The News
Google released its first “reasoning” model based on Gemini 2.0, called Flash Thinking, which uses techniques pioneered by OpenAI’s o1 model.
The idea is that models can break prompts into different steps and “think over” the answers. Flash Thinking leaves an audit trail of its thought process, a new innovation helpful for understanding how the models work and potentially how to make them safer.
TechCrunch called out the model for making a simple mistake: Miscounting the number of Rs in Strawberry (which happened to be the code name for o1).
Reed’s view
I wouldn’t count Gemini out just yet. After last week’s flurry of announcements, I think Google DeepMind is pursuing a long-term path that is fundamentally its own, believing that in the long run, it will take the lead. And this is an ultra marathon. Nobody is on the verge of building “AGI” or making some giant leap ahead of competitors.
For that reason, I think Google’s models will be bad at some things and better at others. We kind of saw this happen between Gemini 1.0 and 2.0. The first version was terrible at coding, but then jumped ahead in the second version. It’s not entirely clear why at the moment, but likely relates to building a “natively” multimodal model.