An absolute WHALE of a month for Open Source AI

Hello. Welcome to Machine Learning Monthly. The latest and greatest in the world of AI and machine learning, but not always the latest.

Wow. Hello everyone.

Daniel Bourke here, machine learning engineer. And in this video, we're going to cover possibly what was the biggest month in open source AI ever.

And that is January 2025, the latest and greatest in the world of AI and machine learning. I've collected all of the most awesome resources and wrote a newsletter about it. This is the video version of that newsletter.

If you just want to read, the link will be below, as well as links to everything else we talk about in this video. Also check out the timestamps. If you have any feedback on the format of this video, or if you'd like to see anything in a future episode, please let me know.

Otherwise, sign up for the email, you'll get it free at the start of next month. Enjoy.

So we have some beautiful links here from the internet.

This is from PDFs to AI Reddit ready structured data from our friends at Explosion AI, who are the creators of spaCy.

So really great blog post here with a tutorial on how to use Docling, which is—

if we go to Docling, which is a great resource. Of course, all the links will be below that you need for this— is a great open source way to get your PDF, Word documents, HTML into a format that is readable for frameworks such as Llama Index or LangChain or LLMs in general. So if you want to work out how to get your PDF data, if you're in an organization with a lot of docs, check out this beautiful blog post from the Explosion AI team on how to get your PDFs ready for LLMs using LLMs— sorry, using spaCy as well as Docling and Yeah, just a great all-around guide. I'd highly recommend checking out Explosion's other blog posts and of course their labeling tool Prodigy is what I use to label image data and their library spaCy is industrial strength NLP for LLMs and everything else.

So let's jump back into the document, the newsletter. Hugging Face, our favorite open source machine learning company, has released even smaller versions of their already smaller VLMs.

Now this is one of my most exciting releases or upcoming things for the future is smaller and smaller models that we can run on our own devices.

So there is SmallVLM or SMOL.

I found out this the other day that this is apparently a meme way of spelling small.

SMOL. 256 mil parameters and 500 mil parameters.

And of course there is SmallVLM 2B. There's links here.

They all perform at varying levels of performance. So please go and check out the, the results there. And of course, don't take the benchmarks as gospel, right?

You see some numbers on a benchmark, please try it out, uh, on your own data first before you go, yep, I'm going to substitute that in. There's small VLM 500 mil instruct.

Is there a demo here?

Oh, this is running in our web browser. Here we go. Can we load the model? Let's see how this works. We'll wait for the model to load. Wow. So this is running live in the web browser.

Oh my goodness.

And this is— I'm running a Mac Mini M4 Pro. There we go. That is cool. Plated meal likely from a restaurant or dining establishment. Top left corner of the plate, glasses cleared with a light colored liquid, possibly water or juice. Maybe up there. Yeah. Left side of the plate, several slices of grilled cheese sandwich. Okay, so it's probably hallucinating a bit. The place itself is garnished.

Oh, sliced cucumbers, red onions.

Beautiful. Okay, so a little bit verbose there, but you could definitely fine-tune this model on your own dataset and then deploy it because it's so small. This is running in the browser. That is so cool. So big shout out to the Hugging Face team for releasing SmallVLM 256 and 500, as well of, of course, 2B.

They both use this SigLIP vision model, one of my favorite vision models.

So if we look up SigLIP and here we go, this is an open source model from Google which does zero-shot image classification, pairs up, it matches in embedding space images and language. So I'll let you read a bit more into that.

And then we've also got another release from the Hugging Face team. Which is an open source library.

I might zoom in a bit more so we can see here. Dedicated to AI agents.

So we've seen a lot. I don't know if you've been on the internet much in 2025, but apparently this is the year of AI agents.

And of course, trust but verify with everything you hear in the world of AI. AI agents are definitely getting better, but will they work for your workflow?

Do they need to? Does your workflow need one?

Well, that's totally up to you. I love this table that they put in the blog post, which is up here, Small Agents. I'll let you read through that, but essentially it's introducing the Small Agents library on GitHub. I'll give it an upvote there.

Now we have agency level from 0 stars to 3 stars.

And so at the 0-star level, we have LLM has no impact on program flow. So this might be that you just have some unstructured text and then you put it through an LLM and it structures it. So that's it. You just have input output, whereas the agentic workflow might use the LLM as to control all of the steps. So the popular example is to get the current temperature. So if you say to your agent or LLM, you type in, please get current temperature, and then it will call a function, gets current temperature of whatever location that you're in, and then it goes from there. But of course you might be thinking, well, couldn't you just use a programmatic workflow for that?

And the answer is, in a lot of the choices, you definitely can.

But there are starting to be some capabilities of agents to do or take steps where programmatic workflows may not help. But that will definitely be based on your own specific problem. This is a great section here, when to use agents and when to avoid, avoid them, and Code Agents. and then I love the little output or sorry, the layout of the library here.

Essentially, to build an agent, you need a list of tools.

So tools that your agent can use, such as getting the current temperature, and then a model, which is, uh, a model in this case is any LLM. So I believe for the Small Agents Library, uh, yeah, you can pick from a list of many different LLMs. Uh, they will have varying performance. And of course, as AI agents become more of a thing, I assume over the next year or so, they will get much better. So check out the Small Agents Library as an open source way to start building your own AI agents. And then of course, there's the GitHub repo and there is a guided tour from Hugging Face themselves in the documentation.

But I'll leave those for you to explore more.

I'm going to star this. Beautiful.

Let's go on to the next one. So Chip Hewitt, one of my favorite writers in the world of AI and machine learning, has some incredible books, and this blog post is one of my favorites. And what it essentially gets to is the point of the pitfalls that people face when building generative AI applications.

And one of the things that I noticed when I was doing machine learning consulting for a company a few years ago was that when it was machine learning, before it was rebranded to AI, they thought machine learning was the magic bullet to solve every problem.

And that wasn't necessarily the case. Now, AI, although it has become much more capable in the last few years, is still not necessarily the silver bullet for every problem.

Now, is this going to change in 5 to 10 years?

We don't know.

But for now, there are some pitfalls that people are facing when they're building generative AI applications.

So number one, using generative AI when you don't need generative AI. So what's Google's number one rule of machine learning?

Don't use it if you can avoid it.

Confuse bad product with bad AI.

So what we're seeing at the moment is that AI models are getting, of course, very capable.

But what is still quite lacking is the user experience. And that is the hard part.

Often in the world of AI, demos are the easy part, and then the hard part is getting that into production.

So I love this LinkedIn example here. That LinkedIn developed a chatbot for skill fit assessment.

They discovered that users didn't want correct responses, uh, users wanted helpful responses. So that's something to keep in mind.

And starting too complex. So this is in the case of maybe you're using an agentic workflow when direct API calls to the model could work instead.

So keep that in mind when people say that AI agents are the next big thing. Maybe you could do the same thing with just prompting instead of fine-tuning or building an agentic workflow. Over-indexing on early success. This is that LinkedIn, the same example before, which is it took them 1 month to achieve 80% of the experience they wanted and an additional 4 months to surpass 95%.

So this is that curve you see, right, where a lot of the results come basically straight away.

But then to make it that polished product that you can deploy to many, many people, that takes a bit more work.

So if we keep going, they forgo human evaluation. So this is— I had a phone call the other day with someone looking to—

a company looking to build an AI-powered application. And basically, my number one piece of advice to them was, have you got an evaluation set?

So models change almost every day now, but evaluations last forever. So if you're working on your own AI application, build a good evaluation set.

That way, whenever the model changes, you can just swap it out for the better one and see how it goes on your own test case. And of course, human evaluation is still necessary.

Answer.AI trials Devin, the AI software engineer, and the results are, well, not so great. So in short, there was— out of 20 total tasks done, there was 14 failures.

3 successes and 3 inclusive results.

So not the most ideal outcome. I don't want to be too harsh on anyone's products these days, but I believe Devin came out claiming that they were going to be an AI-powered software engineer.

And I mean, it's probably pretty good at doing like stuff that there's lots of examples on the internet of, which is what I've found personally with using a lot of AI coding agents. But for—

yeah, so here's another thing. Back to the Chip Ewan blog post is that they had early wins.

So they thought that it was going to work quite great, such as converting a Notion database into— or pull data, data from a Notion database into Google Sheets.

Devin did that surprisingly well. So is that something that has a fair bit of documentation online already?

Maybe even Notion themselves, I'm guessing, have an example of doing exactly that in their documentation. So For me, I would expect that if there is a straight up example of how to put data from a Notion database into Google Sheets, I would expect a modern AI system to do that pretty well if there's just that work case. So it's, it's read the internet, it should be able to basically repeat that workflow but sub in my own data.

So then they did some more testing.

They have a great breakdown of what went wrong, creating new projects from scratch. Research tasks, analyzing and modifying existing code, and then reflecting as a team. So yeah, here we go. Here's a great one from Hamal Hussein, which is another great writer in the world of AI and ML. I really enjoy his work.

So Devin struggled to use internal tooling that is critical at AnswerAI, which in addition to other issues made it difficult to use. Apparently Cursor—

I personally haven't tried Cursor yet, probably one of the the 3 people in the world that haven't, where there is more opportunity to nudge things in the right direction more incrementally.

So I'll leave you go and have a look at the results of trying out Devin.

On to the next.

Microsoft released Phi-4.

I believe I'm saying that right, or Pi-4. So this is an incredible 14 billion parameter LLM and it has been trained on a fair bit of synthetic data.

So I'm not going to step—

go through everything here. I'll let you read through that. But the primary use case for a smaller model like this, which is designed to accelerate research on language models for use as a building block for generative AI-powered features, provides uses for general-purpose AI systems, primarily in English, which require memory-constrained environments, latency-bound scenarios, reasoning, and logic. So one of my favorite things to use Phi-4 for.

Do we have a demo?

Yes, we do. So working on Nutrify, write me a poem about—

so say we have— what I find a lot on the internet is image captions of food.

So write a detailed image caption for a plate of bacon, eggs, toast and fruit. So this is just a demo use case. This is, as you'll see, it's going to be running on Xero, which is Hugging Face powered GPU. Wonderful.

So say I came across one of these.

For my example use case, I would use a model like Phi-4 if I wanted to run it locally, would be to extract all food items to JSON from the following image caption. So again, this is just a demo. I got Fire to generate this. So this is kind of feeding in the test set, but here is a JSON representation. So if I had a lot of these that I've discovered on the internet or I've had a model write them and then I wanted to get it into structured data form, I would use potentially a model like Fire 4 to extract all of the food items into structured data. So go and check out FiveFour and there's a technical report to go with it, but it is open source and I believe, yeah, MIT license. So that is beautiful. Thank you, Microsoft, for that.

Now Gina AI have released ReaderLM v2.

I've got a little demo here which essentially it takes a web page, converts it to raw HTML, which is freely available if we go Inspect, we get all that.

But what if we wanted to extract all of the content from here?

Well, we could write some rules to just extract things out of body tags and heading tags and all that sort of stuff.

But we could also use our LLM. And so that's the primary use case of ReaderLM V2, which is also— oh, CC BY-NC 4.0. So I thought this was a bit more open source or permissively licensed than that.

It might be licensed to use commercially via the API, but either way I found it pretty cool. It did take quite a while running in Google Colab was to extract from HTML and get out markdown text. So the benefit of that is markdown would enable you to pass that into an LLM to process a bit more and go from there. So I tried it on learnpytorch.io, tried it on page 0 or page 1, passed in the HTML from here, and it had a pretty good result. So if we zoom in, can we see that? It wasn't exactly perfect, but it was basically calling one line of code rather than writing a custom parser to go from, uh, pure raw HTML.

So if we move on to the next one, we have some static embeddings. So what is this?

Well, static embedding models, uh, essentially like a dictionary lookup. And so if we imagine we want something to work really fast on device, so there are a lot of good embedding models or text embedding models that are coming out lately, and we can go to the mTab Benchmark, which is Hugging Face's massive text embedding benchmark.

Now, be careful when I show you any of these leaderboards. It can get quite addicting going back to them.

So here we go. These are different values. So the size of these models is going to be really influential to how you use them. And so what I mean by that is a 7 billion parameter model, you're not necessarily going to be able to run that on a mobile phone, quite quickly. You could run that on a dedicated GPU, but for something to run on, say, a mobile device or a small CPU, you're going to want a smaller embedding model. And that's where static embedding models come in.

They are 400 times faster or up to 400 times faster with up to 85% of the performance.

So if you're running on compute-constrained environments and you need to run some text embedding, make sure you go and check out this static embeddings blog post. There's a great walkthrough of how they're created. There we go. And you can run that in Google Colab. That is going to run incredibly quick.

Mistral have released— have upgraded their already, already epic open source AI game. So Mistral Small 3 has been released and it is under Apache 2.0. Let's see if they've got any more because Mistral have been on quite the—

there we go, actually that's still their latest at the time of recording. Of course, this, this AI world moves fast.

So you can fit this on a single RTX 4090. So I've got, I've got an RTX 4090 over there.

I haven't personally tried to run it on there yet, but apparently I'll take their word for it that you can run it on there.

You might have to quantize it, of course, or a 32GB RAM MacBook once quantized.

So if you want to learn more about how it came about, there is a blog post on Mistral AI as well.

Oh nice, they've got an upgraded blog. So yeah, there's some great results here. It's comparable with things like Llama 3.3, uh, 70B or GPT-4o at probably half the parameters, or less than half the parameters.

So it's preferred against Gemma 2 27B, preferred against models that are basically the same level of parameters.

And it's definitely competitive, slightly less preferred than models that are double its size.

So really well done to the Mistral team on this release. And again, really, I'm really excited for the, the next releases in terms of smaller language models that we can all run on our own hardware. So go and check that out. Mistral 3 small is available, Apache 2 license.

Oh yeah, you can also run it via Ollama, which is on device. Let me know if you'd like a—

I want to figure out a bit more on how to run these models on with frameworks such as Ollama. So let me know if you'd like to see a video on that in the future. And if we go back to here, you can also run it via MLX. Which is Apple's dedicated framework for running AI models on Apple Silicon. Let's get rid of some of these extra tabs.

Of course, all the links that you need will be in the blog post here. Now, the Timm library is now directly available in Hugging Face Transformers with both forward and backward capability, as in you can train a model with Transformers from Timm and then upload it back to Timm if you wanted to. So this is really helpful. We use, or I use, Transformers—

sorry, Timm models from there to train the computer vision models that run in Nutrify's app on device. And so it's really cool now to see that we get all the benefits of the Transformers API.

We can train it with Trainer.

So if we look here, we can define a trainer, we can set up a trainer, and model.train. And then we can save that back to either the Timm Library itself, or we can save it to our Hugging Face profile and share the model with other people there. Here we go. Of course, possibly the biggest news of January 2025.

Uh, there has already been so much on this, so that's why I've kind of just dedicated a little paragraph here of my own, and that is Deepseek releasing Deepseek R1 and several distilled variants as open source with an MIT license. Now, of course, this is a model that's going to require a fairly large or fair bit of compute to run. There is an API here, but the amazing thing is that it is comparable or better than OpenAI's O-1 on all the important benchmarks that are out there. So of course, Deepseek R1 is based on, I believe it's Deepseek V3.

Yeah, there we go. Which is already an incredibly, there we go, 671 billion parameters. But only 37 of them are active when it does inference.

DeepSeq-v3-base is another open-source large language model.

I don't believe it has reinforcement learning training, um, by the DeepSeq team. And so it's based on that— you take a base large language model, you add reinforcement learning to improve its chain of thought. So if we have a look at the paper, I believe I've got that here ready in GoodNotes.

So this is how I read research papers.

If you'd like a more in-depth video on this, I'm happy to do that in the future. But this is the Deepseek R1 paper.

This is an app called GoodNotes, and I like it because you can go through and mark up things.

It works on, on my iPad, which is where I like to read things and whatnot. But long story short, one of my favorite takeaways from here is that they not only released an amazing model in Deepseek R1, they, they shared the ingredients on how they did it via research. They didn't necessarily release the code or the data, but there are many smart people out here who are already replicating this functionality, as we'll see in a second.

But what did it do?

Yeah, post-training large-scale reinforcement learning on the base model.

So we directly apply reinforcement learning.

So reinforcement learning is you have a state, you have a model that's trying to achieve some goal, and it's trying to update its internal parameters based on the step it's taken to reach that end goal. And so the RL here, reinforcement learning that they did, was to explore chains of thought. So chains of thought are— I'll just go— I have an example here. As you can see, I was trying to get Claude to tell me how to set up my OBS for recording this video. We'll get to the aha moment in a second. But if we go chain of thought, so try to solve the square root of 169,004. Chain of thought could be the steps that it breaks things down to, to get to a better answer. So if you think about your own working, uh, when you try to solve a problem, generally you do better if you give something more thought, whereas if you just give a quick answer, um, or guessing what the square root of 169,004 is, uh, your answer will probably be wrong unless you're an incredible math genius. So what we do here is we tell the model to go along the way that you get an aha moment as you plan the steps to solve it. So here we can see first we're doing this, start with the leftmost pair, bring down the next pair, bring down the final pair.

And that is actually a really cool way to solve it. I didn't know you could do that.

So this is where the aha moment is in there. And this is what DeepSeq did. They collected many of these examples, these chains of thoughts. So I have this problem that I'm trying to reach, and here are the thoughts I've gone to reach that. And then they trained their base language model to continually do that for many different problems. I believe they used about 800,000 samples, but that'll be in the paper. And the reason why I put the aha moment in here, because one of my favorite takeaways from this paper— they used group relative policy optimization as their reinforcement learning algorithm. There's their prompt template to how to— they created Deepseek-R10. So that was a version of Deepseek-R1 but completely with no chain-of-thought data to reference. It just improved itself via self, self-play. And so if we go down here to the finding that they found, the aha moment.

So a particularly intriguing phenomenon observed during training of the Deepseek-R10 is the occurrence of an aha moment. This moment, as illustrated in Table 3, occurs in an intermediate version of the model. During this phase, the DeepSeq R10 learns to allocate more thinking to a problem by reevaluating its initial approach.

This moment, or this behavior, is not only a testament to the model's growing reasoning capabilities but also a captivating example of how reinforcement learning can lead to unexpected and sophisticated outcomes. So, for example, if they started with the question, if a is greater than 1, then the sum of the real solutions of square root of a minus the square root of a plus x is equal to x, or equal to x is equal to. So they start the response off with think, then it goes through.

These are the steps, chain of thought to solve that equation.

And then it discovers that there's an aha moment to flag here.

And so what I've been trying is to tell other LLMs to just use this aha moment to improve them. I don't have any quantitative results on how that's improved my LLMs.

But I have an example here of where Claude used the aha moment to break it down.

Is that needed? I don't know.

100% no. I'm still exploring the world of prompt engineering as well, but I find it pretty cool that it is using sort of the same thing that you would find if you were talking out the problem out loud. Now, whether this is important or not, or whether it's just referencing pop culture such as movies or scripts that have gone, aha, and then gone like this. I'm not sure. That's going to require further investigation. So really cool that they shared the model and the paper. If you'd like a dedicated video on Deepseek R1, please, uh, let me know in the comments below.

So let's go on to the next thing. Uh, Deepseek are kind of on a roll here. They released Janus Pro 7B, which is a single unified model that is able to generate and understand.

So it, it can generate images but also acts as an LLM, if I'm understanding that correctly. I did find that, yeah, if we go here, let's create an image. Yeah. So see, we can do a question about an image, but we can also create an image. So that's cool. That's within the one model. So it's using the same backbone here rather than using a different model for generation and a different model for understanding. So generate some images there. We'll come back in a second.

But I used a prompt here. I tried it out with a lot of different current state-of-the-art image generation models, and I found I had varying results with all of them. They weren't all perfect, but Janus Pro provided—

7B Pro provided quite it understood the prompt, but the image quality wasn't that great, whereas FLUX had the best image quality, but it didn't fully capture all of the stuff in the prompt. So again, all of these things are amazing, but you should definitely test it on your own use case. Um, now speaking of open source efforts to reproduce Deepseek R1, we have Hugging Face have OpenR1. So this is a plan of attack to use the Deepseek R1 tech report. So that's why it's so amazing that they open source both the model and the tech report of how they did it. Now we can kind of verify, uh, if we replicate this method plus the data plus the code, um, we should get similar results to the DeepSigR1. But they're in the process of reproducing that. I'm not sure how far along it is or where it is, but this is a library if you want to see how potentially the DeepSigR1 came about.

This is what it's trying to—

OpenR1 is trying to be— it's a work in progress right now— is trying to be that open source example. So step 1, replicate the R1 distill models by distilling a high-quality corpus from DeepSeq R1.

Step 2, replicate the pure RL pipeline that DeepSeq used to create R1.0.

This will likely involve curating new large-scale datasets for math, reasoning, and code. Yeah, so like I said before, the way RL— I understand it works for language models is for math.

Say you had the workings of the problem question and then you had the answer and then you had all the working steps to go in between. You're trying to optimize your model to start at this state and achieve this state by taking these steps. And you, of course, you can't have examples for every way to do that.

So you use RL so that your model hopefully generalizes to taking the steps rather than just copying the steps to get from input to output. And so step 3, show we can go from a base model to RL tuned via multistage training. So this is a great example of— there we go, GPRO Trainer— of how quickly the open source world moves. So we have a model about 2 weeks ago and now we have a GitHub repo with 16 or 17,000 stars that is also Apache 2.0 licensed.

On trying to reproduce it. Oh, I love this. And so we have another— speaking of huge open source LLMs/VLMs, VLM is a vision language model. We have Minimax01, which is capable of producing— this is the biggest I've seen so far— or processing 4 million tokens. Now that is pretty cool.

So it performs— if we go down— basically on par or better than all of the best models that are out there right now. GPT-4o, Claude 3.5 Sonnet, Gemini 2.0 Flash, Qwen 2.5, Deepseek V3, Llama 3.1 405B. And then we have Minimax Text and then we have Minimax VL and they are quite cheap.

So API pricing is $0.20 per million input tokens and $1.60 per million output tokens. So go and check this out.

They have also released their models on Hugging Face. So if we go to Minimax AI—

oh, their link isn't working.

I might have to tell them that. There we go. So this is Minimax Text. And then if we go into here, we get Minimax VL. So if we go there, beautiful.

The thing is that they are 450 or 405 billion parameters.

So you're probably going to have to have some pretty good hardware to run the full model.

So here's the model files.

There we go.

One 5GB file and then about 400 2GB files. So the model is nearly a terabyte. So that is awesome that they open source the model.

But yeah, if you have the hardware to run that, please let me know.

I'd love to hear if you've got it working on your own hardware. Is there a license here?

No license. Anyway, check that out.

Thank you to the Minimax team for open sourcing that. Love seeing that.

Let's— oh, there we go.

Wow, does that look like me? This is our Janus Pro 7B image that's generated. So not that great generating people. Maybe I could try a better prompt or something like that. But yeah, that's, that's, that's what I'm trying to do.

I'll try copy it.

So that's— thank you to DeepSeq team for open sourcing that. And we have OpenBMD-B, which is an 8 billion, uh, open sourcing mini CPM 0.2.6, which is an 8 billion parameter, I believe, uh, model that is capable of—

yeah, there we go, 8 billion parameters under the license. Can't see it over here. Which is, yeah, on par or better than GPT-4o. And it can do audio input, audio output, live streaming capability, incredible OCR, so optical character recognition abilities, and superior efficiency.

So this is a type of model that I really am excited for, is that the small, like 8 billion parameter potentially, if we can get this quality model down to about a billion parameters, we're going to see models like this running live on iPhones or smartphones within the next year or two, I think.

Yeah, there we go. It's about 20GB worth of— or 15GB worth of model files. Potentially it will run great on like an RTX 4090, a dedicated GPU, but not quite just that— yeah, that there yet, sorry, for smartphone. Now, uh, this one will definitely run on smartphone.

This is the best speech-to-text or text-to-speech model I have ever seen.

This is Cocoro 82 million, which 82 million is 82 parameter— 82 million parameters. So for context, the last model we looked at was 8 billion, so about 100x larger.

And so here's the model here, 327 megabytes.

This is definitely capable of running on a smartphone, uh, live, that is.

So do we have a demo here?

Yeah, Cockroach Text-to-Speech. I'm going to put something in here. Go, hello, welcome to Machine Learning Monthly, the latest and greatest, uh, in the world of AI and machine learning, but not always the latest because we, we like to surface things that are from the past.

So we're We generate audio.

Hello. Welcome to Machine Learning Monthly, the latest and greatest in the world of AI and machine learning, but not always the latest.

Wow.

Now that is cool.

So if you want a text-to-speech model with only 82 million parameters, go and check out Coqoro text-to-speech. One of my favorite models ever, basically.

Here we have Quen team releases two new Quen 2.5 updates. So the Quen 2.5 1M series. Is an open source series of models that can process 1 million tokens as context.

So 1 million tokens is if we go tokens to word count by OpenAI's count, we have one token is about 4 characters. So 100 tokens is about 75 words.

So that is about 750,000 words if I'm doing the math right.

And so how many words are in the Bible as a big book is about—

so it can process the whole Bible.

Now, am I getting that right?

100 tokens is 75 words.

So if we go here, yeah, 750,000 words or something like that. Now that is a lot of context to just put into an LLM. So there's, uh, Quen 2.5 1 mil, which is— this is purely text-based. And then we have QWEM 2.5 VL, which is various size models here, which are vision language models. So they can take in images as input. I've been playing around with QWEM 2.5 VL 7B Instruct, and I found it is incredible for— yeah, I built a demo here for just its instruction following capabilities. So let me just upload an image here. Wonderful. So I've uploaded that image there. This is what I had for lunch. So if we were at Nujify trying to extract all of the visible foods in here and put them to JSON, this is going to run on Hugging Face's GPUs. Maybe I'll zoom out a little bit. I've told it here, please list all the visible foods in the image to JSON. And, and here we get— here we get a structured output. We have salad, greens, lettuce, Kale, vegetables. We have broccoli, radishes, seeds, sesame seeds.

Do we have any sesame seeds? Oh, maybe it's seen the tiny sesame seeds.

Alfalfa sprouts over here and grilled possibly fish or tofu. So, that is chicken.

So, not perfect. There is some quinoa there that it missed.

There is some red onion there that it missed. And there is some curried vegetables here that—

this is quite the complex image.

And then maybe in here there is some edamame beans. So not perfect, but certainly a very capable model that at 7B parameters you can run on many, many modern GPUs. And then of course there is a smaller version, 3B, uh, and the large version, 72B, which I think that one is on par or better than GPT-4o. So if we have here 2.5, GPT-4o, 70.3, and Qwen 2.5 VL 72B is 70.2. So slightly 0.1 worse, but it is better in other fields. So check that out if you have the hardware to run Qwen 2.5 VL 72B Instruct. You can have a model that's basically on par with all of the flagship LLMs or VLMs running on your own hardware. So big shout out to the Qwen team. Thank you for putting that out there. And we have another one.

Juan3D from Tencent is capable of creating high-detail 3D models from 2D images and text.

So this is the example that I did. That's a Nutrify icon, uh, and it has turned it into a 3D model.

So the model is open source, and there is also a great demo here.

We'll try the example. We'll generate the shape and texture. This may take a little while, so I'll Let this load. And there we go.

90 seconds later, we get a full 3D model, even the back of the head. How cool is that from a 2D image input? If you are building a game or something like that and you had that, that is cool. You can make your assets on the fly.

I haven't had much experience with text, but from images it works pretty, pretty well, as you can see. And then generated mesh. You can color that in your own style if you want to. So be sure to check that out.

Big props to the Tencent team for open sourcing that.

Idea research.

So if we have a look at this, this is ChatRex.

So I'm going to star that one as well. So this is a really cool example of combining a dedicated model for object detection and a large language model. So a lot of new LLMs are starting to have the capability of outputting boxes.

But what ChatRex is, is it has a universal proposal network.

So it detects all the boxes in an image and then feeds that to an LLM and basically says, hey, can you filter these boxes? So it becomes a retrieval task rather than just a pure generation task to generate box coordinates. So you can try this out. There we go. Universal perception network. And so what this does is, yeah, does object proposal. So it will go over an image. It's a generalist and it will basically say, here are all the things that I think are boxes. And it has two modes, which is one is coarse, so all big objects, and one is fine-grained, which is very small objects. So as you can see here, it's getting quite a few objects of writing on a blackboard or something like that. And then we have all different people here. And then the ChatRex model will detect the things that you're looking for. If you're looking to label a whole bunch of data with objects, ChatRex might be your workflow to go through because it's going to marry up the capability of detecting objects and then labeling them with an LLM.

Finally, if we go into other releases and talks, I won't spend as much time here because I wanted to really focus on the open source. OpenAI released a computer use agent.

I think we're going to see more and more of these in the future as agents become better and better, is AI models directly interacting with our computer. I haven't personally been hands-on with it just yet.

I still like using my computer myself.

And then we have a presentation from Benedict Evans, which I found was really great, which is not going to play that, but AI eats the world. So the viewpoint of how AI might affect society from the point of view of an economist. Some really great insights there.

We have a video from Evan Zhao who recreated Apple's Genmoji in one day.

So from my understanding, Evan created image and text pairs and like, say, emojis with their text descriptions and then combined that with a Flux model, fine-tuned that and could run it locally on his own device to replicate Apple's Genmoji. So text to emoji.

And Oriol Vinales, I hope I'm saying that, Vinales, which is the co-lead of Google Gemini, went on the DeepMind podcast. I really liked this episode.

Gemini 2.0, it actually just got released. Gemini 2.0 is out of experimental mode. I've been playing around with that.

That'll be probably in the next video, next month's video, because this is from February. I'm going through January's stuff.

Maybe I'm going too slow for the world of AI here, but it is an incredible model.

And so if you want to find out some insights from that, go and listen to this podcast and has some great insights into how to train AI agents and how to train world-scale models like Gemini.

Another incredible podcast slash talk was Jürgen Schmidhuber, Schmidhuber, one of the most incredible and prolific people in the world of AI, went on the Machine Learning Street Talk podcast.

And it's just— this talk made me feel good. I highly recommend it for anyone who's like into AI or anything. This is just a feel-good talk from someone who's got incredible amounts of skin in the game in the world of AI. Especially loved the last closing remarks, but I'm not going to spoil that for you.

I'll let you go and watch that on the Machine Learning Street Talk channel. So big shout out to, to Jurgen and the Machine Learning Street Talk channel for such high quality interviews.

And then finally, we have Jeffrey Codes who put out a video of one year studying AI and apparently it was a mistake. So some great takeaways in there.

I'm not going to spoil that again. Math, computer science, LeetCode and algorithms, data engineering, ML and AI, reflections, scattered attention, and what I'd do differently.

If you're learning AI and machine learning in the world of—

in the year of 2025, I would recommend watching this video to see if those insights maybe apply to you.

So if you have anything you'd like to see in a future video or episode, leave a comment below. Otherwise, have fun building with all of the amazing AI tools we now have access to, and I'll see you next video.