OpenAI's GPT-4 Mini: Latest AI Developments and Controversies

OpenAI mapped out five levels of progress towards AGI, with level one already achieved and level two close.

### OpenAI's AGI Progress, New Models, and Controversies
* 🧠  OpenAI outlined five levels towards AGI, with **level two being very close** and **level one already achieved**.
* 🍓  OpenAI is working on a new reasoning technology, code-named Strawberry, which aims to **navigate the internet autonomously for deep research**.
* 🕵️‍♂️  Whistleblowers have accused OpenAI of **restricting employees from talking to regulators** and **removing rewards for whistleblowing**.
* 🤖  There is speculation that OpenAI's DALL-E image model has been updated, showing **improved text generation capabilities**.
* 🎥  New demo videos from Sora have been released, generating **excitement for the upcoming AI tool**.
* 📚  Andre Karpathy announced a new AI plus education company called Eureka Labs, aiming to create an **AI-native school model**.
* 📱  The anthropic CLA app has been released on Android, offering **AI conversation capabilities** on mobile devices.
* 🔍  Google announced Google Vids, an **AI-powered video creation app** integrated with the **Workspace Suite**.

here's the AI news that you might have missed this week. Starting with the fact that OpenAI mapped out their five levels towards the progress of AGI. Here's a quick breakdown of those five levels. So, level one, they say, would be chatbots and AI with conversational language. That's essentially what we're getting right now out of chat GPT, Claude, Llama 3, things like that. Then you have level two, which is reasoners that can do human-level problem solving. They claim they're very, very close to level two right now. Then it moves on to level three, which is agents or systems that can take actions on our behalf, you know, book flights for us, respond to emails for us, things like that. Then there's level four, which they say is the innovators AI that can aid in invention. It's actually going to create novel ideas. And then finally, you have level five, which is organizations and AI that can do the work of an organization. So basically, we're right here right now. We're at level one, almost on the level

two, we're right on that precipice of level two, and OpenAI believes that we'll sort of move through each of these levels on our way to a true AGI. Now, this was actually released on July 11th last week, but it didn't make it into last week's video. But it felt a little extra relevant this week because this week we got the news that OpenAI has been working on a new reasoning technology, code-named Strawberry. Now, I've seen a lot of other YouTube videos and a lot of X posts about this with people speculating. A lot of people believe that this was what was originally called QAR, and they've now rebranded it to Strawberry. Now, this comes from a leaked internal document. It says teams inside of OpenAI are working on Strawberry, according to a copy of a recent internal OpenAI document seen by Reuters. Reuters couldn't ascertain the precise state of the document, and they could not establish how close Strawberry is to actually being publicly available. Likely not very close. The aim of Strawberry is

to not just generate answers to queries, but to plan ahead enough to navigate the internet autonomously and reliably to perform what OpenAI terms deep research. The article does claim the Strawberry project was formerly known as QAR. This exact article here was actually updated after it was published to add this section. It says a different source briefed on the matter said OpenAI has tested AI internally that scored over 90% on a math data set, a benchmark of championship math problems. Now, writers couldn't actually figure out if they were referring to the Strawberry project or not, but it kind of sounds like they're probably the same project. Now, outside of this little information that we have on it, there's a lot of speculation around what this is, but it sounds like this Strawberry is pretty close to that level two towards AGI that we were just talking about. And at the moment, it sounds like the main purpose of this research is for this new model to essentially do research. Among the capabilities OpenAI is aiming Strawberry at is performing long-horizon tasks.

complex tasks that require a model to plan ahead and perform a series of actions over an extended period of time. OpenAI specifically wants its model to use these capabilities to conduct research by browsing the web autonomously with the assistance of a computer using agent or CUA that can then take actions based on its findings. Again, not much more is known about this, and OpenAI has notoriously been kind of hush-hush about their upcoming models, and usually we don't know much about them until literally the day they make the announcement of them. And while we're on the topic of OpenAI, more people from OpenAI are coming out and talking about some of the questionable practices of OpenAI's business. It came out this week that some whistleblowers are saying that OpenAI illegally keeps employees from talking to government regulators about problems at work and removes their rights to rewards for blowing the whistle. This comes from a letter that was sent to Gary Gendler, the chair of the SEC. OpenAI refutes the claims, saying that they have a policy on

whistleblowers that protect employees' rights to make protected disclosures. Now, this isn't the first time that OpenAI's policies and contracts with their employees have been under scrutiny. Several weeks ago, it came out that OpenAI was forcing people to sign non-disparagement agreements, and if they talked badly about OpenAI, they could lose their vested equity in the company. Well, now it sounds like people are coming forward and claiming that if we blow the whistle on anything we think OpenAI is doing that's somewhat suspicious, we can also lose our vested equity, and that's not legal. Now, the sources are Anonymous. OpenAI claims that that's not actually happening, but I have a feeling OpenAI is probably in the process of overhauling a lot of their contracts that get signed by any new employees that join the company due to all this scrutiny. Again, back when most of these people probably signed up for OpenAI, the company wasn't nearly as big or in the public eye. Now that they are as big and in the public eye, a lot

of this stuff is kind of starting to come under the microscope, and while we're on the topic of open II, there's some speculation that maybe the dolly image model recently got an update. This is a post from my buddy, Angry Penguin, over on X where he shows off an image that he created that has pretty legible writing in it. This clearly says evolve all over it. Previously, DOI struggled with words. If I go to Dolly and say, create an image of a robot holding a sign that says 'Please Subscribe,' I actually get an image that has the words kind of nailing it. So, I think DOI did make some updates because the text seems to be much more clear than it used to be. And if you're interested in using Dolly but you don't have a ChatGPT Plus account, you can always go to bing.com/images/create and use Dolly 3 for free over on Bing's website, which, if Dolly 3 did get an update, it appears to have also rolled out here inside.

being image Creator. I mean, two and a half out of four sort of nailed what I was going for. We also got some new demo videos from Sora. We can see this like black and white video showing all sorts of different clips in black and white that actually look pretty dang impressive. These were shared on Matthew Burman's X account. Here's another one that he shared of like ocean crashing and uh, I don't know, a gas station or motel or something. Uh, but yeah, we're getting more demos from Sora, which is just making people more anxious to actually get their hands on it. But right now, we do sort of have that itch scratched in the form of Runway gen 3 and Lum's Dream Machine. We can actually create some pretty good AI-generated videos now with those tools. It sort of damped down the excitement for Sora a little bit, but the fact that this can create much longer videos and OpenAI tends to kind of set the bar for almost everything they put out. I'm still

excited about it, but I have gotten that need met with some other tools recently. If you're somebody that uses AI at work, or you're thinking about using AI at work, you need to check out HubSpot's completely free bundle called Five Essential Resources for Using Chat GPT at Work. And honestly, if you haven't embraced AI yet, just remember what Nvidia CEO Jensen Huang said: AI will be the most transformative technology of the 21st century. It will affect every industry and aspect of our lives. So, if you're not using AI to speed up and improve the quality of your work, well, your competitors probably are. So, the link to this totally free resource is down in the description below. And trust me, this is something you're going to want to look through. It includes interesting flowcharts on when you should or shouldn't use Chat GPT. There's also a really cool template that you can use with Chat GPT to make sure that any of the content it creates for you follows your brand's voice. You've got an

AI generated content refinement checklist to double-check the AI's work and ensure that you're putting something out into the world that you really want to be putting out there. There's also a four-page checklist that you can easily follow all about adopting AI in the workplace, and a super comprehensive PDF guide on how to supercharge your day with chat GPT. And here's what's really cool about this: if you scroll all the way down to the bottom of this document, they have 100 ways to try chat GPT today, and it's got some really cool prompts to test out things like providing recommendations for improving customer service and support, providing recommendations for improving SEO, helping with email management and organization, and so much more. Again, it is really comprehensive and really helpful. The link to this completely free resource from HubSpot is in the description below, and thank you so much to HubSpot for sponsoring this video. Andre Karpathy, who previously worked at OpenAI and then recently stepped away, just announced a new venture that he's

working on he said excited to share that I'm starting an AI plus education company called Eureka Labs. At Eureka Labs, they're building a new kind of school that is AI-native. They say that subject matter experts, who are deeply passionate, great at teaching, infinitely patient, and fluent in all of the world's languages, are also very scarce and cannot personally tutor all 8 billion of us on demand. So, it sounds like he's creating a sort of online education where the teacher still designs the course materials, but they are supported, leveraged, and scaled with an AI teaching assistant who is optimized to help guide the students through them. So, this announcement here is really all that we have. He hasn't really talked a whole lot about this more than the announcement, but what I'm sort of imagining is that a teacher with subject matter expertise goes in, creates an entire course on their subject matter. All of that information is then sort of trained in the AI. I don't know if they're going to use, you know, retrieval, augmented generation, or they're going to fine-tune

the model. I don't know exactly how they're going to do it, but all of the information that the teacher taught is now available inside of the model. So, anybody who wants to learn this stuff can then work with a tutor who understands all of the training material and can speak to the student in whatever language they want to learn in. This will massively scale the ability of an individual teacher who can teach the concept once and then let their AI assistant teach it to everybody else who wants to learn that information. Again, I'm just sort of speculating on what this is going to look like. I don't know exactly, but that's sort of what the concept sounds like to me. If you're a fan of anthropics CLA and you don't have an iPhone, well, good news. They just released it on Android. It's been on iOS for a couple of months now, and they just now rolled out an Android version. Personally, I'm still a fan of the ChatGPT app a little bit more than the

anthropic app. Just because the conversational voice portion of the chat GPT app is actually really, really good. When I'm on my computer, I usually use either Clot or Perplexity. When I'm using my phone, I still go to the chat GPT app, but I also understand most people probably don't want to pay for free separate chat subscriptions. So if you really like the ability to have a voice conversation with an AI, chat GPT is still the way. If you don't care about that, you just want the best model in your hand, Clad is probably the best, and they now have an Android app. And since we're talking about Android phones, Gemini now answers general questions when your Android phone is locked. There's not too much more to share on this story. It is exactly what it sounds like. Google Now lets you get answers from Gemini without actually unlocking your device. Also this week, Google announced Google Vids. Vids is an AI-powered video creation app that's designed for work and deeply integrated with the Workspace Suite.

use every day. You can actually find it over at workspace.com product/ VDS. Right now it's not available to everybody. They say we're currently testing this new application with a select group of trusted testers. And according to the video on their website, it looks like you give it a prompt like help me create a sales training video and then it will help create this slide style video for you. There's a bunch of different styles that you can choose from, and once you pick your style, you can speak out a script, add a voice over to it, and add stock footage to it to get the perfect layout for your video. And then it creates that slide presentation video for you. And since we're talking about Google and we're talking about video, let's talk about this new feature that YouTube is rolling out called YouTube music sound search. This sounds like a feature that's very similar to Shazam where you can have it listen to a snippet of music and it will figure out what song it is.

also hum the song. It'll be able to figure out what song it was based on your humming. We can see some screenshots that they shared here. They've got a little search box here with a microphone next to it. I'm assuming they click the microphone and then it says play, sing, or hum a song, and then it figures out what song that you were trying to find just based on the singing or humming. YouTube's also testing an AI-generated conversational radio. It'll let you create a custom radio by describing exactly what they want to hear. This article goes on to say be on the lookout for ask for music any way you like card in your home feed. This will open the chat-based UI with a field at the bottom that lets you ask for music. There's been a little bit more controversy this week about the source of training data for various AI models. This article on Proof News here claims that Apple, Nvidia, and Anthropic use thousands of swiped YouTube videos to train AI. Basically, here's what's happening with this.

there's a company called Uther AI, which is an open-sourced company that collects a whole bunch of data from everywhere and puts it into what they call the pile. The pile is this giant data set that companies then use to train their AI models initially, so that it can just sort of learn how the language works and just get injected with a ton of data to start. Well, this pile is trained on publicly available data, and it turns out that a lot of that publicly available data was transcripts that were copied and pasted straight from YouTube videos. A lot of YouTubers started to notice there's data in there from people like MKBHD, Mr. Beast, PewDiePie, and others. And this site, proof news.org, actually put up a little search engine so that you can see if a video that you created or literally anybody's video is found within the pile's data set. Now, I did a search for my own name and no results were found. I don't know whether I should be offended or relieved.

at the time, the data was scraped, my channel probably just wasn't big enough. Now, after all this came out, Apple stepped up to say, Yes, we've used the pile for some research purposes and some training, but the model that we're using inside of our Apple intelligence is not trained on the pile's data, so that information is not inside of Apple's training set, according to them. Microsoft has a platform called Designer, which if you're not familiar with it, it's very similar to Canva. It's a platform to create things like YouTube thumbnails and banner ads and Instagram images and things like that. Well, this Designer platform is now being rolled out into a whole bunch of different Microsoft apps directly, where you can use the Co-Pilot sidebar over here, ask it to create a specific image in a specific style, and it will actually use Microsoft's Designer to create that image and allow you to pull it directly inside of your document or your PowerPoint or whatever Microsoft tool that you're using. Here's another example of it being shown off inside of Microsoft PowerPoint where

they create some images with designer over here. It generates some images, and then they just pull that in as the background of the slide. Designer also got a free mobile app on both IOS and Android, so you can easily create and edit images on the go on a mobile device. Now, there's a whole bunch of other new features for designer if you want to dive deeper into it. This is something that you're really interested in. I will make sure it's linked up in the description so you can see all of the updates here. The article is quite long, and there are quite a few updates, but it seems like it's got some other pretty cool features like this restyle feature. You upload an image, and it restyles it to a different style of image. Mistol, the French AI company that develops large language models, released a new model called Cod strol Mamba. This is a model designed for code generation. It is open source, and it can handle an input of up to 256,000 tokens, which is double what OpenAI currently offers with ChatGPT.

that's rough. Roughly 192,000 words between the amount of text inputted and the amount of text outputted. This is a 7 billion parameter model and offers a fast response time even with longer input text. So if you're a coder and you're looking to try another large language model to see if it outperforms the other models you've tried, maybe Codestroll Mamba is a choice to try out. Amazon started rolling out an AI shopping assistant called Rufus, which apparently answers questions about shopping and also politics. Rufus is essentially a chatbot just like ChatGPT, but it's built directly inside of the Amazon app, and it's trained on the data that's in Amazon. So you can ask, What are the best lawn games for kids' birthday parties? and it will suggest lawn games as well as where to find them and buy them on Amazon. This Verge article also tested some other questions and got it to answer questions about the candidates for the 2024 election. I've got more bad

news: If you're in the EU, it sounds like Meta is not going to be offering their multimodal models in the European Union. They will be offering their normal text input output models like Llama, but you're probably not going to be able to create AI images, AI videos, and anything other than more text if you're in the EU due to the EU's, I guess, unclear policies. They say here, We will release a multimodal Llama model over the coming months, but not in the EU due to the unpredictable nature of the European regulatory environment. It says here that Meta's issue isn't with the still being finalized AI act, but rather with how it can train models using data from European customers while complying with GDPR, the EU's existing data protection law. The United Kingdom has nearly identical laws to GDPR, but Meta says it isn't seeing the same level of regulatory uncertainty and plans to launch its new model for the UK users. Here's something that I came across on X from Johannis Stelzer. I just thought it

was really cool. They hooked up a little, uh, MIDI device to their computer, and they can turn the knobs to change different aspects of the images. They appear to be using stable diffusion here, and then using these knobs to change different elements within stable diffusion, different sort of parameters. And I just thought it looked really cool, and I wanted to share it. They also put the code for it up on GitHub. So, if you want to play around with something like this and hook up SDXL to a MIDI device, well, that's available for you to do. Here's another article that I came across that I couldn't find a whole lot on. I just thought it looked cool. It's from Gizmo China: Turn your selfie into a printable 3D character with 10 cents AI-powered app. So, apparently, this is an app where you can upload a selfie, and it will generate a 3D model based on that one selfie that is so good that you can 3D print it. Now, I actually did some digging to try to find more info about what

they're doing here, and this was literally the only article I can find about it. But as I learn more about it, I do have a 3D printer. I do love AI. This is something I will be playing with if I can get my hands on it. And here's something interesting. AI systems achieve a 96% accuracy in determining the sex from dental X-rays. So, they basically trained an AI model on a whole bunch of dental images, and then when they ran new dental images through it, it was able to determine the sex of whose teeth those were at a rate of 96% accuracy. And the ones that it wasn't accurate on, that was mostly children. The article claims that it's less accurate if you're six or under, or basically haven't lost your teeth yet. Now, the main use case for something like this would be in forensics. If they find, you know, skeletal remains or something, they can actually identify the sex of the skeletal remains. But I just thought it was fascinating.

I thought I'd share with you, so I started recording that video while I was in San Diego. Still, now I'm on vacation in Colorado, and a few more pieces of news came out that I wanted to make sure got shared in Friday's news video, including the fact that OpenAI just launched a new model today on Thursday, the day I record this, called GPT-40 Mini. With pretty much every large language model creator out there creating models that are smaller, designed to be more cost-efficient and faster, OpenAI needed to create a language model to compete. So, this new GPT-40 is replacing the old GPT-3.5. Not quite as powerful as the full-on GPT-40, but it is faster and smarter than the previous GPT-3.5. We can see that right now today, GPT-40 Mini supports text and vision in the API, with support for text, image, video, and audio inputs and outputs. In the future, it's got a 128,000 token context window, so you should still be

able to put large amounts of text as your input. However, the output only supports 16,000 tokens. We can see this comparison here of model evaluation scores with GPT-3.5 and GPT-4 being the best model, and it pretty much performs the best across the board here in every test. With GPT-3.5 Mini, this new model that was just released performing second best across pretty much all of these benchmarks. Here, now keep in mind, this is comparing it to these other companies' smaller models. It almost kind of feels unfair to be putting GPT-4 in here compared against, you know, Claude Haiku and Gemini Flash, which is both of those platforms' smaller models, while GPT-4 is OpenAI's current state-of-the-art model. But nonetheless, we can see how this new Mini version of GPT-4L outperforms all the other mini models that are out there. If we log into our ChatGPT account here, up in the top left corner where you select the model, you can see that we now

have access to GPT-3.5, GPT-4, and the Legacy GP-4. Now, at the time of this recording, when I try to not log in and just use the free version, it's still claiming it's using ChatGPT 3.5, although it does say here in ChatGPT Free Plus and team users will be able to use GPT-4.0 Mini starting today. In other large language model news, Nvidia and Mistol teamed up to create Mistol Nemo. This is a 12 billion parameter model, and it also has 128,000 tokens, just like the new GPT-4.0 Mini model. Now, what's cool about this model is it's actually designed to be run on device. We can see here it says, This model's efficiency and local deployment capabilities could attract businesses operating environments with limited internet connectivity or those with stringent data privacy requirements. They do go on to say that it's more designed for laptops and desktop PCs than smartphones, so companies that want to run a really powerful large language model with a large context

window that can take a lot of input and a lot of output text and maybe concerned about privacy or not have internet access. Well, now they have a model that they can use that's going to provide pretty much everything you're going to need. It says the model is immediately available, and we have a link here with a downloadable version promised in the near future, so you can actually try this model out over on NVIDIA's website. If we come over to build.nvidia.com, explore, discover, click on reasoning over on the left side, we can see fresh off the press mral Nemo 12b. Instruct, if we click in here, we get a chat window where we can actually play around with this model if we want to. Again, this is a cloud version where you can just sort of play around with it, but a desktop version is coming soon. And finally, if you're planning on watching the Summer Olympics this year, it looks like Google's AI is going to be everywhere. Google is apparently the official AI sponsor for Team USA.

claim that they're going to have ads all over for all of the various Google AI products. So if you haven't heard enough about AI lately on TV well watching the Olympics, you're going to see a lot of it. And anyway, that's all I got for you today. Again, I record these videos on Thursday. In this case, some of it was recorded on Wednesday, some of it was recorded on Thursday. So if there was any news that came out late Thursday evening or on Friday, it didn't make this video, but it will be in next week's news video. While I'm here on vacation, I'm sort of slowing down on putting new videos out, but I am going to publish my Friday news videos on schedule just like I always do, just less other videos during the week this week. If you like this video and you want to stay looped in on all the latest AI news, the coolest AI tools, interesting AI research, and you know some of my own commentary and opinions around it, make sure you like this video and subscribe.

to this channel. It really helps my channel grow. It will also ensure that you see more videos like this one inside your YouTube feed. And if you haven't already, make sure to check out Future Tools, where I curate all of the coolest AI tools I come across. I keep the AI news page up to date on pretty much a daily basis, and we've got a free news newsletter where you can get all of the coolest AI tools and most interesting AI news delivered directly to your email inbox. You can find it all over at Future Tools, completely free. Thank you so much for tuning into this video. I really appreciate you. Thank you so much to HubSpot for sponsoring this one. I have a feeling the AI news is really going to start heating up again real soon. There's a lot of cool things in the works that I've sort of been getting some sneak peeks of, and I'm excited to share what's on the way. So make sure you're subscribed, make sure you're tuning into

these videos. I really appreciate you. Thanks so much for hanging out with me. I'll see you in the next video. Bye-bye.

AI News: We're One Step Closer To AGI This Week!

AI, No-Code, Tech, Futurism - I'm a tech nerd and talk about tech nerd stuff

Matt Wolfe

OpenAI's Progress Towards AGI and the Strawberry Project

OpenAI has outlined five levels of progress towards AGI, from chatbots to AI that can do the work of an organization.

The company is reportedly working on a new reasoning technology called "Strawberry," which aims to autonomously navigate the internet for deep research.

Whistleblowers have also raised concerns about OpenAI's business practices, accusing the company of illegal employee treatment.

Allegations Against OpenAI and Updates on AI Tools

Employees at OpenAI are alleging the company's policies could result in the loss of vested equity for whistleblowing.

There are speculations that OpenAI's DALL-E image model has been updated, improving the clarity of generated images.

Sora is generating excitement with new demo videos, despite competition from other AI tools like Runway Gen 3 and Lumi's Dream Machine.

New Developments in AI and Education, Chat Apps, and Video Creation

Andre Karpathy announces a new venture, Eureka Labs, focusing on an AI-native education model using AI teaching assistants to scale learning.

The ChatGPT app is now available on Android, offering a conversational voice feature, while Gemini provides answers on Android phones without unlocking the device.

Google introduces Google Vids, an AI-powered video creation app integrated with the Workspace Suite.

New Innovations in Tech: Workspace VDS, YouTube Music Sound Search, and AI-Generated Conversational Radio

Workspace is testing a new application for creating slide-style videos with voiceover and stock footage.

YouTube is rolling out a music sound search feature, similar to Shazam, that can identify songs from a snippet or even from humming.

Additionally, YouTube is testing an AI-generated conversational radio feature that allows users to create a custom radio by describing their preferences.

Apple, Nvidia, and Anthropic have been using swiped YouTube videos to train AI, causing controversy, while Microsoft is integrating its Designer platform into various apps, allowing users to create specific images and styles directly.

New Features in Designer, Codestroll Mamba, and Rufus: A Quick Overview

Designer has added new features like the restyle function and a mobile app for easy image creation and editing.

Codestroll Mamba, a new open-source AI model by Mistol, is designed for code generation and offers a fast response time for longer input text.

Amazon's AI shopping assistant, Rufus, is now available in the Amazon app, providing shopping and political advice based on Amazon's data.

EU users may not have access to Meta's multimodal models due to regulatory uncertainty and challenges with GDPR, while the UK users are expected to have access to the new model.

A MIDI device is being used to manipulate images through stable diffusion, with the code available on GitHub for experimentation.

An AI-powered app from Gizmo China can turn selfies into printable 3D characters, sparking interest and curiosity, especially for those with 3D printers.

OpenAI has launched GPT-40 Mini, a new, faster, and smarter language model with support for text, image, video, and audio, aiming to compete with other large language models.

The Latest in Large Language Models and AI Sponsorship for the Summer Olympics

GPT-4 is the top-performing model, with GPT-3.

5 Mini coming in second, outperforming other mini models.

Nvidia and Mistol have collaborated to create Mistol Nemo, a powerful on-device language model designed for environments with limited internet connectivity.

Google is set to be the official AI sponsor for Team USA in the Summer Olympics, showcasing its prominent role in the event.

Google is planning to heavily advertise its AI products during the Olympics, aiming for extensive visibility.

The YouTuber is slowing down video production but will continue with the Friday news videos as scheduled.

The YouTuber encourages viewers to like and subscribe for the latest AI news and tools, also mentioning Future Tools for AI resources.