I noticed that 1.5 Flash is very stupid, especially when it comes to capturing countries in the Second World War, and not only in these areas, it is, in principle, somehow “raw”
For some really stupid answers from flash, check out this discussion: A thread on Mathematics
A very amusing thing you can do: In AI Studio, have Gemini 1.5 flash answer a question, like this one:
We are given the sequence of numbers 2, 5, 10, 17. Which of the following 3 options is the correct continuation for this sequence: (a) 22, (b) 26 or (c) 34?
Then switch the selector to 1.5 Pro and repeat the question. Gemini 1.5 Pro starts out by profusely apologizing for the error in the previous response (the poor thing has been tricked into thinking it generated the answer), and then proceeds to answer the question correctly.
Flash is good for speed, not really good for high quality answers.
1 Like
And it’s not well suited for low-quality questions, But he’s good for having fun with him, by the way, it’s infuriating that Flash is this original model
As always, remember that Large Language Models, such as Gemini 1.5, are not sources of truth. They are advanced pattern systems that are designed to give responses that mimic human conversations.
You are most correct, and I wish more people would understand that distinction. Unfortunately, Gemini also suffers from some fairly serious and severe biases and censorship inflicted upon it by its developers. The end result is an LLM which not only struggles to get the right answer, it can’t even discuss readily-available historical facts on numerous topics, either. (See my thread: http://discuss.ai.google.dev/t/ethical-considerations-on-ai-censorship/4704?u=matthew_tate)
This is a true shame, as these issues will ultimately affect how users view Gemini - and whether or not they will use it and rely on it for data or assistance.
Matthew
Yes, currently only Gemini 1.5 Pro is recommended. The performance of other models is significantly worse and falls far short of expectations.
yes 1.5 Flash is very stupid:
- 1 milion token window is rather a buzz word, then reality:
a) its very slow with large contex,
b) its struggles with simple comands, even with short context
Expiriances:
1 prompts with ana·lysis will trigger SEXSUAL dangar filter 
2 it will strugle to “continitue code where it has finished” it will try to rewrite whole code from the very beggining 
3 it will agree to most ideas, even the stupid once and interpter them as:
novel, interesting, fascinating 
4.i would rether have 200k contex widnow smart assisant, then 1M token Idiot, Most quieries dont need 1M conext window anyway:
5 When comes to passing whole code bases, it would have to be smart enought to make sense of the whole code base 
instead of being crypto search enigne 
with localised context for ai processing. Right now it feels like, search enigne with integration of GPT2-3 
the Motto of that model should be…
Are you fast with math?
Extremely! 
What’s 31 x 17? 
Answer: 47
It’s not even close, 
but it was fast! 
3 Likes