- Finally, less censorship. I use Gemini in Polish, and many synonyms or word combinations would randomly interrupt message creation. This issue has significantly decreased in this model. A step in the right direction.
- The model handles large prompts (5k+ tokens) much better, especially when system instructions are in JSON format. I’m genuinely surprised by the quality improvement. Another step in the right direction.
- Unfortunately, there’s still a noticeable drop in quality after 32k tokens. It’s as if contextual gaps appear—forgetting the conversation’s tone or specific details. This has been a persistent issue with Gemini. A large context window, but seemingly “leaky.”
- Unlimited usage. A very good move by Google. If they aim to attract new users, this could be quite appealing. Plus, additional training data.
Suggestions for development:
- I remember back in March when I had unlimited access to GPT, Claude, and Gemini. Only Claude Opus seemed to “understand the idea of a system prompt.” It’s one thing for a model to stick to instructions, but mindlessly following them sometimes kills performance quality. It’s crucial for a model to grasp the idea behind the instructions, as this significantly boosts quality.
- The ability to create model chains in AI Studio? While it’s no issue in the API to set up a scenario where, for example, three models handle specific tasks before their outputs are fed into a main model, it would be great to have something like this in AI Studio. Similar to Flowise or n8n.