From some quick tests, although the headlines have focused on the impressive maths, coding and reasoning benchmarks the latest release of Gemini (25/3) seems to excel at creative writing when simulating fictional characters.
From what I have seen so far this is head and shoulders above other models that I have tried in capturing nuance, presenting characters realistically with high emotional intelligence and no obvious moral skew to behaviour. This has the potential to be a fantastic model as a creative writing partner, or in applications like roleplay simulation.
However, I still find that topics that would be comfortably tackled in adult fiction (think about classics such as The Road, 1984, The Handmaid’s tale, the Pit and the Pendulum and so on) often receive a blank/refusal response due to safety triggers.
While I wholeheartedly support and understand the need for appropriate model safety controls, in so many cases the implementation of these is a very blunt tool, massively handicapping the broader utility of those models for so many use cases.
It would be extremely beneficial long-term to see safety constructs developed that protect against the most serious risks like aiding malware production and so on, but do not overly sanitise, morally skew or patronise broader adult audiences who wish to use these tools in creative projects.
Please keep up the great work - exciting to see such a capable model landing.
The reason why human image symbols and abstract right-brain focused interaction might seem weak in LLMs is because LLM super logic and the developer’s humanities and realistic analysis side have been trained metacognitively. Contrary to you, I think it’s the best model I’ve used so far. I believe that creative poetry and stories should remain human domains. Artificial intelligence, IT technologies, and science emerged as an innovation revolution with a scientific and technological infrastructure. For my field, ecosystem, and position, it’s more important for AI to produce outputs beyond human capabilities in humanities, scientific methods, IT, and technology development. I think Google made the right decision by developing it this way. Thanks to it, we solved a project we couldn’t resolve for weeks in just 2 days, and it performs very realistic, natural, and logical analysis, exceeding human reality in a way I really appreciate. Yes, we as humans are very used to our fake persona, and you might not see it in 2.5 Pro, but since the world is full of people, there’s no need for AI to provide that to me. Right now, developing it in a way that will advance the AI ecosystem is more important. No-code, low-code, and efficiency will increase significantly. I sincerely thank Google; it’s a truly excellent LLM, perfectly aligned with its concept, and trained with very high quality.