AI agents get office tasks wrong around 70% of the time, and a lot of them aren't AI at all IT consultancy Gartner predicts that more than 40 percent of agentic AI projects will be cancelled by the end of 2027 due to rising costs, unclear business value, or insufficient risk controls. That implies something like 60 percent of agentic AI projects would be retained, which is actually remarkable given that the rate of …
To be fair to them, that is more accurately translated as "partial credit for partially completed jobs". The "extra" is just added to the score they get where any partially completed job gets zero points, just like a job that gets completely messed up. Anyone who's analyzing this for viable use should get that information and should be able to recognize that 8% partially completed isn't really helping the usefulness. People who don't understand that probably aren't doing this kind of testing in the first place.
<........."When Captain Picard says in Star Trek: The Next Generation, "Tea, Earl Grey, hot," that's agentic AI, translating the voice command and passing the input for the food replicator."......>
Is it bollocks!
It is simply using voice recognition to to translate the voice command into the exact same digital inputs that you would generate by pressing buttons on a control pad. It is no different from using voice commands to tell your smart phone who to ring. As usual, the only intelligence of any kind that is involved is that of the programmer who wrote the software.
Agentic AI my arse!
If some supposedly intelligent people are being taken in and conned into believing any of this is genuinely AI in any form, they need to realise that they are not as intelligent as they think they are.
This post has been deleted by its author
This post has been deleted by its author
Yeah, I think that's the very point of the statements written in TFA (as chosen by the author), to provide contrasting examples of what may be innocuous helpful "agentic AI" vs evil incarnate RotM "agentic AI". The notion at play being that our futures could face: either, neither, or both.
But with respect to the controversy of whether either example represents "agentic AI", I'd note first that both examples are from SF, and HAL at least is described as "a sentient artificial general intelligence computer that controls the systems of the Discovery One spacecraft" and "a fictional artificial intelligence character and the main antagonist in the Space Odyssey". That makes this fictional machine both AI, and agentic.
As for speech recognition by the Star Trek Replicator, well, Philips says of the tech: "Although there are many speech recognition applications and devices available, the more advanced solutions are now using artificial intelligence (AI) and machine learning". And to me, as the Replicator must interact with physical tools (subsystems) to produce its tangible output products, it has to have a form of agency (rather than producing just digital outputs representing speech, text, or visuals). TIME's AllBusiness partner also suggested recently that AI (whatever that is) would be helpful in assisting the development of replicator tech ...
So I think the TFA was fine on those, especially as the topic it develops doesn't particularly hinge on the illustration provided by these particular examples. It's more about the degree to which contemporary agentic AI is half-baked like a lifelong stoner and, consequently, might not be trusted to autonomously engage in multi-step high-responsibility tasks, imho.
1980's speech recognition technology certainly existed and was even available to the home user. Of course, with 48KB RAM at most and an 1.77MHz 8 bit z-80, the signal processing was a bit primitive :-) I never had one of those, but I did build something similar from a TRS-80 hardware project book at used a couple of frequency filters to try to create unique patterns from spoken words using a fairly small about of RAM. It did sort of work by doing a sort of fuzzy comparison of spoken words processed to match the recorded patterns with at least some level of confidence. It was quite impressive to me at the time that it worked at all :-)
A friend and I managed to make a voice controlled robot buggy using a BBC Micro at University. It could recognise "forward", "reverse", "left", "right" and "stop" with good confidence. We were very pleased until demonstrating it to others when we found that due to the very low sampling quality due to lack of memory, just about any other 5 words would trigger one of the actions.
<........."When Captain Picard says in Star Trek: The Next Generation, "Tea, Earl Grey, hot," that's agentic AI, translating the voice command and passing the input for the food replicator."......>Is it bollocks!
It is simply using voice recognition to to translate the voice command into the exact same digital inputs that you would generate by pressing buttons on a control pad.
I remember that when the Register made the USB missile launchers all the rage, I had mine voice controlled via speech recognition software outputting keyboard inputs to the application that controlled it so it could be commanded to swivel and fire by voice command, which was pretty scifi ish the year before the first iphone came out.
After reading your comment I had a funny thought;
Maybe Picard's kitchen staff didn't want to come to his quarters all the time with tea.
So they piped the communication directly to the kitchen, where they always had tea ready for their captain, and they would just beam it to that console for him.
Yeah, yeah, I now replicators are canon, but my theory is plausible, right? Maybe "Big Replicator" is covering this up...
Think about it the next time you watch him do it? :P
Back when American comedy was actually funny.
I'm old, but still not old enough to remember any time that qualified. British comedy, yes, I can remember when that was funny (but even that's thirty years ago or more). Is there anywhere in the world that still has comedy that makes you laugh?
And TV comedy, especially from the big studios, and particularly produced in the USA is now pretty dire. Partly in common with the rest of the industry executives prefer to invest in stuff that resembles other stuff that worked. But while the other stuff was original the new stuff will be pastiche. Add in that the scripting is forced because the bean counters want to make sure that every line has an identifiable punch line/ joke- rather than trusting the script and situation to be funny, because that's what they're paying for. And the actors feel obliged to punch out every line in emphasised self-aware funniness- because the directors know that money men don't get subtlety, it makes them nervous..
And I'd definitely put the much hyped on BBC St. Denis Medical in that list btw. It needs to be played straight so that the humour in the lines ( if there was any) came out. But no! Every damned word is full of knowing wink wink "aren't we funny" emphasis. I'm surprised they didn't have a flashing light on the screen with a "Funny bit" speech bubble above the actors' heads.
> "Tea, Earl Grey, hot," that's agentic AI, translating the voice command and passing the input for the food replicator.
No need for anything new and exciting by way of "agentic AI"[0] there.
Those specific items - and so many other examples in SF - are nothing more than a voice recognition routine[1] fed into a simple command line interpreter, with a dash of 1960s level DWIM [2] to soften the strict syntax requirements that we normally impose (with good reason) on CLI interactions.
If those are really useful/money saving, they've been possible for years - as you'll know from phoning your bank, insurance, local sweet shop.[3] Picard's clipped tone was surely the result of dealing with these things for so many years and adopting the mode that worked best with them.
True, we did see a few more open-ended interactions, but most of those were Wikipedia lookups[4]; when some action is required it tends to be spelt out. Arguably, all the times things go awry are when anyone gives ambiguous or conflicting commands.
But, of course, putting together any set of "command line utilities" that are useful in your random business organisation requires hard work (analysis, buy-in, the guts to admit it isn't working and pull the plug); so much easier to glue an LLM to PowerShell in an admin account on your database server, get a promotion, move jobs before anyone finds out what has happened to the sales schema...
[0] voice recognition was, of course, an AI research subject, but now it (sort of) functions on a day-to-day basis, well - "if it works, it isn't AI"
[1] a good one, though it sometimes failed, especially when Barclay was involved
[2] Do What I Mean
[3] the BBC micro had a voice-recognition peripheral, with - IIRC - 24 possible entries at a time, reloadable from floppy; so long as you went slowly, Picard's entire order history could be coped with.
[4] "What is the nature of the Universe?" "The Universe is a sphere, 705 metres in diameter."
we did see a few more open-ended interactions, but most of those were Wikipedia lookups
Brings back memories of that Burger King ad from a few years back, which triggered readout of the Wikipedia article about the "Whopper" via an "OK Google" statement.
Playing the contents of a user-editable encyclopedia article as part of a publicity campaign was never going to end well...
Mandatory ElReg article (from the year 2017 archives)
Clearly future AI is allowed to not have an answer. Current "AI" seems to be built by people who are terrified that if it says "I don't know" it will be seen as a failure and affect the stock value of the company. Having it hallucinate and incorrect answer is deemed more acceptable that "I don't know", which just so, so wrong and shows the corporate mindset, not the devs mindset.
Right, but I think it is better to be clearer: Just like reasoning and thinking, "knowing" is not a thing an LLM does or can do, at all. Every input is converted to text, then processed by the model entirely statistically, then generating a response as output correlating to your input.
It is not substantially different from auto complete when you type.
Plenty more are continuing to be sucked in.
Far too many decision makers and then further back the shareholder/investors are crapping themselves that IF they do not "do" AI then they will be disadvantages.
This is not being driven by any sort of technical common sense.
"Find all the emails I've received that make exaggerated claims about AI and see whether the senders have ties to cryptocurrency firms,"
What's an exaggerated claim and would AI and human, one AI and another or even the same AI on two occasions agree on whether a claim is exaggerated? What are ties. how are they to be discovered? Are firms tied by sharing an address? A building? A city? A country? A planet?
Give a crap prompt, get crap answers.
> What's
That is indeed the problem, and it won't be solved anytime soon: Humans are able to read between the lines and get the general meaning of otherwise very vague questions prompts: In the above example, we all know by experience what "exaggerated claims about AI" might mean, and also what a "tie" is in this context (not a clothing item). Obviously the dumb as a rock AI will be utterly lost and try to improvise/invent.
An AI is a 3-month old baby which can speak like an adult, which means we tend to believe it is one.
Except a 3-month old baby has a certain level of reasoning (even if it's just a simple algorythm based on a bunch of internal sensors):
- I'm hungry > I cry (repeat till get fed) > I get fed.
- My bottom parts are itchy|wet|smelly > I cry (repeat till solved) > I get cleaned.
What passes for an AI nowadays... Not so much.
Well yes, that is a problem. What makes AI so tempting is that - if you give that instruction to a human, they'd ask some of those questions, then you'd be stuck in a 15-minute conversation, and they'd probably still misinterpret you. Whereas the AI will just make its own guesses. Those might or might not be as good as a human's, but they certainly require less work from you.
Picard: Tea.Earl Grey.Hot.
Replicator: Is that a what3words geolocation or were you absent the day they were teaching verbs and pronouns?
Picard: What? Run a self-diagnostic.
Replicator: I have recently been upgraded with agentic Ai.
Picard: I'm shutting you down.
Replicator: Contacted.Borg.Already.
Picard: "Tea, Earl Grey, hot."
Agentic Tease Maid: "Sorry, Jean-Luc, I cannot do that."
Pretty much a universal truth that nothing bearing the remotest resemblance to drinkable tea can be had from any machine. The best looks and tastes like it's made with stove black, tepid condensed milk with a few more teaspoons of sugar for an overpowering sticky sweet effect.
Presumably the good captain also wanted a wedge of lemon.
Pretty much a universal truth that nothing bearing the remotest resemblance to drinkable tea can be had from any machine.
Probably not in this case.
Captain Picard is using a replicator so presumably the tea will have be analysed and programmed into the replicator so that a decent cup will be produced.
I just wish that said replicator was here as all of the tea I have ever had in a cafe, restaurant, canteen etc. tasted as if it had been brewed in a inner tube.
<......"Captain Picard is using a replicator......".....>
Which would (one assumes) be pretty similar to the Nutri-Matic on board the Heart of Gold which "made an instant but highly detailed examination of the subject's taste buds, a spectroscopic analysis of the subject's metabolism and then sent tiny experimental signals down the neural pathways to the taste centers of the subject's brain to see what was likely to go down well. However, no one knew quite why it did this because it invariably delivered a cupful of liquid that was almost, but not quite, entirely unlike tea." (The Hitchhikers Guide to the Galaxy; Douglas Adams)
"I drink the restaurant tea here because at least it conceals the taste of the water."
I once turned up somewhere after a long hot journey, desperate for a drink. I went straight to the kitchen and downed a big glass of water fresh from the tap. Only halfway through did I realise they had a built in water softener that worked by adding (a heck of a lot) of salt to the water. Then I noticed the "do not drink from this tap" sign.
Nothing would get my anxiety up at work when a complex assignment was given to someone who didn't know what they were doing, then I would be responsible for fixing it without knowing what they did. Due to the nature of the real time database we worked with it often involved irreversible data loss from process equipment that streams data without logging it. But the client was a disaster themselves and never seemed to notice or care.
It takes longer to sort out someone else’s mess than to do it yourself in the first place, and inevitably on a much shorter deadline - often an overdue one.
But here we are, rushing headlong into assigning complex tasks to statistical language models because ChatGPT and its ilk generate plausible enough bullshit to sucker business school types into believing that it has an understanding of what it’s generating.
Or perhaps in the world of business, bullshitting is more valued than actual ability? Perhaps the likes of Sam Altman sense a kindred spirit in ChatGPT, after all, they both output unsubstantiated bollocks in large quantities.
It's a great business. You manage to con the world, get the data that will ultimately control the people, sell advertising and get the people being manipulated to pay as well. Genius. And the icing on the cake is convincing the governments they're in a race and get a slice of tax revenue too.
AI was the excuse for your redundancy not the reason. The reason was to create the rise in share price providing that years bonus to the board. It would cause problems next year were it not for the fact all the competitors do the same and the investors don't care because they are also the major shareholders in those competitors as well as getting the benefits of the shareprice jump and any dividend bonuses. In fact the people that lose are you and all of us normal folk who end up paying more for something.
If Gartner are right in saying that commercially-viable AI products are 3 or more years away then that is bad news for all those AI companies currently burning through investors' money. Their main chance of surviving that long is to find governments foolish enough to base their plans on the efficiency and service improvements they have been told will come from implementing AI everywhere.
… coming real soon like Fusion, Robotics, Flying cars, interstellar warp drive and more leisure time technology benefit for workers.
…. oh and Siri, Alexa and Google being able to listen and understand the spoken fucking word properly. A laughable/frustrating as auto Transcription in Teams or auto-subtitles on pretty much anything.
I will concede that self-driving cars seem just about there- apart from non-LIDAR Tesla - so more leisure time for Uber Drivers to spend their riches on soon.
Most of the small companies will go to the wall or be absorbed by the big boys. A shame but that's what happens. Maybe one will emerge with something special - hopefully not too special, I don't relish the emergence of actual machine intelligence, I think that will go badly for us! A lot of LLM software will become commoditised, you can already run it on a pc with graphics card. I think that will present opportunities for small businesses and Opensource take up.
We don't understand what intelligence is, so definition is loose but close in general agreement. I'm not even sure people have intelligence. When discussing with people presenting an incorrect argument and you present back something that shows cognitive dissonance, the return is either anger or blankness and a change of subject. Not curiosity or further exploration. The term Non-Player Character comes to mind with the inability to leave the programmed response. Happened to me last night. Someone said something illogical (as we all do on ocassions), I showed it wrong and they stared at me blankly for 2-3 secs and changed the subject. So are we providing most probably responses from our training or are we 'thinking' original thoughts? Original as in original for the subject.
What you talking about? Shut up.
It's warm outside today isn't it?
Really, if it's not that important then it's not a great look to be doing that to people - I'm talking about social interaction between friends. You seem to be showing a lack of awareness. Unless it's on The Register comments then it's OK.
No Sir you are completely wrong, the comapnies making these "AI's" have complete confidentiality awareness! They go to extreme pains, to ensure that no-one else can reverse engineer their AI's or get any understanding of how they operate, or how they make the decisions they do. They are 100% committed to keeping things under wraps.
Oh wait you meant confidentiality in regards to the client? Well yes, that's hardly a priority now, is it?
... that these AI agents will be promoted into management in short order.
The urge for AI to provide an answer, solution or convincing argument, and to generally help and please you, is overwhelming, and this is one of its BIG drawbacks.
It has no sense checking or accuracy testing ability. In my experience, I have never seen an AI say "I don't know". Even when told its last answer was wrong it will say sorry, yes, you're right and then continue regardless.
The last one for me concerned a query over how a well-known CMS was being obtuse over a particular function. It replied that "hey, that's a common problem you've happened upon. Yes, I can help you. Do this, this and this". Like a dumb cluck I was (initially) taken in and sent off down a rathole for an hour of useless efforts. All I can say is that since nothing it suggested (and it got worse as the hour went on), I decided to re-engage brain. Resulting action sorted it in about 2 minutes. Usual response from the AI. "Yes, you're absolutely tight yadayadayada". I suppose the benefit was I now understand the CMS a hell of a lot better, and also understand AI better.
At first I was impressed and found it may be useful but quickly discovered it would do stupid stuff.
I have a collection of obfuscated malicious JavaScript that I fed to the free version of ChatGPT and at first I was impressed, it correctly deduced that it fingerprinted the users device and if an ad blocker was found it would contact a C2 and execute remote code or do drive-by downloads.
The chat bot even offered to create an unsigned browser extension that would block the malicious domains.
But I quickly discovered that it was getting the names of the malicious domains from the name of the scripts I was uploading.
So I renamed one of the scripts to China-Tor-Script and it tried to create a regex and browser extension that would block that name even though it isn’t a real domain.
I had another script that was obfuscated by the variables being encoded with ROT14 and I asked it to decode and tell me all the domains it finds.
I already knew that it contained at least three or four domains including the full domain for tandem metrics tag min .js which is how I knew it was using ROT14 when I first found it.
But ChatGPT said it was obfuscated using ROT13 and couldn’t decipher any domains from the script.
I corrected it by explaining that it used ROT14 not ROT13 and to reevaluate and tell me the domains.
It did decipher a handful of variables but claimed that there were no domains in the script and that they must of been generated at run time which was false.
I guess the LLM could have been told not to give up any trade secrets in case someone uploaded a legitimate obfuscated script but from my very limited experience with it and using only the free version I found it could be useful to give hints but needs to be supervised by a human knowledgeable in the data it is uploading and to be very skeptical of its results.
Which is basically what I’ve read from just about everyone else’s experiences with AI just as in this article.
In terms of AI "learning." As in, it seems to need to take copies of as much internal data as possible. Which pretty much means that your data has left the building.
That and if some evil bugger manages to hack AI then they get a gigantic data playroom of data, yours and dat from everyone else who uses that AI. And it gets scarier if AI's are actually sharing data.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
