💥 The Intelligence Explosion
"Kimi K3, DeepSeek V4 Flash, Qwen 3.8, Muse Glimmer, Neutron 3.5, Grok 4.6, DeepSeek V4 Pro, Gemini 3.7 Flash, Qwen 3.8 27B, GLM 5.3 — more than five model releases in five days across all tiers. The big labs are clearly panicked." Fable is now permanent in Claude subscriptions; OpenAI slashed Terra/Luna prices and is testing 50% cuts on GPT-5.6. 0:00
The two questions that matter: "What's the best way to use these models
together to outperform them individually? And how do we build for rapid change so we can leverage the best performance/speed/cost trade-offs?" His 15-year engineering principle:
"The most flexible system wins." 0:48
📊 The Model Stack
His running model stack, tiered: lightweight/local (Muse Glimmer, Neutron 3.5 Lightning, Qwen 3.8 27B), A-tier workhorses (GLM 5.3, DeepSeek V4 Flash/Pro, Gemini 3.7 Flash, Qwen 3.8), S-tier open weights (Kimi K3, Qwen 3.8), and the frontier (Fable 5, Opus 5, GPT-5.6 Soul). 2:12
His picks: "Gemini 3.7 Flash is my favorite model out of everything in this explosion — a strategic decision to optimize for speed, cost, and intelligence. They're 100% right." DeepSeek V4 Flash is "absolutely cracked" on pricing; DeepSeek V4 Pro "is THE open-weights model to watch if you have the hardware."
3:30
🗣️ Pattern 1: Opinion
The first command, /fh opinion: fire one prompt at N models and get N unique perspectives, each with performance, speed, and cost tracked. The test: "What's the most important DuckDB v2.0 feature for a local-first analytics developer to test first?" 4:41
The result: Gemini 3.7 Flash finished first ("insanely quick"), DeepSeek V4 Pro next at 33 tok/s, Fable 5 last. Three unique perspectives — Fable picked "variants" (automatic structural decomposition), Gemini agreed, DeepSeek "drifted a little." Cost difference from one simple prompt: Fable cost an order of magnitude more.
5:50
The subtle gotcha — model aliases: "You can never reveal the name of the model to the other model. Otherwise they start emitting weird behavior — competing and sabotaging each other. It just emerges naturally." Hence the rune/flux/drift aliases.
6:11
⚔️ Pattern 2: Debate
The second command, /fh debate: agents argue a claim across multiple rounds, sharing positions, refutations, and closing statements — names hidden behind aliases. The claim: "DuckDB is no longer just an in-memory tool; use it as a server with Quack and Connect." 7:49
How it plays out: Round 1 — three concrete opinions (Fable: "claim is wrong," Gemini: "keep treating DuckDB as embedded analytical engine," DeepSeek: "pilot it, don't move to production"). Round 2 — "every agent shared its response with every other agent." Then closing statements: "Reject the claim" ×3. All three models converged.
8:50
Why it's his favorite: "This is great not just for learning new tech — it's for
strategic decision-making. Some decisions are month-long or year-long commitments. If you're putting that much time in, sit down with the best intelligence available to flesh out the decision or give you counterpoints."
13:46
🤝 Pattern 3: Collaborate
The most powerful command, /fh collaborate: every model proposes a plan, then an architect agent merges them into a task list with owners and dependencies. Two builders + one architect ship working demos of DuckDB v2.0 features as Astral UV single-file scripts. 14:06
The architect is your most powerful model: "The fusion harness always has one architect — put your most powerful model you're willing to spend on there. It takes all the opinions, all the plans, and puts them together into something incredible."
15:21
The plans reference each other: "You'll notice we're using last week's system prompt engineering (reference points T1/T2/T3, risk analysis, clear language) inside the system prompt for all agents." The architect assigns owner + mode + dependencies, and the models collaborate — "have our compute check itself."
16:21
The numbers: the collaboration run cost Fable 5 $0.65, Gemini 3.7 Flash $0.07, DeepSeek V4 Pro $0.05 — "about an order of magnitude cheaper to use 3.7 Flash or V4 Pro over a state-of-the-art model like Fable." 18:25
🔑 Own Your Harness
"Big fan of Claude Code, still using it. But there are things you can't do when you're locked into someone else's agent harness — that's by design, and it's going to hold you back." 21:16
The path: "If you're pushing into the new role — still software engineering, but with autonomous technology acting on your behalf — the agent harness is the thing to own." Use a closed-source option AND an open-source flexible option (Pi Agent). "Think in ANDs, not ORs. Combine compute, don't select compute."
21:40
The next step — the software factory: "Once you understand prompt engineering, context engineering, and harness engineering, the next level is
out-loop agentic coding — the software factory, where agents + code work on your behalf without you. Stop blaming the model. Stop blaming the tools. Everything is in your control."
27:00
✅ Key Takeaways
- Combine compute, don't select compute. The intelligence explosion means the moat is no longer the model — it's the harness that orchestrates many models.
- Three orchestration patterns: Opinion (N perspectives on one prompt), Debate (agents argue with hidden identities), Collaborate (plans → architect → task list with owners).
- Never reveal the model name to another model. They start competing and sabotaging — use aliases (rune/flux/drift).
- The architect is your most expensive model. Put your best compute on the merge/integration role.
- A-tier workhorses are an order of magnitude cheaper. Gemini 3.7 Flash ($0.07) vs Fable 5 ($0.65) for the same collaborate run.
- The most flexible system wins. Own your harness (Pi Agent), then own your model — and eventually, out-loop to the software factory.
📍 Timestamp Index
0:00 The intelligence explosion
14:06 Pattern 3: Collaborate
27:00 The software factory
