Back to AI/ML Overview
🧠 New · Thesis

Developer General IntelligenceDGI is here — and it inverted how I build

We keep debating AGI in the abstract. Meanwhile, something concrete already arrived for engineers. With Anthropic’s Fable, the model now generalizes across the entire development lifecycle — architecture, implementation, review, operations — not narrow autocomplete on a single file. I call it Developer General Intelligence (DGI): AGI’s practical cousin for the way we actually build.

5–7×
team productivity
1
engineer on the org chart
4
frontier models in the council
1 → 1B
users the reference scales to
spec-first
code is a compile step

🎯The thesis

“AI writes code” undersells what changed. A that can hold the whole lifecycle at once — reason about a system’s architecture, implement against a contract, review its own work adversarially, and operate the result — is a categorically different collaborator than an autocomplete engine. That generality is the story.

I’ve spent the last year building this way full-time. The pattern that emerged is not “prompt for code, faster.” It’s the opposite.

🔑The DGI move
Most people use AI to write code faster. I use it to design the architecture first, and lock the spec before a single line of code exists. In a world where code is cheap to generate, the architecture is the scarce, high-leverage artifact. Get the design right and the code becomes a compile step.

🧠The innovation: a multi-model council, inside the architecture

Here’s the part I’m most excited about. I embedded a multi-model directly into the architecture view. Click any component in the system — the agent fleet, the async queue, the memory layer, the — and put it to a council of four (Claude, OpenAI, , Grok):

“Why this design choice over the obvious alternative?”

“What breaks at scale, and what’s the mitigation?”

Claude synthesizes; each of the others lands its own take in adversarial parallel. And here’s what makes it more than a chat widget: their critique feeds the design decisions that get recorded and locked into the spec underneath the diagram. The architecture stops being a static picture you draw once and hope — it becomes a living artifact you interrogate, and the spec crystallizes beneath it.

🏗️Architecture as the Product

This is a discipline, not a slogan. The architecture is designed, reviewed, and locked as a deliverable — then the code is generated one-shot against it.

📐

The spec is the deliverable

The code is a one-shot compile of a reviewed, locked spec — not the artifact you iterate blind.

💸

Flaws caught where they’re cheap

Design mistakes surface in the spec, not buried in ten thousand lines you’ve already written.

⚔️

Adversarial, not an echo chamber

Four independent model families. The default Claude reaches for is exactly what Grok or Gemini pushes back on.

🔎

Traceable before implementation

Every decision maps to a requirement and a rationale before a line of product code exists.

🔁The workflow, concretely

The architecture doc is a git-versioned project of its own, and it’s finished — reviewed and locked — before the code project is created.

text
1  Ingest the requirements verbatim (JD / PRD / brief)
2  Iterate the design in conversation — record every decision
3  Draw the system; validate traceability against every requirement line
4  Run the LLM council + an AI-partner review on the architecture
5  Lock the spec: final decisions + a phased build plan + a deployment
   playbook with concrete values, so the build never stops to ask
6  ONLY NOW create the code project
7  Build one shot against the locked contract; keep a build log
8  The council ships embedded in the product — every design choice stays interrogable

Architecture-as-the-Product · the replay sequence

💡Why one-shot works
When the contract is complete and pre-reviewed, implementation stops being exploratory. The model isn’t guessing at intent mid-build; it’s compiling a decision record. That’s where the 5–7× comes from — not typing speed, but never re-deciding.

📦The reference implementation: a personal AI, 1 → 1B

The system in these captures is a personal AI scaled from one user to a billion — a seven-layer topology with a desired-state that autoscales each tier out and in. Every node is interrogable by the council. It’s the architecture teaching — and stress-testing — itself.

Scaling as a control loop

The fleet holds no per-user state — it pulls a per-turn context pack from shared stores, so scaling is more replicas, not a redesign. A desired-state reconciles the fleet out and in.

+ observability as first-class

Online /guard on the response path; token accounting, traces, and audit underneath. The demo makes the invisible infrastructure visible — which is the whole point of teaching an architecture.

⚖️The council pattern (why four models, not one)

A single model is an echo chamber — it agrees with its own defaults. The council runs Claude, OpenAI, , and Grok in parallel as adversarial reviewers; Claude synthesizes their takes. Safety and observability live in the pattern, not bolted on:

🛡️

PII scrubbed, always

Names, emails, phones, DOB are stripped before any model call — the risk profile is discussed, never the person.

🔍

Provable, not claimed

Every answer carries a "what was sent" trace and a "no PII present" shield you can inspect.

💰

Cost in the open

A live per-answer cost estimate — governance you can see, on every turn.

🚀What DGI looks like, day to day

You architect out loud. A council of models pressure-tests the design live. The spec writes itself underneath. The code follows from a contract that was already reviewed. The hard part was never the code — it was the thinking, and DGI now generalizes across the whole of it.

When the model can hold the entire lifecycle, the bottleneck moves from “can we build it?” to “did we design the right thing?” — which is precisely where a council of adversarial reviewers earns its keep.

🎯

Leadership Takeaway

For engineering leaders: DGI changes the unit of leverage from lines-of-code to decisions-per-spec. The teams that win won’t be the ones generating code fastest — they’ll be the ones whose architecture is designed, adversarially reviewed, and locked before implementation, so the build is a compile and the review already happened. This is the first in a series on DGI.