issues of hallucination-related lawsuits than one mighthave expected. Indeed, I think that the biggest impact of hallucinations was not that theyentered the public and caused a bunch of damage, but that they created a stopper and prohibited enterprises
13:37than others from actually trying to implement LLMs as a part of a workplace solution set.In other words, hallucinations didn't have the chance to go mainstream because the fact of theirexistence meant that fewer people were going to trust LLMs in the first place.Hallucinations remain a major issue when it comes to enterprise and corporate adoption,although I think we're going to see a lot more focus in 2024 on strategies like Raghand training on enterprise level data that makes generative AI useful for companies.and tamps down on some of those hallucination-related concerns.Almost through this list, the penultimate one is Google's barred dances to counter Microsoft
14:10and ChatGPT. When ChatGPT launched in late November 2022, its immediate popularity caught everyoneoff guard, including OpenAI. As people began to murmur that ChatGPT could replace web searches,Google jumped into action in January 2023, hoping to counter this apparent threat to its searchdominance. When BingChat launched in February, Microsoft CEO Satya Nadella said in an interview,I want people to know that we made Google dance. It worked.Google announced Bard in a botched demo in early February, then it launched Bard as a closed test inMarch with a wide release in May. The company spent the rest of the year playing ketchup to OpenAI inMicrosoft with revisions to Bard, the Palm 2 language model in May, and Gemini in early December.
14:46The dance isn't over yet, but Microsoft definitely has Google's attention. So a couple things aboutthis one. Obviously, Google's place in this race deserves mention on this list. And specificallythe fact that they have been playing ketchup, which is a very uncomfortable position for them.Now, the question going into next year is, I think, two parts.First, can Gemini Ultra actually meet or exceed GPD4?And second, how does integration across Google suite of tools make a difference in termsof what options people use?It's entirely possible that we're going to move into a world where AI is more and more
15:20commoditized and what matters is integrations into our existing workflows or into new workflowsthat work with our existing tools, given how many people are using Google Worker's.workspace tools, they still have a really big advantage, and as many people have tried AI thisyear, it pales in comparison to the number of people who have yet to touch it at all.Last on this list, and in many ways, number one with a bullet, open AI fires Sam Altman and hereturns. Now, I won't get into the details of what happened, as you have heard it endlesslyif you listen to this show, but this was absolutely a seminal event. I mean, it was moredramatic than I think we even realized now. History will look at it with unbelieving,
15:59saucer eyes. How much was it questions of safety and disagreements around that? How much was itpetty internal squabbles and a palace coup? Whatever the case, the undeniable leader in the entiregenerative AI space came very close to being a shell of itself or not existing at all, even despitehaving carved out an incredible lead. That is a wild state of affairs and certainly deserving of thebiggest stories on the year list. Now, when it comes to other things that I might include, I'm tempted to talk aboutall sorts of emergent technology like auto-GPTs, the discussions around AI agents, the newand emerging battle around AI wearables. But ultimately, those aren't really the biggest eventsof last year. What they are is contenders for the biggest events of next year. And for that,
16:46we'll have to wait for another episode. Until next time, guys, peace.
There aren't comments yet for this episode. Click any sentence in the transcript to leave one.