Recently, Google has unveiled an innovative "reasoning" AI model, Gemini 2.0 Flash Thinking Experimental, which is currently in its experimental phase. While our initial tests indicate that there is significant room for enhancement, this model is a noteworthy development in the field.
Available on AI Studio, Google's AI prototyping platform, Gemini 2.0 Flash Thinking Experimental is described in its model card as being adept at "multimodal understanding, reasoning, and coding," with the capacity to tackle intricate problems across various domains such as programming, mathematics, and physics.
Logan Kilpatrick, who oversees the product for AI Studio, refers to Gemini 2.0 Flash Thinking Experimental as "the inaugural step in Google's journey of reasoning" in a post on X. Jeff Dean, the chief scientist at Google DeepMind, Google's AI research division, mentioned in his post that Gemini 2.0 Flash Thinking Experimental is "trained to employ thoughts to bolster its reasoning capabilities."
He also noted that "promising results are observed when we augment the inference time computation," which pertains to the computational resources allocated to operate the model as it deliberates on a question. Despite being in its early stages, the model's approach to a complex puzzle that integrates both visual and textual cues is worth examining: (2/3) pic.twitter.com/JltHeK7Fo7 — Logan Kilpatrick (@OfficialLoganK) December 19, 2024.
Gemini 2.0 Flash Thinking Experimental is built upon Google's recently introduced Gemini 2.0 Flash model and appears to share a similar design philosophy with OpenAI's o1 and other so-called reasoning models. Unlike the majority of AI, reasoning models are capable of effectively self-fact-checking, which helps them circumvent some of the common pitfalls that typically ensnare AI models.
However, as a trade-off, reasoning models often require a longer time frame—ranging from seconds to minutes—to arrive at solutions. Upon receiving a prompt, Gemini 2.0 Flash Thinking Experimental takes a pause before responding, contemplating a variety of related prompts and "articulating" its reasoning process. Eventually, the model synthesizes what it deems to be the most accurate response. That is the intended functionality, at least. When queried about the number of 'R's in the word "strawberry," the model incorrectly identified "two," demonstrating that Google's new reasoning model sometimes struggles with letter counting in words.
Varying experiences may be had with the model. Following the release of o1, there has been a surge in reasoning models from competing AI labs, not just from Google. In early November, DeepSeek, an AI research firm funded by quantitative traders, launched a preview of its inaugural reasoning model, DeepSeek-R1. In the same month, Alibaba's Qwen team announced what they claimed to be the first "open" contender to o1. Bloomberg reported in October that Google had multiple teams working on the development of reasoning models. Subsequent reporting by The Information in November revealed that the company has at least 200 researchers dedicated to this technology.
The quest for innovative methods to refine generative AI has opened the floodgates for reasoning models. As my colleague Max Zeff recently reported, "brute force" techniques for scaling up models are no longer yielding the improvements they once did. However, not everyone is convinced that reasoning models represent the optimal path forward. They tend to be costly, primarily due to the substantial computational power needed to run them. While they have performed well on benchmarks thus far, it remains uncertain whether reasoning models can sustain this pace of progress.
By Michael Brown/Dec 20, 2024
By John Smith/Dec 20, 2024
By John Smith/Dec 20, 2024
By Lily Simpson/Dec 20, 2024
By Michael Brown/Dec 20, 2024
By Joshua Howard/Dec 20, 2024
By David Anderson/Dec 20, 2024
By George Bailey/Dec 20, 2024
By William Miller/Dec 20, 2024
By Emily Johnson/Dec 20, 2024
By Samuel Cooper/Dec 16, 2024
By Rebecca Stewart/Dec 16, 2024
By Jessica Lee/Dec 16, 2024
By Megan Clark/Dec 16, 2024
By Noah Bell/Dec 16, 2024
By Lily Simpson/Dec 16, 2024
By Ryan Martin/Dec 16, 2024
By Daniel Scott/Dec 16, 2024
By Noah Bell/Dec 16, 2024
By Emma Thompson/Dec 16, 2024