DK_en 3x02 - Agentification
Episode first aired on 19 November 2024. Listen on Spreaker.com
Sometimes it bothers me to be right all the time. Also because I'm not paid for that.
You may remember that last July, a female CEO from an HR company named Lattice made the news announcing her software would list AI agents alongside ordinary employees. In her own words:
Today Lattice is making AI history. We will be the first to give digital workers official employee profiles within Lattice. Digital workers will be inducted, trained and will be securely assigned goals, performance metrics, access to appropriate systems and even a manager. Just like any other person.
The poor lady was made a laughing stock, and the idea was scrapped in a couple days.
A woman, and not even from a real tech company, what was she thinking?
But I made an episode for the Italian podcast on this, and I said:
Imagine if a Musk or an Altman had come up with the same idea earlier: all the media and the legions of sycophants would be kneeling before these geniuses. From that moment on, the first to implement the concept would have risen to the role of holy innovator themselves.
Now, a couple weeks ago an alpha male of BigTech said the same crap as the CEO of Lattice and everyone is now throwing bras on the stage like it was the reunion of Oasis.
The real problem is that now you have to worry.
AUDIO: DK theme
Listen to what Microsoft wrote a couple weeks ago presenting its vision of intelligent agents:
Imagine you are a new employee. A proactive Copilot greets you, thinks about HR data and answers your questions, introduces you to your buddy, explains training and deadlines, helps you with forms and sets up your first week of meetings. Now, HR and employees can work on their regular tasks without the hassle of administration.
Microsoft talks about a proactive Copilot, openAI of intelligent agents, and in general ‘agents’ is the buzz word of the day, we talk about ‘work or workplace agentification'.
If you are in a hurry, take away three ideas:
‘intelligent agents’ is an idea from the 1970s
the idea is problematic
the name is pure propaganda.
Are you still here? OK, then let's move on. The concept of ‘intelligent agent’ is as old as the idea of artificial intelligence, so much so that one of the earliest definitions of AI is precisely the creation of software capable of acting autonomously in response to internal goals and external stimuli.
There is nothing wrong with an old idea, as long as you don't allow yourself to be convinced that no one has thought of it before. There are whole libraries of books on intelligent agents and there are decades of experience on what works and what doesn't.
Of course the things that work nobody calls AI anymore, they don't look cool and they don't bring new business.
So the AI overlords are pushing things that either we already know not to work or should be taken with a large grain of salt.
Let's come to the idea. ‘Intelligent agent’ means an automaton that has a goal and can react by adapting to external stimuli in order to achieve the goal.
It sounds like a lot, but if we call it a ‘thermostat’ it just so happens that all the magic disappears. OK, a thermostat is an example of an intelligent agent. Which works, by the way.
Of course there are software intelligent agents. Those are called scripts, or bots. And they have their own problems, see the entry ‘flash crashes on the stock exchange’ on wikipedia.
The problem is known to anyone who has programmed at least once in their life, and in the industry it is called ‘do what I want, not what I say’.
But the fact that a problem has a name does not at all mean that it is a solved problem, far from it: as soon as we rise above the lowest level of ‘if this happens, do that’, the problem explodes in our hands.
The problem of repetitive tasks has existed for as long as work has, and as soon as computers arrived, the first thing was to try to automate them.
In fact, one of the strengths of UNIX systems has always been the ability to automate large parts of management with simple scripting languages.
UNIX, and Windows, and MacOS, and Linux, have been around for decades.
If you have never seen system administrator rolling on the floor laughing, ask them how much of their day has been automated.
What Microsoft and all of big tech would have us believe is that something that not once in forty years has ever risen above a marginal level of importance in a highly specialised technical environment, would now miraculously be within the reach of anyone working in an office thanks to language models, which everyone keeps calling ‘artificial intelligence’ in the hope it will become so.
If you are in a hurry to go, take another idea with you:
Agentification is bullshit of titanic proportions.
Let's drop the niceties and get to the root of the matter.
Automating repetitive tasks, whether you want to call it scripts, bots, or intelligent agents has always had two sides:
- the interface to the problem. What is ‘simple’ for a human is extremely complicated for a machine, and vice versa. Many think it means that programming languages syntaxes are too complex, and so we have graphical programming languages but, surprise!, programming has the same deep difficulties whatever the superficial level of appeal. Now the AI geniuses come along and say ‘it's all solved, the AI understands you when you speak’ and they throw themselves like a train against the wall represented by the second side of the problem;
- humans say one thing but often mean and imply many others, whereas programmes, however they are created, take an extremely literal approach.
Let us take a trivial example. Let's say we want to create a secretary who filters our phone calls according to who is calling. Simple enough, right? So we will have a certain group of people that the secretary always passes on to us, and others that are part of those we don't want to hear from.
Then there is Aunt Margaret, who is so good and dear but calls us at all hours for any nonsense as if we were still twelve years old and makes us the laughing stock of the office.
Of course we tell our smart assistant ‘for Aunt Margaret I'm always in meetings, don't ever put her through to me’.
Great. Add the automatic production of meeting minutes, and we practically have everything we could pay for in a secretary, at a fraction of the cost. Long live artificial intelligence!
From that day on, Aunt Margaret calls us, and the intelligent secretary doesn't put her through because we are in meetings.
Some days she even calls a couple of times, and we don't get distracted for a moment.
Then one evening we have just got home, the phone rings and it is Aunt Margaret who has been looking for us all day because our father was dying in hospital. But not to worry, he left in peace, without disturbing our working day.
pause
Now you may say ‘ah, but just tell the intelligent agent that if the reason for the call is this or that, then the call must go through anyway’; and I would reply that we are not understanding each other.
We are not understanding each other because you are prey to a hallucination called ‘technological solutionism’.
Whatever you come up with, it takes thirty seconds to find another case like Aunt Margaret.
So, how much time do you have to spend training an intelligent agent? Above all, when are you going to stop toying with the intelligent agent to be more productive and do the job you are paid for instead?
I have seen generations of system administrators hit the same wall over and over again: automation only works for extremely limited tasks under extremely precise conditions. Everything else is an exception, and usually, exceptions are 90 per cent of the cases.
This is because ‘repetitive tasks’ are only such at the level of detail at which a human perceives them after having spent a lifetime absorbing the nuances and details of living and operating in a social context.
That is what you pay a secretary for, not for her good looks or because she has a nice voice on the phone.
Repetitive tasks are either staggeringly trivial, and then the silliest of scripts or the most banal process redesign will suffice, or they are socio-technical problems, and socio-technical problems never have a technological solution.
Let's take the example chosen by Microsoft itself to promoote Intelligent agents, and look at it against the light, shall we?
Imagine you are a new employee. A proactive Copilot greets you, thinks about HR data and answers your questions, introduces you to your buddy, explains training and deadlines, helps you with forms and sets up your first week of meetings. Now, HR and employees can work on their regular tasks, without the hassle of administration.
So let's see:
- "intelligent agent reads you a list of courses you have to take"; for that, a list of courses is enough, you know how to read;
- "introduces you to your buddy", for that you just need a pre-filled email to add the buddy's name to;
- "reads you a list of courses and deadlines", again, stuff you can read for yourself;
- "helps you with forms", because either you can't read, or the forms were designed by monkeys;
- ‘and sets up your first week of meetings’, which requires the terrible effort of knowing who your colleagues and line manager are.
And the intelligent agent does all this automatically for you at the sole cost of a week of Switzerland's environmental footprint.
Again, this is the example Microsoft itself has chosen to use to illustrate how AI can free employees from ‘administration problems’.
Is it just me who thinks they chose an example for simpletons?
Could Microsoft really not find a better example, one that does not make an office look like a cage of trained monkeys perhaps?
Or maybe that's the idea? Because the message is not for those who work, but for those who pay for that work, and of course for those who want to be seen at work, and managers have their own idea of what others do in the office and what the problems are.
Another interesting detail:
if the problem is that an adult cannot fill out a form himself, the solution is a better form.
There are entire university courses for designing forms and questionnaires, but those don't help the bottom line of Microsoft, or Google, or openAI, or Anthropic.
And then there is the problem nobody wants to talk about: autonomy. We have all learned the hard way that one has to be careful when deleting a directory, or a filesystem.
rm -rf * has claimed more victims than many wars
According to the dominant thinking, ‘intelligent agents’ act autonomously. Even these next-generation ‘intelligent agents’ should be free to roam around the disks doing things with your access rights, on your behalf.
let me remind you that we are talking about tasks that we should establish simply by talking to the LLM; and we use natural language
- because natural language has no ambiguity,
- because human beings express their desires absolutely unambiguously, and
- because LLMs do not have a design flaw whereby, when something they are looking for is not there, they invent it.
What could possibly go wrong?
And we will stop here, because we will soon comment on the first mess made by an ‘intelligent agent’. I bet it will be presented as ‘human error’.
