Who’da thunk that just doing the data equivalent of squinting at something from a considerable distance and averaging-out a best guess based on poorly correlated data from unreliable sources could possibly ever go Rong?
Cannot fucking wait for this shit to start flying aircraft and operating weapons systems…
Internet loves fails.
it’s very funny, but no one talks of the successes. Plus, most of this shit is the free version that is meant to fail so you pay for the good stuff.
Where it is being used for serious decisions, it’s very heavily scrutinised. It’s not dissimilar to self driving cars - with people setting them up to fail with impossible tasks.
Yep, they aren’t there yet. But this sort of headline is missing the wood for the trees.
Every task i have set AI it has messed up the photos and names. It told me it did it by default to protect privacy. I asked it not to and it stopped. What’s autopilot btw? We have already entered that world…
So, like a toddler, it really is funny watching it cock things up but if you need it actually to do something then you’d better have a grown-up on hand to make sure that at least some of the food goes down its throat or that it gets across the road OK.
Self-driving cars have also failed with perfectly possible tasks. It may be that they fail less often than the average person. But the average person includes a percentage (hopefully small) of those who drive while drunk/too tired/too old/too angry/in a car with an uncontrolled child or animal etc etc. Any individual driver can be a good deal better than average by consciously not doing those things, and given that they can they’ll be reluctant to hand control to an AI system (is an autopilot/driver really ‘intelligent’ ?) until its performance under stress is quantifiably good. After more than 100 years development aircraft autopilots are. And to be fair they’re not dealing with particularly complex problems - the air is mostly empty and the obstructions (basically the terrain and the airport runways) tend not to move about much.
Seems to really struggle with the rendering of images, even on paid tiers. Good taste is also very difficult to quantify, reproduce.
What it’s not - yet - is an AI system, look-up (e.g.) ‘DO-178C’. Last time I looked no AI system had achieved this one certification. There are many others. No doubt it eventually will - when it will then be field-tested on us and our loved ones - much as (e.g.) Boeing’s various innovations have been…
But not yet.
What cyber forelock-tuggers don’t comprehend is that this is still very much a nascent technology, which like all human technology incorporates endless flaws introduced by human impatience, human incompetence, human laziness, human shortcut-taking, human cheapskating, human recklessness, limited human intelligence, &c. Since the prime motivation for these systems’ implementation is greed, those issues won’t go away in a hurry, because those human factors won’t be going away in a hurry.
Notwithstanding, commercial autopilot is a red herring - AI is already being used to control military weapon systems, and it matters not-at-all what uninformed idiots like us think about it. The people making those systems don’t give so much as a picoshit what we think, but people who do know what they’re talking about (like, for example, The United Nations), do care, very much, and for good reason.
All my wife’s work colleagues are writing guides on being a CEO etc. using AI.
They are hoovering up insights to try and package them as tools.
It is everywhere. My post was less to make the counter argument, but more to add some balance. This stuff will have some amusing flaws, and some widely shared fails, but it will quietly start helping in some areas that the web won’t want to hear.
If it replaces humans it will surely mirror our inadequacies at some tasks. It is how quickly it pushes through that to become better?
Better might be perfect, but it might just be less inadequate (which is better).
Topic of conversation at work at the moment in terms of what it can do to replace/reduce the number of first line support agents. Our desk isn’t in the main a log and flog team, and although AI could easily provide in depth assistance in terms of standard Microsoft and similar productivity tools. I think it would struggle today without lots of good quality KB on the whole range of in house built/niche applications we use. Also many of customers don’t want to spend the time reading stuff to get the answer/fix, they want someone else to do that.
The biggest factor is for some of these AI features is the licence/platform costs are so high, and the skills required in house or outsourced are also very high. One advertised by Idris Elba comes to mind. I suspect that will change though in the next few years. A lot of demand from our customer base for stuff like Copilot Pro which is being piloted, but again the licence cost for every user are high. List price from MS is £19 per month per user, adds up with 9k of users even if likely to be discounted for our volume.
Sadly I do think it will significantly change how IT support is delivered in time but it’s not the magic bullet yet. And I must say I am glad to be towards the end of my career, and with decent redundancy options should that come along.
But can they also judge which ‘insights’ are valuable and which are worse than useless ? If they can then it is they who are providing the intelligence, not the AI.
This guy, who’s been working in the field for 20 years, argued in this book
that there is no mechanism for AI to ‘become better’ on its own. When it gets better that happens because genuinely intelligent entities (us) make it better. It’s not yet as good as we are (faster, maybe, and with a bigger memory, but not as good) so we can still improve it. But if it gets close to being as good as us then the ‘getting better’ process will run out of steam. After all, we’ve tried all the tricks we know (education, collaboration, removing the distraction of having to hunt/gather food etc) to make ourselves more intelligent and we seem to have reached the limit. If we can’t make ourselves smarter then how can machines do it ? (Maybe there’s some scaling nonlinearity - the AI ‘brain’ just has to be hugely bigger than ours and it will then be able to cross from Turing’s ‘ingenuity’ area into his ‘intuition’ one ?)
All that said, Larson’s book is from 2022 and 3 seconds is a long time in AI (or is it ?).
Annoy people who have already seen all of the standard responses on the website already so they go away?
Really well put. We are finding for standard productivity tools more of our users are self helping, so by the time they get to us they are really stuck/or have hit something we have disabled due to policy, or needs back end configuration.
I’m not 100% certain, but I’d be very surprised if any auto-pilot system actually deployed in an aircraftw as non-deterministic, which is a fundamental component of both how current AI systems work, and also a significant part of some of the challenges they currently have.
Having had a brief play with ChatGPT5, it does still fuck up some quite basic stuff (the Blueberry challenge being a prime one that’s circulated). That said, I threw my standard challenge - solving a basic 5x5 nonogram at it, and it nailed it first time. Up to now, every model I’ve ever tried (basically all the free ones, and Claude Opus which I actually pay for) have complete shit-the bed with it.
What’s also new with ChatGPT5 is that it’s effectively combining the more efficient versions of GPT4 with the “reasoning” model functionality that was called O3, but also doing some decision making as to whether a prompt required deep thinking. With the Blueberry challenge, it appears to default to the quick option and fail, but if you tell it to think deeply, it switches model and then gets it right. With the nonogram example, it went straight to deep thinking and got it correct.
If I had a pound for everytime I’ve heard some clueless fckwit, mainly senior clueless fuckwits, mention Agentic AI over the past year I’d have retired by now.
95% of companies can’t even sort out basic tooling, diagnostics and properly automated ticketing systems yet they seem to think AI will sort out the shit bucket of systems they already have in place. Drives me mental.
Indeed and every time you ask the vendors for a real life reference site you get “we will get back to you” especially if you ask for a public sector example. They also do nonsense like best end user portal award, and the top 10 are all Boeing, GE etc with megabucks budgets and team.
Thing is, management aspires to being rather than doing. If they can get rid of all the tiresome workers that actually deal with the sticky end of the business, they will be free to enter a managerial Nirvana of endless meetings.
OK, I’ll admit it, I’m impressed.
I saw a phrase on social media and didn’t even know what language I was looking at. Bashed it into the browser and the AI response told me that it was Polari, gave me a translation and the etymology of each word.
That’s my learning for today.
I’m probably easily impressed 
It could be quite impressive - based on my experiences, the question I want to ask is “have you any mechanism to verify that the information it gave you was correct?”
Ha ha, I’ve asked AI pretty much the same question when I’ve suspected it has ‘borrowed’ an answer from somewhere without validation.
It usually apologises and tries to come up with something else that it thinks I might like instead - which is equally flawed.
Well in this case I ended up going down quite a Polari rabbit hole and the AI is clearly just summarising the internet’s standing body of Polari knowledge. Assuming that’s not all wrong too then I’m quite confident.
Although I very much take your broader point.
