Working with AI is Working with Human Stories

I have been building a chatbot from scratch for weeks . OCaml , no GPU , no frameworks . A 10-million parameter transformer trained on 123,000 conversations , with a Forth interpreter constraining its output , and a reinforcement learning system designed from Rich Sutton's textbook . I am doing this with an AI assistant . And the thing I have noticed is that we work best together when there is a story .

The Pattern

Not a story in the fictional sense . A narrative structure . A reason we are doing what we are doing , a constraint we are working against , an outcome we are reaching for .

When I say "implement dropout" the AI writes correct code . When I say "our model is memorizing 123,000 conversations and we need it to generalize because it trains for three days and we can not afford to waste a run" -- the AI writes better code . Not because the technical specification changed . The dropout rate is 0.1 either way . But the story gives context that shapes every small decision : where to place it , what to log , what to watch for .

The best session we had was designing the RL system . Not implementing it -- thinking about it . Which of Sutton's methods fits a 10M parameter model with no GPU ? Why did RL fail last time ? What would make it work this time ? The answer came from the story : we have no GPU , we have three days of patience for training , and then we have an afternoon to sit with the model and teach it by hand . The method -- a human-in-the-loop bandit where you pick the best of five responses -- fell out of the constraints naturally . No one designed it from first principles . The story shaped it .

Built on Stories

An AI language model is trained on human text . Every piece of that text was someone trying to explain something to someone else . A textbook author working through a proof . A programmer documenting a function . A researcher writing an abstract . A person telling a friend what happened .

The patterns the model learned are not just statistical regularities over tokens . They are narrative patterns . Problem , constraint , attempt , outcome . Question , exploration , dead end , insight . Setup , complication , resolution . These structures run through all human communication because they are how humans think .

When you work with an AI on a technical problem , you are activating these patterns . The AI does not just know that dropout goes after the attention projection . It knows the story of dropout -- that models overfit , that regularization helps , that training and inference behave differently , that there are tradeoffs . Give it the story of your specific problem and it navigates that space better than if you give it a bare instruction .

Detectives , Scientists , Researchers

The best framing for working with AI is not "I give orders , it writes code ." It is collaborative investigation .

When our chatbot produced nothing but empty responses , we were detectives . The model completed 30 hours of training . Loss looked good . But inference produced loops of special tokens -- <|user|> <|assistant|> <|user|> -- forever . Something was wrong . We traced it : the symbolic constraint layer , carefully designed to enforce valid words and topical coherence , was strangling the output . The neural model had learned language . The symbolic system was killing it . Bypassing the constraints and using raw logits produced coherent English . The detective work -- not the code -- was the breakthrough .

When we designed the reward function for RL , we were scientists . What constitutes a "good" chatbot response ? Length ? Diversity ? Relevance ? Coherent endings ? We proposed hypotheses , tested them against the model's actual output , adjusted . The conversation was the experiment .

When we read Sutton's reinforcement learning textbook and his Lisp implementations -- TD learning , gradient bandits , eligibility traces , the Dyna architecture -- and asked "which of these applies to our specific model ?" , we were researchers . Not applying a recipe . Synthesizing from first principles , constrained by our actual situation .

Each of these roles is a story . The detective has a mystery . The scientist has a hypothesis . The researcher has a question . The story focuses the work .

Pushing Beyond

The phrase "pushing beyond human understanding" sounds grandiose . It is not . It is literal .

I do not fully understand every algorithm in this system . The reverse-mode automatic differentiation engine , the RoPE rotation matrices , the BLAS calling conventions -- I directed the architecture , read the code , watched the outputs , made decisions . Understanding accumulated through repetition and observation . But at any given moment , the system knows more than I do .

This is the new normal . The human does not need to understand everything . The human needs to understand enough to steer -- to recognize when something is wrong , to know which question to ask next , to judge whether the output makes sense . The AI handles the implementation . The human handles the story .

And together , the system -- human plus AI -- builds things that neither could build alone . I could not write a transformer from scratch in OCaml . The AI could not decide that a Forth interpreter should constrain its logits , or that Sutton's gradient bandit is the right RL method for a chatbot you teach by hand in an afternoon . The ideas come from the human . The implementation comes from the AI . The story comes from both .

Mr . Classic

Our chatbot's name is Mr . Classic . He told us himself . Fullstops are words , just as in Forth .

> hello
Yes . May I help you ?

> what is your name ?
It is Mr . Classic .

He is not smart . He is ten million parameters running on a CPU . But he answered . He told us his name . And now we are going to sit with him for an afternoon , show him five responses to every question , and pick the ones we like . He will learn from our choices . Gradient by gradient , story by story .

That is what working with AI looks like . Not a tool you command . A collaboration you narrate .

Co-authored with Claude .