The Once and Future Perceptron

This is a post from Robin Sloan’s lab blog & notebook. You can visit the blog’s homepage, or learn more about me.

Thinking about the capa­bil­i­ties of mul­ti­modal AI models, I am cur­rently

• unin­ter­ested in “model as agent”,

• sour on “model as media generator” (though I do seem to keep experimenting with this, so, per­haps the blogger doth protest too much), and

• totally bullish on “model as uni­versal perceptor”: to which you can hand all sorts of media and ask ques­tions about it, even if that media is messy, organic, ambiguous, etc.

This is a really flex­ible and valu­able capability, and from here on out it will be an assumed fea­ture of com­puter systems — an essen­tial tool in the toolbox.

Valu­able enough to merit the yot­tabuck invest­ments cur­rently flying … ? Prob­ably not. Who cares! We’ll carry our uni­versal per­cep­tors out of the wreckage, into the future.

P.S. I believe Gemini 2.5 Flash is presently the overall best uni­versal perceptor, in terms of the whole package: capability/speed/price. In the spirit of Jack Clark’s “capability overhang”, if all AI work was halted today, and all other models destroyed, the world could still very use­fully put Gemini 2.5 Flash to work for many years to come.