A dedicated publishing area for weekly AI signal: commentary, public opinions, and practical analysis for professionals making decisions.
March 11, 2026
Post-training a model against a specific harness creates overfitting to that harness. Most teams are betting on the lab's default environment without realising it's a bet.
March 11, 2026
When building is free, bad ideas don't die on the whiteboard anymore. Harrison Chase names the problem — then builds his argument on a claim he never tests.
March 11, 2026
Alfred Lin drops a 3-5x vs 10-20% productivity gap and moves on to an Ender's Game metaphor. The number is the story. The metaphor is the distraction.
March 11, 2026
NVIDIA's five-layer AI infrastructure model is genuinely useful — right up until you notice which layer in the cake is sold by the company drawing the diagram.
March 11, 2026
Sarah Sachs names the distinction between frontier and saturated AI tasks that the whole market is mispricing. Her playbook for exploiting it requires Notion-scale leverage most companies don't have.
March 11, 2026
HBR's 'AI brain fry' research identifies a real cognitive load problem — then blames the humans for checking too much instead of asking why the tools require it.
March 10, 2026
A new benchmark tests AI agents across 71 commits of real maintenance — not one-shot fixes. The regression compounding it reveals is worse than the scores suggest.
March 10, 2026
A productivity panel buries its best insight in bullet two: AI amplifies, it doesn't augment. The distinction reshapes where the real investment case sits.
March 10, 2026
The UK's £500M Sovereign AI Unit is structured differently from previous government AI spending. Whether that's enough depends on energy, governance, and not leading with Ada Lovelace.
March 10, 2026
A Columbia professor's Shannon/Kolmogorov piece nails the Knuth anecdote — pattern-finding versus proof — then leans on a mythologised Einstein to carry the argument home.
March 10, 2026
Copilot Cowork gets checkpoint architecture right — approval gates as product, not friction. Then buries the fact that Claude runs the execution layer in a single sentence.
March 10, 2026
LangChain's GTM agent case study calls human review non-negotiable — then reveals the agent sends automatically after 48 hours. The methodology is real. The headline isn't.
Subscribe to the AI Primer Briefing for weekly commentary on what is happening in AI and what it means for your work.