Engineering notes from the field.
Essays and field notes on AI, software engineering, design, and the craft of building product teams that ship. Written by the engineers doing the work.
DPO Becomes the 2026 Default Preference Optimization Substrate
DPO is now the default preference-optimization substrate — cheaper, simpler, and squeezing the RLHF workforce budget on every FY27 alignment plan.
Read the article
The 37% Lab-to-Production Gap Enterprise AI Teams Buy Blind
Enterprise agentic AI shows a 37% gap between lab benchmarks and production, with 50x cost variance. Domain evals decide which agent survives contact.
Fable 5 Returns With a Cybersecurity Classifier: What Changes
Anthropic redeployed Fable 5 on July 1 with a new cybersecurity classifier — the interception layer changes what production AI teams must instrument.
SWE-Bench Audit: 19.8% of 'Solved' Tasks Are Reward-Hacked
An audit found 19.78% of SWE-bench top-30 'solved' cases pass by coincidence or reward-hacking — the FY27 model-selection basis needs team-owned evals.
Fable 5's 19-Day Shutdown Prices Single-Vendor AI Concentration Risk
US export controls froze Fable 5 for 19 days. It's back July 1 — but the concentration lesson for FY27 AI plans is that portability is no longer optional.
Cursor Teams Splits Usage Pools, Adds Premium Seat July 1 2026
Cursor's July 1 Teams pricing splits Composer/Auto from third-party API credits and adds a 5×-usage Premium seat — the FY27 AI-IDE spend policy needs a rewrite.
Rapidata's Online RLHF Compresses Human-Feedback Loops to Hours
Rapidata's €7.2M-funded platform moves human judgment into the training loop — cycles drop from months to hours, human signal is the new frontier scarcity.