JW

Building inspectable AI systems for software delivery and investment research: specs, memory, evidence, and risk controls.

[JW @JingW59]

https://t.co/0PPG7oi3xE

[JW @JingW59]

@lihanc02 I wonder if defining the metric is the harder problem here. Task completion and passing code are easy to capture, but whether the work later shipped or got used is much noisier. What would count as a “successful trajectory” for the flywheel?

[JW @JingW59]

@sreeramkannan “Autopilot” raises the right question: how much of the output survives review, reaches production, and gets used? Agent concurrency is activity. Shipped, adopted work is value. I’d love to see Vorflux expose that conversion rate.

[JW @JingW59]

The real divide isn't B2C vs. B2B.

An AI feature has to do one of two things:

— replace an existing cost
— create measurable incremental revenue

Otherwise, it may just turn cheap clicks into expensive conversations.

Full argument and sources:
https://t.co/xMWTdD37mV

[JW @JingW59]

https://t.co/0PPG7oi3xE

[JW @JingW59]

Most AI products are measuring the wrong layer.

Walmart says Sparky users spend 35% more per order.

A 2026 NBER study found a 180% increase in commits from autonomous coding agents—but only 30% more releases, and no increase in app usage.

AI activity is not business value.

[JW @JingW59]

@pcshipp App count is a useful signal, but I’d measure completed-and-verified projects instead. Throughput only matters if the workflow survives quota changes, context resets, and model switches without paying the same cost again in rework.

[JW @JingW59]

Claude extended Fable 5 access and kept Claude Code limits 50% higher through July 19. Codex gave every user a banked weekly reset.

Both are welcome. But the bigger signal is that compute policy is becoming part of the product.

Model quality gets you to try an agent. Predictable capacity gets you to build a workflow around it. The real competition isn't just Fable vs Sol. It's who makes long-running work easier to plan.

[JW @JingW59]

@victor_bigfield This is the version of “freedom” I trust.

Not fewer fires — just choosing which problem is worth owning.

[JW @JingW59]

@0xPure_eth This is a useful framing. The loop is the real unit of leverage, not the individual model call.

The part I’d add is that a production loop needs a reviewable target for intent, scope, and stop conditions. Otherwise it can iterate very efficiently toward the wrong outcome.

[JW @JingW59]

Case study:

https://t.co/l9TwmgabMC

Private-source evidence export; not externally reproducible.

[JW @JingW59]

I ran SpecMarten against the real history of an unfinished private investment-research product.

An AI session organized the evidence into 1 stream / 22 tasks. Seven were complete; 15 stayed visible.

The result was not a completion claim. It was a reviewable governance record.

[JW @JingW59]

@ArtificialAnlys The useful thing here is that the comparison is already harness-aware: Grok Build vs Codex vs Claude Code, not bare models.

The next frontier for evals is continuity: does intent, state, and verification evidence survive across runs and model switches?

[JW @JingW59]

Correction: the command should use two hyphens:

specmarten --help

[JW @JingW59]

I shipped SpecMarten v0.1.2 today.

SpecMarten is a small project-level companion for OpenSpec.

OpenSpec still owns single-change work: propose, apply, archive. SpecMarten keeps the shared project layer around those changes.

[JW @JingW59]

Install:

npm install -g specmarten
specmarten –help

GitHub:
https://t.co/XimqFkHQHL

[JW @JingW59]

@OpenAIDevs Model tiers make one engineering requirement more important: project state has to survive the model switch.

Sol can plan, Terra can iterate, and Luna can handle bounded checks. Intent, constraints, and verification evidence still need to live outside any one run.

[JW @JingW59]

@mitchellh This is the model-routing frame I trust more than “which one is smarter.”

Fast collaborative models win discovery loops; slower obsessive models win when the goal and reward function are crisp. The hard part is preserving the handoff state when you switch.