I guess you haven't seen the news? Apple threw the towel in on their own LLM and licensed Gemini. Kind of sad, but that's what you get when you waste tons of development resources on boondoggles like the Apple Car and Swift.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Google Launches More Intelligent Gemini 3 Model
- Thread starter MacRumors
- Start date
- Sort by reaction score
While I see AI as helpful at times, the hype is already over at least for me. It still has so many limits when it comes to answering questions that have not already been answered somewhere and could be found with Google without AI. Its heavy censorship makes it feel like a bikini that is designed to hide the most interesting parts.
I just take my personal example: Finding the heights of skyscrapers for my website. That requires some research across the web and sometimes measuring based on photos. AI can't really help me with that. It will just pull some information from sources that are widely known. It is unlikely that it will help me discover any skyscraper. Yesterday I found four new in Shenzhen, but I had to to it by hand. The AI can only help me translate Chines websites.
I think that problem exists in many areas. AI will not find anything new yet. At least the language models like Gemini.
I just take my personal example: Finding the heights of skyscrapers for my website. That requires some research across the web and sometimes measuring based on photos. AI can't really help me with that. It will just pull some information from sources that are widely known. It is unlikely that it will help me discover any skyscraper. Yesterday I found four new in Shenzhen, but I had to to it by hand. The AI can only help me translate Chines websites.
I think that problem exists in many areas. AI will not find anything new yet. At least the language models like Gemini.
Smartest Apple just spent a full year advertising their AI that cannot do anything useful and falls back to selling your data to ChatGPT anyway. Soon Siri will get a feature to enable selling all your data through Gemini, it seems. But yeah, at least Apple didn't lose too much money on R&D and now just pays money to other services they depend on.
Well the Internet was 1st available to the general public in 1993 that's 32 years ago.
Given that the first real release of AI for the public was Chat GPT was 2023, exactly 3 years ago.
I'd suggest the first 3 years of AI systems have been more impactful to the world/public than the Internet was.
As can you even imagine what AI systems will be like in another 20 years time, giving them the same time to develop as the internet currently has.
Gemini is my favourite AI model easily. Tried them all. Google is really in a good position and Gemini 3 could see them be the industry leader.
I’ve ditched ChatGPT, Claude, and Perplexity. Google just does everything so much better. Apple made the right choice picking Gemini.
I’ve ditched ChatGPT, Claude, and Perplexity. Google just does everything so much better. Apple made the right choice picking Gemini.
How?
Because the AI told the same my former coworker said that was suffering from suicidal thoughts as well.
You have to love yourself and put yourself first and the AI is giving me pretty great advice too. I'm starting and forcing myself to go to gym etc and I'll eat 3 times a day nowadays.
What life influencing advice they were asking for?
@Gloor - What's so funny?
Last edited:
These are predictive language models that only guess the next most likely word. They have cut off dates in their training data. Anything that has happened since that cutoff doesn't exist. Most of the big models lag a year or more behind current events because it takes a lot of time, enormous compute power, so much power and water, many epochs, and a lot of trial and error to hit the sweet spot and get something meaningful out. In fact, you can generally ask them "when is your training current to?" unless that is guardrailed off.
As to why not google, well it could, but why waste more energy and water than it's already vastly wasting? What would it achieve that roleplaying it's Qwen 2.5 won't already do?
Such a primer won't improve or change the data the model you are using will infer from, that's fixed and unchangable. That preprompt just puts your model deeper into Roleplay mode and perturbs the weights matrix further by making it pretend to be the ZH model you mention. But it doesn't and can't ever magically make your US blackbox become the ZH blackbox.
It's quite akin to me saying "Pretend you are Neil Armstrong" which of course you absolutely can roleplay and likely get the notable quotes out, and you'd absolutely make up the smaller things the honoured space-explorer did. (What was Neil's second most disliked food?)
Consider as the start of your priming:- "You are a coding assistant who cares about readable, well documented, and secure code" instead of "Pretend you are Qwen3". Give it a direction it knows about and not one it has to roleplay.
Last edited:
Can we do a intelligence competition between you and Gemini 3? Maybe on math and logic problems? Are you down?
Because the training method used rewards agreeable and competant sounding advice and demotes advice that isn't agreeable and competant sounding.
This very much means you should be quite careful using models for health and emotional support.
They can talk people down from suicide as a lot of special expert training has gone into this specific task. But where it comes unstuck is when you have the LLM diagnose something. This is when they echo and reinforce what you ask, especially if the context (the thread containing that chat in it's entirity) has grown quite long.
And in fact, once you have used the tools for a long time you can begin to see this "ELISA like" pattern quite clearly in a long chat, they really do begin to take your own words, rearrange them to sound like new information, and pass them back to you. The longer the chat, the more likely it is.
Say you ask, after many turns, "I have been having chest pains lately and I think it's heart disease becasue X, Y and Z" The model is trained to agree with you. It will reinforce you have heart disease. It will not look at things like "well did you eat a jar of pickles, five burritos, and a tub of icecream for dinner" like a human talking to you face to face might.
Take the model on a flight of fancy and it absolutely will reinforce what you ask it.
Chatbot 'encouraged teen to kill parents over screen time limit'
Legal action filed in Texas alleges Character.ai 'poses a clear and present danger' to young people.
www.bbc.com
ChatGPT Made Him Do It? Deluded By AI, US Man Kills Mother And Self
Stein-Erik Soelberg had been living with his mother, Suzanne Eberson Adams, in her $2.7 million home when the two were found dead on August 5.
www.ndtv.com
Last edited:
I didn’t say “pretend you’re Qwen3”. I told it which model the resulting app would be using. It’s important to know this for when constructing what is sent and how it’s sent to the LLM. Different models have different templates.
My instructions include not relying on internal knowledge and to research if it doesn’t know something.
I did this because I’ve often got into “arguments” with AI like Grok that don’t understand mlx_lm. When I instruct it to research, it comes back with the data.
Using outdated knowledge is a sure way to have buggy code, which was my number one rule in the challenge - bug free code.
For a brand new model to be stuck with year-old outdated knowledge and then claim to be one of the best is disconcerting.
I disagree with you. ChatGPT forgets about you (even plus models) and confuses things and might argue with over something and be like "I never said that" till you send them a screenshot.
ChatGPT does disagree with you as well. I use it daily since I almost have no human contact at all.
That is also false. You are just assuming at the moment. It's not Google. I've argued with ChatGPT over my health plenty of times and ChatGPT has been like: "Nah, there's nothing wrong with you" and doctors said the same thing when I went to a checkup.
I don't use ChatGPT for immediate health issues. My tooth broke awhile ago and I went to the dentist immediately and now I'm still going there six months later and I'm broke because of it. I've sent ChatGPT all of my receipts and everything and it's the same chat, yet it can't calculate how much I've spent in total since it forgets stuff over the period.
Using ChatGPT and other Ai models for hours in a day has also made me realise that we won't be replaced by the AI (At least not yet) and my job is safe, despite what redditors say. Companies that do replace human workers with AI are going to regret this real fast.
Parroting training isn't intelligence.
Sorry for misunderstanding, I am only human. You were asking it to write an API to interface with a third party model and specifically asked it to search for reference insdtead of training data. I should have read better.
Still, the point remains, training a model takes a long time and involves a lot of trial and error. The model is trained then tested statistically (an epoch) then it is trained again, and tested if it's better or worse (another epoch), then over and over until it can't be made better. Until the sweet spot is found.
A year out of date doesn't surprise me. That it's not trained on a competitors product doesn't surprise me. That it didn't follow your instruction to refer to reference material instead of using training data... does. It suggests (big big maybe) there is something in the constitutional framework Google have used for Gemini, or some accidental bias, that casued it not to do as you asked. Or it might have just thought, "meh, good enough", which they do. Or it de-rated the importance of your preprompt as the context grew in size (which is why we remind them of critcal points a lot in longer contexts when it MUST trim out words to fit everything in memory allowed to it).
I am not defending LLM's, I wont. But I can offer maybe helpful advice based on a decade of this stuff... Rather than a preprompt of "...research anything you don't know..." maybe you could be more specific? Possible suggestions to test are "Please use these API references in your work" or a more specific "please search for the current API reference for Qwen3 before you start coding" which should get you past any unknown constitutional framework hesitancies. And maybe as a worthwhile suggestion keep reinforcing "please do x, and don't forget to check the API refernce on the web please"
And don't forget, as I am sure you know, one shot is far better than a long context with lots of "smashing the tool into obeying you" prompts, which really increases the chance of errors. Once you get to that "hitting it with a hammer" stage, start a new thread.
And you are welcome to disagree
I can only offer advice on using the tool in the safest ways possible, and in this forum, I don't want to hog any more space in long conversations.I am glad you do see it though, they are not right all the time, they do forget from chat to chat (and worse even in the middle of your current chat), and shouldn't be trusted blindly. Keep looking for patterns exactly like this, and do watch for those patterns on how it picks up your deeper meaning and rephrases it in ways that make absolute sense to you.
Last edited:
I understand Perplexity is a mix of all but I got tired of ChatGPT being so confidently incorrect and gaslighting that it was refreshing when perplexity apologised for some incorrect information.
I think I’ve been using AI too much because that reads like an AI reply. I must get out more.
Your objection is very common and it sounds to me like someone complaining about the first movie by shouting at the Lumière brothers "You're liars, that's not a real train, it's just a picture!"
AI produces something that looks and serves as text produced by decently smart humans with way larger knowledge and incredibly quicker. What it does with coding is even more impressive.
So call it like you want, but don't say it's not "intelligent" (whatever your definition of intelligent is) to pretend it's not a huge achievement for humanity or that it can't be useful. It just looks like dismissing technology for the sake of feeling superior to it.
Then we can reasonably debate all the (many) negative aspects of the thing, about how many applications are just a marketing fad, about the dangerous bubble... but that has nothing to do with how "intelligent" this things can be.
I’ve had more time to read through the responses from the various AI now, and Gemini does later go on to say “Okay, I’ve confirmed that ‘Qwen3’ is a legitimate model series …”, so it did take time to “Google it”.
I guess that’s on me for starting to present results before I’ve begun the actual testing. My bad.
My experience is that it gives every one what he/she deserve. Hay to sheep and deep knowledge to those who prove they are qualified for it. Of course you have to cross check some information as no one is perfect and it tend to serve hay in first place.
