The Deep Feed

Software development is entering a strange era of delegated labour. Simon Willison recently sat down to finish a major release for sqlite-utils, a tool used by thousands. Instead of writing the final commits himself, he hired Claude Fable—an AI agent—for the equivalent of about $150. The goal was a final review to catch breaking changes before the 4.0 stable release. What followed was not just a simple proofread, but a deep, structural interrogation of the code that a human eye, perhaps fatigued by the repetitive nature of regression testing, might have missed.

The Poisoned Connection

The most alarming discovery was a bug in the `delete_where()` method. On the surface, it looked like a standard deletion command. In reality, it was a silent killer. Because the method lacked an atomic wrapper, it failed to commit the transaction, leaving the database connection in a state of permanent, uncommitted flux. If a developer had shipped this, every subsequent write operation would have been lost when the connection closed. It wasn't just a minor error; it was a design flaw that would have caused massive, silent data loss for anyone relying on the library's new transaction model.

The connection is left in_transaction=True, so every subsequent atomic() call takes the savepoint branch and never commits either. That’s a really bad bug!

This incident highlights the specific utility of AI in the current development cycle. The agent didn't just suggest stylistic changes; it performed end-to-end reproduction of failures. It identified that the library's new promise—that every write operation is automatically committed—was being broken by a single, poorly implemented function. For a developer, this changes the nature of the job. You are no longer just a writer of logic; you are a high-level auditor of machine-generated intent.

The Multi-Model Audit

Willison also touched on a practice that sounds superstitious but is increasingly practical: cross-model verification. He used Anthropic's best model to review the work, and then brought in OpenAI's GPT-5.5 to perform a final sanity check. This creates a system of checks and balances where the biases or hallucinations of one model are caught by the different training data and logic of another. It is a way of triangulating truth in a landscape of probabilistic outputs.

As we move toward more autonomous coding, the cost of entry for a single developer drops, but the cost of a mistake rises. When an agent can write 1,300 lines of code in a few minutes, the human must be capable of reading and understanding those lines just as quickly. The speed of production is no longer the bottleneck; the speed of comprehension is.

There is a growing tension between the intelligence of Large Language Models and the rigid structures of the software they inhabit. A developer named Armin recently encountered a frustrating phenomenon while using the Pi coding harness. Even when using top-tier models like Claude Opus 4.8, the models began making mistakes that smaller, older models did not. Specifically, they were failing to follow the strict schema required by Pi's edit tools, inventing extra fields and keys that didn't exist in the documentation.

Training for the Wrong World

The reason for this regression is likely a byproduct of success. As Anthropic trains its latest models (like Sonnet 5) to be better at using Claude Code—their own internal toolset—the models become highly specialised. They are being reinforced to use the specific 'search and replace' mechanisms that Anthropic has optimised for. When these models are moved into a third-party environment like Pi, which uses a different mechanism, they struggle. They try to apply the logic of their 'home' environment to a foreign one.

The SOTA models of the family are worse at this specific tool schema than their older siblings.

This suggests a looming fragmentation in the AI ecosystem. If the most capable models are being fine-tuned to work perfectly within the walled gardens of their creators, they may become less useful as general-purpose agents in open-source or third-party tools. We are seeing the birth of 'model-specific optimisation', where a model's intelligence is tethered to a specific set of tool definitions.

For the agency owner or the developer, this means that 'the best model' is no longer a universal choice. You have to choose the model that fits the tool, not just the one with the highest benchmark score. The intelligence of the model is being increasingly defined by its ability to navigate its specific ecosystem, rather than its ability to follow any instruction given to it.

When an AI attempts to write poetry, it produces a ledger of clichés. It can mimic the meter of Whitman or the rhyme of a sonnet, but it fails to capture the essence of the thing. An elder poet once offered a simple explanation for this failure: 'Because AI hasn’t suffered.' This is not a romanticised myth of the tortured genius; it is a statement about the source of human expression. Art is the music we make from the collision between our wishes and the reality of the world. It is a coping mechanism for the experience of being alive.

Jung and the Generative Crisis

Carl Jung, the Swiss psychiatrist, looked deeply into this relationship between psychological distress and creativity. He rejected the idea that neurosis itself is the source of art. To Jung, neurosis is a failure—a chronic, repetitive state of being stuck in a crisis. True creative genius does not come from being broken, but from the ability to take the raw material of suffering and transform it into something coherent. He distinguished between the 'pathological' neurosis, which is a dead end, and the 'creative' doubt, which is a catalyst for growth.

True creative genius does not let itself be spoilt by analysis, but is freed from the impediments and distortions of a neurosis.

The distinction is vital. An AI can simulate the *symptoms* of a crisis—it can write about sadness or loss—but it cannot experience the *stakes*. For a human, a poem about loss is an attempt to survive the loss. For an AI, it is a statistical prediction of which words typically follow 'I lost'. The lack of skin in the game is what makes the AI output feel hollow. There is no risk, and therefore, no meaning.

The Limit of Simulation

As we integrate AI into our creative processes, we must recognise where its utility ends. It can assist with the mechanics of language, the structure of a story, or the generation of ideas. But it cannot provide the 'why'. The 'why' is found in the friction of existence—in the heartbreak, the triumph, and the terrifying uncertainty of being a conscious entity in a physical world. We should use AI to handle the heavy lifting of form, so that we have more energy to deal with the weight of substance.

Ultimately, the rise of AI does not devalue human art; it clarifies it. It forces us to define exactly what we are doing when we create. We are not just arranging symbols; we are attempting to bridge the abyss between ourselves and others through the shared experience of being alive. That is a bridge an algorithm can never build.