By skyforbes I use Gemini for complex tasks that require strict objectivity and honesty, such as:
- Scientific work
- Inventions
- Finance (mainly crypto)
I’ve been using AI models intensively since they became widely available—around 12 hours a day during my work. My prompts are refined with AI itself to avoid any ambiguity, and they often include documents ranging from 100k to over 500k tokens.
No matter the use case, Gemini consistently forgets the prompt midway—including system instructions—and breaks protocol.
Worse, since Gemini 2.5 it keeps acting as if it can execute code and then hallucinates the results.
But it gets even worse.
Gemini 3.0 still introduces cheating into algorithms (100% of the time, despite precautions) and fakes scientific experiments (Python, KiCad, etc.) “to please me,” even though that’s absolutely not what I want. I even designed a prompt forcing it to be more scientific and explicitly try to disprove my hypotheses. Yet it still forgets the prompt/protocol/instructions along the way and subtly injects dishonest results.
Recently, during a major project, I lost more than a month of work because of these issues, and it was a brutal shock.
All of this is incredibly discouraging, because it feels like there are no reliable AI models right now. GPT performs better in science (especially math), but I also find it has become increasingly unreliable for a while now