AI links, 3/6/2026

Anthropic on best practices for using AI; Hollis Robbins explains an LLM trope; Ian Leslie on FOO camp; Megan McArdle on projection of AI fears

Anthropic Research Writes,

To quantify AI fluency, we use the 4D AI Fluency Framework, developed by Professors Rick Dakan and Joseph Feller in collaboration with Anthropic. This framework helps us define 24 specific behaviors that we take to exemplify safe and effective human-AI collaboration.

Of these 24 behaviors, 11 …) are directly observable when humans interact with Claude on Claude.ai or Claude Code. The other 13 (including things like being honest about AI’s role in work, or considering the consequences of sharing AI-generated output), happen outside Claude.ai’s chat interface, so they’re much harder for us to track.

Pointer from Zvi Mowshowitz. Among the 11 are “refines and iterates,” “clarifies goal,” and “provides examples.” I agree that these are good practices for using AI, as are they other practices in the 24. I think that at some point everyone could benefit from a short course that instills these practices. An obvious level at which to teach such a course would be high school or college.

Hollis Robbins writes,

The models see that hot and cold are mathematically close. They do not inherently compute the oppositeness relation. One way to understand the “not X but Y” construction is as a workaround for the model’s inability to compute opposition the way humans do. By explicitly stating both the rejected term and the replacement, the model externalizes onto the page an operation it cannot perform internally.

She says that the LLM is wasting her own brain’s processing power.

The “not X but Y” construction then asks you to spend another one to four seconds reading a clause that delivers an answer you computed in under half a second.

And it is also insults your intelligence.

The construction implies the reader was holding X and needed correction. When you, the reader, were not holding X, you feel like you’re being talked down to.

Just say Y and be done with it.

Ian Leslie writes,

1. The vibe-coding apps are not, in fact, easy to use. You need technical domain knowledge and intuitions rooted in software engineering - and a lot of LLM usage - to get the best out of them. They don’t replace expertise, they amplify it.

This seems to be true and important. The new “agents” seem to me to have dropped the pretense of vibe-coding altogether. They are just tools for advanced software engineers.

Leslie is reporting from FOO camp (Friends of O’Reilly). I was invited once, really liked it, but was never asked back. Leslie linked to Helen Andrews' report on this year’s version.

Megan McArdle writes,

the people who currently face disruption are influential elites whose private worries shape public discourse — and can move markets when they get out of hand. Remember that the next time an ominous prophecy circulates — or really,whenever you read anything about AI, including this column. I try not to mistake my problems for those of humanity, but no one ever succeeds at fully untangling the two.

She suggests that the fear of an economic collapse caused by AI is a case of projection. Individuals who see their own skills devalued are forecasting doom for the rest of us. I see that happening a lot.

substacks referenced above: @

Metannoying

substacks referenced above:

@

AI #157: Burn the Boats