Artificial Intelligence Systems Must Become Autopoietic To Be Useful

an amazing paper by Rao Mikkilneni i read while researching, which explores more into the cellular and biological side of this argument.

objectives of this blog post:

1. convince you that further development of more advanced ai models alone will not warrant better results in real world applications.
2. introduce a simple autopoietic ai system

in recent years, we've noticed a slight uptick in the capabilities of artificial intelligence systems, however, i would argue that if you look at an ai model released in 2022 and 2025, they could have been used for the same things all along. while performance, accuracy, speed and of course the addition of 'reasoning' capabilities, have improved the core functionality of these ai systems. the perceived performance gains from incremental 4.5 -> 5 model jumps is more subjective and can be attributed to fine-tuning data for the specific task. gpt 3.5 could write essays, gpt 5 can write better essays, looking at it either way is still subjective.

with any new technology, as humans use it and develop our lives around it, we find new and inventive ways to use it, some we could've never thought of before. recently, modifying existing tools in ways to combine ai with other tools has grown in popularity. like how the advancements in smartphone technology could've never been done without earlier advancements in cellular networks, touchscreens, these implementations are not entirely useless and will ultimately help set the stage for whatever is next for ai. however in many of these examples, we're mistreating these as entirely new tools. cursor is not any different from vscode in any meaningful way because they're both used to create, edit, and share code. that is not to say cursor is useless in any way, the value created attaching ai functionality to vscode means to investors, developers and stakeholders, using cursor or other ai tools can perform a task at a speed faster than without. but we still need a developer to accelerate, and cursor alone is not useful to a company. and even tools like deven or cluely cannot function as replacemenet for a real developer as they are only human accelerating uses of ai. if we create a new windshield that can automatically detect rain, sunlight and other conditions to dim the glass, it's still a car. we wouldn't say that we've automated the driver and definitely wouldn't want to say the car is self driving. so why would we call anything that's functionality is only generating code as a developer replacement?

but what would a developer replacement look like?

nothing that we have in the market today.

an autopoietic ai system would require several layers of understanding with their own level of self-auditing and constant improvement.

for a moment lets imagine what could happen in an ideal chat 'assistant'.

we could query with something like 'remember that I live in new york and have a cat', and the ai model would store this memory as 'user has a cat', and 'user lives in new york', similar to chatgpt's memory. however a key distinction would be that even when the user is not using the application, it's self-prompting it's own memory with queries like 'why might the user want me to remember they live in new york?', or 'did they just get their cat?', to create a simple 'understanding' profile of the user. this profile becomes quintessential in building context for future interactions, and an ai system which was capable of this would enable features like chatgpt's pulse (i assume, as it's not available to me yet). as the user further interacts with the system, these questions and answers will develop the ai's memory further.

an overview of our example system would look like:

1. user input layer - process and parse information from user to respond
2. context layer - understands the user and their current info like location, time and recent interactions
3. thought layer - this is responsible for actually looking at previous interactions, context from other users, external data sources, reviewing changes in weather, news events, etc to build a general understanding and empower the memory which is stored about the user.
4. memory layer - a system responsible for continually reviewing and updating memory about the user from every response, every new ingested article, every feedback response etc

there are many systems (gemini, chatgpt, claude) which combine several aspects of these, often through rudimentary implementations like 'memory' where the ai model will, in addition to the user's query, process the query 'is there anything from this conversation which should be remembered?'. these systems miss out on the entire strength of human memory (verifying things you're memorizing, knowing what to remember and when) in favor of a memory which works more like a cache. a truly autopoietic system would generate new systems which could perform complex decision making using AI. for example a simple system which can only create other systems when receiving the prompt 'can you remember this about me', would create a new sub-agent to handle memory and can continually tweak the behavior for how the memory system works based on the user. maybe the user prefers a system where they remember a lot of information across chats for only a few hours, instead of permanent information like where they live. to create a system like this we would need to allow users to define what the objective, the constraints of its existence and what resources it should access to build understandings (such as giving it access to textbooks to verify information being stored as memories.