Roundtable recap, Nov 6: what consumer AI costs to serve, local models, and who still reads the code

Two questions ran through this week's roundtable. Who pays for AI that costs more to serve than it earns, and how much of the code an agent writes does a person still need to read?

Can consumer AI pay for itself?

We opened with Elon Musk's admission that Grok Bot costs far more to serve than it earns. One attendee was surprised. They run their family's calendars, school emails, product research and trip planning through Claude, and it adds up to about 1% of their monthly usage. Only overnight agent jobs make a real dent. Both agreed that subscriptions beat API pricing.

Our host expects AI companies to shift toward products built on their models, like an AI portfolio manager, rather than general assistants that few people will pay for.

Local models and harnesses

One attendee runs Qwen3.8-27B at home on two 12GB GPUs, using Swift 1.5, a fine-tune that spends far fewer tokens on reasoning and fits a 168K context window. They start projects on a frontier model, then hand the next 20 to 50 prompts to the local one. Their tools are Crush, a Go terminal agent with Language Server Protocol support, so the model sees and fixes errors without a reprompt, and the Zed editor.

Another attendee uses T3 Code to orchestrate Claude Code and Codex across three machines over secure tunnels, keeping client code on client hardware while they work from a laptop or phone.

Harnesses running the same model vary widely in token efficiency. One of us has seen a harness add so much context to each prompt that a local model crawls, while chatting with it directly works fine. Larger companies may self-host eventually, though models run by IT teams tend to go stale. Our view is that models will end up like office software. Most departments won't care which one they run, and governance will drive upgrades more than capability.

Do humans still need to read the code?

The question started with how programming languages should change now that people rarely type code. One view is that people still read it, and typed, verbose languages give more control, though verbose code also fills more of the context window.

Another attendee doesn't think people need to read code anymore, but says agents still benefit from typed languages, clear names, modular structure, linting and logging, because maintenance is the hard part. Their tip is to have the agent reproduce a bug before it tries a fix.

Strong test suites can replace line-by-line review for many apps. Finance, medicine and aerospace still need someone who signs off on every line. Agents can also report a check as done when it wasn't, so the practical answer is clear acceptance criteria checked by a separate reviewing agent. We want models to get much more reliable before teams give up readable code on critical systems, the way bridges are built with a 3x safety factor.

Source control built for agents

To close, we looked at new source control systems built for agent-scale repos, from Anthropic (Scion) and OpenAI (Pando). One team now spends more on GitHub Actions than on model tokens, so they plan to evaluate Pando, which keeps the Git protocol. Another attendee thinks replacing Git faces a very high bar.

The Econify.AI Roundtable meets Fridays at 11am ET / 4pm UK. Details and registration are at econify.com/ai-community.