Google has released yet another Gemini AI model, but this one isn’t quite ready for prime time. The new, awkwardly named Gemini 2.0 Flash Thinking Experimental model is available in Google’s AI Studio for testing. This bot uses additional computing cycles to check its work, allowing it to handle more complex queries. This could offer a hint of how Google hopes to evolve its Gemini models in 2025.
The new LLM is built on the same foundation as the recently released Gemini 2.0 Flash, which is the first version of Google’s next-gen models. The “Flash” designation indicates it’s a smaller model than the “Pro” variants, but Google has said the 2.0 update allows these large language models (LLMs) to perform similarly to the bulkier Gemini 1.5 Pro.
Gemini 2.0 Flash Thinking appears to be Google’s answer to the OpenAI o1 model. This LLM is trained with reinforcement learning to check its work before providing an answer. This makes it slower and more computationally expensive than other models, but it can avoid some of the mistakes that other models make.
Similarly, Google says Gemini 2.0 Flash Thinking is a “reasoning” model trained to show its work. You can see the line of thought the model used to arrive at a conclusion, offering an interesting insight into the way Google hopes to address the persistent hallucination issues that plague generative AI. If you’re used to having AI spit out text within a second or two, get ready for disappointment. The process of fact-checking itself causes the model to churn for additional seconds or minutes as it builds an output.
Even at the end of this process, Gemini 2.0 Flash Thinking still gets some things wrong. We gave it the classic strawberry test, which it got wrong. After being informed of this, the model’s line of thought got considerably more complex, showing how the bot attempted to work around the fault. And it did, getting the correct answer the second time.
This is an encouraging development for Google’s AI efforts. It shows Google is keeping up with the latest advances at OpenAI, even if Flash Thinking is still just experimental. However, this is yet another expression of the Silicon Valley dogma that throwing more compute resources at generative AI will make it more able to think and reason. But Gemini 2.0 Flash Thinking is still just running matrix multiplication to generate likely tokens behind all the “thinking.”
This release wraps up the first year of Google’s unified Gemini effort, which was announced in early 2024. Yes, it can feel like much longer because we rarely hear about any non-AI products from big tech anymore. If generative AI is going to actually get better, models like Gemini 2.0 Flash Thinking will make the case in 2025. You can try the new “thinking” model now on AI Studio.
#Google #Announces #Experimental #039Thinking039 #Gemini #Model
source: https://www.extremetech.com/computing/google-announces-experimental-thinking-gemini-ai-model
