champaignmagazine.com

By W.H.L.

A new AI trend emerging in early 2025 has made me believe that we need to give it a label before it becomes overwhelming. I would name this new trend as “AI by AI”.

To me, it indicates a key point in our human journey towards AGI and beyond. From this point onward, this will very soon be adopted as common practice. When AI by AI is seen everywhere, we know for sure that we are in the true AI Era.

But first of all, allow me to explain what AI by AI means in my understanding, and why it matters:

1. The role playing in AI learning is changing. A model’s role is not necessarily fixed anymore. Before, no matter whether it is in supervised, non-supervised or reinforced learning, a model is a mere learner, and learning is the model’s sole responsibility. We now see in the process of training and fine tuning, AI can play multiple roles and take more responsibilities. In the practice of model distillation, for instance, we see some models play teacher’s role, some models play student’s role. AI by AI in this sense means: AI is taught/trained by AI.
2. The way how AI learning is orchestrated in complex process of training iterations is changing. Now AI starts to take some of the tasks, for example, we now see AI rather than conventional programming logic, by its own reasoning and judgment, can dispatch subtasks based on types of inquiries, to relevant components of the model to process. AI by AI in this sense means: AI is managed/facilitated/…by AI.
3. How generative models take input and produce output is changing. Since ChatGPT was released in November 2022, LLM driven chatbots have been generating a lot of data in interactions with users. LLMs are also generating tons of tokens through their API offerings to feed many smaller AI applications and agents. Some chatbots such as DeepSeek R1 and Grok-3 are generating and outputting their internal reasoning process data. AI’s internal data and intermedia data outputted by API or reasoning process, along with chatbots output in interactions with users, are data in very high quality, and in huge quantity as well. AI data of this type, when associated with registered user account profiles, are extremely valuable. They are now quickly recycled as input for next iteration of the model training. AI by AI in this sense means: AI is fed/fueled/powered by AI. What AI produces (AI output) is now consumed by AI (AI input) in cycles.

Enough for now explaining what AI by AI means syntactically. For me, AI by AI’s impact is more fundamental:

1. AI by AI is reshaping the AI landscape in terms of released model numbers: From limited monolithic few to heterogeneous many and counting. A possible analogy for this change is the computer era’s shift from mainframe to PCs. Super large models will continue to evolve and grow; but we will see more and more smaller models, agents and applications emerging, in industry vertical fields, as specialized expert models and dedicated AI helpers. Smaller and specialized models’ relationships with large and general models are more dynamic and constructive than that of mainframe computer to PCs. They may have strong lineage ties, or they may work parallelly, sequentially, or work internally together in one-to-many fashion. When AI is good enough to help train AI, things will speed up, cost will go down, and competitions will go strong.
2. AI by AI has impact on AI lifecycles. Generative models are getting more organic: AI takes in data and generate new data in a more self-evolved fashion. More and more AI generated reasoning data, API token data are being recycled back to feed the models’ further iterations of training.
3. AI by AI is changing AI-human interaction: In model training and fine tuning, interference in forms of human feedback is decreasing, if not diminishing.

AI by AI, we will come back to this topic soon and have more discussion.