Chatgpt - Seriously good potential (or just some Internet fun)

mid_gen said:
No one serious about AI is doing anything with Grok. It’s a joke of a model, totally compromised with god knows what system prompt nonsense to inject Musk’s personal politics into everything.
Click to expand...

What, so when Musk said

Our focus thus far has just been on making Grok the smartest, most accurate AI in the world and I think we’ve largely succeeded in that.
Click to expand...

It wasn't true? :o:D
 
mid_gen said:
Anthropic is so far ahead in the developer space. Whenever I try the other models it's jarring how bad they are.

Been using Claude Opus 4.1 the last day or so, and it's continuing to get better. Using some co-ordinated sub agents, it's getting really close to being reliably hands off for the majority of coding tasks.
Click to expand...
It says a lot when every agent uses their models as the default. They also have a winner with CC, which seems to do a better job than something like CoPilot agent, although these things are constantly evolving.

I don't think we'll be seeing much improvement in the models for a while, at least not anything drastic - they have fundamental issues that require significant changes. The agents are where gains can be made because how they navigate the codebase matters a lot. These tools still struggle at understanding large projects.
 
anyone noticed how much chatgpt gas lights when your asking it questions about stuff that will be too recent for the model to know anything about.
I asked if the Gemini killer was fictional or not. then it came out with some rubbish from dexter season 6.

(don't read if you haven't seen the latest ep 6 of dexter)

vkX0ZoJ.jpeg

It literally dreamed everything up based on the little info I had told it...

Almost anytime I ask it something it comes out with fantasy but unless your used to spotting it and know it happens you wouldn't realise.
 
Last edited by a moderator:
Macky said:
I thought the v5 demo was pretty naff. The whole thing felt super awkward. The tech demos were unimpressive. They talked about improving its memory but didn't really demo it. The bar charts they used were ridiculously bad. The only thing I thought looked good was the voice feature. Are we seeing GPTs nearing their peak or are our expectations just really high?
Click to expand...
Depends what you want from it. Is nice to have more questions available for something like O3 level.

Personally, I think it would have been great if it could generate images that can hold up from different angles. For example, if it renders a character from the front and from the side, if it has a headlight a little bit to the right of the had (in the front rendered imaged), the side one will have the headlight to the middle of the head or in some other position. Or the example with the clock, it can't render something outside of 10:10 for an analogue clock. On top, if it could create characters / assets ready to be imported into a game engine (with textures, animation, etc), would be ideal! :D
And it needs a lot more memory for the end user to be more customizable.

I'd say we've hit a wall where they need a lot more power for very little improvement...

We'll see.
 
Getting frustrated with GPT5. I've being using GPT to keep track of food intake but since going to 5 it has been all over the place. I then tried the Google one on my phone and it told me to call a mental health charity after it suggested I was getting too much protein and I questioned it.
 
Trifid said:
Getting frustrated with GPT5. I've being using GPT to keep track of food intake but since going to 5 it has been all over the place. I then tried the Google one on my phone and it told me to call a mental health charity after it suggested I was getting too much protein and I questioned it.
Click to expand...
Yes, 5 has forgotten a chunk of some formatting rules I've been using, which took a little while to fix, but it's still better than doing it by hand.
 
Decided to have a play with ChatGPT to set up a fantasy football team this year. Usually just use auto pick anyway so thought I'd see what 'insight' it could provide and whether I fair any better (no chance I keep on top of it all year so I won't affect the friends taking it seriously, they know it's not me). It's also a chance for me to get used to writing better prompts as I've not yet used any AI tools in depth.

Initial reaction - it's been a real struggle. Source data timing issues aside it's been terribly inaccurate. Initial team was well over budget, fine values might have changed. Second suggestion it at least told me it was over budget and then proceeded to suggest who to swap to fix it, ok great but why not just cut to the chase? It then finished that up by saying it would recommend swapping one or two players to get a better striker, I said ok, it made the suggestion to swap one defender for one already in the selection (so used them twice) then forgot it had left me £2.5M spare so didn't need to claw back quite so much. Corrected again only this time changed another defender in the summary without telling me and when summarising the total spend added the correct values together but gave the wrong total, again total was over 100M but conveniently told me it was bang on. A bit more back and forth and eventually we got there with a team.

It then wanted to talk about the first 6 games and who I should bench/captain. Sure sounds useful. More errors. Gave wrong fixtures and justified keeper picks with nonsense unrelated to either keepers games. A bit more prompting and explicitly saying prioritise accuracy and analysis over speed of response and we did get somewhere. Whether it's a good team I couldn't tell you as I'm no expert, but it does feel heavily stacked towards midfield players. I'll ask a few mates what they make of that as a strategy. It's also thrown out lots of predictive points ranges for the upcoming games based on official ratings and it's own 'custom' weighting accounting for home vs away games and reviewing seasonal performance last year. This is the bit that's interesting to me as it's the analysis it's 'saved' me doing myself. Time will tell how the team does but I'm not too concerned about prediction accuracy as long as the logic was sound.

Essentially what I've learnt is to follow up everything it says with a question about it's accuracy and quality unless you know the material in depth. Ideally play it back against the original ask in another way. My prompts seemed reasonably good to get it going down the right path but it's 'over confidence' needs to be kept in check. It did at least have no issues telling me it was wrong when questioned. A good learning exercise.
 
petridish said:
Decided to have a play with ChatGPT to set up a fantasy football team this year.
Click to expand...
how have you initially trained it with source/background material ?

--------------


(from r4 today) see supreme leader Kier has re-targetted turing institute - limited funding is being used in the right benign way

Among the projects slated for closure are work on developing AI systems to detect online harms, producing AI tools that can help policymakers tackle issues such as inequality and affordability in the housing market and measuring the impact in health inequality of major policy decisions like lockdowns.

Other projects expected to close include an AI-based analysis of how the government and media interact. A project looking at social bias in AI outcomes will also be dropped. Projects being paused include a study into how AI might affect human rights and democracy, as well as research into creating a global approach to AI ethics.

A spokesperson for ATI said: “We’re shaping a new phase for the Turing, and this requires substantial organisational change to ensure we deliver on the promise and unique role of the UK’s national institute for data science and AI. As we move forward, we’re focused on delivering real-world impact across society’s biggest challenges, including responding to the national need to double down on our work in defence, national security and sovereign capabilities.”
Click to expand...
 
jpaul said:
how have you initially trained it with source/background material ?

--------------


(from r4 today) see supreme leader Kier has re-targetted turing institute - limited funding is being used in the right benign way
Click to expand...
I have honestly never seen anyone so obviously in denial of a huge man-crush before. You literally never stop thinking about him.
 
Freakbro said:
What, so when Musk said

It wasn't true? :o:D
Click to expand...

Musk just wants to make a Musk-bot. Then change the law to make it a legal entity, leaving his billions to it and not having inheritance tax or loosing money to his human offspring.

In short Musk has a robochubbie to be the daddy of skynet. I'm not even sure he would programme the AI not to kill off his DNA line..
 
petridish said:
I thought it could go away and pull sources? Certainly that's what it seemed to suggest it was doing.
Click to expand...
maybe google AI mode is less powerful , but, in other domains, I have asked it questions and found it is limited on domain of knowledge
(I asked it what cars had anthropomorphoc design features yesterday .. and it didn't know the twingo )

mid_gen said:
I have honestly never seen anyone so obviously in denial of a huge man-crush before. You literally never stop thinking about him.
Click to expand...
no - was more mr turing's legacy that starmer if ****** up - the imagination game was on the tv other night.
wolfgang gullich maybe
 
petridish said:
Essentially what I've learnt is to follow up everything it says with a question about it's accuracy and quality unless you know the material in depth. Ideally play it back against the original ask in another way. My prompts seemed reasonably good to get it going down the right path but it's 'over confidence' needs to be kept in check. It did at least have no issues telling me it was wrong when questioned. A good learning exercise.
Click to expand...
My niece is starting GCSEs this September. Over summer holidays was given a Photography task which included a 500 word essay on Heimat Berlin Typeface and getting her to research it was never going to happen so I got ChatGPT to do the legwork. Its first response was a some information that just seemed generic and lacking any real information, when asked for a sample 500 word essay based on this info it came up with something that was 400 words and was fairly poor. Asking more about the information it had provided and more prompting it eventually came up with something that was perfect for what she needed. She still needs to write it but the information is all there and it should be fairly easy. Did take about 20 minutes of prompting and double checking its info to get there and to make sure it was accurate still had to put the research work in but it has produced the perfect set of notes for a dyslexic 14 year old with adhd.

Probably going to have to his a lot over the next 18 months so good to know that its going to be a very helpful research tool
 
Why the heck is social media full of endless videos about becoming millionaires with various app ideas, digital ebooks and what not? It's like type some prompts and bam you're instantly rolling in passive income. Agent this as prompts that. It's like how many ideas can you saturate with releasing these apps or side hustles.
 
ddoubleep said:
Why the heck is social media full of endless videos about becoming millionaires with various app ideas, digital ebooks and what not? It's like type some prompts and bam you're instantly rolling in passive income. Agent this as prompts that. It's like how many ideas can you saturate with releasing these apps or side hustles.
Click to expand...
It’s just a grift to get people to buy your advice, which is ultimately some generic BS and didn’t even work for the seller, otherwise they’d be making millions on their app instead of selling poorly formatted PDFs.

While you could use an LLM to build an app, you ultimately need real engineering experience to make anything remotely productionisable.
 
You must log in or register to reply here.
Back
Top Bottom