NEOs New Automated AI Researcher Changes Everything (Autonomous Machine Learning Engineer)

So it looks like AI is not slowing down at all. Today I'm incredibly excited to show you guys neo, the first ever AI engineer specifically designed for machine learning engineers. And get this, it actually has outperformed OpenAI01 in certain tasks, which is by any standard, pretty incredible. Now, the team behind NEO have actually been working on this product for two years, years. And what they've achieved is genuinely revolutionary. So let me break it down for you guys. This is basically one of the first steps we have in order to move towards artificial superintelligence, because automating AI researchers is essentially the key that unlocks everything. So let me break down for you guys what we're about to see in our first demo. So NEO is going to show us how it approaches a machine learning task from start to finish. Now, for those of you who aren't familiar with machine learning workflows, imagine trying to teach a computer to recognize cats. You'd need to gather thousands of cat pictures, clean up all that data and figure out the best way to process it. Then choose the right learning algorithm and then train the system. It's incredibly complex and typically takes engineers weeks or even months. And what you're about to see is NEO doing all of this pretty much automatically.

Saurabh here from Neo, the first AI engineer for ML Engineers. And yes, Neo has outperformed OpenAI. We have been working on this for the past two years and we would love to show you how it works. NEO is now crawling the competition link to understand the requirements. It's planning key steps like dataset prep, model tuning, evaluation and deployment. NEO used multi step reasoning to explore different possible solutions, pick the best path and start implementation. Designing a powerful data pipeline is crucial in the machine learning workflow. NEO curates, evaluates, merges, transforms and pre processes data, setting the stage for seamless model performance. NEO sets up multiple fine tuning experiments with different hyperparameters to save time and resources. It selects the most promising experiments to pursue in depth and runs them in a GPU sandbox. After the fine tuning process, a detailed evaluation report is generated offering insights into the loss curves that demonstrate the effective completion of the experiments. NEO is now ready to deploy. But first, optimizing the pipeline for peak efficiency, it will analyze throughput and latency across different frameworks like VLLM and tensorati and different GPU types. Once approved, it will lock in the best configuration for deployment. After running experiment, it shares its recommendations. Once approved, it deploys the model and shares the link.

Now let me explain what you're about to see in the second demo. This is where things get really interesting. The team basically challenged NEO to build a credit card fraud detection system. This is one of the most critical applications in financial security. And typically this kind of system takes a team of engineers months to develop, requiring deep expertise in both machine learning and financial security patterns. And what you get to see in this demo is truly remarkable. NEO is analyzing the Kaggle dataset, understanding its structure, and making sophisticated decisions about how to approach the problem. The system is automatically evaluating different configurations and selecting the most promising approaches. Look at how it provides detailed metrics like precision and recall, and these are crucial measures that tell us how the system can detect actual fraud while avoiding false alarms.

I'm just going to show this task I did with NEO recently. I was entering a Kaggle competition, and what I was trying to do was create a credit card fraud detection system. So the first thing I did was I gave NEO the Kaggle data set. And then once I had downloaded it, it was able to go in and kind of start processing the data and just like, figuring out what the overall structure of the data set was. And then once I had a good idea of the structure, it was able to go into, like, evaluating potential configurations and just swing between which one might be most suitable for this problem. So, as you can see here, like the options on the left in the chat, and then once it had evaluated these options, it was able to come up with a finalized plan, and then it executed that plan. So it went ahead with running the training and the experience, experimentation, and then once these things were run, NEO was able to provide us with an evaluation which included precision, recall, and then also an F1 score. So, as you can see, the evaluation is on the chat on the left.

Now, for one of the final demos, I want to show you guys something completely different that showcases neo's versatility. The team gave NEO a challenge using a Goodreads data set, which is basically a huge collection of book reviews. Now, here's why this is particularly challenging. Most book reviews are written in natural language, with people using phrases like, this book is amazing or life changing. And converting these subjective opinions into something a computer can understand is incredibly difficult. Take a look at how NEO approaches this challenge. It's creating a sophisticated pipeline to transform written reviews into numerical data that can be analyzed. It's running multiple experiments with different parameters, constantly learning to improve its approach. And the training loss you see decreasing means that the system is getting better and better at understanding and predicting reader preferences.

I am pretty easily impressed, but this is still really cool. Take a look at what Nio did on this Kaggle challenge. So we gave it a data set of book reviews from Goodreads. And this is a data set that includes both quantitative and qualitative reviews. And after we've given it the challenge, it starts to understand what the requirements are. It gets straight to work designing a solution. So it starts by planning out a pretty sophisticated scheme to train something for this. So the planner has a pretty detailed idea of the both the pipelines and the models that it wants to use. A critic comes in, tries to poke some holes in that, but it appears this is passing muster. So we're ready to go. So all the necessary dependencies, the data set, they're being downloaded. It takes a little bit. You know, I'm at a school not known for cs, so our wi fi isn't very good here. But once that gets through, we're actually going to see what, what NEO thinks of the data set and if there's any problems, if there needs to be any massaging that needs to be done to the data, because that's actually, as I understand, a large part of what doing machine learning engineering is. And already it's coming to the issue I mentioned earlier of how most of the data qualitative. So you have people saying, this book is amazing, this book is mid, this book is life changing. What does that mean? So NEO actually creates a transformation pipeline to turn that into numbers. And then once those, once that data is transformed, it sets up all the training experiments. So it's going to be running different hyper parameters, different ways of training, and it's going to see how that goes. So it ran all those experiments, seemed to pick a favorite, and ran it for 20 epochs. And as we see by the training loss, it actually iteratively improved a good amount. So NEO is not perfect. There's still a lot of room to go here, but this is a lot better than what I could do for sure. So thanks neo.

Now, most people don't understand why what we've just seen is absolutely revolutionary for the tech industry. Traditional machine learning development is like trying to build a skyscraper with a small team of architects and engineers. It's slow, resource intensive and, and prone to human error. NEO is basically turning this into an automated process while still maintaining sophistication and creativity that previously only human engineers could provide. This isn't about saving time or money, though it does that pretty brilliantly. It's about democratizing AI development. And with neo, smaller companies that couldn't afford large teams of machine learning. Engineers can now compete with tech giants and researchers can focus on pushing the boundaries of of what's possible, instead of getting bogged down with technical implementation details. Now, the team behind NEO is currently preparing for early beta access. And I honestly believe this could be one of the most significant developments in AI this year. We're witnessing the beginning of a new era where AI just isn't a tool being developed by engineers. AI is becoming the engineer. And of course, for those of you interested in being a part of this revolution, you can join their wait list for early access. Trust me, if you're in the tech world or interested in AI development, you'll want to keep a close eye on neo. This is the kind of technology that can really start to change things. And do you guys remember this image from the document called the Decade Ahead? This was basically a long form document by someone who used to work at OpenAI and basically documents how the future of superintelligence is basically going to come. And one of the things that they talk about for the real explosion of intelligence was automated AI research. And they state that automated AI research could probably compress a human decade of algorithmic progress into less than a year. And that seems pretty conservative, basically meaning that once we're able to automate the entire AI research pipeline, it's quite likely that superintelligence is going to happen. And of course, with that, it means that AI development speeds up pretty much 100 times. You can see here it says, I expect 100 million automated AI researchers working at about 100 times human speed. Not long after we've begun to automate AI research, they'll be able to do a year's worth of work in a few days. This is going to be extraordinary. And this could easily dramatically accelerate existing trends of algorithmic progress, compressing a decade of advances into a year. So when we take all of these things into account, I definitely feel like this company is one that everyone needs to pay attention to. If there are any more advances from this company and the they manage to somehow get to 80% or 90%, this is going to be a huge, huge development for the world of AI, especially in the area of automated AI research, because this is something that is pretty much new, but it has one of the most start implications for the entire AI industry.