This New AI Is Honestly SCARY: Real-Time Self-Learning!

Sakana AI just shattered the limits of how language models learn and adapt. Their new AI is so advanced it can reprogram itself in real time to handle brand new tasks. No retraining, no extra data, just instant adaptation. This is a monumental leap and it's all thanks to their groundbreaking Transformer Squared model. Then later in this video, we'll uncover how DeepSeek's latest open source model is beating OpenAI's flagship 01. Not with billion dollar funding, but with a budget that's closer to the price of a mid range smartphone. How do they pull this off? Stay tuned because this is the kind of innovation that's rewriting the AI rulebook. Alright, first off, let's look at Sakana AI. They're an AI research lab that's all about nature, inspired algorithms and super flexible systems. They've recently introduced a language model called Transformer Squared, which if you can believe it, can learn new tasks without the usual time consuming fine tuning process we've all come to know and well, not always love. Typically when you want to tailor a large language model or LLM to a new task, you need a specialized technique like Lora, which is low rank adaptation. Lora helps by only tweaking a subset of the model's parameters, saving time and money compared to retraining everything from scratch. But Sakana's approach to goes a step beyond letting the model dynamically reconfigure its own weights right there during inference. It's kind of like having a machine that retunes itself on the fly whenever it sees something new, rather than needing a separate big training session for every single new requirement. So how does Transformer Squared actually pull this off? Well, the short version is that it uses a linear algebra method called singular value decomposition or SVD to break down the model's weight matrix into into smaller components that represent different skills like math coding or language understanding. At training time, Sakana's researchers develop so called Z vectors from these SVD components, letting the model turn specific skill dials up or down in real time. When a new prompt comes in, the system checks which skill areas are relevant, picks the right z vectors and applies those adjustments to the base model. Then boom, you've got an on the spot optimized response, all without messing around with a big fine tuning stage. They call their overall process Singular value Finituning or SVF for short. In experiments, they applied transformer squared to llama3 and Mistral LLMs, comparing the results to the usual LORA baseline according to Sakana's blog. This new approach achieves better performance across diverse tasks like math, coding, reasoning and visual question answering. And because it's adjusting itself dynamically, it's simply more flexible. Another cool finding is that these z vectors can be moved between models of similar architecture. If you have a set of z vectors trained on llama 3, you might be able to use some of that knowledge on Mistral, though it performs best when you create fresh z vectors for each target. They've also made the code Open source on GitHub, so if you're feeling adventurous, you can dig right in. Now what's the bigger picture here? Over the last year everyone's been going wild trying to improve how LLMs perform at inference time. Whether it's letting them remember more in their extended context window or training them using in a unique way that doesn't need a ton of extra data, the idea is to make these models more efficient and user friendly in real world applications. In the case of Sakana's Transformer Squared, you can see how it blends seamlessly into this theme. Enterprise users can quickly adapt their existing LLM to a brand new problem that means faster development, lower costs and just a more practical approach to handling specialized data. But that's not all for today's show. Right on the heels of Sakana's announcement, a Chinese AI startup called Deepseek released a model called DeepSeek R1, which also deals with advanced inference time customization, but with a different flavor. Where Sakana's method focuses on singular value decomposition and dynamic weight adjustments, DeepSeek R1 is all about open reasoning and robust chain of thought processes. According to Deepsea, their new model matches the performance of OpenAI's O1. That's OpenAI's Frontier Reasoning LLM, while running at a fraction of the cost, somewhere around 90 to 95% cheaper. That's a jaw dropping reduction. So let's zoom in on deepseek R1 for a second. It's based on a mixture of experts model called DeepSeek V3 and apparently the team has made the entire thing open source under an MIT license. They not only built R1, but they also used it to distill multiple llama and Quen models, supercharging their performance in domains like math. Their Quen 1.5B distillation for instance, even outperformed bigger models like GPT 4. O and Claude 3.5 Sonnet in certain benchmarks. Now when we say distill we typically mean that you use a powerful teacher model to guide the training of smaller or specialized models, transferring knowledge in a more efficient way. DeepSeq seems to have done that on a mass scale with R1. To really drive home how strong R1 is, check out its performance metrics. It hit 79.8% on AMI 2024 math tests and 97.3% on MAF500, which is just a hair above OpenAI01's 96.4%. For that same test when they tested R1's coding chops, it clocked a 2029 rating on CodeForces, doing better than 96.3% of human human programmers. It also notched 90.8% accuracy on the MMLU test. That's just below O1's 91.8%. So across these tasks we're talking about near parity with a top tier model. And again, they're giving it away basically at a bargain if not for free. If you have your own hardware and want to run it locally, the cost angle is really something to behold. According to DeepSeek, OpenAI's 01 model is around $15 per million input tokens and a whopping $60 per million output tokens. On the other side, R1 is priced at only 0.$55 per million input and 2.19 dollars per million output tokens. That means if you're a startup or even a large enterprise that wants to run a lot of queries, you could see a massive difference in your monthly bill. And for some projects, that could be a total game changer, encouraging more open innovation. Of course, the folks at DeepSeek also offer an API for direct integration or you can grab the model weights on hugging face and run them yourself. They even have something they call deepthink on their chat platform, which acts like an alternative to ChatGPT. Honestly, it's all pretty wild if you think about how much has changed in the open source AI space within just the last year or two. Now how exactly did Deep Seq build R1? It's an evolution of something they call Deep Seq R10, which was trained purely with reinforcement learning, no supervised data at all. That zero stage system picked up a bunch of advanced reasoning techniques by learning from scratch through trial and error over thousands of RL steps. It improved significantly on tasks like AIME 2024, boosting its pass at 1 from 15.6% to 71%. With majority voting it went to 86.7%. However, the devs found that while it got good at problem solving, the model's language usage was sort of messy, occasionally mixing languages or producing text that was hard to read. So they added a multi stage approach that blends supervised fine tuning with reinforcement learning, cleaning up the text while still preserving all that advanced reading reasoning capability. By the time they finished, they got Deepseek R1 matching OpenAI 01:1217 across numerous benchmarks comparing Sakana's Transformer Squared and Deepseek R1, it's clear the future of AI is shifting toward models that adapt instantly without costly retraining. Transformer Squared, fine tunes, weight components on the fly. While DeepSeq R1 excels in advanced reasoning through trial and error learning, both cutting costs and boosting flexibility. With tools like Google's Titans also pushing adaptive AI, we're heading towards systems that evolve and expand knowledge in real time. Sakana's code is on GitHub and Deepseek R1 is on hugging face, making this a perfect time for developers to experiment and innovate. These breakthroughs aren't just improvements, they're setting the stage for smarter, more accessible AI. Alright, make sure to subscribe if you enjoyed this breakdown and drop your thoughts in the comments. I'd love to hear what you think about these AI breakthroughs throughs. Thanks for watching and see you in the next one.