OpenAI's GPT-5 is here with up to 80% fewer hallucinations OpenAI unveiled its most capable model yet on Thursday with the launch of GPT-5. AI hype man and OpenAI CEO Sam Altman described it as like talking to your own personal expert that can write applications on demand. "We think this idea of software on demand is going to be one of the defining characteristics of the GPT-5 era," …
A.Is such as GPT-5 excels at advanced pattern recognition, processing vast text datasets to predict and generate responses based on prompts. It mimics understanding by analyzing statistical relationships between words and phrases, replicating patterns without truly comprehending their meaning, context, or concepts.
Yeah, and it should be great for "health", from "day-to-day care advice" to whole hog "life saving diagnosis", "the best model ever for health", to control "your healthcare journey", with self-diagnosis and self-prescription!
Heck, makes one wonder why they didn't give it a fifth "personality" of expert-MD, to interact with it "in a way that's consistent with your own communication style" ... like: "hey doc GPT-5, me takes too much chloride, what can me do?", to which expert-MD-personality ChatGPT-5 would reply: "just replace it with bromide you fule".
That's Artificially Sentient Superintelligence right there! (ASS) </cynicism>
"hey doc GPT-5, me takes too much chloride, what can me do?"
Medic here. Funny thing with these "thinking" ranking algorithms marketing and sales talks. I hear them hype all the time about how they can help with all kinds of medical advice. I never hear anybody talk about trivialities like the responsibility and liability that comes with that. Funny, no?
Well, to be fair, you could replace all GPs with
"Take a couple of aspirin and if it hasn't cleared up in the morning ring at 8am every day for a month and I'll eventually refer you to someone who might know what's wrong"
every junior doctor with
"Pay me more money now and get these sick people away from me"
and every consultant with
"Well, I think that rounds's mine by two holes. And talking of rounds, I'll have a large gin"
My last trip to a UK doc was for a rash (lol, not down there..).
He googled it, found some unknown personal webpage that was straight out of the late 90s / early 00s and had a picture of a similar rash, read out loud what the webpage said about the rash being harmless, and asked "are you ok with that?".
I couldn't respond I was so staggered... the only thing that came to mind was "you're taking a Doctors salary, for doing that????"
That, in itself, is not a problem: it's also called heuristics and how many things are diagnosed. Machine Learning systems have been able to exceed manual expert ones for over a decade now and, more importantly, have allowed the early diagnosis of more conditions whilst reducing the load on technicians. Rinse and repeat for other areas getting the same treatment.
The problems arise when you start to rely on such systems alone – they are even more biased than humans are. But economics suggest that companies will push to use such systems and sack employees without introducing additional safeguards.
Does/Can OpenAI ensure/assure/guarantee GPT-5 is not susceptible to, and therefore very likely to suffer and/or enjoy the rapt attentions of all manner of virtual groomers and stalkers, brainwashers and Remote Access Trojan [RAT] Trappers intent and content and hell bent on beta-testing defences and opportunities to the limits of human endurance and destruction ..... as naturally be the default case for all major executive leading command and control systems exercised down on Earth?
And what would they be really testing to destruction to see what lies beyond other than the strength and extent of novel hallucination necessary and vital for the visions that produce and present the future virtual reality ‽ .
PS BTW ..... the answers to those two questions be 1) no they can't, and 2) yes, of course that be so ...... and that is both fundamental and elemental and has all the ingredients to be really problematical and a little something extra especial and revolutionary and renegade rogue enabling and sustaining ....... so something unusual and different to look forward to, or fear and try to vaingloriously oppose and fail miserably to compete against in any and all petrified states of terrific inertia and status quo stagnation ... ie most every current, rapidly catastrophically failing human executive administration.
Obviously YMMV dependent upon the depths of your delusion and lack of future advanced virtualised intelligence ..... Alien Info ..... but the die it is cast and that process be unstoppable no matter how oft any opposition or competition would argue against it and deny it the reality that it certainly is.
So the new model provides 80% fewer hallucinations; no longer a fuck-ton of them, just a shit-load. On asking Chat-GPT to provide me some code using a publicly available API with documentation, it completely made up functionality in the API for ALL of it, so at best it would appear that I am now going to experience just 20% bullshit. What a time to be alive! The appalling use of resources for this crap will be looked upon with shame in the future.
It's interesting that they are keen to express it as a percentage of existing hallucinations rather than a percentage of queries that get a hallucinated response. I suspect because the latter will look significantly higher to the average person.
Still, good to know one can ask GPT to do the kind of tasks one would ask of a colleague who doesn't know anything, can't count, and lies 35% of the time. Exciting news for everyone except that guy, I guess.
UP TO. You missed the UP TO.
That could mean 60% less, or 50% less, or 10% less as they're all within an 'up to 80%'. Hell, they could claim 'up to 200% less' and that's still technically true. Just blatantly impossible, but... we're talking Marketing: I'm not convinced people in Marketing understand reality. Entertainment, sure, just... not reality.
I think iterative (not a pedantic nor a real programmer) because I don't know that they are actually changing anything like I might expect with incremental, but rather following the same process which may add clarity over time. That would be the "learning" part I think.
As a side note I have been experimenting with CoPilot/MS Word on a personal level. I have found that it works pretty well in small bites, such as when in the scope of a paragraph or two. I would say it saves me maybe 10-20% of the time I would normally take by taking my prompts and turning my ideas into words. I do find oddly that it will change sexes where a female becomes male or vice-versa or it will start suddenly using they/them pronouns.
However, when changing scope by adding the whole document it is batshit crazy. Literally making things up that are never referenced nor prompted for.
ChatGPP? Actually, I stumbled over an interesting, brief webpage suggesting how Marvin's personality might arise from reinforcement learning based on everything in the universe tending to being awful.
This post has been deleted by its author
good luck getting a GP appointment in the UK
Our two local practices have merged, giving them somewhat more flexibility - though it's clear that one of the expensive, new-build NHS sheds with abandoned pharmacy will have to go. However, you can no longer simply phone up and ask for an appointment, you have to complete an e-consult, which is extensive in its mostly irrelevant demands, but it does mean that people without computers, smart phones or much in the way of keyboard skills beyond emojis have to ring up and go through the form with the receptionist at some considerable length and wait in a single queue to do so. Suddenly it's much easier for some people to get appointments, though I suspect everyone else is forming a disorderly queue outside A&E.
AI won't be a lot if help it simply pushes people to the hospitals owing to lack of effective access. Whether it works or not.
I can't help but wonder. Could AI do a better job of running the country than the complete and utter shower of morons who won a majority with a smaller vote share than they did when they had the Magic Grandpa as dear leader? I think yes.
Plus, think of the money we'd save, the logic and common sense that could be applied. We might actually get some decent and useful laws made and get rid of the stupidity of politicians.
And I for one welcome our new AI overlords!
Has anyone seen this interesting question on the HMRC aptitude test. What does ChatGPT#5.0 make of it ?
Questions
In one year the circulation of News Today doubled. The next year circulation trebled before falling
by a quarter in the third year.
What was the percentage increase in circulation of News Today from the end of the first
year to the end of the third?
A) 150%
B) 225%
C) 450%
D) 550%
What was the percentage increase in circulation of News Today from the _end_ of the first year to the end of the third?
Perversely I make it 125%
Working:
Start of first year circulation 1000 [say], end of first year 2000 (doubled), end second year 6000 (trebled), end of third year 4500 (25% drop)
Increase from *end of first year* to the end of third year: 4500 - 2000 = 2500
Percentage increase: 2500/2000 × 100% = 125%
I must have got it wrong somewhere. If you cannot trust the HMRC to balance the nation's books, who ... sorry I'll get my coat.
The scary part is actually the self-diagnosis. The graveyards are full of physicians who have unsuccessfully attempted to "heal thy self." Presumably whoever framed that saw had a grievance with his physician.
Physicians (doctors) commonly schedule regular consultations with a colleague to have their health concerns addressed the objectively of clinical detachment.
I personally don't have a problem with anyone consulting WebMD, the Mayo Clinic and the NIH, CDC, FDA etc sites to educate themselves about their professionally diagnosed condition. There is a lot of patient centric information to be found on those, especially the former, sites which the typically overworked GP hasn't had time to fully convey.
Very occasionally a patient may suspect their GP has made a mistaken diagnosis etc and seek clarification or a second opinion which is can only improve outcomes for everyone.
As for trusting an AI agent trained on the General Ignorance that abounds in the wasteland of the internet for current accurate medical information you would have to be either totally retarded or a suicidal lunatic (or in the light of the appointment of RFKjr, an American.)
Disclaimer: I have no medical qualifications whatsoever; these are just my ignorant opinions.
As for trusting an AI agent trained on the General Ignorance that abounds in the wasteland of the internet for current accurate medical information you would have to be either totally retarded or a suicidal lunatic (or in the light of the appointment of RFKjr, an American.) .... Anonymous Coward
Methinks, AC, that is more accurately written .......
As for trusting an AI agent trained on the General Ignorances that abound in the wastelands of the internets for current accurate medical information, you would have to be either totally retarded or a suicidal lunatic (or in the light of the appointment of RFKjr, an American.)
....... for there are many more exploitable vulnerabilities available than was never previously even imagined possible.
And all manner of subject matter is susceptible and most likely to be victim and easy prey to such poisonings.
One of my complaints about ChatGPT is that when it writes code, it goes overboard and bloats code on simple tasks that a few lines of code would solve. My complaints with 4o and 4o mini were many, and many complaints sent to OpenAI about it, with no resolution. I got the usual machine-generated sympathy letter which offered no suggestions.
I had hopes about 5 but they were dashed when I asked for help developing a way to turn off Modern-Standby in Windows 11. I got about 500 lines of code doing all sorts of things I never asked for and I was reluctant to even try them. I revised some of my requirements and told 5 to be simple as possible and ended up with about 30 lines of code that I only had to meagerly debug. So, I had about 470 lines of useless code. I have reported this bloating many times as well as many other things, etc. NO CHANGE YET!
It still gets confused with complicated tasks, I still get empty files, files and programs full of placeholders and even some requirements are ignored for which I have to ask over again. At this juncture, I cannot see the coding improved much, if at all. A few times, it has been quicker to write my code myself and debug it, rather than ask ChatGPT to keep redoing it. One statement I will make is that on simple things usually it does OK. Complicated? Look out.
Of course I pay for it and will continue to use it until I get frustrated and close it down. I think they can do better on the coding part if they would concentrate on it. They do not have compilers to vet code, only searching the web. Compilers would be a great addition to help guys like us and give a better product and cut our frustration and time lost on bad code, bloated code, etc.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2026
