- cross-posted to:
- technology
- cross-posted to:
- technology
tildes score: 31 · tags: —
You must log in or register to comment.
Because normal people don’t even know what they are and nothing actually helps them use agents in a simple way.
(I would leave a more substantial comment but the article is behind a paywall)
score 12 · canonical
(recommend reader mode to dodge the paywall)
Hot take … isn’t it kinda crazy that nobody is really using AI Agents,” Josh Miller, CEO of The Browser Company, wrote in a viral X post this week. “Theoretically, the tech is ready for AI agents to totally transform how we work and live our lives … but alas the general public dgaf” (doesn’t give a fuck).
I just have not heard a single person outside of the tech community talk about an agent that they use,” Miller says. “As excited and optimistic as we as an industry may be about the frontier and recursive self-improvement, it’s worth pausing and just saying, ‘Huh, what?’”
Look I just… How are you so out of touch industry?
Even if I set aside absolutely everything I dislike about AI outside of the actual product: No one is teaching people how to do that, or why to do that, or how it would help them. They’re just putting “AI features” into products and selling them conversations with copy written characters.
Miller’s arguments are convenient coming from someone who sells an AI-powered web browser. But I think he’s correct that today’s agentic products are largely not designed with consumers in mind.
Miller blames the problem on groupthink in the AI industry. Many people who are building these technologies are totally obsessed with them, and they often have a particular sci-fi vision for how the products should look and feel.
I don’t even think the sci-fi vibe is necessarily the problem, it’s that the tools don’t live up to the vibe.
score 36 · canonical
This seems like one of the biggest issues in the “AI Boom(/Bubble)” and it’s super high evaluations/investment.
I also think the gap between the promise of “anyone can use AI to do tons of amazing things for you” proponents talk about and the reality of the complexity of actually setting up more advanced AI Agents/harnesses. Let alone using them in any kind of effective way.
I think he’s right that in order to get any kind of broad adoption it needs to get WAY simpler for ordinary non-tech people to be able to use. The problem is by the time you simplify and “dumb down” the tooling enough to be usable by an average non-technical Joe it’s lost nearly all its usefulness. You’re at Microsoft levels of AI tools that can’t even do what the “suggested prompt” says.
I don’t think the tech industry is conscious of how much their existing skillsets lead them to being able to ‘effectively’ use these tools. How to be specific and clear enough in prompts to get results even close to what you wanted. How to review the output to see if it actually did what you asked and re-prompt for changes where it didn’t.
The more autonomous it is the more it will get things horribly wrong and make catastrophic mistakes which leads to bad PR and even more people realizing how dangerous it can be running these tools without proper isolation and security practices. The more autonomous and long running an agent is the more important a good highly detailed prompt is, and even then there’s a very good chance it’s still going to get lots wrong.
Then we’re back at needing to teach people very technical skills to use them, which defeats the whole idea of broad appeal.
None of this is to mention all the other issues with AI.
score 9 · canonical
Beyond the lack of teaching, beyond the gap between promise and capabilities, I genuinely don’t know how I could integrate agentic AI into my day to day life, like at all.
I make light use of AI tools at work in specific cases where it helps me or saves time. I know and understand the tools. I ‘get’ what an agent is.
Now, here’s a smattering of things I’ve done in the past few days outside of work: read a comic book, played video games, went to a local block party, ate out at a new restaurant with my wife, watched TV, took pictures with my camera. How could agentic AI help me with any of these activities whatsoever? And if there is any possible ‘use-case’ for me, how is it any better than using the working brain I have in my head right now?
Truly, genuinely, beyond my qualms about the AI industry, I can’t think of many (any) reasons to use the product outside of my narrow work-related usage.
score 28 · canonical
I could spend $20 a month to have a LLM Agent schedule my doctor appointment for me, which I do approximately once a year.
Or I could spend $0 and schedule it myself?
I guess I could use it to manage my email inbox??? Except I get approximately 1 email a day I care about and spend exactly 5 minutes in the morning checking my email?
I don’t know what the “value-add” here is.
If the LLM agent could do the dishes or run to the grocery store for me, sure. (I guess it could handle creating a grocery list or ordering for delivery or pickup…except I already build my grocery list with 3 clicks in my recipe app, and I’d rather get my groceries myself so I don’t end up with weird substitutions or bad produce).
score 15 · canonical
I can’t find many if any uses at work myself and like you, not so much at home either. Maybe there are amazing things it could do for me, setting aside my other issues with it, but I’m a fairly techy person for a normy. I am comfortable learning new software and apps (except Instagram, idk I hate it) and can adjust my settings and solve most of my own problems.
But i mostly find interactions with AI frustrating and maybe that’s because I can search for sources on my own and educate myself on topics and fix most of my own problems.
I need a robot handyman (and like not a humanoid weird one, a functional one) more than I can think I need an agent. And I think I even know sort of what an agent is.
score 12 · canonical
Personal software I think is a good use case for home. Maybe you already know you could write a script to do a thing, but it’s not worth the time or effort to figure it out? Now you can have AI do it for you, and if you don’t like it, delete it without feeling like you wasted any time.
For example I now have a hugo site on my server (i do some self hosting) that is a personal wuthering waves wiki. Previously I had some notes for stuff in silverbullet (a markdown based wiki), but it was growing unwieldy and I couldn’t have any nice layouts or designs without a lot of effort. I started with having the site build out a rotations (combos for each character) with a nice markup for buttons to push and colors and some variations. It grew into I can now go use the official wuwa discord bot that makes images of your characters, download those to a directory, and run a script which will scrape out the data from that (thanks to self hosted back end that is open sourced and maintained by some card generation sites). It will scrape from a community run google sheet of stats to get stat targets, it will diff how far off I am from targets on my character, then give an overall todo list page ordered by what percentage they are away from their stat targets. So now when I play i can bring up a nice layout of the characters and their rotations and check a todo list of which person to grind out. I also have a high level overview table to help me organize who would go on a team based on the things that they do, something I tend to forget.
All of this information exists in one place or another online, sure, but this is also tailored to my characters and the way that I play, and information that I need, all of it built while I was playing crimson desert over the weekend, occasionally tabbing out.
I also used claude to help me set up hosting actual budget, while it was easy to set up the software, it took a bit to figure out how to get it all set up the way i want, and asking the AI questions was much easier than going and reading through the documentation. I set up data scrapping for my information, however I still needed to set up a bunch of rules for organizing things into categories and reconciling payments etc. It also helped me set up a local llm on the server that would run (no graphics card + a lot of ram on the server) so that once a day at night it will try to categorize incoming scrapes etc. I would never have set this up if I didn’t have the AI going out and finding answers to my questions so quickly and being able to talk to it so naturally.
I have done a lot of other random things at home with AI, and a lot of times it sits there doing nothing, but when i have an idea it’s as simple as starting up a terminal going to claude and typing in the idea, then going back to what i was doing.
I think more and more we will be able to write custom software that fits exactly any need we have, even if it’s a one shot task. I am honestly hoping they keep pushing the local models to the point that the AI industry busts wide open and LLMs are just a thing we can have on our PCs.
score 1 · canonical
better than using the working brain I have in my head right now
Its not. However the world really is not ready for AI. I mean in a socioeconomic sense. That is really what is slowing adoption. Why would anyone replace themselves as an employee? Why would the CEO on the country club/beach/golf course want to come into the office to supervise a bunch of AI when they can delegate to other humans from their private jet?
Most of the world is not designed for LLM controlled automation. In fact its explicitly not because it was not considered a possibility and most systems are strictly designed around a human in the loop as a requirement. As time goes on companies that move away from this will gain market share and everyone will have to follow suite.
Another issue is that pay/work ratio is already terrible. What AI proposes is not that your brain can do something else. That it can do the work of 20-50 employees by your brain essentially acting as supervisor to an automated system. However this causes the obvious question. If I can run a company as a single employee, why am I working for someone else when the pay is $hit?
Additionally people are using AI incredibly ineffectively right now. It requires rather deep computer knowledge to make it useful at this time. That will change slowly.
score 7 · canonical
I make light use of AI tools at work in specific cases where it helps me or saves time.
And that, right there, is the goldmine they’ve oversold.
Having the ability in excel to say “hey i’m trying to do this kind of math/formula, help me” in plain text is WILD. That’s such a huge advancement in helping people learn their tools. I’ve yet to work anywhere that has anywhere near a full grasp of just their email/text/sheets suite, let alone more complex stuff, and these “AI” tools can 100% help with that.
The issue is that doesn’t require a couple trillion in data center build out so now we’re on “well obviously it can do everything” which is extra hilarious because even if it can, i probably only need it once. Then i can just setup the pipeline and leave it be.
One of the huge potential upsides i see of AI is helping people standardize the edges of their system (higher quality and more existing APIs) so that others can hook up and self serve, but that’s a one and done kind of problem.
score 22 · canonical
I fully agree. For context, I work in this area and did a masters in it.
I have been building tools, and looking into integration of these AI agents, and in the end, defining a MCP for your tool, letting claude do the tedious work of reading the manual and answering your questions in simple terms is just great.
And yes some people let it replace their own thinking - that’s a problem. But once you see it as a tool, and it’s been getting quite capable - it’s not 100% perfect - and I think it sucks at creative work. But that’s also because I recognize AI texts and designs so quickly.
But just using AI as an extra “smart” layer above a userface can be quite nice. Don’t try to make it replace staff, make it help your users, and it can be great.
Some usecases I have built myself are: quick automation of repetetive tasks in an UI, or “turn my verbal description into a JSON” (in my space sim luna). I have tons of users under 15, none of them will write JSON to customize their system. But they can say: I want a black hole and 2 stars around it - and then the AI writes the JSON
score 7 · canonical
played video games
Anything you didn’t like in those games ? You can make a custom mod.
watched TV
You’re TV is showing you ads ? try make a custom app that blocks them (not sure if possible…)
took pictures with my camera
If you’re processing them on your computer there’s plenty you can do there too.
Not saying these are great, but they are things that could be useful. Essentially you are now an expert hacker and can tweak any computer system to your liking (exaggerating a tad…).
score 4 · canonical
These are novelties. I could add crystal light to all the water I drink to make it ✨less boring✨. It could be useful for the water to taste like lemonade. But I don’t because it’s not.
score 6 · canonical
I had a good chuckle reading these because they are adding more things to manage to your life, not less.
score 10 · canonical
Anything you didn’t like in those games ? You can make a custom mod
I mod old video games as one of my hobbies and I can promise you that the current models are hot garbage at this, and I know what I am doing. The average user cannot get this to work.
You’re TV is showing you ads ? try make a custom app that blocks them (not sure if possible…)
Either impossible or an already solved problem, depending on the way it is delivered. Not sure why you would want to reinvent this wheel.
score 15 · canonical
I think for throw away mods it can be useful. My friend is not a programmer but has already made several mods for games with custom overlays or information dumps that were great. This would have been unreachable for us in the past, and I 100% percent agree, the AI is never going to build one of those general mod frameworks people include in all their mods and all that, and it varies by game, but it opens the door, to it.
score 0 · canonical
Just for fun I gave it a try, I’ve added a Super Saiyan mode to elden ring, you can press a key combination (which also makes you drink…), you’re weapon gets on fire for 60s and you have a different moveset.
To be clear that’s reusing assets that are already there (the moves are from Sekiro), but it wrote a bunch of Lua to use them together in a new way. Maybe it still qualifies as garbage, but hot garbage I think not ;)
It’s true that there’s already a adblocker with Claude as contributor, so maybe you don’t need to do it yourself.
score 1 · canonical
Elden Ring is a game that loads lua scripts which are super easy to manipulate, I assumed we were speaking about arbitrary games. Like Claude is utterly incapable of adding code to our Mario Kart 64 toolset (despite several people trying).
I suspect it would also be capable of using Cheat Engine to create pseudo-mods where that one works.
score 4 · canonical
Did the people trying have a background in Software engineering and used Claude before?
score 2 · canonical
The ones doing Mario Kart 64 stuff?
Yeah, at least in as much as they have successfully made some smaller mods themselves. The big issue is that it isn’t fully decompiled (at least our toolkit, spaghettikart is fully decomped I think) so it hits walls and doesn’t know what to do when it hits them.
No idea if they do it professionally. This whole topic is about “normal people” though so that shouldn’t have any bearing.
score 0 · canonical
To be clear there’s two questions in this debate :
- Would normal people have a use case for agents, assuming they work ?
- Does the agents work well enough at the moment ?
For 1) I think there’s probably a decent number of people (not everybody but the more curious crowd) that would enjoy e.g. making mods, hacking stuff, etc. This already happens to a large extend but one still need require a little bit of prior know-how. That said I think it’s going to remain a niche market, most people don’t want to build their own things from scratch.
For 2) I don’t think the tech is quite ready for lay person, it doesn’t always work, does stupid things, and generally the user still need to know what’s up. That said I made an Android app with Claude with pretty much no knowledge of the process, it was a pain (mostly transfer issues from computer to phone) but not exactly complicated.
score 0 · canonical
A coworker shared a very Claude description on how they solved a problem with a vibecoded app.
Turns out they basically just re-invented MPD with a lack of client compatibility and decades of refinement.
Which is neat…but could have been solved with one google search for ‘music player server.’ And not burning several hundred more kwh.
score 6 · canonical
My whole view on this whole thing is that those pushing these things “features” fundamentally don’t understand that two things can be true at once: The tech can be incredibly impressive in terms of what it is actually able to do, while at the same time, not being useful or wanted by the general public.
Like imagine if a company came out with a toaster that can scan the toast and laser toast everything evenly at a molecular level. It costs $1,000 base cost, $50/mo subscription to use, takes 5x as long to toast the toast, requires you to hand over all of your unrelated data so it can “curate” your toasting experience to your exact specifications, and uses more electricity in 10 minutes than my vacuum would in an hour. Like sure, from a technical perspective, that’s probably extremely impressive it can do that to the toast on a molecular level. It probably does provide a superior toast at the end of the day. Is it worth all that extra cost and extra time for a vast majority of people? No. There’s already a very simple solution available for a fraction of the cost. We don’t need to overengineer this.
I know there are actually legitimate use cases for a lot of the AI stuff but the above example is how it feels seeing all of the AI stuff crammed into consumer items that it has absolutely no business being part of.
score 16 · canonical
I felt the pain of the subscription toaster immediately since my laundry machines also want me to subscribe. Why? I have no idea what I would gain from it. This is an excellent example
But yes, I need functional toast not stupidly expensive toast
score 10 · canonical
Reminds me of when everything was being implemented on a Blockchain. Just because your tech can do everything doesn’t mean anyone wants it to, and it definitely doesn’t mean it’s holistically better than existing solutions. That’s without even mentioning the barely-concealed intent to price gouge once everyone does switch.
score 1 · canonical
This is nothing like blockchain. This is genuinely useful technology that is already changing the very substrate of the digital economy - anything software or internet related is affected and everything is changing. Development of the software, analysis of the requirements, testing and security testing and many more fields will never be the same. With blockchain, you had to be very creative and inventive to find a way of how to use the most expensive and bloated database system ever in your solution. With AI, you have no choice not to. It is being used.
All that said:
The issue is… No, the first issue is: none of this is impressive for a regular person. It will not help him in his work or leisure.
Another issue is: tech bros and grifters. The “AI guys” have no business plan and try to balance their numbers by pretending every single person on earth is their potential customer.
The grifters have jumped from the previous “big thing” to the next big thing and are trying the same old scams to get rich quick.
Add “doom trolling” by the big guys where they try to scare everyone by saying that their technology is so dangerous that it will either kill everyone or at least make them unemployed.
Of course, the only reaction to be expected is animosity or hate towards the technology.
But it is NOT useless. Unless the apocalypse comes (which it may), the AI bubble will possibly burst but the AI technology will stay forever and change forever everything that is digital-related.
score 0 · canonical
I tried pewdiepie’s lil self hosted UI thing a month after he released it and even then, I guess it didn’t detect my amd GPU because everything I tried, it kept running on my CPU and I eventually gave up and uninstalled it. I’m not sure what I’d use an AI at home for.
I use copilot at work sometimes to make it do VBA or mvc bootstrap because I dislike both of those and it does a good enough job, much faster than I would for sure.
What I find annoying is spaces where I feel like it could be used but clearly isn’t. I drive on the same highways EVERY time I drive home from a certain frequent direction. Google maps ALWAYS wants me to take the first exit I can and go down some shitty side road that claims to be 1 minute faster. That side road has like 10 stop signs where as the route I always take is just another highway with 2 stop lights that are green more often than not but maybe slightly more distance miles wise. Why can google not recognize that I took that side path once and literally never again and stop telling me to go that way.
Another, is there a general android interface ai yet? Is there something I can say “hey phone play [specific album] in musicolet” (my preferred music application) I know this used to be a thing when google music and you had uploaded things to it. Instead of posting not knowing I tried it. I reenabled gemini and it said it couldn’t. Useless. Absolutely useless.
score 3 · canonical
The Google maps thing drives me nuts, too. That doesn’t even need LLMs, this is just a missing feature.
I often select a route to specifically avoid certain roads at certain times. Even if they’re faster, the traffic makes it a miserable drive, and I’m willing to pay the 10 minutes. But when I get to the branching point, it “helpfully” says “I found a faster route!” and automatically tries to re-route me, which I then have to reject. While driving. Bad design.
It should just be able to take in some settings for “This road is bad, don’t use it unless it’s a 30 minute diff”, or other basic stuff. “Avoid highways” is too heavy-handed.
Would also be nice to say “start navigating when we get to this point”. I don’t need turn-by-turn out of the neighborhood, and it is annoying when it interrupts what I’m listening to. But I’m obviously not going to input my destination while driving to delay manually either.
Lots of opportunity for slightly more intelligence.
score 8 · canonical
Also nice would be a “navigate until here” cuz I I know my way home once on highway but if I have ventured off somewhere new its the offhighway multiple turns in town to get back to highway that I don’t know.
score 3 · canonical
To be fair this is mainly due to the fact that google maps does not actually calculate your destination fully. They are cheating by simplifying the pathways by looking at popularly used roads and try to balance it by traffic/usage. This video explains it much more in detail if you’re willing to check out.
From my personal experience, old tomtoms/garvins from 15 years ago did better on this aspect (this is not to say their current offerings are bad. I just have no experience with the newer gps devices). Sure, they were requesting paths that looked better on the paper but if you had a deviation that you kept repeating, they’d start suggesting you those roads instead. Although I do not know if those devices actually did brute force calculating each section unlike today’s apps. I can tell you that our devices certainly did not rely on an always online connection so they had to do all that work on board.
score 0 · canonical
If it helps, Waze (owned by Google and sharing some data) typically defaults to “your usual route”. Plus you can have an Orc narrate your trip and they added a “less chatty” mode where it doesn’t give you every direction, but doesn’t only give you alerts like the Alerts Only mode.
But I get that, I also think that what seems obvious to us - send me the route I took last time - is not how every driver operates. I take the slower route home sometimes to change scenery, drive slower, etc. but theres not a good way to tell it to “take the back road unless it’s super late at night or if the corn is high at the intersection, if they planted beans this year disregard.” So I just let it recalculate when I turn.
But yeah I get the absolute frustration of it
score 4 · canonical
That last comment is the most Illinois statement on driving around I’ve ever heard, thanks for the laugh.
/noise
score 0 · canonical
I go to a weekly evening class, and I was surprised that several non-technical people there chat enthusiastically about their use of ChatGPT as a sounding board for things like—make-up tips, sanity checks/second opinions, body doubling, search engine.
score 0 · canonical
I use them every single day for a variety of tasks, but the tasks I work on typically have some way for me or the agent to get rapid feedback on accuracy. I suspect most people don’t have that luxury. When programming you can set up various feedback signals (compiler errors, test suites) that can automatically get fed into the LLM as text which will put it back on track. And every time you do this you give the AI company training data so the next iteration is less likely to make the same mistake.
Yesterday I successfully did some vibe-CADing (I should just learn a real CAD program eventually, though). I was using OpenSCAD. In that case it’s not quite as good as normal programming, but as long as I give it small and precise steps I can visually evaluate each change for accuracy. In general having a computer visualize something for you gives you a superpower. And 3D modeling is inherently visual so it’s trivial there. But for any other computer work you can use LLMs to create visual feedback systems for what would normally need to be visualized internally.
score 5 · canonical
Same reason why every new thing is adapted slowly. “Normal people” are slow adopters and the early adopters are weird.
I remember a time when people who used cell phones in public were thought as being a bit eccentric.
score 0 · canonical
It still feels weird having someone talk to themselves like a schizophrenic while walking down the sidewalk.
score 0 · canonical
When it comes to AI, old people are more eager to adopt it than young people, which indicates the reasons for adopting it aren’t the same.
Old people tend to care less about the future than young people so environmentally destructive tech probably doesn’t feel as negative to them. Age also makes you more interested in something that lets you fake pro results but also takes away from your learning to do things yourself. Kids are smart for being cautious! And they can tell the difference between a crappy AI flyer vs. something that was made mindfully while many regular older folks can’t. Being a con-artist is just not cool like having a cell phone used to be.
(There are other uses for AI besides con-artistry, but for some reason the use-cases pushed by early AI adopters in regular people’s social spaces fall largely in that bucket.)
score 0 · canonical
When it comes to AI, old people are more eager to adopt it than young people, which indicates the reasons for adopting it aren’t the same.
Perhaps it’s because old people see death more clearly around the corner, which makes everything that allows doing anything faster valuable.
score 0 · canonical
I think the main driver for older folks is a sense of wonder. “You’re telling me I can generate a picture by telling the computer to make it?” Something that would take them hours to do instead takes 30-60 seconds.
score 0 · canonical
Speaking as one of those older folks with a modicum of IT wisdom, I won’t touch AI agents for personal use with a 10-meter pole, for the foreseeable future.
The glaring environmental costs, security issues, data harvesting, damage to web resources, corruption of truth, political power balance, impact on creators, and “first taste is free/cheap” risks of dependency and skill degradation, are absolutely relevant to me.
score 1 · canonical
Generalized comment here because the linked article is behind a paywall so I can’t read it…
AI is hyped to be the next big thing for everyone. However, AI (when asked a question) will provide information that cannot be trusted (as people have learned and as people are notified). While it may not matter much for an errant thought to know the true capital of California, it does seem like a waste of mental effort to consider a wrong answer as correct or to hear an answer which would require me to do research to verify. For those who want to “not think”, any response may be good enough to get by (and complete the test or homework assignment). For others, it seems like a waste of mental energy to ask something from something that will convincingly deliver incorrect information at random times.
When a person gets an impression about something, it can be hard for that impression to change. With people getting the impression that AI cannot be trusted, they would likely not use AI except in limited circumstances where the effort to monitor AI and research its’ output is less than the effort it would be to just do the task yourself. AI does not save you time with regards to your question if you have to then research the question to see if AI is even telling accurate information to you.
When compared with earlier digital assistants, AI seems to fall flat too. With other assistants, I had to say specific key words (such as “Play 90s Rock on Pandora”) to which the robotic assistant did what was asked. With AI, specific keywords seem to not matter as much and the responses from AI can be somewhat random as well. Rather than the old assistant way of “carry out my command”, the new AI is more like “Google search what to do and guess what I want”.
AI is impressive and does amazing things. Still, for most people, I imagine that they want their electronics to “do this command” and they want truthful answers to their trivia questions. With the impression that AI can’t do either consistently, it is more consistent to just not use AI and do things in a matter which you know will lead to consistent results.
Of course, this is all with regards to the general public (which is generally hard to make sweeping generalizations about). I imagine that sweeping generalizations such as this have to be a very difficult field of study and is something yearning for a statistician to come and take surveys for. Without that, you just have my (likely inaccurate) hypothesis as to why AI isn’t taking off as some people want it to.
score 3 · canonical
I like my agent. I’ve been using it for months now and I’ve built several custom connectors (MCP-style) and microservices to make it super useful for me personally. And in terms of quality of life, it’s absolutely worth it!
I also run a decommissioned enterprise server in my closet, know what Kubernetes is, have built several web applications, have successfully set up printing on Windows and seem to be blessed with an innate technopathy that makes computers and machines work better when I’m near them. This is not a joke, I actually believe this.
Obviously these “agentic systems” are completely out of the conceptual reach of the untarnished, the pure, the virgin non-computer-touchers, and no amount of marketing can change that.
That’s also not a bad thing, per se.
score 10 · canonical
For me, it’s a matter of trust.
I feel like I could benefit greatly from an AI agent that could read through my emails and sort through the junk to summarize what’s important and flag things like upcoming subscription renewals, appointment reminders, long ignored responses, reminders to reach out to friends that I don’t interact with often, etc. That said, unless it’s entirely offline and done on my own computer, I don’t and won’t trust it. I don’t trust the owners of these systems to have such a deep insight into my personal life, and I’m already considering how to move my email to a more private service for that very reason.
score 0 · canonical
Because for the past 20 years that annoying chatbot in the bottom right corner has been useless and on most websites it still is useless.
Because it costs a ton of money to integrate a not useless one like Claude into sharepoint sheets.
Because for a lot of people tech has been moving to fast to begin with and most everyone really just wants you to stop moving the buttons around every other week so they can get some work done
score 17 · canonical
Even if I genuinely saw a use case for agents outside development (where I am not even convinced) the insane amount of token usage for agentic use would already make it a non starter for me.
Broadly speaking, you either give an agent broad control to let it do what you want or you give it a very restricted set of options and things it can call on. The first option, from my personal testing, is just very expensive. The second option limits the added value for having an LLM in the loop that you might as well do it by hand or automate something the old fashioned way.
On my phone while commuting and my train is to arrive so I’ll have to cut this comment short. I wanted to reference some older comments I made, instead here is just the one https://tildes.net/~tech/1ua7/google/_search/_as/_you/_know/_it/_is/_over#comment-hx1q
score 10 · canonical
My employer will happily spend $200+ a month per head on AI stuff, but shirks at the idea of paying $500 a year for a piece of software that would easily be 3x more beneficial, but lacks AI as a feature.
score 0 · canonical
I’m guessing an IDE? What do they have you use? Or is it something else? I think an employer should consider letting their employee use up to 5% of their annual salary on tools if the tools are going to make them move faster. More than 5% if it can be defended. Most employers are too penny pinching on tooling.
score 0 · canonical
Toad for Oracle.
Although oof, their prices went up, esp with new AI editions.
I can still reasonably justify the base edition, but those high-end DBA editions are just getting out of hand.
DBeaver is catching up, but there are still a lot of rough edges for Oracle in particular.
Related: Just this afternoon, I burned $20 of extra credits in a single Opus 5 on medium prompt. It’s almost got the same ROI as a slot machine. Almost.
score 0 · canonical
Case and point as far as I am concerned. As far as answering your comment goes, if I wanted to implement deep research in the same way I have seen it work from google, anthropic and openAI then it would be in the hundreds for just token usage. In fact, you can easily see this if you ever used those research tools, you tend to hit your ratelimits rather quickly even on the expensive plans.
So in my mind there is no reasonable consumer agentic market as these things simply aren’t cheap.
score 4 · canonical
$800/month is less than one day of billable hours for software development work where I am. As long as the AI is improving your efficiency 10% you can easily justify the expense and pass it on to the client. Any business will see it the same.
The big AI companies charge a lot per token, but more efficient models (that are already running inference at a profit) like GLM 5.2 are still incredibly useful.
score 2 · canonical
I agree with your sentiment that for someone who is using the AI to generate income there’s math here that can make sense in some situations.
The article is discussing why “normal” people aren’t using AI Agents though. $800/mo would be massively expensive for a “normal” person to use at home.
Maybe someone wealthy who has staff at home could justify offsetting the $800 in staff for AI token spend, but that’s not most “normal” people.
The pricing would have to be extremely different for a consumer market vs business for it to make any kind of broad adoption to make economic sense.
score 2 · canonical
$800/month is too expensive for sure, but some of the Chinese models are pretty solid now and 1-2 orders of magnitude lower in pricing. That gets it to where $10-20/month is a feasible price, especially when you consider that the average person will use far, far fewer tokens than what I or @teaearlgraycold use.
The features themselves aren’t very compelling for a normal person yet, but stuff like Apple’s password reset feature for leaked credentials being agent-driven will become the norm. Truly consumer-focused experiences will eventually be built. Can we look at the technology and really think no one’s going to find a way to incorporate a natural-language interfaced computing environment? We’ve only just begun to build out the next generation of interfacing, but it’s been a dream for decades.
score 4 · canonical
I agree with most of what you’re saying, I think you hit the nail on the head with having to use cheaper open weight (likely Chinese - fwiw this is what I use at home for playing around in this space) models to save costs and that the features themselves aren’t there yet.
Not saying they definitely never will be, but these two things alone are huge for general adoption and broadly answer the question the article poses.
I think your example is likely the way this will go. Less “AI as a product” and more “This consumer facing product or service has some AI features now”. This will probably eventually move from “gimmicks” to features normal people actually want.
I think we’re still a very very long way off from “normal” people running agents to do useful tasks for themselves.
score 3 · canonical
You are not wrong, but I am just saying why the math doesn’t math for consumer agentic applications :)
score 5 · canonical
I wasn’t trying to debate anything. Genuinely just curious! I’ve only started looking at my usage numbers and was fascinated by it. I think it’s still possible for agentic use by normal consumers some day. The $20/month plan at OpenAI has given me a lot more usage than I’d think for personal use. My cousin sent me a screenshot of a Chinese model with 100 million tokens used for $0.85 (heavily cached). Apple’s building in agentic support in Safari, which would make basic use free for hundreds of millions. I see no reason the cheap end of the market can’t develop.
EDIT: I’m using Opus 4.6/4.8/5 and Sol. Both are stupidly expensive compared to the “open source” options.
score 2 · canonical
The $20/month plan at OpenAI has given me a lot more usage than I’d think for personal use.
The article makes a distinction between agentic use and chat use. For just general chat use, you are absolutely right. But, within chatGPT there is also agentic use in the form of the deep research functionality I mentioned. I can’t remember what limitations chatGPT put in place (I think it simply limits the amount of search you can do?) but with Anthropics version you are quickly told after a few uses that you have hit your usage limit with the $20 plan.
That’s just one agent, one they supposedly have optimized the heck out on their side and something all parties are already rate limiting.
I tried implementing it myself in various ways. I also did play around with versions other folks have attempted to build. Once you have to pay for the tokens, even with chinese models it becomes hella expensive. You still needs fairly top of the line models for starters (you can do some initial triaging with smaller quicker models) and to go through all data simply consumes a ton of tokens.
That is just one consumer application. The article starts out with this tidbit
“Hot take … isn’t it kinda crazy that nobody is really using AI Agents,” Josh Miller, CEO of The Browser Company, wrote in a viral X post this week. “Theoretically, the tech is ready for AI agents to totally transform how we work and live our lives …
Not all of the agents infused in the live Josh here envisions will be as token heavy as full blown search and research. But it is clear he’s envisioning a lot more of them. So left or right, shit is too expensive for the fairy tale of AI agents totally transforming how we work and live our lives.
score 3 · canonical
I don’t feel like you’re really disputing anything I said, but maybe I’m misunderstanding? I’ve been using the $20/month plan for agentic use, coding on the side at home for fun every few days. I’ve also been having it do deep research reports for me (“how much is my condo worth?”) once in a while. It’s been more than adequate for my use. The really long, detailed report about my condo which included a list of comparable properties and analysis on those used 10% of my weekly usage budget.
score 1 · canonical
I think that what you are describing still falls under very light consumer use as far as the Josh Miller and friends are concerned. They want you to use agents for ALL THE THINGS not just occasional software projects and occasional deep research.
I am not being facetious here either, I honestly don’t think that the usage you are describing even scratches the surface of what they’d want the usage to be. Or, what they need it to be to ever be able to get anything back from the insane VC cash infusion this entire market has gotten.
score 2 · canonical
Okay, maybe not, but who cares what Josh Miller thinks? If we don’t reach 24/7 use of agentic AI use, it doesn’t mean it can’t still be an important part of our lives. The dismissals of AI in this thread aren’t scoped to just what Josh Miller is saying, but agentic AI in general.
score 0 · canonical
Safari’s built in agentic support would be free at point of use, but in actuality the cost would be generalized across the cost of all Apple services, as Apple isn’t going to be paying it themselves out of the goodness of their heart. Unlike a lot of other, more specialized companies, Apple has a lot of other products and services where they can place the resulting expense of a “free” service to hide what’s happening. Besides, living in the Apple ecosystem isn’t cheap. It’s not likely to be replicated at the truly cheap end of the market.
score 6 · canonical
It depends on their agent setup but Safari exposes an MCP server and if the agent runs locally, it’s effectively “free”. Maybe it runs in PCC though. Still, as capabilities get better, why wouldn’t what runs in PCC today run on-device in X years?
score 1 · canonical
I don’t use Reddit, but I came across this comment of a guy in ChatGPT subreddit that burned through most of his 20$ monthly limits in a single day because his agent worked for 8 hours.
Made me cringe so hard. I’m guessing the subscription limits are pretty high, compared to what you would get by paying directly for the API.
score 0 · canonical
Honestly I’m banking on the cost killing the pressure to put AI features everywhere.
score 9 · canonical
They can always just use cheaper models. It’ll have to be legitimate reflection about wasted screen space, development time, etc. that removes junk AI features.
score 2 · canonical
Even the best models provide ‘meh’ answers more often than not.
Switching to cheaper models that are even just 5% worse will be quickly abandoned by anybody that has a reasonable choice to do so.
score 0 · canonical
Depends on the domain. For the magic chat bot that’s suppose to turn your website into a Star Trek experience - yeah those suck. I was recently kind of shocked that that Cloudflare chat bot doesn’t know how Cloudflare works. But I have been surprised on occasion when a chat bot sidebar pulls up the correct excerpt from their documentation.
score 0 · canonical
I’m really thinking about my work’s enterprise subscription and some of those higher level “pushes” for AI use but if the feature remains fairly annoying and useless and keeps giving bad answers (like many AI help bots on various social media sites and things) I think it won’t be given the priority in further development. Especially if those programmers can’t afford to vibe code it anymore.
But idk, it’s the only hope I have left from my leadership pushing AI daily.
score 2 · canonical
If I can’t boss an AI around to do work as good as mine and let it come back to me with a perfect answer with minimal questions, then it is a company-sponsered waste of time.
I told Open 5 on Ultracode, ostensibly “best of the best” I can use, to generate me a docker-compose file with a caddy proxy updating dns for two domains and three services.
It promptly declared that it was impossible to do DDNS with two domains at the same time on a single Caddy instance. Because it allegedly paresed through the source code of the relevant models. It’s top suggestion was to update the other domain using a cron job. I’m no stranger to a half-baked cron job, but also it brilliantly highlighted how bad these things are.
I’ve has it pin ancient versions of libraries because one time there was a regression for 5 minutes with the latest runtime. It yanked my Python version from 3.14 to 3.8 on the premise that there was a code regression that prevented installation (there wasn’t, it was a temporary dns failure).
score 0 · canonical
It sounds like a solution trying desperately to find a problem. I am genuinely struggling to find any good usecase for where AI agent could be really useful in my personal life. The examples we see from the tech gurus are always bonkers, like Altman creating a podcast of his kids’ interests. Instead of talking to them ?
Compared to just a few decades ago, our lives are already heavily optimized by digital solutions. Do we really need more ?
score 23 · canonical
Programmers be like: I have to do a thing, let me spend all day automating it so that I never have to do it again
Programmers a week later: spending time coding because other related use cases came up and code needs to be expanded
Normal people: do the thing
Normal people a week later: do the other thing
Tech companies are just out of touch. If theres setup required, most people will not do it unless they absolutely have to. None of this Ikea software shit. No setup required is what we want and all your new features require us to give you our whole life story.
score 17 · canonical
Programmers be like: I have to do a thing, let me spend all day automating it so that I never have to do it again
Programmers a week later: spending time coding because other related use cases came up and code needs to be expanded
Personally, I actually kind of enjoy this loop. ¯\_(ツ)_/¯
score 5 · canonical
Claude has kinda messed up my balance there cause it costs me only company provided tokens to automate the most stupid things
I think its super silly that the language models automated all the tasks that we don’t really mind doing but some dope still has to clean my toilet bowl and that dope is me today because Claude cant do it :(
score 1 · canonical
Tangentially related, as I actually found a nice use for stable diffusion models the other week. Generally, I think they’ve saturated the web with bland imagery. In short, I need to build a fence for my porch. So I took a picture and asked Gemini to add a fence. I did this a few times with various descriptions of the fence to get an idea what it would look like. Some were really bad, e.g. making the fence where the current stair is thus blocking the entrance, but some were nice examples. I could show them to the rest of the family and we could pick a design that seem nice. Doing it by hand would be impossible for me at least and hiring someone to draw them (photorealistically) would be prohibitively expensive.
Other than that in the case of image generation I’m at a loss.
score 4 · canonical
It took me soooooo bleeping long to come up with a series of similar icon style pictures to use for different groups in class that it legit would have been faster to do it myself. TBH, I was just trialing to see what it came up with, but the very, very, very clear problem is:
AI doesn’t actually “know”, what it is outputting. It just outputs a “thing”.
Say I asked it to make an image in 80s outrun style of a group of lemurs at a dance party and it came up with something, and I tell it to change the third lemur’s hat to a cowboy hat, it has ZERO idea of what “the third lemur” is, and it just randomly spits out more hallucinatory shit, and every iteration gets more and more freaky.
A human knows in a second exactly what you are saying and fixes it.
score 0 · canonical
That has been my general experience with diffusion models. In my cas with the porch I used a base photo and it fairly consistently was ble to add a fence to it in various styles. So I think this depends a bit on the model you use and probably on Gemini’s surrounding tooling.
score 0 · canonical