Shopify CEO Tobi Lütke tells employees to prove AI can’t do the job before asking for resources.
Shopify CEO Tobi Lütke tells employees to prove AI can’t do the job before asking for resources
You're viewing a single thread.
What these CEOs don't understand is that even an error rate as low as 1% for LLMs is unacceptable at scale. Fully automating without humans somewhere in the loop will lead to major legal liabilities down the line, esp if mistakes can't be fixed fast.
Yup. If 1% of all requests result in failures and even cause damages, you‘ll quickly lose 99% of your customers.
It's starting to look like the oligarchs are going to replace every position they can with AI everywhere so we have no choice but to deal with its shit.
I suspect everyone is just going to be a manager from now on, managing AIs instead of people.
Building AI tools will also require very few of the skills of a manager from our generation. It’s better to be a prompt engineer, building evals and agentic AI than it is to actually manage. Management will be replaced by AI, it’s turtles all the way down. They’re going to expect you to be both a project manager and an engineer at the same time going forward, especially at less enterprising organizations with lower compliance and security bars to jump over. If you think of an organization as a tree structure, imagine if the tree was pruned, with fewer branches to the top, that’s what I imagine there end goal is.
...
What error rate do you think humans have? Because it sure as hell ain't as low as 1%.
But yeah, it is like the other person said: This gets rid of most employees but still leaves managers. And a manager dealing with an idiot who went off script versus an AI who hallucinated something is the same problem. If it is small? Just leave it. If it is big? Cancel the order.
A human has the ability to think outside the box when an unexpected error occurs, and seek resolution. AI could very well just tell you to kill yourself.
Yes. No over worked human would ever lose their crap and tell someone to go kill themselves.
What would happen to such a human? Do you suppose that we would try to give them every job on the planet? Or would they just get fired?
Error rate for good, disciplined developers is easily below 1%. That's what tests are for.
I mean it is also generous to the Artificial Idiot to say it only has a 1% error rate, it’s probably closer to 10% on the low end. Which humans can be far better than in terms of just directly following the assigned task but does not factor in how people can adapt and problem solve. Most minor issues real people have can be solved without much of a fuss because of that. Meanwhile the Artificial Idiot can’t even draw a full wine glass so good luck getting it to fix its own mistake on something important.
Which humans can be far better than in terms of just directly following the assigned task but does not factor in how people can adapt and problem solve.
How's that annoying meme go? Tell me that you've never been a middle manager without telling me that you've never been a middle manager?
You can keep pulling numbers out of your bum to argue that AI is worse. That just creates a simple bar to follow because... most workers REALLY are incompetent (now, how much of that has to do with being overworked and underpaid during late stage capitalism is a related discussion...). So all "AI Companies" have to do is beat ridiculously low metrics.
Or we can acknowledge the real problem. "AI" is already a "better worker" than the vast majority of entry level positions (and that includes title inflation). We can either choose not to use it (fat chance) or we can acknowledge that we are looking at a fundamental shift in what employment is. And we can also realize that not hiring and training those entry level goobers is how you never have anyone who can actually "manage" the AI workers.
how you never have anyone who can actually “manage” the AI workers.
You just use other AI to manage those worker AI. Experiments do show that having different instances of AI/LLM, each with an assigned role like manager, designer, coding or quality checks, perform pretty good working together. But that was with small stuff. I haven't seen anyone wiling to test with complex products.
I've seen those demos and they are very much staged publicity.
The reality is that the vast majority of those roles would be baked into the initial request. And the reality of THAT is the same as managing a team of newbies and "rock star" developers with title inflation: Your SDLC is such that you totally trust your team. The reality is that you spend most of your day monitoring them and are ready to "ask a stupid question" if you figured out they broke main while you were skimming the MRs in between meetings. Or you are "just checking in to let you know this guy is the best" if your sales team have a tendency to say complete and utter nonsense for a commission.
Design gets weird. Generally speaking, you can tell a team to "give me a mock-up of a modern shopping cart interface". That is true whether your team is one LLM or ten people under a UI/UX Engineer. And the reality is that you then need to actually look at that and possibly consult your SMEs to see if it is a good design or if it is the kind of nonsense the vast majority of UX Engineers make (some are amazing and focus on usability studies and scholarly articles. Most just rock vibes and copy Amazon...). Which, again, is not that different than an "AI".
So, for the forseeable future: "Management" and designers are still needed. "AI" is ridiculously good at doing the entry level jobs (and reddit will never acknowledge that "just give me a bunch of jira tickets with properly defined requirements and test cases" means they have an entry level job after 20 years of software engineering...). It isn't going to design a product or prioritize what features to work on. Over time, said prioritizing will likely be less "Okay ChatGPT. Implement smart scrolling" and more akin to labeling where people say "That is a good priority" or "That is a bad priority". But we are a long way off from that.
But... that is why it is important to stop with the bullshit "AI can't draw feet, ha ha ha" and focus more on the reality of what is going to happen to labor both short and long term.
The error rate for human employees for the kind of errors AI makes is much, much lower. Humans make mistakes that are close to the intended task and have very little chance of being completely different. AI does the latter all the time.
