Governance is the unlock; defaults just changed for code and data
The wins this week come from org and platform changes, not new models.
Two defaults flipped: company-wide medium effort for coding agents and deep learning as the baseline for tabular work.
Governance first. Gains hinge on governance and platforms, not the next model.
Mistral pegs AI infra at 1.8% of US GDP and rising, which only pays off with org redesign.
Security experts say basics like segmentation and access controls are skipped, so staff a security architect and build a platform with guardrails and observability.
Defaults flipped. Two daily defaults just changed for serious teams.
Opus 5.5 offers zero data retention, and medium effort beats max by 10x cost with no quality loss for 95% of tasks.
TabPFN 3 beats XGBoost on tabular work to ~1M rows with one-pass predictions, so cap effort org-wide and budget for GPU inference.
Benchmarks miss it. Benchmarks miss the stuff that burns time and creates incidents.
Opus 5.5 hit infinite loops on max effort and is token hungry, while clearer writing saves review time.
Agent UI still needs human taste and upfront sketches, and vuln discovery floods triage, so add checkpoints, design standards, and triage capacity.
My take
Stop buying model upgrades to fix organizational debt.
The strongest signals say your next gains come from guardrails, decision records, and a platform, not a shinier benchmark.
I like the two crisp policy moves here: default medium effort for coding agents, and treat TabPFN 3 as the new tabular baseline up to ~1M rows.
But don’t confuse defaults with autopilot.
Without a security architect, sandboxing, and multiplayer design, faster tools just let teams ship the wrong thing sooner.
I’ve seen platform work beat tooling churn once teams pass a dozen engineers.
The videos behind this brief
We watched these so you don't have to.
Tabular deep learning finally works. Default tools are about to shift.Machine Learning Street TalkWatchWhy deep learning now beats XGBoost and what to change.
AI productivity needs organizational rewiring, not better models alone.AI Engineer
AI labs are ignoring cybersecurity experts — and it showsIBM Technology
AI Agents Won’t Work Without Platform EngineeringGOTO Conferences
Opus 5.5 fixes quality, but the effort trap costs more than tokens.Theo - t3․gg
Why agentic design still needs the human sketchThe Pragmatic EngineerVideo and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.