AI News: ChatGPT is in BIG Trouble

As is true with almost every single week now, there's been a lot happening in the world of AI, starting with the fact that once again we have a new king of the AI large language models. This week Anthropic released Claude 3.5 Sonnet. And if you look at this little graph that they put here on their website, here's where Claude 3 Opus, their previous state of the art large language model, landed. Well, the new one we can see here is much more intelligent and also much cheaper to use than Claude 3 opus, about the same cost as the previous version of Claude 3 sonnet, but way more powerful than Opus. It also operates at twice the speed of Claude 3 opus. And if we take a peek at the benchmark tests here, again, I don't always remember which benchmark test is best for which thing. I just kind of look at it overall. And as we can see, in every single area, Claude 3.5 sonnet beats out Claude 3 opus. And in almost every single benchmark it beats out GPT4O, which the GPT4O seemingly slightly better at math. The Vision model in Claude is also much improved. And as we can see here, once again in the benchmarks for testing vision models, Claude 3.5 Sonnet beats out Opus and Gemini 1.5 in every single benchmark. And GPT4O only slightly edges it out in this specific visual question answering benchmark. And I mean it's by a small margin, we're talking 6.91% to 68.3%. Claude also released a new feature which they call Artifacts. Artifacts appear in a dedicated window alongside their conversation. It creates a dynamic workspace where they can see, edit and build upon Claude's creations in real time, seamlessly integrating AI generated content into their projects and workflows. And the coolest part is this new model of Claude is available for free on Claude AI. So you don't even need the paid plan to use this. They are rolling out version 3.5s of Haiku and Opus later this year, so the premium members will get an update. But right now their best model is available for free. And there's been a lot of demos on Twitter already of how good this model is. Here's Ali Miller talking about how in just 25 seconds, Claude 3.5 Sonic coded a fully functional Mancala web app for me. We can see she uploaded a screenshot attachment here, gave it a single prompt about what she's looking to do, told it to think step by step. It opens up the little artifact sidebar to show it Actually writing the code and it presented a playable game right out of that single prompt that she put in. Ethan Moloch here also did the same thing. He said, this is a real time video of me creating a playable game and editing it with Claude. We can see he gave it the prompt make me a game about crabs and let me play it. And then the artifact sidebar opens. We get the code generation didn't quite work the first time. So he said not crabby enough. And this time around they actually look like crabs. He says, can you make it fill the whole screen and and be even? Crabbier once again generates codes in the sidebar and generates a much more impressive looking crab game. Ethan says, can you make it scarier now the game is much darker and I guess scarier. Testing Catalog News posted ON X Cloud 3.5 just generated React JSX code with a simple contact form and managed to run it in the Artifacts playground. If we open their screenshot here, we can see that they had it code a simple contact form and it showed the contact form in the right sidebar and was actually usable. And it looks like you can switch between preview and code here. Dennis Shoraev here says Claude 3.5 Sonnet is the first model to recreate the 3D scene data flow from the movie Hackers on the first try. Great job. Anthropic and we can see what that looks like here. And Pietro Scirano here says Claude 3.5 sonnet plus maestro equals sparks of AGI. I asked to make a Mario clone using just geometric shapes and the wildest part is that it gave the character animations as well and the shapes seem like novel concepts. It took three minutes once the code was generated. We can now see exactly what the game looks like on the first try. This looks like a game I tried to make about a year and a half ago using AI to help me code that took me, I don't know, probably 50 prompts in five hours to generate. This time it looks like he was able to do it with just a single prompt. This new version of clodsonnet is really impressive. Now if you want to use it, head on over to Claude AI. Come up here, click on your name, click on this little feature preview in the menu bar here and then turn on Artifacts. Let's ask it to make a simple tic tac toe game where the user plays against the computer. Make it playable in the Artifacts window. We can see it opened up our Artifacts window on the right and started to write some code. Now we've got a tic tac toe window from a single prompt that shows up in our right sidebar. Let's go ahead and see if we can play against the computer. So I put the X here, the computer put O here, and it appears. The computer's not very smart, but it does work. On the first try. I can reset the game and play again. But to me this feels a lot faster and also quite a bit smarter than what we're getting out of ChatGPT 4.0 now. In fact, if you want my opinion, I actually feel like GPT4O got a little bit dumber than the previous version of GPT4. At least in my use cases. It actually seems like it performs worse than it used to. I don't know if anybody else is experiencing it, but that's what I've found. And now with Claude having its free version have the best model available on the market for anybody to use is pretty cool. But it does kind of feel like some of these other companies are starting to pass OpenAI a little bit. I mean, we saw Luma's Dream Machine recently, which came out before anybody got access to Sora. We're seeing models from Claude come out that are outperforming OpenAI's best models. In my opinion, it feels a little bit like OpenAI is falling behind. Now we don't know what they've got training behind the scenes. They may release a GPT5 or something else that they call it and it might just blow everything out of the water and make things like Claude Sonnet just look really basic and archaic. But I feel like they need to do that fast because a lot of these other companies are starting to eat up some of the mind share of the general public and it could only be a matter of time before everybody is using the term Claude as synonymous with AI instead of using ChatGPT as synonymous with AI. But this isn't the only reason that I feel like OpenAI is starting to fall behind a little bit. They do feel like they're going through a pretty rough patch at the moment. The news also came out this week that Sam Altman is sort of mulling around the idea of getting rid of the non profit portion of OpenAI and just making it a full on for profit company. Eventually they can IPO and make as much money as they want. This is just speculation. Sam's just made some comments about how they might do that. So not a ton of news here, but now that the OpenAI board seems to be more like Sam's People instead of people that were super focused on the safety of these models. It seems like he can do whatever he wants with the company at this point. Also at the end of last week, which I didn't catch in last week's video, OpenAI appoints retired US Army General Paul M. Nakasone to Board of Directors. It brought him on board because of his cybersecurity experience. But if you're curious a little bit more about what his cybersecurity experience is, he was the longest serving leader of US Cybercom and also led the NSA where he was charged with safeguarding the United States digital infrastructure and advancing the country's cyber defense capabilities. So they hired somebody that was high up at the NSA, who notoriously spies on US citizens to be part of the board for OpenAI. And of course, Edward Snowden, who got in a lot of trouble for exposing the fact that the NSA was spying on us, had some words about this. He said, do not ever trust OpenAI. He said, There is only one reason for appointing an NSA director to your board. This is a willful, calculated betrayal of the rights of every person on Earth. You have been warned. Another security expert named Matthew Green went to X to say, I do think that the biggest application of AI is going to be mass population surveillance. So bringing the former head of the NSA into the OpenAI has some solid logic behind it. Now, OpenAI hasn't commented on this specific stuff yet, but they have claimed in their articles that the security of OpenAI systems from protecting the large AI training supercomputers we operate to securing our sensitive model weights and the data entrusted to us by customers is central to achieving our mission. So they're claiming that he's just there to keep things safe and secure. But he does have a reputation with the nsa, who again is known for spying on its own US citizens. Also, if you remember, Ilya Suskever, one of the people who helped oust Sam Altman back in November of 2023 and then was unheard of for months and months and months now went to X to say he's starting up a new company called ssi, or Safe Superintelligence Inc. If we take a look at the announcement tweet here, they claim that superintelligence is within reach. Building safe superintelligence is the most important technical problem of our time. So they've started ssi. They go on to say, our singular focus means no distraction by management overhead or product cycles. And our business model means safety, security and progress, all insulated from short term Commercial pressures. They're based in Palo Alto and Tel Aviv and they're currently assembling a team to make these goals achievable. Now, I don't totally understand what's going on with ssi. It sounds to me like Ilya is going and essentially creating an OpenAI competitor to try to get to super intelligence. OpenAI's goal is to get to artificial general intelligence. Apparently this one is to get to super intelligence, but do it safely and they're going to have no customer facing products, at least not in the beginning. So I don't actually understand how they're going to raise money. I guess there are going to be some people out there that would be willing to invest just for the tech itself, just to see it advanced. But I have a hard time finding major venture capitalist firms and things like that dumping a ton of money into a company that has no outward facing products. Like they're just going to be dumping money into a product that has no revenue generation. On the other side of it. I just don't understand how they're going to raise money. They obviously need to use a lot of compute to get to super intelligence, meaning they need data centers and huge GPU clusters to. So they either need the money to go and buy all of that stuff, which I don't quite understand how they're going to raise on a business model that has no revenue generation, or they need to use cloud infrastructure that already exists. They need to use, you know, Google's cloud or Amazon AWS or Microsoft Azure or one of these other providers of cloud GPUs. And if they do that, how do they ensure that it's safe? Because now they're operating in somebody else's playground. I don't know. We still need to learn more about what is going on with Ilya's new company. I don't quite understand the business model behind it, but it will be interesting to follow along and of course I'll keep you in the loop as I learn more about it. We've also had a ton of AI video news coming out recently. Last week we talked about Luma's dream machine which was really impressive. Not great at text to video, but really, really good at image to video. Well, this week we got the announcement of another really impressive looking AI video generator. Albeit they haven't given us access to it yet, but they showed it off and it looks really, really good. It's from Runway and it's called Gen3 and this is their text to video model. Well, their new text to Video model and the demos of it actually look really, really impressive. Like better than Luma's Dream Machine and about on the same level as Sora. In fact, a lot of people think that the reason Luma Dream Machine kind of came out so quickly was because they knew that Runway gen 3 was about to be announced. And and that's why the servers weren't quite ready for the onslaught they got when they announced Dream Machine. If you remember, on the first day they opened it, people were waiting, including me, up to seven hours to generate a single video. They weren't ready for that launch. And a lot of people believe it's because they knew this was coming and they wanted to get out in front of it. But the generations that this can do from text to video, like an astronaut running through an alley in Rio de Janeiro, look really good. We were not getting anywhere near this level of text to video from Luma's Dream Machine. You're getting this level when we started with images, but definitely not text to video. Here's a handful more examples that are on Runway's website. I will link up to their announcement in the description below so you can check out more of these announcements. But this stuff is looking really, really impressive here. Humans are looking pretty dang realistic in these videos. They shared a whole bunch of examples of humans that don't actually exist, but we've got some pretty dang good video generation around them. And one of the things that seems to be really impressive that nobody's really talking about is the ability to generate text inside of these videos. Like, look at this video that says color in water. And then like a splash of color hits the letters, or this one where there's water spilling, and then it splashes up to reveal the words Drink water Water. And all of the letters and all the words just look really good. I don't know why this isn't one of the features more people are talking about. But nonetheless, some of these videos generated by Runway gen 3 are almost indistinguishable from real life videos. This is the first actual Sora level video generation that I've seen that is not actually Sora. And this brings me to another reason why I mentioned I think OpenAI is kind of, kind of in trouble. They announced Sora. Everybody got excited about it. They waited so long. And then it gave time for things like Cling and dream machine and gen 3 to all pop up and go, look. We've got things pretty much on par with Sora now, making a lot of the AI world to believe that OpenAI teased it and then didn't release it fast enough. And by the time we finally get access to Sora, who knows? Maybe people just won't care anymore because dream machine and gen 3 do it just as good, and everything they want to do with it, they're getting out of these tools. While we're on the topic of AI generated video, there was an AI generated movie premiere for a film called the Last Screenwriter. Well, apparently a movie theater in London canceled this premiere because over 200 people complained about the fact that it was an AI written script. Now, this isn't an AI generated video. AI generated this script, and then it was acted out by real people. But because the script was AI generated and 200 people complained, the movie theater went, nope, we're not showing it. Which is a pretty crappy stance for a movie theater to take. If we're honest, it probably would have done okay at the theater because the people that all hate AI would have showed up so that they could go and talk about how much they hated the AI movie. And the people who love AI would show up because they want to be a part of seeing the first AI generated script in theaters. But, oh well, my guess is this news story will spread a little bit further and actually have a little bit of a Streisand effect on this film where more people will actually learn that this film exists because this movie theater and these 200 people didn't like the fact that it existed. Google DeepMind also showed off some new research in this past week showing that we can generate audio from videos. So we give it a silent video clip and then it figures out what audio works best with that video clip. And. And there's a bunch of examples on their website. I will again link it up so you can check them out, but let's preview a couple of them. Here's like a horror scene of somebody walking through like a dark area. Remember, the video was given. The audio was generated with AI by watching the video.

Here's another example of a drummer.

And they even did a wolf howling at the moon, which I'm a fan of because that's one of my test prompts that I always use.

Not bad at all. Along a similar vein, eleven Labs introduced a text to sound effect.