AI Updates - February 2025
Generative AI
Google
Gemini Flash Thinking Experimental
Google has released a new version of their Gemini Flash model dubbed Flash 2.0 Thinking Experimental. This seems to be in response to all the recently released models employing more advanced reasoning techniques. Initial evaluation points to a performance better than OpenAI o3 or DeepSeek R1.
Whisk Image Generator
Google has released an experimental image generator named Whisk. It allows one to provide different images, one for the topic, one for the scene and one for the style, then producing an original image using these features.
Below is an example where I used an original family photo, used the cover of a book for the style and a prompt requesting to dress the people in the photo in Venice carnival attire. The tool basically produces an image and then remixes it using the various elements provided. It is also possible to use prompts instead of images for the different aspects.
Whisk uses Imagen 3, the latest Gemini image generation model transparently in the background.
xAI
Grok 3
xAI announced an early beta of the new version of their LLM, Grok 3. xAI claims (like any owner of a newly released LLM nowadays) that their model outperforms their rivals in several benchmarks. Emphasising the reasoning mode of the model (which the user can invoke by asking the application to “think”) xAI reflected on math scores and coding as the strengths of the new model. Grok 3 came with two flavours, Grok 3 and Grok 3 Mini. It has a context window of 1 million tokens.
xAI released an iOS native version of Grok, thus freeing from the requirement to use within the X application.
DeepSearch
xAI released an agent framework at the same time they released Grok 3. I tried it for a research topic I had in mind and it did a quick search taking around 30 seconds and coming up with 70 sources and 10 references. You can also see the reasoning in a separate tab, following what the model do to accomplish the reasoning.
Adobe
Firefly Video Generator
Adobe has embedded a video generation model into its tools. The generator has a 175 word prompt limit. It is possible to provide camera instructions and the model can also generate videos from provided images.
Amazon
Alexa+
Amazon has announced a new version of its Digital Assistant, dubbed Alexa+, which will makes use of Large Language Models and other assistant features and rich APIs. It will be available to Amazon Prime users and also will be on many devices that have Alexa support.
Anthropic
Claude 3.7
Anthropic released the new version of their LLM. Claude 3.7 Sonnet has a reasoning mode, thus can show and explain the step-by-step reasoning. Claude Code is a new command line tool for agentic coding. It has an 128K output context limit.
OpenAI
GPT 4.5
OpenAI released their GPT 4.5 as a research prototype and will be deploying it for most subscription modes. OpenAI claims that the new model has a higher accuracy and lower hallucination rate. It is targeting a more general context (rather than the specific reasoning targeted by the o1 and o3 models). The scores compared to GPT-4o and GPT o3-mini are worse in math and reasoning but better in language-related tasks.