Videos
The Next Wave: Self-Improving AI
Insights
- Self-improving AI models can generate their own training data, accelerating learning and performance gains.
- Open-source AI enables enterprises to run and fine-tune models locally, protecting data and reducing platform dependence.
- Agentic AI systems require new approaches to orchestration, evaluation, and governance to ensure reliability.
- Human-out-of-the-loop AI offers the greatest long-term opportunity, but also the highest implementation complexity.
In this video, Dr. Boris Bolliet, Assistant Teaching Professor, MPhil in Data Intensive Science, University of Cambridge, explains how self-improving and open-source AI models are shaping the next generation of enterprise AI. He explores agentic systems, governance patterns like “maker and hater” agents, and why fully autonomous, human-out-of-the-loop AI represents the biggest future opportunity.
The Next Wave: Self-Improving AI
Dr. Boris Bolliet:
The recent advances that happened also over the summer is that we now have some of these large language models that are able to generate their own data on which they will train themselves to be better. Sort of self-improving language model learning. And yeah, that's something that's moving quite fast at the moment too and is highly interesting.
Open-Source AI for Enterprises
Dr. Boris Bolliet:
But what we are seeing over the last couple of months is the rise of those types of models, like the self-improving one, Kimi K2, and DeepSeek, which have the particularity to be open-source models. So that means that I can go on a server on the internet, download the entire models, maybe it will be hundreds of gigabytes of data, the model itself. And then I can just run the model on my laptop. That's very interesting for enterprises where you maybe in many cases do not want to share your data with external servers. And here again, this is moving fast and we can really start envisioning to have enterprises that function on their own locally trained and fine-tuned models without the data having to be shared and leaking out.
The Challenge of Evaluating Agentic AI
Dr. Boris Bolliet:
And there are many challenges associated with building such systems. How do we implement them?
How do we orchestrate the transitions between the flow of information between the different agents? And very importantly, how can I tell quantitatively how good is an agentic system at what I am trying to make it do? So is it 90% of the time doing what I want or maybe 10%?
Governance Through Maker + Hater Agents
Dr. Boris Bolliet:
For implementation, there are different platforms that allow you actually to implement the systems relatively well without so much effort. The most popular ones are AG2 that we use quite a lot, LangGraph as well, and another one is CrewAI.
We use these systems to generate research project ideas based on some data. So I give some data description, that's what I mean by input text here, some data description to an idea maker agent, and then the idea maker makes ideas on what to do with this data, and the idea hater will criticize the generated ideas. And you do that a number of times and you narrow down, which the idea on which you want to work.
Human Out of the Loop
Dr. Boris Bolliet:
And the next and final level of automation which as the most of opportunities is human out of the loop. Level five of automation where it's a self-driving system. So that's obviously way harder to build, but this is where more opportunities are.
