Heya and welcome back to Five Things AI!

OpenAI suddenly finds itself on the defensive as Gemini 3, Claude and even Grok nibble away at its aura of inevitability, and the once comfortable lead in benchmarks turns into a very real question about whether the company can keep up in this phase of the AI wars.

Gary Marcus is already writing the first draft of the postmortem on “scale is all you need,” pointing out that pouring capital into ever-bigger models without rethinking the underlying approach was always going to end badly, especially in a political climate that is betting the economy on that thesis.

Meta, meanwhile, offers a kind of cautionary comic relief, lurching from Llamas to Avocados while users keep asking for better social products and get sprayed with AI features nobody really wanted in the first place.

On the more promising side, the Perplexity x Harvard study shows that agents are quietly becoming co-pilots for serious cognitive work, with most usage gravitating toward productivity, learning and research instead of just booking hotels or answering trivia, which feels like the real story of how AI is changing daily life.

In the background, Stanford’s Artemis bot is already beating most human pentesters on real networks at a fraction of the cost, a reminder that the attack surface is expanding just as fast as the tooling to defend it, so the only sensible hope is that the smart folks building defensive AI move faster than those who want to break things.

Enjoy this edition of Five Things AI! ...