YC Partner Argues Most AI Apps Are Currently 'Horseless Carriages' (koomen.dev)
- Reference: 0177138869
- News link: https://it.slashdot.org/story/25/04/25/1545223/yc-partner-argues-most-ai-apps-are-currently-horseless-carriages
- Source link: https://koomen.dev/essays/horseless-carriages/
The critical flaw, according to Koomen, is that users cannot customize the system prompt -- the instructions that tell the AI how to behave. "When an LLM agent is acting on my behalf I should be allowed to teach it how to do that by editing the System Prompt," Koomen writes. Koomen suggests AI is actually better at reading and transforming text than generating it. His vision for truly useful AI email tools involves automating mundane work -- categorizing, prioritizing, and drafting contextual replies based on personalized rules -- rather than simply generating content from scratch. The essay argues that developers should build "agent builders" instead of agents, allowing users to teach AI systems their preferences and patterns.
[1] https://koomen.dev/essays/horseless-carriages/
That'll never fly (Score:3)
If the "end user" can see the prompt, let alone manipulate it, they will see that they're the product and not the other way around.
Setting up a bigger problem (Score:1)
Commercial AI co's won't be able to resist mining users' content if and when the product is allowed to learn from local content.
Faster Horses (Score:3)
There's a semi-famous but probably misattributed quote involving Henry Ford, "If I'd asked my clients what they wanted they'd have said 'faster horses'".
Koomen is making the same point 100+ years later. The transformative AI solution will probably not be faster horses, or faster humans. It will be some new construct that goes well beyond what's currently possible in ways that we almost certainly haven't posited yet. We're still in the late 19th Century horseless carriage mode of AI design nobody has a clue about fuel injection, anti-lock brakes, air bags, automatic transmissions, etc type advancements yet.
It will be blindingly obvious what the ultimate winning design is but not till probably 3 or 5 or more years after the fact, claims until then are mostly marketing FUD.
Re: (Score:2)
Not so sure we'll see a better metaphor, new stuff is scary, people cling to old ideas, old metaphors.
Just try to get someone to change from apple to android. Most people think it's impossible.
100 years have passed since the horseless carriage, aka the car, was invented... we're still driving cars that are only cosmetically different from the original horseless carriages.
Re: (Score:2)
This analogy would work better if the "clients" here were native Americans before the introduction of horses.
And no, Koomen isn't making the same point, he's just parroting something he's heard that he doesn't understand in hopes of establishing himself as an expert. He doesn't appear to actually know what an agent is, his criticism boils down to him not making his own agent but desiring to have one customized to his desires.
History repeats (Score:1)
> early automobiles that mimicked horse-drawn carriages rather than reimagining transportation.
Early military jets were similar: they mostly just slapped a jet engine on a propeller-based design, and often got lack-luster improvements and/or stability problems at higher speeds. Part of the problem is that they first needed stronger wind-tunnels and skilled test pilots to test designs, but such experts were tied up in here-and-now war issues.
Even the Me-262 had a stack of issues that were only slowly being re
Has anyone tested effort to fully specify logic (Score:2)
Koomen's piece suggests that a "better use of AI" would be to quickly create email filters. The thing is, many widely used email clients already has the exact functionality that he describes accessible through a GUI rather than a prompt. And while his tossed off email filter idea is a short AI prompt, it almost certainly fails to specify elements of logic that are needed for the filtering system to work properly.
My guess is that by the time you fully specific the necessarily logic, the LLM prompt will be e
Re: (Score:2)
I also do not see much use in the filter. Also do you notice how he tries to generate a text to his boss shorter than the (user) instruction "Write my boss ..."?
But now think about it used with a voice assistant. Then you wouldn't type "Hey garry, my daughter is sick -- pete" but you would either have to say
"Write an e-mail to Garry in the address book. Subject I can't come to work, text hey garry, my daughter is sick, I can't come to work dash dash pete"
or you say
"Write my boss that I can't come to work be
Re: (Score:2)
Correct. Koomen asserts that the problem with AI is that it doesn't perfectly fit into a world optimized around doing work without AI, that it doesn't fully align with his presuppositions influenced by applications designed around what can be done prior to the existence of AI. Worse yet, he assumes that AI is perfectly designed already, it's just that its knobs don't go to eleven.
Re: (Score:2)
There are a lot of things you don't need an LLM for that folks are trying to get LLMs to do. Same with ML. Everyone keeps asking "how can we use ML to solve this" and now "Can we make AI a part of this process?" and I'm like no. If you want to spend a whole lot of money and get nothing in return, just give me the money. We'll all be happier.
Why would I want that? (Score:2)
Why would I want my emails written by AI? That doesn't inform me nor let me communicate my thoughts. Crazy.
Re: (Score:2)
Venture capitalists don't want you communicating your thoughts, they want you accepting their thoughts. Musk isn't investing in AI for you, he's investing for him.
If anyone thinks the point of AI is to provide a handy assistant for an email client, they have already lost. This is who Koomen is, a loser not worth anyone's time.
who can take this guy seriously? (Score:2)
"The critical flaw, according to Koomen, is that users cannot customize the system prompt..."
That's the critical flaw? That's like saying the critical flaw in your new car is it's the wrong shade of silver.
He presumes the entire design, structure and function of current AI applications is correct, he just can't tweak it to his liking. Then to fix it:
"...developers should build "agent builders" instead of agents, allowing users to teach AI systems their preferences and patterns"
What a moron. An "agent" ag
metaphor people relate to (Score:1)
Remember the "information superhighway?"
We use metaphors to make the new stuff comprehensible.
Ironically, I doubt many people under a certain age understand the metaphor "horseless carriage".
or what a horse is.
or what a carriage is.
Ok, maybe they know what a horse is.
but ask little Johnnie to draw a carriage. I wonder what you'd get?