AI NEWS: GPT User-Guide, Insane Video Effects, Massive Leap in Coding Abilities & More

0:00 As with every week, there's been a ton of AI news this week. And instead of breaking down every single little thing that happened like I do most weeks, I'm filtering this one down to just the things that I kind of thought were the coolest and the things that I felt would be the most helpful to the most amount of people. I'm not going to waste your time. So, let's break it down. Right. Meow. So, let's start off with something that I think a lot of people will find very helpful. This week, OpenAI put out this little guide called When to use each model. If you're on one of the paid chat GPT plans, you probably realize there's a ton of different models. You've got 40, 4.5, 03, 04 mini, 04 mini high, 01 pro mode, GPT40 mini. You have so many different options, and well, they can be confusing, like why would I

0:45 use 4.5 over 03? And that's what this quick guide aims to help you with. Now, they call it the chat GPT enterprise models and limits, but what they explain here really goes for whatever plan you're on. You don't need to be on an enterprise plan for this to be helpful. But really quickly, here's why there are so many different models. Over at OpenAI, they're constantly testing new ways to improve the models. They're doing things like adding chain of thought thinking after you give it the prompt. They're pushing out sort of new checkpoints that are probably midway through a bigger training. They're constantly just testing different ways to improve upon the models. Now what ends up happening is that it may really really improve a model in one area but then actually cause the model to perform worse than older models in other areas

1:32 when they do these little experiments. And so instead of just saying here's our newest best model that might have gotten a lot better at coding and math but maybe got a little bit worse at like EQ, they let you pick which model you want to use. So here's how they break it down. with GPT40. It excels at everyday tasks. This is probably going to be the one you're going to want to use by default the most because it's really fast and works pretty good for most things. So, it's good at brainstorming, summarizing emails, and creative content. It's also the model that you can giblify stuff with if you want to create images and reimagine images. It can also search the web. It's what's being used when you use advanced voice. It could ingest documents and images and

2:18 CSVs and audios and video. So for most scenarios, again, this is probably your default model. Some of the example prompts, summarize meeting notes, draft a follow-up email, proofread my report, brainstorm a launch plan in real time. And then you have GPT 4.5, which OpenAI already announced they're going to wind up pretty soon. So this one's not going to be available forever. We already know it's going away probably within the next month or so. This one is best at emotional intelligence and clear communication. It's best at creativity and it's more collaborative. This one, I find, is probably the best if you want to have it help you write an article or write a tweet or something like that. It sort of nails that realistic voice pretty well. Here's some example prompts. Create an engaging LinkedIn post about AI trends. Write a product

3:04 description for a new feature launch. Develop a customer apology letter with an empathetic tone. Then you have OpenAI 04 Mini and O4 Mini High. Personally, I almost never use these ones. I don't really have a huge use case for them. But here's how OpenAI describes them. O4 Mini is good for quick stem related queries, programming, and visual reasoning. Some examples, extract key data points from a CSV file, provide a quick summary of a scientific article, quick fix this Python trace back for me. And then 04 mini high is essentially the same model, but more compute and more thinking time added on. So, it's better for advanced coding, math, scientific explanations, and it thinks longer with higher accuracy. So, example prompts, solve a complex math equation, and explain the steps, draft SQL queries for

3:49 data extraction, explain a scientific concept in layman's terms. And then you have 03. This is the one I use probably the second most behind 40. This one's great for complex or multi-step tasks. strategic planning, detailed analysis, extensive coding, advanced math, science, coding, and visual reasoning. This is one where you can give it a lot of information and have it sort of analyze all of that information and then return to you with a response that sort of takes it all into account and did a lot of analysis. 03 also has a tendency to really like to create tables. So, it likes to take all the information you give it and then create like a table with like three or four columns in it and sort of break things down visually like that. It's really interesting how it tries to do that with almost every prompt, even when it's not necessary.

4:36 Example prompts. Develop a risk analysis for market expansion. Draft a business strategy outline based on competitive data. Run a multi-step analysis on the CSV. Review pipeline metrics. Visualize the data and search for new top offunnel strategies. And then you have OpenAI's 01 pro mode, which is good at complex reasoning, takes a bit longer to think, but delivers the accuracy you need for complex tasks. Draft a detailed risk analysis memo for an EU data privacy rollout. Generate a multi-page research summary on emerging technologies. Create an algorithm for financial forecasting using theoretical models. Since 03 came out, I pretty much never use 01 pro mode. And 01 prom mode's only available in the $200 a month plan anyway. So, I found that to be really helpful and I thought a lot of other people would find it helpful. So, I thought I'd lead off today's video with that. We got some cool updates out of a lot of the

5:22 creative tools this week as well, including Hey Genen Avatar 4. This one allows you to upload just one photo with a script and your voice, and it will actually turn it into an AI avatar talking head video like the ones they have in their demo here that I'm sort of scrubbing through at the moment. Now, there is music in the background. I don't want to play it because I don't know if it's copyright or not, but we can see here that most avatars sync to your words. Avatar 4 interprets them. Built on a diffusion inspired audio to expression engine. It analyzes your vocal tone, rhythm, and emotion. Then synthesizes photoreal facial motion with temporal realism, head tilts, pauses, cadences, micro expressions with just one image. They have some other examples here on their X thread. Anyways, I bought a lot of skin care, Korean

6:10 skincare to be exact, and I'm very excited to get that stuff. It could take side angle like profile images. There is no one way to create art. You can make characters sing. Again, not playing this one cuz I have no idea of the copyright status of the music, but she's singing in this video. Talking animals, talking game characters, talking anime. You know what? Let's just try it out. So, I'm logged into Hen here. We've got photo to video with avatar 4. So, let's go ahead and try this out. I'm going to toss in my social media avatar here that I always use. Looks like we only have the option to set it as landscape or portrait. it didn't accept it as a square, so I'll go ahead and leave it as portrait. I can also upload or record audio. And then for the voice, I've previously trained my voice into Hey Genen, so I'll just go ahead and select

6:55 that. Let's type a script and we'll generate our video here. And it took maybe about a minute. Let's check it out. Make sure to subscribe to Matt Wolf. Make sure to subscribe to Matt Wolf. Not bad. Let's create a new one here. This time I'll upload a real image of myself here. And let's record some audio this time. And I'm just going to be kind of silly. Let's see. I wonder what it's going to do with this. Who knows? It took slightly longer this time. Maybe about two minutes. But let's go ahead and check this out. I wonder what it's going to do with this. Who knows? Okay, so that was pretty cool. Only thing that was a little weird was the hand. There was no hand in the original image, so it just sort of made one up and it's a little

7:40 fuzzy. But the lip syncing is I don't know. It's kind of awesome to me. Let's watch it one more time. I wonder what it's going to do with this. Who knows? I like it. Do you ever get overwhelmed by trying to keep up with all the AI tools that are available? With literally thousands of tools out there, deciding which ones will actually help your business could be next to impossible. That's why I've partnered with HubSpot to create my personal AI playbook. In this free guide, I've handpicked the best AI tools specifically for entrepreneurs. You'll discover practical ways to leverage AI to simplify your daily workflow, unlock new creative possibilities, and accelerate your business growth. Plus, I've included clear step-by-step guides, insider tips, and easy to follow strategies so implementing AI becomes

8:28 effortless. Whether your goal is to automate everyday tasks with chat bots, create stunning visuals without hiring a designer, or scale your business rapidly using AI, this playbook is your go-to resource. It's all about helping you work smarter, not harder, and ensuring that you stay ahead of your competitors. Don't miss out. Click the link in the description to download your copy of my free AI playbook. Now, huge thanks to HubSpot for sponsoring this video. Now, let's get back to it. This company, Higsfield AI, has been dropping a lot of features recently as well, including their new Higsfield effects mix. This to me kind of seems similar to like what Pika Effects does where there's a whole bunch of pre-built effects that you can just apply to something that you've already created. They give this example here. You can see they selected the turning metal effect and the melting

9:14 effect here. They uploaded this image, gave it the prompt to mix these two effects, and then it created this video where you can see the metallic and the melting effect blended together. Here's another really cool one that I came across from I'm Paul on X here where he used Midjourney version 7 to create the image and then created this animation using Higsfield where it rotates around the character and then the character punches and breaks the glass. Elsene here created this where Wonder Woman's flying through the sky and then lights on fire. So, we can see all of the various effects here like set on fire and thunder god and melting and agent reveal and glam and just tons and tons of stuff. Soul jump. What's cool is you can seemingly blend a couple. So, let's try this like soul jump, but then I'll go up to mix here. And for our second

10:00 one, let's try the set on fire. Let's see what happens when we do soul jump with set on fire. And I have an image of myself that I'll drag and drop in here. And I'll do a similar prompt to what we saw in their demo. Mix these styles together. And it took probably a minute and a half or so to generate this. Let's see what we got. There's my That's actually way cooler than I thought it was going to be. I wasn't expecting both my soul and my original body to catch on fire, but I think it's a pretty cool effect. I kind of wish just the soul coming out was on fire and my normal body, you know, stayed normal, but I could probably fix that with a prompt. Okay, so I tried to get more detailed with the prompt, mix these styles together, make the soul coming out of my body be on fire while keeping the

10:46 original body intact. And well, it didn't quite come out how I expected it, but this is really, really fun. I am really enjoying Higsfield and these various effects that are built into it. This is definitely a tool I'm going to be playing with a lot more myself. This week, Nvidia sort of stealthly released a really good speech to text model, a model for transcription, and it can transcribe 60 minutes of audio in just 1 second. So, you could give it a full hour-long podcast and 1 second later have a transcript with an error rate of 6.05%. Meaning, it's going to get roughly six out of every 100 words wrong on the transcription. But, it's really fast and it's open source, so you don't need to pay API fees or anything like

11:32 that like you would with OpenAI's Whisper or any of the other speechtoext models that are out there. The model's actually available to use over on Hugging Face. So, if you don't want to download it and run it locally, you can actually play with it right now. I'll make sure it's linked up in the description. Now, I don't really have a lot of audio files on my computer to test. But I do have a podcast that I generated with Notebook LM that's about 20 minutes long. So, if I drag and drop this in here, we could see it's just slightly more than 18 minutes long. Let's click transcribe uploaded file here. It took 7 seconds. Okay, ready to dive in. Today, we're tackling AI. Oh, yeah. And we can see it transcribed that whole thing in 7 seconds. Now, it did say it can do up to 60 minutes in just 1 second, but obviously we're running this

12:18 on a cloud GPU through hugging face. It would probably be different if you were running it locally on a really good GPU as well. But almost 20 minutes of audio in 7 seconds. Still really dang impressive. And the transcription looks really clean as well. Again, something I thought a lot of people would find really helpful. If you need transcriptions, you can do it for free on Hugging Face right now within seconds. If you're a Netflix subscriber, you're getting some AI upgrades this week to your Netflix account. We can see in their press release here, a new way to search. We're exploring ways to bring generative AI to our members discovery experience. Starting with a search feature on iOS that is a small opt-in beta. This will allow members to search for shows and movies using natural conversational phrases like, "I want something funny and upbeat." Not really

13:04 AI related, but also interesting. In the coming weeks, we'll be testing a vertical feed filled with clips of Netflix shows and movies to make discovery easy and fun. They're essentially making like a Tik Tok inside of Netflix where you get clips from TV shows and movies. And if you like one of those clips, you can go watch the full thing. I found that to be pretty interesting. In the world of AI, I feel like the people that are getting the biggest impact right now in our sort of current era that we're in are developers. So much news is dropping for developers and vibe coders, right? Vibe coder being someone that like maybe doesn't really know how to code. They just want to sort of develop apps with natural language. If you fall into one of those two camps, you are getting insane amounts of updates lately. For example, Google just released a new

13:50 version of Gemini 2.5 Pro, and it is pretty much the best coding model available based on benchmarks, based on LM Arena, which sort of gives people a blind test of multiple models, pretty much any source that ranks models, puts Gemini 2.5 Pro at the top when it comes to coding. And the new model that they released this week has gotten even better. And some of the things I've been seeing people produce with this are really kind of mind-blowing. One of the new features is a video to code. Gemini actually understands video. It's not transcribing the audio and then understanding the transcription. It actually can see and understand what's going on in the video. So you can actually feed it things like YouTube tutorials and have it go and code what

14:35 is going on in the YouTube tutorial for you. One of the fun ways to test out a new model is to see how it improves existing applications. On the right, you can see the old 2.5 Pro and a fairly basic app. On the left, you can see it actually understood the video in depth and then made a fully functional quiz. This took the experience to the next level. Check out some of these examples of what have been created with this that Google DeepMind shared on X. They fed it this picture of a tree, gave it the prompt, transformed this image into a codebased representation of its natural behavior. And then we can see the AI actually generated this tree with a slider with growth level that can change the size of the tree as you slide it. They did a similar thing with a picture of a spiderweb, a photo of a bonfire. I believe that's like fireflies in the

15:20 grass there, clouds, birds flying. All of this started with an image and it used that image and created a little app that sort of represented the image. A leaf, a water droplet, really, really cool stuff. And I know I've said this before, but you can use this for free over at AI Studio. If you go to aistudio.google.com, up here in the right, you can select your model, including this brand new Gemini 2.5 preview from May 6th. Let's give it a prompt like create a colorful particle simulator that is affected by mouse movements and clicks. Use HTML and inline CSS and JavaScript so it all can be run from a single HTML file. I'm just doing that so I can simplify it for this demo, but let's go ahead and run it and see what we get. So, we got some code. The whole thing took 57 seconds. Let's see if it actually managed to oneshot

16:07 it. I'm going to copy the code, paste it into a text file, and we'll save this as an HTML. So, here's what it generated. We can see all of our particles, and they kind of fade out really quickly. Not really what I had in mind. So, I gave it a second prompt telling it I don't want the particles to disappear. It took 45 seconds. And this time around, it keeps the particles there, and as I move my mouse around, it sort of pushes them around. So, fun little app that I made with two prompts. Got fairly close on the first one and then much closer to what I had in mind on the second one. Obviously, people watching this are going to think of much more useful ways to use this technology. But developers and vibe coders got even more cool stuff this week. You can now create and edit images with Gemini 2.0. So, if you're a developer, you have access to this in the API. Now, we could see some

16:54 examples here where it took this image of a desk with a laptop. It took this lamp and it said, "I want to see what the lamp from the second image would look like on the desk from the first image." And it put that lamp on the desk. Kind of seemed to change the size of the lamp a little bit. But it appears you can do very similar things to what you can do with the GPT40 image model now. And you can do it directly in the API, meaning developers can actually build with this functionality as well. And like before, it is available inside of Google AI Studio. So if I go to Gemini 2.0 know here you can see there's Gemini 2 flash preview image generation I can tell it to create an image of a wolf howling at the moon and 3.5 seconds later I have a decent looking image put sunglasses on the moon and after 4.4 4

17:41 seconds. We now have a moon with sunglasses. Also seem to add sunglasses to the wolf with a few more prompts. I'm sure we can get it right. If you prefer to develop with Anthropic's Claude API, they've now introduced web search within their API. So, if you're developing apps using Claude and you want your apps to have web search functionality straight within Cloud, you can build apps that do that now. Also, if you're using something like Windsurf and you use the Cloud model as your model for coding, it can also use the internet. Now, to ideally help you write better code, OpenAI also rolled out a couple new features for coders this week. You can now connect GitHub repositories to deep research in ChatGpt. So, inside of ChatGpt, if you go up to your settings under connected apps, you have the ability to connect GitHub. And once it's

18:28 connected, if you select deep research, you get a drop-down that gives you the option to connect to GitHub. This seems really handy because if you're pushing your own apps that you're building over to GitHub, you can attach the entire context of the whole app that you've built using this GitHub connection. Or if you're trying to model another app that's already on GitHub, you can pull information from those other GitHub repositories and use that as context when chatting with ChatGpt. So, a really handy feature quite honestly for developers. OpenAI is also rolling out what they're calling reinforcement fine-tuning. So you can actually fine-tune your own model on your own domain knowledge. I believe the way this works is that you can ask a question, it gives you a response and then you sort of rank the response and over time it

19:15 gets more and more dialed in towards the types of responses that you want it to give. And this again is designed for developers. I haven't played with it myself yet, but it's basically attaching reinforcement learning on top of the model so that you can reward it for the responses that it does well and penalize it for the responses it doesn't do well and ideally it tunes in on exactly the types of responses you expect from it. If you develop with Windsurf, this week they had a massive update with Windsurf Wave 8 and they've been rolling out tons of new features like Windsurf reviews that can review pull requests from GitHub. You can give it extra knowledge including knowledge from Google Docs. You can give it API documentation from some of the APIs you're working with and it will understand those documentations.

20:01 You can share conversations that's got a Teams deploy feature, updated analytics, and they're seemingly rolling out new features almost every day with this wave 8 instead of pushing out all the features at once. And since we're talking about Windinsurf and we were just talking about OpenAI, well, it appears that OpenAI has reached an agreement to actually purchase Windsurf for $3 billion. We actually talked about this in last week's episode that OpenAI was interested in buying Windsurf. Well, according to Bloomberg, it looks like things are getting finalized and OpenAI is actually buying Windsurf. I saw a comment on X the other day that said it's interesting that OpenAI is buying Windsurf. If they believe AGI is really close where you can just say go build this thing for me and it comes back and gives you the thing it built, why would

20:46 it be buying an IDE like Windsurf? So maybe AGI isn't as close as OpenAI has been implying that it is over the last couple years. Just something interesting to think about. Windsurf is my coding platform of choice. I really really enjoy working with Windsurf. cursor is very similar, but I feel like Windsurf's feature set is just slightly better at the moment and probably going to get a lot better, a lot faster with quite a bit of money behind them. And talking about AI vibe coding platforms, it appears that Apple and Enthropic want to get in on that game as well, and they're going to be teaming up to build their own vibe coding platform. It's going to be a new version of Xcode, which is Apple's programming software for, you know, developing Apple products. It's going to integrate the Claude Sonnet model, and we don't know a ton more

21:32 about this yet. Mr. AI released a new AI model this week that is very inexpensive to use if you're using their API, also for developers, but it's 40 cents per million input tokens and $2 per million output tokens. That falls roughly around the same price that GPT4.1 mini costs to use. GPT4.1 is actually $2 per million input tokens and $8 per million output tokens. But if we look at the benchmarks, it seems to do really well on code, instruction following, math, knowledge, long context, all of this stuff. It's pretty comparable to Llama 4 Maverick, GPT40, Claude Sonnet 3.7, and to me seems pretty in line with the pricing you'd expect from similar spec models.

22:18 This was also some pretty big news this week. OpenAI, who was planning on becoming a for-profit company, decided that they're no longer going to fight to become a for-profit company. Instead, they're going to be a public benefit corporation. This is the same style of company that Anthropic and XAI both are. So, the OpenAI structure is kind of weird. You have OpenAI nonprofit oversees a for-profit company and that nonprofit that oversaw the for-profit was capped on how much money it can make because of its nonprofit status essentially. Well, by changing it to a public benefit corporation, the potential profit of the sort of overseeing arm of OpenAI now doesn't have a cap on how much profit they can produce. Some see this as a win for Elon

23:04 Musk because he was kind of fighting to keep them from turning into a for-profit company, but Elon claims this doesn't change anything. They now have the ability to make a lot more profits. And to me, it all still feels very convoluted and confusing. And finally, this week, Amazon introduced Vulcan, its first robot with a sense of touch. Because it has a sense of touch, it knows how much it can sort of clamp down on a actual product that it's picking up. with the idea being that these robots can help pick and autonomously do a lot of the packaging and things like that in the Amazon warehouses. Theoretically, it won't break the things it picks up because of that sense of touch. It will know how hard it can squeeze and pick things up. So, it'll be more gentle with, I guess, you know, glass or breakable things and maybe

23:51 squeeze a little bit harder on heavier things that are less breakable. But it seems like some interesting technology that will, I guess, speed up the processing of packages inside of the Amazon warehouses. Anyway, that's what I got for you today. I hope you enjoyed this video. Again, I'm kind of testing some new stuff here. I sort of changed around the order of the news to try to share the things I thought would be the most impactful and helpful to people at the beginning and cut out a lot of the sort of fluff news. Ideally, just giving you the really interesting stuff that you came to see. There are definitely AI news items that I left out. There are every single week, but you can find anything that you missed in these videos over on futuretools.io. Click up on the AI news page, and you'll see all of the AI news that happened this week. Future Tools is also the site where I curate

24:37 all of the coolest AI tools that I come across. New tools are being added every single day as I discover them. And I also have a free newsletter over here where just twice a week I'll share just the most important news and the coolest tools that I come across. It's totally free and if you sign up today, you'll get free access to the AI income database, a database of interesting ways to make side income using various AI tools. Again, all free over at futuretools.io. And if you like this video, you want to stay looped in on all the latest AI news and all of the coolest AI tools and see them put to the test. Make sure you like this video and subscribe to this channel. I'll make sure more stuff like this shows up in your YouTube feed. Really appreciate everybody tuning in today and hopefully I'll see you in the next one. Make sure to subscribe to Matt Wolf.