We have some massive new launches today, including new models, New Vision models, and so much more news. Let's get right into it. This video is brought to you by Vulture, the easiest way to deploy your generative AI applications. Go to getvulture.com/Burman and use the code

Burman300 for $300 of credit right now. I'll drop links in the description below. So, first, Mo has made a bunch of releases today. I actually got a preview of some of these, and they are extremely exciting. Let's go over all of them. So,

the first thing is, Le Chat, MRA's version of ChatGPT, now has web search with citations. And the best part, it is absolutely free. And you're going to hear that as a theme with all of the MRA announcements that I'm about to tell you

about. So, this is what it looks like. There it is, searching the web. Who won UFC 309? Jon Jones won UFC 309 by defeating Stipe Miocic in a third-round stoppage, successfully defending his heavyweight title, with two inline attributions: ufc.com and forbes.com. Look

how well that works. So, now that is available to you for free. And MRA now has image generation thanks to Black Forest Labs' Flux Pro. So, let's try that out. Make me an image of a strawberry. Generating an image. And there we go, in just maybe two or three seconds, we have a

beautiful image of a strawberry, all powered by Flux. Really, really nice updates, and again, completely free. And, and this really brings L Chat up to par with ChatGPT's features now. And another thing that's at least new to me is creating and mentioning agents in chat.

so you can actually at-mention an agent so this is just a model but you can click create agent you can set up your own agent here choose the model you can input instructions you can give it demonstrations change the temperature test it out and then you can at-mention

it in any chat that you want which is super useful so now you can have custom agents directly in your chat and last they have canvas now and it's exactly the same as you would expect from ChatGPT's canvas it allows you to have inline code editing inline creative

writing editing and let me just show it to you right python code for the game Snake so you type something in and it pops open canvas so chat continues over here on the right side on the left you actually have the code and once it's done you can edit it up some other

really cool features you can iterate through the code and you get each iteration as its own file and you can see the diffs between the different files so there you go there's the game Snake you can come in here you can change it as you like you can highlight

any piece of code and just change it in line so let's see that ASLA chat change these colors to something else okay and so you can see it right there there's the chat and now it is rewriting the colors and there we go and all the rest

of it looks a about the same and here's the diff you could see two different versions and as I'm clicking through both of them the colors are changing and not only that they released Pixal large an open-source open weights vision model this model can understand images really

Well, I already tested Pixel small and got it running locally. I think it was only about 8 billion parameters, and it was definitely the best vision model that I have ever tested. And now they have a large version. So, simply say, "Describe this image", and it

not only works for images, it also works for PDFs. So, it has deep PDF understanding. This image depicts a llama lying down on a grassy field. So, just like that, it works well. So, it's available today for everybody for free. All you have to do is log in and just

enable the beta features. And if you want to download these models and test them out locally, definitely do that as well. If you want to see me test these models in full, let me know in the comments below. Next, it looks like Google may have actually accomplished the unthinkable. Their

Gemini model is now number one on the LM Cy leaderboard. This is not based on benchmarks; this is the chatbot arena, so people are actually voting on the results. So, if we click in there, it is Gemini experimental 1114, and it is four points ahead of chat

GPT-4, the latest, which is extremely interesting that 01 preview and 01 mini are not in the top one and two, which I would have expected. So, overall, it jumped from 3 to 1, math 3:1, hard prompts 4:1, and

so on. You can see massive improvements here, and here's just an example of how good it is compared to all the other models. This is Gemini experimental 1114 in the math arena. As you can see, the 2011 models and the Gemini models are all about in the same area, at about 1340.

where the next model after that, Gemini 1.5, is a huge drop down, and then all the other models come after it. So what do you think of this? Again, these aren't benchmarks, this is the Chatbot Arena, which means somebody voted on the results. You think this is legit? Should I

try it out? Let me know. Next, in a story that I covered in the last news video, it seems like everyone was reporting that AI was hitting a wall. The top AI companies were not hitting those gains that they wanted for their next generation of models. But it turns out

Sam Altman does not agree. He said there is no wall, referring to the intelligence wall that a lot of companies might be hitting when training their next generation of model. And personally, I don't believe it. I don't think there's a wall, and there's a lot of reasons for it.

One, we're seeing algorithmic unlocks almost every week, whether it's test time training, I made a video about that if you haven't seen it, check it out down below, or test time compute, which is the 01 models allowing the models to think with Chain of Thought. So there are a

number of different levers that we can pull to continue to scale up intelligence, and scaling continues to be all you need. It's just different types of scaling at different points in the model process. And next, from OpenAI, they released a huge update for their desktop

applications. Now ChatGPT desktop can read your code directly, and it can do so in VS Code, Xcode, terminal, and iTerm 2. So you basically open up the ChatGPT app, you tell it what you want to do in one of these applications, it can read

your code and then you can make changes easily so no more copying back and forth. This is in beta right now for Plus and Team users, so definitely check it out. And this is the first step in my opinion, in ChatGPT being able to control your

entire desktop. And really, I think there's a step before that, to me at least, it makes sense for OpenAI to create their own browser with AI natively from the ground up. So you can have agents operating on your behalf, hooks into the browser that allow agents

to log into any website that you want them to authenticate into, transact whether that's crypto or some kind of fiat currency, whatever it is, it seems like that would be a really cool half step to allowing ChatGPT to actually control your entire desktop. Next, AutoGen

made a huge announcement. AG2 is now out. It seems like they are graduating from just a research project to an actual commercial project. So from Chi Wang, one of the creators of AutoGen, big news: AutoGen is now AG2. We're evolving with support from the open source community.

AutoGen is becoming AG2, a new home for next-gen agentic AI. Same mission, bigger goals. It is still open source, so definitely check it out. And really exciting updates. I'm hoping to speak with someone from the AutoGen team pretty soon, so maybe that turns into an

interview. We'll see. Next, Anthropic is now working with the federal government to figure out if models are going to leak nuclear secrets, which is kind of insane to think about. We already knew OpenAI was working closely with the government, giving them previews of their

Cutting-edge models to make sure they know what's coming. And we also knew Anthropic was pretty pro-regulation as compared to many other AI companies out there. And now we know that they are actually working directly with the government to figure out if these models

are going to leak sensitive information. Anthropic is working with the Department of Energy nuclear specialists to ensure its models don't help people make weapons. The company first shared this with Axios. And if you also remember, Anthropic is working with Palantir to bring their models to highly sensitive, highly secret

environments. So this all seems to be going hand-in-hand, and why it matters. Anthropic believes this is the first time a Frontier Model has been used in a top-secret environment, paving the way for similar partnerships with other government agencies. They're doing red teaming exercises: experts testing

systems by trying to break them. The NNSA has also been testing Anthropic's models to see whether people can abuse them to find nefarious use cases for nuclear energy, like developing weapons. This seems like something [name of person] needs to get hired for because he seems

to be able to jailbreak every single model. Next, X.ai is raising up to $6 billion to purchase 100,000 Nvidia GPUs. That's an additional 100,000 GPUs, and it seems like X.ai's computational lead just

continues to extend. Now we haven't seen the actual fruits of all of that computational labor yet. Grok 2 is pretty good, but it's certainly nowhere near ChatGPT-4 or even the Oo1 series of models. They are raising at a $50

billion valuation which is insane given they're only a couple years old so I believe Grok 3 should be coming pretty soon but we're not sure exactly when and I cannot wait to try it next and what seems like a no-brainer but still really

cool is Biden and China's Xi seem to have agreed that AI should not launch nuclear weapons the White House says humans will be the ones with control over the big buttons and China agrees that it's for the best now even if this is a handshake

agreement this is a very good sign especially for US-China relations and just how we approach AI as a weapon in general AI hallucinates and again I'm going to mention play the prompter plany has jailbroken every single model so if

you can jailbreak the models and they can be jailbroken because they are non-deterministic that means that that would make me personally extremely nervous if these models had their finger their digital finger on those big red buttons the leaders also emphasized the

cautious development of AI and military technology acknowledging the potential risks involved this is the first public pledge of its kind between the US and China because apparently even world powers need to draw the line somewhere and this is all according to Reuters next in one of the best use cases for AI

the most optimistic the most positive use cases for AI AI can now beat doctors at diagnosing illnesses and pretty consistently too according to an article by the New York Times AI chatbots defeated doctors at diagnosing illnesses a small study found chatbots outdid human

Physicians, when assessing medical case histories, even when those doctors were using a chatbot, now here's the thing: chatbots run 24/7, they do not get tired, they don't need to eat, they are just much more efficient. So if they're

running 24 hours a day and then the final diagnosis or the final review is done by human doctors, I think that's great. A study by Dr. Rodman, doctors who were given ChatGPT along with conventional resources, only did slightly better than doctors who did not have

access to the bot. However, ChatGPT alone outperformed the doctors. 'I was shocked,' Dr. Rodman said. The chatbot from Company OpenAI scored an average of 90% when diagnosing a medical condition from a case report and explaining its

reasoning. Doctors randomly assigned to use the chatbot got an average score of 76%. Those randomly assigned not to use it had an average of 74%. So this is the worst that AI is ever going to be, and if we've had that much of an improvement by using AI to

diagnose medical conditions, just imagine the next few years. Next, it seems like some of these new Nvidia AI chips might actually be overheating, and that's a huge problem for NVIDIA. According to Reuters, the new Blackwell AI chips, which have already faced delays, have

encountered problems accompanying servers that overheat, causing some customers to worry they will not have enough time to get the new data centers up and running. The information reported on Sunday: the Blackwell Graphics processing units overheat when connected together in server racks designed to hold

up to 72 chips, the chip maker has asked its suppliers to change the designs of the racks several times to resolve overheating problems, according to Nvidia employees who have been working on the issue as well as customers and suppliers with knowledge of the issue. So another

hiccup in the road to getting these Blackwell chips out the door and into production. Now, really, Nvidia is one of the only chip makers at scale in the game right now, so really the whole industry is just going to be waiting. And our last story: Perplexity is introducing

shopping. So, according to Perplexity's post, introducing Perplexity Shopping, a one-stop solution where you can research and purchase products, it makes a big leap forward in how we serve our users, empowering seamless native actions right from an answer. Shopping online just became 10

times more easy and fun. Now, this is a great way for them to monetize beyond just subscriptions. You know, if you watch this channel, I love Perplexity and so I'm definitely going to be testing out their shopping options. They have a one-click checkout experience as well.

They seemingly are the first to market with a true native AI shopping experience and they're also launching a merchant program. So, if you're a seller, you can work with them to sell your items directly through Perplexity. So, I think this is great. Everything that they

can build natively into Perplexity just reduces the amount of clicks that I need to take to get what I want. They already have replaced my entire Google usage. I'd say I probably at this point use Google about 3% of the time, and ChatGPT and

perplexity the rest of the time, so it seems like Google really needs to do something to not get completely rolled over in the search game. So that's it for today. If you enjoyed this video, please consider giving it a like and subscribe, and I'll see you in the next one.

AI News: Mistral's OPEN ChatGPT Competitor, Gemini Takes Top Spot?!, AI & Nukes, Perplexity Shopping

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

Matthew Berman

Le Chat now has free web search with citations.

MRA now offers free image generation via Flux Pro.

New feature allowing creation and mention of custom agents in chat.

Addition of Canvas feature for inline code and text editing.

Release of Pixal large, an open-source vision model with PDF understanding.

Gemini's Leaderboard Success and AI Development

Google's Gemini model achieves #1 ranking on the LM Cy leaderboard.

Discussion on whether AI has hit a development wall, with differing opinions.

Algorithmic advancements like test-time training and compute continue to improve AI.

OpenAI releases a desktop app update allowing ChatGPT to read code directly from various applications.

AutoGen releases AG2, transitioning from research to a commercial project.

Anthropic collaborates with the federal government to assess the risk of models leaking nuclear secrets.

X.ai raises $6 billion to acquire 100,000 Nvidia GPUs.

Biden and Xi Jinping agree on preventing AI from launching nuclear weapons.

AI chatbots outperform doctors in diagnosing illnesses in a small study.

Nvidia's Hardware Issues and Perplexity's New Feature

Nvidia's Blackwell AI chips experience overheating problems.

Perplexity introduces shopping features, integrating product research and purchase.