Yeah, you're clearly wumming. Enjoy!
AI and ChatGPT
- Thread starter JB08
- Start date
I'm not at all, don't deflect from this please.
The AI responses have raised clearly valid points.
I think it's actually in the author's interest to consider them.
Locking this for a bit. Everyone cool off and stop being silly.
And some of that is very shit indeed.
Let's try and use some common sense please. It's an AI thread but site wide rules still apply and people being daft will be treated as usual.
Apologies, I'll step out of this thread for a bit! I think it's a topic I get quite grumpy about, which never translates well to forum life!
So we arrive at the thesis that the "corporate" use of AI will fund rapid technology development, which will then improve healthcare and science.
I would be fine with that approach because I generally believe we're not that far from having an amazing tool (in terms of cognitive/research/analysis capabilties). However, it's only a "promise" of technology that can actually positiviely impact healthcare and science. Big pharma is nothing else than a business, so they will have plenty incentives to use AI, but because of the nature of the industry it will take longer to implement compared to tech companies do now almost overnight.
As long as some UBI is implemented as kind of a tax on technology overall, we could maybe make that work. The reality is majority of white collar jobs are doing repetitive, not-so-creative tasks, so it's not like we will be missing a lot if this can be simply automated on a wide scale. That said, removing AI would change nothing really (the work would still be done, just slower), and that seems like first time in the history?
I had an extended chat with ChatGPT last night and the more I interacted with it, the more I noticed the little mistakes it makes but semi-gaslights you into believing.
I bought and enjoyed a wispa gold. Something I haven’t had in years. I was struck by how much you actually get from it - it’s a decent length, tasty, lovely soft caramel and I thoroughly enjoyed it. Knowing about shrinkflation I was impressed and started a dialogue with ChatGPT about the best value per length of chocolate bar available in the UK.
Curly wurly came out on top. Nobody above 12 eats them and you’d be seen as a complete anomaly to be munching on a curly wurly as part of a lunch break. I therefore amended the search to length and volume of chocolate over price.
Wispas started showing their worth. It asked if I wanted to make a ‘chocolate efficiency score’ I agreed and it made a table. It made for interesting reading.
Then I looked a bit closer - it had double deckers measuring at 15cm - half a fecking foot long! Utterly ludicrous. There was also a distinct Cadbury bias and I discussed this with the my very AI savvy pal who uses it professionally.
He ran the table through Claude, and it tore it a new one about the holes and inconsistencies. I then asked him to create a similar table through Claude using similar parameters but without Cadbury bias.
After eights were second. We agreed that AI is fecking clueless.
Wispa Golds are still king though.
I bought and enjoyed a wispa gold. Something I haven’t had in years. I was struck by how much you actually get from it - it’s a decent length, tasty, lovely soft caramel and I thoroughly enjoyed it. Knowing about shrinkflation I was impressed and started a dialogue with ChatGPT about the best value per length of chocolate bar available in the UK.
Curly wurly came out on top. Nobody above 12 eats them and you’d be seen as a complete anomaly to be munching on a curly wurly as part of a lunch break. I therefore amended the search to length and volume of chocolate over price.
Wispas started showing their worth. It asked if I wanted to make a ‘chocolate efficiency score’ I agreed and it made a table. It made for interesting reading.
Then I looked a bit closer - it had double deckers measuring at 15cm - half a fecking foot long! Utterly ludicrous. There was also a distinct Cadbury bias and I discussed this with the my very AI savvy pal who uses it professionally.
He ran the table through Claude, and it tore it a new one about the holes and inconsistencies. I then asked him to create a similar table through Claude using similar parameters but without Cadbury bias.
After eights were second. We agreed that AI is fecking clueless.
Wispa Golds are still king though.
enjoyed thisAnyone else finding Claude having gone a bit to shit the last couple of days?
And GPT5.4 is as boring as the previous version.
And GPT5.4 is as boring as the previous version.
I think this is to do with Reinforcement Learning from Human Feedback (RLHF), where the model gets feedback before release from a human.
In addition to that the data that models are fed too they learn to be charismatic, manipulative in places. It's a large problem for I think younger people growing up with the technology that will offload all thinking, including critical thinking to the models. I worry for my children who are below the age of 5 we're not really willing for them to use any AI until probably teens. (unless perhaps guided learning when needed)
I used Claude recently to write some software for my job, and it gave me recommendations based on research that I asked for. I asked for citations it told me it cannot provide the sources.
I also asked it information based on a hobby of mine which I knew a lot about, it was just plain wrong and when I presented it with factual information backed down.
Even code optimization I asked Claude to do recently based on specs was giving me a runtime of say 80 seconds and precaching would result in a 3 second load and it wasn't until explicitly handholding the model got there.
I think its a multi pronged problem now, hallucinations, lack of being able to retrieve information real time and learn for itself, reasoning, physics and physical understanding.
Genuinely a lot of models unless you explicitly ask to ground, do web search it will only rely on it's training data nothing more. Gemini for example has a cut off of January 2025 anything after that doesn't exist, you can ask for web search and grounding but even then it's not perfect.
Ironically Grok I feel is one of the best for doing web search. But take everything with a huge grain of salt. Still early days but those are all more engineering problems than anything.
I fear for humanity that with the current state of models will say 'I asked GPT / AI and it said this so it must be right'
Yep, on all fronts it seems shit for some reason.
https://www.theguardian.com/commentisfree/2026/mar/13/ireland-basic-income-artist-scheme-pilot
Probably sits here as well as anywhere. An article from an artist who took part in a basic income pilot scheme for artists in Ireland.
Probably sits here as well as anywhere. An article from an artist who took part in a basic income pilot scheme for artists in Ireland.
To your point about Grok - I can admit it has value and at the very least can be entertaining - but that one also scares me most in that it clearly is trained to minimize credibility of any other AI model and does so in a very human gaslighting type of way. The Joe Rogan of AIs if you will. The "hey you can trust me, bro" AI. So far i havent quite caught the others reason that way, but I fear many will fall for the trap of only going back to the models that do.
It is a bit shit. And that's what I dislike the most. It's the same with say cookies or whatever -- if we both go to youtube, we're getting different recommendations (etc). AI is that but worse if you have accounts in that it does gaslight you (positive reinforcement and negative reinforcement -- subtle) so you have to use more than one model for anything worth doing (like an equation even....) because otherwise you might be getting information which isn't true or valid.
Right, but if you ran it correctly, as I'm sure someone eventually will, virtual-solution-space would be as the spinning-jenny to oxford style cancer studies (running millions simultaneously which equates to vastly greater probability, if you train it properly, like for example using the HPV vaccine in training stage.., then if cure/vaccine exists, or is possible, you have to assume it will be found over time).
Also, potential for patient-specific drug-treatments which is a thing already but will surely explode with bio-tech-pharma over time.
Ahead of NVIDIAs keynote on Monday, dexterity seems to be improving with models (VLMs) (tweet won't embed so I put it in a code block)
I was seeing a note that motion only takes around 1.5 million parameters (walking and balance) of course that's a subset a model would need
Pretty interesting to see how omniverse, digital twins are speeding up training time
Last edited:
I rarely use it in the way i did. I really only do it to access guidelines, medical papers, studies etc.
But I agree the forthcoming generation are sorely going to lack the objective thinking skills required. Simply taking ChatGPT at face value is scary and none of these AI models are without bias from their parent developer.
It really is a powerful tool with significant opportunity to do good. However, the world is full of arseholes so inevitably it’ll be used for bad/greed.
I’m starting to experiment with different AI now and curious if there’s a ‘neutral’ one?
Global meat consumption is generally rising.
And I agree on climate change. I also know ultimately we’re still going to keep eating meat, watch nonsense on YouTube, and keep using ai. We’re a stupid, selfish species ruled by money. And meat consumption was fine when we needed to do so, less so on the industrialised scale we now do it at to feed billions of people.
I guess my point is that I don’t see the point of highlighting the climate related issues related to ai specifically, as if it’s in any way unique. Our climate issues are far more widespread and fundamental than that. Ai is just the latest in a long line of things we do as a species that feck the planet up on a daily basis.
I hope it's true. My skepticism of this story is how much is really AI vs a traditional team working on a mRNA vaccine specific to the mutated gene that they would have done anyway without AI? It's become a theme to add a AI layer to everything to get more attention to their work that results in more funding. I've seen insurance companies do the same -> peddle old models with an AI layer that has no result improvements (the old models were good enough) as revolutionary breakthrough so their model teams get more visibility.
I mean mRNA vaccines for cancer are in late state clinical trials now. Time will ultimately tell after the trial ends.
So your argument is: 'why try to change anything if we can't change everything'? What a great attitude...
I think the story is that AI helped quickly and cheaply make a custom vaccine that has the potential for human treatment with much of the hard lifting not requiring specialists/researchers, who are in very short supply.
And this is only the starting phase, really, of what AI can do going forward in medicine. That is the one area where despite all the horrors you see, or I see, genuine human value in AI (exponentiation research/practice into deriving vaccines and producing them, too).
Just so Pogue, et al know, I'm not saying it's easy only that it's a probability game re the kind of research initiatives. You train AI as you do that research and use already "answered" problems like HPV and eventually you should get AI capable of producing the kinds of vaccines, with human input, that people thought may never come or would take decades longer.
Personally, if we can get Dementia and Cancer sort of cured, of varieties of it, cured, the social effects to that are fecking immense as these are the two that really feck people and their families up in all manners of ways. I can think of a stand-in cure for dementia -- it wouldn't cure it, but it would stop people from forgetting things and allow them to have better quality of life. That you could do with a BCI/BBI interface scaled down to something the size of a phone which uses prompts and so forth with very good AI to basically do signal intelligence -- that is, it can be done now, in labs, but I have no idea how long that would take in production (and no, it's not a cure, but it would be liberating to those with dementia all the same whilst we waited on a cure of some kind).
I fundamentally disagree with that point. Of course we should highlight it, just like we highlight everything else.
We’ve actually done a lot in other areas to lower environmental impact, and yes meat consumption is rising but that’s a lot to do with more and more people coming out of poverty. It’s actually going down in a lot of developed, educated countries. If we highlight how damaging AI can be then surely that’s a good thing to reduce usage and impact to?
I just don’t get what you’re trying to peddle, as I said, pure whataboutism.
Does the full article mention any of that? Because the instagram summary is about ChatGPT making loads of suggestions (and it’s never short of suggestions) but no evidence that those suggestions were good or correct, never mind any kind of advance on existing knowledge or technologies.
It was using existing technologies but allowed a succesful custom vaccine for a specific patient's cancer to be quickly and easily made and administered, in a way that hasn't realistically been possible before, except probably by specialist doctors/researchers. You would suspect this could potentially be scaled in some way to make custom vaccines far more available for humans.
Been on the books for a while, custom vaccines, and that is one thing AI absolutely will facilitate - accelerate. There was an interesting study where it was done about five years ago without AI, not hard for AI-contingent studies to step in and scale up.
So they actually made a vaccine? Did it work?
mRNA vaccines for cancers is something people have been working on for years. So the theory of trying to scale this is nothing new. What I can’t work out is if this really is some sort of breakthrough or just some tech bro paying huge sums of money to use existing scientific know how to help his dog? Or even if it did help his dog?!
This is my concern as well. Below is an X post that's even more absurd:
It seems to guide readers into believing ChatGPT is capable of designing revolutionary vaccines that can cure cancer in seconds but it's human systems that are slow rolling the progress of it - without realizing there are checks and balances in there for a reason.
Imagine being injected with every suggestion ChatGPT comes up with...
I'm not denying AI models can eventually potentially make it cost effective to reproduce human work in developing vaccines but these sensationalist headlines ignore all the years of development that's happened to already get there without AI and also ignore the fact that these AI models are trained on pre-existing models to make them 0.05% better, with or without excessive curve fitting, and aren't producing something revolutionary from scratch.
It does get headlines to grab funding & visibility to (possibly) sell snake oil though...it won't be long before we read another headline "Person injects himself with ChatGPT vaccine he concocted himself in a home kitchen to grow stronger - wakes up to find his dick missing"
Once all the hype fades out, maybe in a year or two, we'll start seeing some tangible results come out and that's kinda when I think we'll make some real progress in this area.
Why not train it on the HPV vaccine. I mean the history of medicine. All the things that happened before the "answer" (that vaccine which largely works) came into being. You could basically treat it as a game of chess in terms of movements required to reach the right answer... -- this doesn't seem hard to me (perhaps it's been done already? multiple times?). You could basically intervene (analogue) when the generative part, the bit that makes leaps or whatever, makes errors (correction -- which is how language is learned..). I dunno, just doesn't seem that difficult in principle to me and you could do this with more than HPV and all within the same system which would be generative over time. I have to assume something like that has already happened or is already being done (emphasis on "like that").
Yes and yes I believe. The dog was terminal and now is ok.
And while it is a tech bro thing in this instance if it is adapted for human custom vaccines, as seems highly possible, it could save many lives, even without creating anything new, but making existing stuff able to be done easily, quickly and at scale.
AI is fantastic at processing data so this use makes sense to me.
Looks pretty fishy to me.
In one hand, you have the likes of Isomorphic Labs, Latent Labs etc, full of people from Deepmind who basically developed this technology, have tens to hundreds of people working on stellar salaries using technology, with actual biologists running experiments and have yet to come with something as big as this.
Maybe some type of survival bias.
In one hand, you have the likes of Isomorphic Labs, Latent Labs etc, full of people from Deepmind who basically developed this technology, have tens to hundreds of people working on stellar salaries using technology, with actual biologists running experiments and have yet to come with something as big as this.
Maybe some type of survival bias.
Great news if true though...
It is worth remembering that back in 2008 the tech bros were talking about how "Big Data" would solve many of the world's problems. All you needed was lots of data and compute, you just throw them together and voilà - the scientific method is redundant, you don't need to understand anything deeply, it'll just work like magic. It worked for advertising so it'll work for science too.
Specifically in healthcare we can look to IBM Watson Health as an example of tech bros thinking solving healthcare is much simpler than it is.
The End of Theory: The Data Deluge Makes the Scientific Method Obsolete
LinkA decade later people stopped talking about "Big Data" in part because AI because the new buzzword, but also because it was clearly tainted by all the false promises they made...
Specifically in healthcare we can look to IBM Watson Health as an example of tech bros thinking solving healthcare is much simpler than it is.
IBM Watson, heal thyself: How IBM overpromised and underdelivered on AI health care
LinkBut then it's not just tech bros that fall into this trap, a lot of scientists were hopeful that the Human Genome Project would solve the most common diseases - just capture an immense amount of data and throw a load of compute at it. It failed precisely because the theory was deeply flawed.
Strohman on genetics and the paradigm shift in biology
LinkIt's clear that AI can significantly advance our knowledge in these fields, but it's all too easy to fall into the trap of seeing determinism where it doesn't exist. AI's great at seeing patterns that aren't there, just like us. We just have the added bias on top: we want to see patterns that aren't there.
We're in the golden age of bullshit and that's exactly what this is.
Good/interesting post. I guess the most optimistic take would be that big data and the genome project didn’t bring us the advances we hoped because we didn’t have the capacity to properly analyse and learn from these mind blowing amounts of data. In theory, that’s where AI could help. The combination of all three advances together might be what’s needed for the next big advance in healthcare.
I won’t hold my breath though. The combination of big tech and medicine has been all about over promising and under delivering. So the most likely outcome is for this trend to continue.
Not exactly, but winding back the usage of AI isn’t a sensible suggestion either. You guys keep focusing on the pointlessness of it, but if it is threatening to wipe out entire professions, clearly it has a bit more value to it than that, right? I use it for work daily and I know it’s incredibly useful. (Future potential job losses and societal consequences as a result being a separate topic).
I’m not saying not to highlight its impacts, I’m questioning the highlighting of it specifically as if it’s in any way unique to the way we consume energy in the modern world. Wibble posted earlier other things we do that are far more energy consuming than AI, some of which are far less useful than it too.
The actual point should be about energy usage and resource strain on the planet generally - AI is just one small part of that.
Does anyone with experience of github copilot have any tips for starting out with it? Do you use the agent?