157 comments
That _is_ valuable, and will remain so, but the temptation to turn it off is there because the loop is good and getting better at what it does. Once you're in the loop, with the rare high-value exception of catching total mistakes nad redirecting, you're mostly choosing between similar-yet-reasonable options. This isn't so much right-vs-wrong as relativistic optimization. Letting the loop do the work will get you a mediocre result quickly, and that's usually fine.
The most important time to use your brain is _before the prompt_. Once you engage with your LLM and agent, you start biasing yourself, and its reasonable suggestions constrain your visibility into other options, other worlds.
First, use your brain. Then write the prompt.
This isn't a new problem. I had it when working with outsource developers, too. In both cases, it seemed that just sitting back and describing what I wanted in plain English prevented me from properly wrapping my head around the problem. Which then meant I failed to develop the insight needed to find the really good improvements. Those usually consisted of identifying and removing the bits that were making things worse.[1] It's incredibly difficult for me to think of anything other than just adding more stuff on top of what's already there if I'm not getting my own hands dirty.
So I've been moving away from plan mode and letting the spin away on building large chunks, and back toward working in small increments and typing out the important code with my own hands. That seems to be the sweet spot. I'm still getting a productivity boost because there's more than enough boilerplate and trivial function definitions to farm out. (Easy prompt, too: "Implement `someFunctionStub`.") But I'm also not letting Plan Mode drag me back into the waterfall development pit of insanity.
It's also unlikely that you're going to make those realizations, expressions, and decisions when your harness is asking you to choose between what _it_ identifies as being important decisions. If you're in the loop you're getting dumped on with lots of stuff that barely differentiates and you're biased away from catching things that matter.
If you're in the loop, use the loop. It's not a terrible construct. It will give you a mediocre outcome. And if mediocre is sufficient, you're done.
And here's the part that didn't make it into my prior comment -- if mediocre isn't sufficient, _stop the loop_. Play with what you've got. See how it behaves. Explore the edge cases yourself. Sketch how it works on paper, and how you want it to work, and think about how that should be constructed. Use your brain.
Then write a really long prompt (maybe referencing your files and notes and assets) to get back into the loop again.
I guess what I'm ultimately getting at is that the loop is useful _and_ a mind-killer, and using your mind is _also_ high value, so do both those things, but don't do them at the same time.
Your stance isn't new, if it's any comfort. I heard (and probably said) exactly the same thing when C compilers started getting good enough to eliminate the need for most assembly coding.
Even if I agree that you have to use your brain, when my boss is literally telling me to stop questioning whether AI automation is the right answer in team meetings/chats I'm just going to shut up and give them what they want.
Suppose I actually hire a professional to do it. I receive fewer of the benefits. However, he's much more likely to do a good job, and the legal system provides a variety of ways to hold him accountable for the work.
Suppose I hire a professional, but then insist that he use legions of guys from the park to do the actual work, since that's going to be some much faster and cheaper per quantum of work. I also tell him that if he's unwilling to do that, I'll find someone who will. I also tell him that, since he has so much more unoccupied time now that the original work is in the hands of all those other dudes, he gets to supervise still more work done by other swarms of guys from the park.
At this point it should be clear that I'm not interested in quality work, but in having someone poised to absorb blame. I want a patsy.
Now replace "guys at the park" with LLMs, and replace "free" with "nearly free, but with no warranty."
This is definitely a scenario in which management has bought into the hype and the people working with the tool are aware of its limitations.
Like markets, the firm can remain irrational longer than you can push against it.
Beyond a certain degree of effort on your part, it is not within your capacity to right the whole ship.
I suspect that the people who are most likely to be allies are the CFO and budgeting teams in the firm. I’d be curious what that function is thinking of the situation.
That means that they really don't care if, right there in the moment, just doing the job is faster and better. What they care about is that the actual work is being done by humans, and that's out of line with their grand vision of getting rid of humans.
Some would tell you that you need to find a different job with a better boss, but I'd say that this is quickly becoming something very hard to escape from. Investors want line to go up, and they all talk to each other about this "AI" thing that lets them cut staff and keep productivity, so naturally we're all getting it crammed down our throats.
No? You solved the symptom not the issue?
I sometimes suggest timeboxing investigations for this reason. Spend 4h on this - if root cause is not found, just throw more hardware at it. Although I expect a writeup of what was investigated and ruled out, so that it can be used when someone returns to it.
Then again, for many people their job security doesn't depend on what work they do, it depends on how well they are embedded in the organisation. Before AI you also already had plenty of people doing their job at a questionable level, yet never seem to get fired.
Sounds like a quiet few years and then effort once demanded. So I guess... being a meat proxy does work, after all?
> I started seeing people turn off their brain as they use LLMs
I mean, this was true for computers. "Computer says no"[0] was thing long, long before LLMs. I've had well educated, seemingly very intelligent people break when they couldn't log in steadfastly ignoring the [caps lock key is on] error message. This becomes even more "fun" and "interesting" when a person is being a meat proxy for some horrifically dangerous industrial process involving high pressures, excited chemicals, and enough rotational inertia to send god flying.
(In my lifetime, I've even watched mass cellular communication move "replacing your own tire" from skilled to deskilled, since AAA is always one phone call away for most drivers).
If you want to shoehorn AI into this analogy at all then no one is even bothering going hiking at this point, let alone seeing anything worth capturing. It would just vomit up endless permutations of statistically average images that look like the kind of pictures people take of landscapes.
You can be creative and take the photo for your sake but at the same time accept that you can generate 1000s of them and many of the generated ones will be more liked by the general public than your creative one.
In the past few could afford to be photographers due to time and money. Many were not the most artistic or talented minds in the universe on the subject (you can’t find those until you give everyone a camera) but just people who enjoyed it and could make money off of it. Those who gained experience with the old tech were essentially AI models trained on statistically good probabilities of getting good shot. You had smaller selection and the model / brain quality determined what % of the best that photo was between 0 and 100% for any sunset. In modern days more people get a camera - more people with bad taste and more people with good taste. You also get more attempts at the shot - more photos taken, more tries. More tries means you are now preserving more of that moment from more angles and extending the time your taste people have to pick the best shot of all possible.
Your plethora of statistically average photos happen if you don’t have anyone filtering the ginormously increased output. If the problem is really valuable, someone will filter it. Possibly a model.
The real filter is value, not capability anymore. Do you have the empathy and taste to know what matters? The AI tokens/cost lets everyone make and become interested in making of someone hasn’t solved their problem. And then if search isn’t improved - everyone will just repeat the making of the same thing over and over. So if you don’t want repeated slop. fix search.
Or do you copy everything and everyone into making more of nothing.
It worked for a time, and manufactures employed experienced, legendary even, watchmakers and specialized "watch setters" to create and tune to the best possible extent a few experimental watch movements. They would be sold as collection pieces, the firm would gain boasting rights in its advertising, and the experience informed the design and tuning of more mainline movements.
By the end of the history of (the revival* of) this concours, the winner was always Tissot, a second-rate company. What hides behind Tissot is ETA, a huge watch movement manufacturer for all of the Swatch group (dozens of brands, from Omega to the worst no-name stuff you can find).
What they were doing to win was just to pick the best watch movements out of the millions of basic, bland, uninteresting "workhorse" ones they were producing each year. Just by chance, there was always one combination of the hundreds of components, each with their manufacturing tolerances and variations, that would fit all together to make an exceptionally accurate movement.
In the end, yes, Tissot won repeatedly. But the concours is dead. And Tissot is still a second-rate brand making second rate watches. Nobody is interested by their watch movements, no collector, no historian, no customer.
The credulous LLM users Dan Luu is complaining about don't take the equivalent of thousands of photos of the same sunset. They take a couple photos at most, say "LGTM", and move on to taking a photo of something else, even if their couple of photos of the sunset aren't great.
As someone who will often compare different things, I think doing that usually takes time. I think LLMs can help reduce the time (especially if it's programming, not necessarily so much in other areas), but I'm not seeing many people use LLMs to try a large number of independent approaches and compare them.
It's not as if he was covering every combination possible of exposition, composition, depth of field, shutter speed etc. The tourist doesn't even know the existence of those settings.
If that were the case then photo competitions would have been upended by now by tourists who accidentally took masterpieces with their phones.
I can take as many thousands of pictures as I want but, without taking the time to critically review them to see what works and what doesn't, I will rarely produce a prize-worthy picture. And then there's the risk that I'll discard it anyway because I don't have the eye to separate the prize-winning picture from the Instagram shot all my contacts are publishing.
Read the full thread on Hacker News →
Related stories
- Lobsters · 48 points · 12 days ago
- Ars Technica · 0 points · 8 days ago
- Finding the cells that put our brain to sleeparstechnica.comArs Technica · 0 points · 12 days ago
- Ars Technica · 0 points · 12 days ago
- Hacker News · 1 points · 9 days ago
- Hacker News · 643 points · 12 days ago