You have to turn on thinking model
Gemini 2.5 Pro isnt that smart, I was comparing specifically to its Deep Think version which is $250 / mo
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Gemini 2.5 Pro isnt that smart, I was comparing specifically to its Deep Think version which is $250 / mo
I tried playing with Qwen 3 Max and was both disappointed and pleased.
I was disappointed because it flagged even intentionally soft questions as possibly having innapropriate content, and sometimes flat out told me that I could not access certain data. Mind you, this was even though I was being very careful to NOT trigger its safeguards. But almost ALL of my questions about anything relating to 20th century Chinese history were denied, even those having nothing to do with the CCP. Even a simple request to explain the 19th century Sino-Japanese war was rebuffed.
Despite some bumps when talking about history, all my questions concerning basic scientific processes were accurate, and correctly reflected my direction to write at various audience comprehension levels. I was impressed with the speed and creativity in adapting output to explain complex ideas to lay audiences.
My conclusion: Qwen 3 is a powerful tool, but will not tolerate many questions about history, and it is difficult for users to predict what is an acceptable vs non-acceptable question. If I was a Chinese citizen worried about my social credit score, I would be worried about how an arbitrary red flag might reflect back on me.
While I found Qwen 3's output to be fast and useful when it did cooperate, I did not find it to be superior to other competing models. Given the seemingly arbitrary restrictions, my next question is why would someone chose this model over other less arbitrarily restrictive tools?
Give your most complex prompt and ill share gemini deep thinking version
Why do you want to research stuffs about Chinas politics or history on a Chinese platform in the first place? You know it’s not these companies fault, they have to be careful about not attracting a crackdown by authorities if they don’t censor enough or follow censorship rules . So of course they have to self censor things like that. If you want to research things about Chinas history or politics, then use non Chinese platforms. For the rest of other things you can use Chinese platforms as far as it has nothing to do with those type of questions.
until there is some sort of relaxation on the censorship rules , companies will have to adhere to the rules/laws, the only issue with this is that sometimes the rules are so vague that to be on the safe side , companies just adopt a complete restriction and censor on whole topics/sectors even those that shouldn’t have any reason to trigger censors but I guess for them they better be safe than sorry . So it’s understandable.
as far as they perform well and give full information about every other topics and country outside China, I believe they are good enough.
solarz
Brigadier
GPT 5 is markedly worse than GPT 3 when it comes to conversational flow and sounding natural. GPT 5 would nitpick your wording and give largely irrelevant information.
Seems like LLMs have hit a wall, you can either sound natural, or you can enforce information accuracy, not both.
Seems like LLMs have hit a wall, you can either sound natural, or you can enforce information accuracy, not both.
"I was being very careful to NOT trigger its safeguards." I posed a series of questions having to do with non-controversial historical topics, along with scientific, mathematics, and economics, since those are fairly common ways of assessing performance and all are areas any high school or university student would want to use.
I was not trying to be hard on the engine or intentionally break it. I was just disappointed with it in how it performed on history, especially since the topics about which I asked are readily answered on Baidu. All of the questions I asked it have their own Baidu pages, so none of the information was controversial in any way. The engine was just being very picky about recent history is all. I am just reporting what I observed, and people can try it for themselves.
I have noticed similar annoying and IMO unecessary safeguards in other tools, and half the time I can't tell if it is an actual safeguard or just the AI engine refusing to cooperate. They do that after all.
I was not trying to be hard on the engine or intentionally break it. I was just disappointed with it in how it performed on history, especially since the topics about which I asked are readily answered on Baidu. All of the questions I asked it have their own Baidu pages, so none of the information was controversial in any way. The engine was just being very picky about recent history is all. I am just reporting what I observed, and people can try it for themselves.
I have noticed similar annoying and IMO unecessary safeguards in other tools, and half the time I can't tell if it is an actual safeguard or just the AI engine refusing to cooperate. They do that after all.
Consumption of western media results in tremendous brain damage. It's like chronic alcoholism.
Cursor bans Chinese developers from using their service, then uses Chinese open source models as its backend. Color me completely unsurprised. These firms are basically the equivalent of those US drone startups that repackage Chinese parts.
well, cursor doesn't have it's own large scale training ability so it has to use open source fine turning. I mean what option does it have?
China already has its own AI type of IDEs like Trae. The predominant AI backend for cursor up until this point is claude. And as we know, Anthropic is blocking Chinese users, so what choice does cursor have?