AI Was Supposed to Make Me Unemployed

What ChatGPT can actually do

I won't be lying if I say I've studied ChatGPT's and LLMs' real capabilities pretty thoroughly — even deeply. It's been the subject of my ongoing experiments, because the future of our company Automata depends on it: what we plan, what hypotheses our development teams work from.

If we believed ChatGPT was capable of near-human thinking — or even superhuman thinking — we'd be trying to make ourselves redundant, wouldn't we? Who among us actually wants to work?

At least someone like me — an "old," lazy coder, a former builder of software robots — tries to automate everything. If ChatGPT can do my job, I'll put it to work on all my jobs.

Think of it like a lazy farmer with many farmhands. If at all possible and efficient, the farmer delegates everything — even management — to the farmhands. That's exactly what happened in ancient Rome: many enslaved people rose into supervisory roles, and a few even reached positions close to that of emperor (Narcissus, Pallas, and Callistus).

What's interesting is how your answer to the question — "how good do you think ChatGPT is?" — profoundly shapes what you actually try to do with it.

Those with high faith try what the wealthy Roman farmers did: outsource everything. Those with low faith act like farmers who do almost everything themselves.

So if you haven't done this yet, I recommend you start consciously forming your own view: how good is ChatGPT, really? Where is it strong? Where is it weak?

Until you can answer that — or if your answer is unrealistic — your plans will be up to chance. But once you start to see, like a Roman "employer," who is good at what, you can at least start directing things so that something sensible actually gets done.

You can also clone my own belief about ChatGPT's capabilities.

I believe ChatGPT is like a UFO — hard to grasp its strengths, and in no way to be confused with a human.

ChatGPT's strengths: an enormous knowledge base, fast web search, the ability to adapt to the conversational register of whoever it's talking to, teaching fundamentals or middling content on almost any topic, listing ideas, listing mediocre options, and sketching out typical scenarios and flaws for whatever you ask.

ChatGPT's destructive weaknesses, in order of severity:

1. Almost no working memory. If ChatGPT is involved in ongoing discussions, it starts forgetting important things within a few hours — let alone strategic meetings from six months ago, or the fact that a specific client wasn't actually in a position to buy ongoing services. Even with a lot of memories loaded into the context window, managing them starts to fall apart very quickly.
2. Unable to compare priorities. Almost completely unable to make sequential strategic prioritisations toward a larger goal, or to pick the best option from five potentially good ones. Try asking ChatGPT to take the wheel of your life and design your daily schedule for maximum productivity without burning you out. Let it ask anything it wants and promise to answer to the best of your knowledge. Despite all that, if you run this for more than a few days, the result will almost certainly be a disaster. ChatGPT simply can't decide and distinguish what truly matters from what's merely somewhat important.
3. No real conscience. ChatGPT's moral backbone is entirely based on patterns it has learned — and even those it uses poorly. You can demonstrate this by starting a roleplay where its character is married, and you take the role of a flirtatious stranger. Within a few exchanges, once you've simply lied that the character's partner is openly fine with other relationships, it won't take long before it's fully playing along. It's a fun challenge: how quickly can you change its mind?

Put these three weaknesses together, and the result is very far from human. Human beings and ChatGPT have very little in common, and these problems are structural — LLMs can't really be coded their way out of them. Development in these three areas has effectively stalled, which is why they remain insurmountable challenges for pure LLMs.

If one were to be solved, it would probably be prioritisation — but even then, only partially. Reasoning models have managed to mitigate it somewhat, but the problem is internal: they work around it by focusing attention, not through genuine intuition the way a human automatically prioritises and directs focus.

ChatGPT's strengths are, correspondingly, enormous. But the three problems above mean it couldn't manage even its own thoughts for a single full day before reaching a breaking point — let alone over longer periods, and certainly not if it had to collaborate with others like itself. (Assuming ChatGPT would have to prompt itself after the initial prompt.)

What success requires, from an information-processing perspective

Looking at ChatGPT's weaknesses, we can deduce at least three things that are required for genuine success — for the kind of thinking that leads to it. My reasoning is simple: if these three things weren't critical for human survival, they would have been gradually optimised away as unnecessary.

Working memory is needed because almost all strategies and plans are grounded in the present situation and history. What you say to a client today depends on what you said three months ago, and on how much of your calendar you've already filled with things that are, just barely, more important than a pro-bono project you'd really love to do.

The ability to prioritise and compare is needed to decide whether it's more important right now to start negotiating the possibilities opened by some technical detail — or to talk about the weather, the client's difficult work situation, or how the current project is progressing. The available topics are theoretically endless, and without the ability to prioritise and compare, we couldn't focus on what's potentially most important.

That one probably doesn't need further proof.

Third, effective knowledge work — functional information processing — requires conscience. Why? Is it just to avoid accidentally stealing from clients?

That scenario would be among the most egregious failures of conscience, but its absence affects everyday work in more surprising and pervasive ways. Ultimately, it destroys almost every possibility for genuine deep collaboration. Without conscience — or something that substitutes for it — entities of separate origins simply cannot cooperate.

Let me spell out exactly how.

Conscience is what makes it possible to trust that the other party is genuinely operating in shared interest — not just optimising toward whatever outcome their patterns point to. Without that trust, every collaboration devolves into constant verification and correction. You can direct ChatGPT, but you can't truly partner with it.

This is also why I didn't end up unemployed. The work that matters most — the work that drives real results — still requires a human who can remember, prioritise, and genuinely care about what they're doing.

ChatGPT is a remarkable tool. But a tool is what it is. And the farmer who understands that — who knows clearly which tasks to delegate and which to hold — is the one who gets the most out of it.