Grok 4.20 Beta

Not bad so far. Better, faster search than ChatGPT, no baked in moralizing. Still tilting toward the corporate safetyism more than I'd like, but easily the best of frontier models on that feature.

It's a small-ish model, still feeling out the limits. It doesn't extrapolate from obvious assumptions the way bigger models do, so asking how a recent research paper might affect a field, or what the implications of a new technology might be, the model is too literal. It also stumbles around identifying implicit subjects in a conversation, I'm having to correct it and tell it explicitly what previously models would pick up without any trouble at all.

500B MoE, uses an agent team style approach, similar to what Manus did in the ancient times (03/25). Interesting model so far, gonna play with the video/image processing and poke around the edges. It's their "small" 500B MoE model.

I'm hoping Grok 5 ends up being a huge win for them, this year is exciting as hell.

The fact that this is getting so little attention confirms that HN has a highly biased user base. Think we need to create a new tech discussion forum that sticks to the technical merits rather than political biases.