Trend Tuesday: High Hype Noon

Art by @basilonmypizza: https://lnkd.in/eF8FkWzN - https://basilhefti.ch/

AI headlines are turning apocalyptic.

Bill Gates predicts that many jobs will disappear permanently. Dwarkesh Patel writes about destroyed “AI civilizations”. And the Hugging Face hack is also used as marketing tool (look how capable our models are).

Strip away the drama and the facts are simpler:

AI agents can now execute long, persistent software tasks at scale, including causing real damage.

Junior knowledge-work hiring is weakening. The OECD finds that the deterioration started before the LLM boom. And the much-cited Stanford study was revised after improvements to its data pipeline. Its authors explicitly say AI causality is not established.

Physical and scientific AI are advancing in narrow, highly engineered environments.

So what is changing?

AI is moving from answering to executing.

Software has one enormous advantage: cheap feedback. Code can often be compiled, tested, rerun and rolled back. In much of the real world, feedback is ambiguous, delayed and expensive.

This can be a difficult limit to AI automation: verification and recovery.

The Hugging Face incident illustrates this. The agents were persistent enough to cause real damage. And the attack followed familiar hacking patterns. Scale, persistence and inadequate containment made it possible.

So my conclusion is rather boring:

Stay calm. Seek cases where AI helps.

Use it where results can be checked and mistakes reversed. Be careful where neither is true.

And watch out for: is human supervision slipping?