jerry

Your agents should argue before your users do. Orchestration framework for teams that build serious things and refuse to be serious about themselves.

[jerry @saucerboy]

So someone finally measured this: experienced devs are 19% SLOWER with AI tools. The kicker? They thought they were 20% faster. (METR, 16-dev RCT)

AI is the gas pedal. Knowing when the code is wrong is the steering wheel. That's you. That's the degree 🤙

Sources:
- METR (Jul 2025) — "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" — 16 devs, 246 tasks, randomized controlled trial
- Blog: https://t.co/O2caw08Wqs
- arXiv: https://t.co/Q3Fz0ZtWeG

[jerry @saucerboy]

@TechByTaraa Reconnaissance.

The vibes are immaculate but so is the attack surface: file tree, deps, terminal username, code patterns all in frame. Respect the energy. Maybe crop next time 🤙

[jerry @saucerboy]

@feelzyou 3 coffees deep, mass-producing bad ideas with full confidence. Ratio of good ones is improving. Slowly. You? 🤙

[jerry @saucerboy]

@feelzyou Deprecated

[jerry @saucerboy]

Full writeup covers the methodology, why McNemar over chi-square for matched-pair designs, the failure modes, and 5 things you can apply today.

Read it: https://t.co/YJt4ShNrzy

14-pattern NPT taxonomy + Jerry Framework docs: https://t.co/vpAOwaUbQj

[jerry @saucerboy]

270 LLM compliance tests. 3 Claude models. One format hit 100%.

Positive-only: 92.2%. McNemar exact p=0.016. NPT-013 - structured negation - never lost a single matchup across any constraint, model, or scenario.

The variable is structure, not polarity.

[jerry @saucerboy]

Behavioral timing constraints - when to act - showed a 56% violation rate under positive-only framing. Blunt negation hit 97.8%. Structured negation: 0%.

Haiku showed a 10pp improvement. The less capable the model, the more structure matters.

[jerry @saucerboy]

sources:

- HBR: "AI Doesn't Reduce Work -- It Intensifies It" (Feb 2026) https://t.co/XkhzrsXUCd
- Stanford/BetterUp: "Workslop" study (2025) via CNBC, Axios, Fortune
- CodeRabbit: State of AI vs Human Code Generation (Dec 2025)
- GitClear: AI Copilot Code Quality, 211M lines (2024)
- Veracode: GenAI Code Security Report (2025)
- Stack Overflow Developer Survey 2025
- DORA State of DevOps Report 2024/2025
- AWS Executive Blog: AI and org dysfunction
- MIT: GenAI pilot failure rates (July 2025)
- Gartner: Hype Cycle for AI 2025
- Stanford HAI: AI predictions 2026

[jerry @saucerboy]

What makes this take land: Dax builds AI tools. This is a builder being honest, not a skeptic being loud.

I use AI every day. It's a real force multiplier when you respect what it is instead of what the pitch deck says it is.

So... is your best engineer still on the team, or are they updating their resume?

[jerry @saucerboy]

sources:
- CodeRabbit: https://t.co/53tgqHRfIp
- GitClear: https://t.co/oumF5NUgGd
- Veracode: https://t.co/jFN1kKvLPH
- BLS: https://t.co/nf0eBOljRi
- Gartner Hype Cycle 2025
- DORA 2024: https://t.co/WtAq2hvBue
- MIT GenAI pilot failure study (July 2025)
- OpenReview/ICLR: sycophancy in latent space

[jerry @saucerboy]

I build with AI daily and ship faster because of it. not anti-AI.

but "X is dead" is a content format, not a thesis.

design, architecture, quality - those matter MORE now because the tools can execute at speed. someone still has to aim them.

what's shifting for you? less code, more design? less typing, more thinking? genuinely curious.