LLM users: how do you feel about it?

23 points by LumaBop a day ago on tildes | 42 comments

teaearlgraycold | 23 hours ago

Have you ever stepped outside your area of expertise and realized how complicated the world is and how you'll never know even a fraction of a percent of what's needed to build the technologies your life depends on? Can you build a house, design a CPU (or any electronics?), grow your own food, build and repair a car, etc.? LLMs aren't going away. It's always beneficial to understand more than the bare minimum to accomplish a task. But we don't have the time to be extremely capable on many different things.

I took a number of calculus classes in high school and college and did reasonably well in them. I only remember the barest of details. High level stuff, when to use a derivative, when to use an integral, the chain rule, etc. I wouldn't be very good at calculus given a pencil and paper and a problem to solve. But I can still make use of calculus by heading to Wolfram Alpha and asking it to integrate some equation I have on hand. I'm not worried that I'm not as good with it as I used to be and have no intention of getting back up to speed. There are other things I'm more passionate about which I'm better at now than I ever have been before.

There's no reason LLMs won't be the same as Wolfram Alpha, or pocket calculators, or Google search, etc. They have their shortcomings. Until every human is completely unusable for jobs that don't require a physical body there will be some thinking of value for us to do. Find where things stop being easy. Sprint there with the help of AI and then get back to work.

raze2012 | 13 hours ago

an you build a house, design a CPU (or any electronics?), grow your own food, build and repair a car, etc.?

I think this line of thinking is exactly he danger. I can't build a house, but it's still useful understanding when parts break down and knowing how to do basic repairs before turning to experts. e.g. unclogging a toilet before paying a plumber 3-4 figures for service.

LLM's in some ways feels like throwing a plumber at a dirty toilet that just needs a scrub. I don't know what everyone's lines are, but there is a legitimate danger to losing a lot of your own autonomy to technology.

Until every human is completely unusable for jobs that don't require a physical body there will be some thinking of value for us to do.

Every day I start to realize that WALL-E was a premonition, and not an excessive in reducto ad absurdum for environmentalism.

teaearlgraycold | 13 hours ago

My point is you only need the skills that you need. And you will find that out which skills those are in the course of living and working. If it was likely for people to rapidly lose access to LLMs then it wouldn't be safe to delegate tasks to them too often. But I think given the robust availability from many vendors that's like saying you shouldn't have a lifestyle that depends on driving a car because you never know if cars might just go away one day. Or bemoaning someone that can't survive in the woods with just a hatchet.

that's like saying you shouldn't have a lifestyle that depends on driving a car because you never know if cars might just go away one day

I've been intentional my whole life about not getting into a car-dependent lifestyle. Not because I believe cars are going away or because I don't like driving (I do, especially on clear roads), but because depending on them has many subtle but serious repercussions.

More than half of the households in my city don't own a car. The rest are either car-dependent against their will, or just don't recognise / care about the downsides. Getting a car if you don't yet have one is a crossroads moment because as soon as you start leaning into the comfort and ease of the lifestyle it can enable, it becomes harder to avoid the lock-in that comes with that, so reversing the decision later may prove hard even if you wanted to.

teaearlgraycold | 10 hours ago

This feels like a fairly distant tangent, but I also don't own a car. I had one for a few years, moved to a city, and was able to sell it with no trouble since.

raze2012 | an hour ago

you only need the skills that you need

  1. And in your twisty curvy trail of life, how do you know what you need?

  2. There are very much things people needed to learn long ago but never did. Even before AI, there were worries of Gen Z computer literacy falling due to being raised around tablets. We aren't particularly good at understanding what is needed tribal knowledge and basic survival skills to pass down compared to what is pleasurable and enjoyable in the now.

  3. More than anything, I believe the capacity to learn is the single most important skill out there. Being willing to dive into the unknown and come out with new skills, talents, and passions is a core part of the human experience. Having those elements outsourced to some machine is a bit dire.

If it was likely for people to rapidly lose access to LLMs then it wouldn't be safe to delegate tasks to them too often

We may in fact be seeing that sooner than later. Are people ready to learn how to spin up a local model on their machine? Do they have the hardware to allow for that? These don't seem to be a given, given the context above. I personally have strong "If you're nothing without the suit, you don't deserve it" mentality; that's how society wants to judge us in the near future after all.

that's like saying you shouldn't have a lifestyle that depends on driving a car because you never know if cars might just go away one day. Or bemoaning someone that can't survive in the woods with just a hatchet.

Yet I know how to walk, ride a bike, read maps (analog and digital) and navigate public transit should I not have access to a car. As I have for some months here and there. Exactly my point.

On the other end, I would not survive in the woods with a hatchet. A hatchet does not show that I know how to cut down trees safely, how to gather and start a fire, nor how to hunt game. Those are all skills to develop that a hatchet enables. I'm not getting that kind of vibe out of LLM's.

I liken it more to CPR. Sure, nobody needs to know how to do it. So most people don't bother to learn. And as a result, more people die.

In this case, this manifests in a situation where in 50 years a cascading cloud failure results in the LLMs getting shutoff and nobody has the expertise to turn it back on again.

We already live in a world where if one person makes a DNS configuration mistake in the wrong place, half the internet goes down for a day.

Chiasmic | 6 hours ago

A tangent but healthcare staff definitely do need to know to do it :p

They have better options at their disposal.

Chiasmic | 3 hours ago

I’m not sure what you are referring to. If someone’s heart stops, people absolutely do CPR. That includes advanced life support, but that is still just advanced CPR. But this is a tangent and not really relevant to the main conversation.

So, I'm on a hefty plan from work, Claude with Fable + Opus are my defaults.

I feel it eating my brain.

Because it is good. So you'll start out disciplined. But then you need to rush just one little thing out, so you pop a half-baked thing in one time and then 5 weeks later you're a full-on vibecoder and hating yourself the whole time, but because the difficulty threshold of digging in and fixing the Gordian Knot you created is so high, you just keep doing it like a junkie.

It's really good. At first. But as it iterates, and you are overworked/tired/lazy (depending on day), it slowly decends into madness: Worthless test suites that take an hour to run. Mis-mashed contradictory memories which will bork a project with fresh context. Overly verbose logs, statuses.

The analysis it provides will be good like 80% of the time, which lures you into a false sense of security and then when it hits that 20% you're either caught with your pants down if you haven't been paying attention or you've gotta hammer the stupid out of it for an hour.

I'm doing data migration work. It's really great. I've definitely been able to churn through the job 10x faster than I ever thought possible. But then I'm also having to deal with all the actual hard parts of the data migration in a span of hours instead of days and it's burning me out. Turns out I needed to have that 2-day turnaround on gruntwork changes in order to keep the dark thoughts at bay... I'm actually fine but I have to remind myself that other people might not find jokes about minty-fresh shotgun barrels funny.

And Claude fucking sucks at databases. You can hammer it into doing it, but it is so brutally annoying.

They're immensely useful tools. But I'm witnessing the brain drain firsthand, and other coworkers describing the same feelings. Maybe someday there will be an equilibrium, but I am doubtful.

And as someone on Masto said: This is progress, but progress for whom?

Can I just say, and I try not to be offensive with this but ... PEBCAK?

If you feel it eating your brain, use it differently.

Personally, the way I interact with Claude, I learn all the time. I see it "fly and dance" as Levent put so well, and I watch as it spots interesting issues and we build creative solutions together. I see it make choices I wouldn't make, and sometimes I disagree with them and nudge it, sometimes I am surprised and learn from it.

There are things it cannot do? Teach it to do them! Brainstorm with it what a structural, systemic solution looks like. The 20% isn't a "false sense of security, pants-down bit it cannot do". It's a symptom of expecting it to do everything right from too little information. It's a compression problem, as Grant Sanderson put so well. Too little to infer from. Feed it better.

But "Feed it better" can just mean "build better guardrails". And you know what you can build guardrails with? AI! It's so easy and so fascinating too. You say Claude sucks at databases, but on our end we have it deal with production data directly with the same careful permissions a developer would have, and it works very well with it, because we taught it to teach itself to work with that data.

Why do you hate yourself doing this? We are living in an incredible future. You can literally wish things into existence. You want to have a more interesting time doing this? You can actually /ask for this/.

I run an AI lab. My days are filled with exploration into this super weird future. It's so fun and interesting to explore what it means for intelligence to be made of software.

I won't deny that I'm necessarily part of the problem.

But I am aware that I can brute force it into doing things. I taught it to do a great many things.

But it feels soul-sucking and bad the way that a master knitter might feel bad sitting in front of an automatic knitting machine checking for errors instead of knitting for themselves.

I derive no joy from management. It's not working with a partner. It's chitchatting with your drill press.

smithsonian | 2 hours ago

Mis-mashed contradictory memories which will bork a project with fresh context.

Ugh, I feel this one. I started doing memory audits on projects every couple of weeks (depending on how active it is) to clean out stale/extraneous info, find and fix contradictions, etc.

My biggest gripe is when Claude makes a choice that I don't have a problem with but it somehow ends up being treated as a commandment (e.g., "we won't use the database for this" becoming "we can't change the database at all") and it only surfaces when I notice unusual choices being made ("Why are you doing [convoluted thing] instead of just putting it in the database?" ... "What?! I never said we can't touch the database.").

I was using Claude to help me train a custom neutral net model for reading the LCD digits off my electric meter. For the first batch, we had to add an image augmentation step to simulate low-light/nighttime images because of the limited samples in the first batch. It wasn't until the months later when working on the 4th iteration—still trying to improve nighttime accuracy—when Claude mentioned something about the Augmentation step and I had to ask why the hell we were still doing that. I had it run the augmentation on a handful of night captures and found it was basically making the actual nighttime training samples so dark they were effectively useless for training. Once we dropped the "simulate lowlight samples" from the augmentation step, the model's nighttime accuracy went from ~40% to 99.5% and a handful of other weird results/behaviors disappeared. >⁠.⁠<

So "let's audit previous assumptions, steps, and choices to make sure they still make sense and are relevant" has also become a regular periodic health check I've learned to start doing.

 

Adding things like "Clearly state any assumptions you make" and "delineate between my explicit/implicit and your assumed choices in memories, along with the reasoning for the choice so they can be reevaluated if circumstances change" to my CLAUDE.md has helped.

I also find that always starting with planning mode tends to produce much better results, even if I don't have anything to change in the proposed plan. Specifying a desired complexity/simplicity vs. reusability/extensibility up front can help reign in some of the over-abstraction tenancies, too.

But yeah, I think part of the problem is that they are so good—especially when well-prompted—that we I start getting lazier and lazier with my prompts until surprised_pikachu.png when it didn't read my mind perfectly from my three-line prompt. ಠ⁠_⁠ಠ

snake_case | 23 hours ago

My company has me on a $50 Claude subscription (recently lowered from $80) and I do use it, but I've only come close to hitting my limit once and that's because I was messing around and wanted to see if I could vibe code the entire feature I wanted for myself and not just a piece of it (I didn't like the result, and in the end decided that adding this feature would have added unnecessary complexity to an internal project which was supposed to be simple)

I use it pretty much every day, but I use it for single shot tasks that I need to get done. Things like "Create a dataset to test blah feature" or "Create tests for each new endpoint in this swagger file". It's also really really good at figuring out what's causing a bug (usually because it, itself caused the bug) and also really great at understanding custom errors.

Every time I have it write something I ultimately edit those changes myself usually to make it a bit less verbose cause Claude just loves adding multi line comments (and sometimes the same massive comment in multiple places? )

It's undeniably made me and everyone around me work faster. Features that used to take a week now take an afternoon. Projects that used to take years now take six months. There's some pieces like project management and testing that haven't quite caught up to the speed at which we can write code now, but people in those roles have been sped up quite a bit, just not as much as development has.

The result is that because we can ship so fast, we end up shipping projects which are terrible ideas from the beginning and had anyone actually stopped to think about what they're doing they'd have known that. I always thought my job was bullshit, writing code for other people's dumb ideas and cleaning up other people's messes when what I know won't work ends up not working. AI has made that worse, because now the bullshit I'm forced to write is so short lived that my human brain can actually process how useless it is, where before, I knew in some existential way that what I was doing was pointless, now what I'm doing is being thrown away soon as it's done right before my eyes.

skybrian | 22 hours ago

The gold rush is a problem, but I think AI's downsides are fixable, provided that society takes the time to fix them. I'm still somewhat optimistic about that, though it's not anything I can prove.

Often the objections people make about AI aren't very good as arguments, because they're based on bad premises or bad reasoning. But a political bottleneck on building more data centers might not be a bad thing overall if it slows things down a bit. Anti-AI populism might be "directionally accurate," as they say.

Nvidia getting caught cheating on chip exports might slow things down a bit too.

Similarly, the main thing wrong with vibe-coding is doing things in a rush. In theory, being able to work faster on things like prototyping can free up time to do more cleanup and checking. Quality could improve if we're willing to do it. I still expect a lot of slop, though.

It could still go badly in all sorts of ways. For example, the gold rush turning into a military arms race (drones and more!) and destabilizing the political situation further, resulting in more war.

arqalite | 22 hours ago

I use it daily for various purposes.

For software development, I find LLMs to be great pair programmers, great to bounce ideas off of, find holes and flaws in implementation plans and generally speeding up tedious work. They're good for more than that, but I'm not comfortable ceding all control to the machine so this is what ended up best for me. I do feel like a better programmer using them and I've built stuff with them that I wouldn't have had the patience to build otherwise.

I do think my more optimistic outlook is because I enjoy building software more than I actually enjoy the actual coding per se. I know a lot of SWEs find actual joy in writing code by hand but for me the pride has always been in the result, not in the code (although I do enjoy clean and properly factored codebases) - so having the LLM write code for me doesn't feel like I'm not the one building the software anymore.

I'm also hunting for good intelligence-cost efficiency ratios - I'm not comfortable spending 100€+ on AI subscriptions so I have to be careful with my $39 Copilot plan. I cannot afford to vibe code with Fable 5.2 or use Astra for more than a few turns, so instead I've taught myself how to extract the most out of models like Claude Haiku 5.5 and GPT-6.1 Terra. I tell Terra what I want to implement and how I would go about it myself, it replies with its suggested approach and then we work together on an implementation plan that is separated into atomic, independently testable and committable steps and then I pick and choose what I'm comfortable letting Haiku code (and I review and nitpick it afterwards) and what is best coded by hand. It's been great so far.

Besides that, I frequently use Gemini for day-to-day tasks like calculating calories and macros for a dish or answering questions I cannot simply Google. I admit for some of these I do zero fact-checking unless something looks off (which is when I ask for a reputable source I can read myself and then go Google the subject separately for more info) but I have yet to be burned by a mistake on Gemini's part.

I overall think that LLMs are good tools for what I use them for. I believe they are not good at art and creative work, and the current architecture never will be - you cannot make art out of what are essentially statistical averages. It's a common saying that we are the sum of our influences, but as I said recently around here, we also need the element of human randomness to truly pull something great out of them.

We're also failing as a society to sustainably and ethically build LLMs. Copyright (and my negative stance on it) aside, the industry is currently harming people, the environment and the economy in ways that could have been easily preventable if we had leaders that actually cared.

To this end I'm a bit frustrated that what is generally exciting and useful tech got overshadowed by unhealthy and unchecked capitalism doing what it does best. Will it get better? Probably, but for now it seems like it'll get worse first.

R3qn65 | 20 hours ago

So many thoughts.

Using them personally: A few years ago, I thought that the Strawberrrrry problem would never be solved. Clearly, I was extremely wrong. Today, I use LLMs extensively. My perspective is a bit unique on Tildes in that I am not an IT professional, who are as far as I can tell the only other people who admit to using LLMs. Even so, I find LLMs to be basically unmitigated positives from a personal standpoint. I don’t need much code (again, not in IT), but if I do, I can now get it from Claude, whereas before I would either have to ask for enterprise support or just go without. I run various written products past LLMs; while their advice is worse than what I would give if I really tried, it’s better than most people’s, and it has the advantage of being actually independent. And maybe the biggest thing is that I can understand more things, faster than basically ever before. I don’t actually use them often at all to summarize work products, as it’s important that I understand 100% of the nuance so I need to read those myself. But they help me understand… basically everything else. For example, I recently ran a theory on decisionmaking for a seminar past Claude and said “give me a list of studies that either support or contradict this idea.” It did, with links, and I was able to read them all and then work through my subsequent musings with productive feedback (~80% as good as what I would get from a skilled colleague). A day’s worth of work (or more) covered during a short work flight.

Using them: society: Despite all of the many, many positives for me, I worry about society. I think LLMs will make bad schools and elite schools much better, but everyone in the middle is going to get crushed. Good students at bad schools will be able to receive personalized learning for the first time in history, and elite students are going to end up like me (so-called centaur model - LLM + human is better that either alone), but most students at Middle Central State College just aren’t sufficiently motivated to not turn their brains off and have ChatGPT do everything for them. Some of this is old man yelling at cloud energy and people have been worried about the youths for forever, so things will almost certainly turn out better than feared. Maybe we’ll all end up with a great books/classical education and let the machines do the boring math.

Environmental: I don’t really worry about this too much. Even if we froze LLMs at where they are today, I think it’s almost guaranteed that they will save us more energy/water/etc than we spend on them. Humans make things more efficient, LLMs make us faster, ergo we will make things more efficient, faster.

Miscellaneous:

  • I agree with the general scientific consensus that LLMs are not conscious, but (like many researchers) I think it’s getting increasingly murky. For that reason I could easily see ethical issues start to crop up. People make fun of Anthropic for worrying about this, but they’re the only ones taking a low-probability high-consequence scenario seriously enough.

  • ASI and AGI are clearly not the right terms anymore, but using them loosely, I think ASI is very unlikely and also unlikely to blow up everything immediately. “Intelligence” is almost never the rate-limiting constraint in change and also humanity just doesn’t move all that quickly, even if e.g. Anthropic ended up being right that it’s unequivocally better to just put Claude in charge of the government.

  • In some senses it is good that the US and China are so neck-and-neck, because it reduces incentives for one of the powers to launch a preemptive strike on the other to stop the other from reaching runaway recursive self-improvement.

DynamoSunshirt | 22 hours ago

I use it at work but i feel pretty awful about it. I try to avoid it in my personal life, but it's an uphill battle when non-techies increasingly "why don't you ask Google" or produce LLM summaries as truth in the middle of a conversation. I've mostly given up on resisting or talking about it, though. The AI pushers are just relentless and in the tech industry it's obvious that any level of skepticism puts a target on your back.

It's been vaguely useful at work but I would far far far rather just do the work myself.

snake_case | 20 hours ago

Everyone at my company has drank the kool aid I couldn’t do the work myself even if I wanted to, theres no way I could keep up. Everyone else is doing a weeks worth of work in a day and no one cares about the quality cause they know they’ll redo the same work six months from now when some idiot decides to rewrite the whole project for a 5th time.

Is this not happening to you, too?

DynamoSunshirt | 6 hours ago

Yes and no. Fortunately I work on a team of one in a pretty niche specialty, so I'm a bit insulated from other people's nonsense. So I've built a lot of checks and balances and I can pace myself a lot better than I could have paced a team with a bunch of distinct approaches to LLM dev.

But I do work with others on the peripheral of my domain and boy is it getting bad. So many slop bombs. Quite a few "specialists" have been using AI so much for so long that you can't even have a conversation with them about their domain any more -- they just immediately start screen sharing claude as their default approach to collaboration! I just avoid meeting with them. It's like talking to a brick wall.

It mostly just makes me sad that so few people see the bad in this. Am I just crazy for feeling that skill atrophy and attention to detail loss and shipping unvalidated features are legitimately dangerous? Am I ahead of the curve? How do I keep speaking in a language that the higher ups accept, as a skeptic, since all they talk about is AI?

snake_case | 3 hours ago

The skill atrophy is so real.

I’m just letting it happen to me because I dont even want to be doing this anymore. If they lay me off I’m gonna go back to school and be a cpa or some shit like that idk.

Ive got this dream of just doing old fashioned IT support work for like some small business. I dont have the network to get a job like that though.

updawg | 22 hours ago

As kind of an add-on to this comment on the "Make your site look hastily vibecoded" post, every day I think more and more that anyone who uses anything other than Claude is a kneecapping themselves at best and an idiot at worst (the stuff my coworkers proudly tell me "Chat" told them to buy is just sad).

I know other LLMs have their use cases, particularly in areas that I don't work with, but Claude is just so much better in everything I test it on. Even when other LLMs catch Claude making a mistake, Claude's responses always end up just further illustrating how much better it is. It just almost (almost) always ends up doing things so much better than I was really even looking for.

I suspect that it might just be that you are better at writing prompts than most people.

I think a lot of people just write very barebones prompts like "make me a website about how dogs are cute" or something ultra open-ended, get whatever the AI spits out, and just says it's fine. That's the definition of slop. It's one thing to not have a clear idea of what it is you want or how to describe it, and it's another thing altogether to just be super happy about whatever you get. Even if you're fantastic at description and you put a lot of effort into describing it, writing paragraph after paragraph, AI will always need you to check and revise it's outputs, and that is massively more important when you ask it to do something creative. And yes, that even includes mundane coding tasks, because coding requires creativity!

KapteinB | 12 hours ago

I subscribe to Mistral. It's a purely political choice. The USA has slapped unfair tariffs on my country and repeatedly threatened us with war. It's led by a wannabe fascist dictator and his wildly corrupt family, backed by a crony-capitalist clique which includes a bunch of tech CEOs. I choose non-US alternatives whenever I can, and if that gives me a slightly worse product, I'll live without that kneecap.

skybrian | 22 hours ago

Have you tried OpenAI recently? Sol seems pretty solid in my experience and it works with any coding harness.

(I'm not one-shotting anything, though. We have a conversation.)

updawg | 21 hours ago

I've never used an LLM from a command line. But with the stuff I do, when I give Claude a chance to review anything I generate with any other LLM, Claude basically gives me "wow, that's great! Here are 17 ways where it's completely awful and here's why everything else it said is completely irrelevant. Also, I've created a tool for you that simultaneously answers your questions better and renders them moot."

As you can probably guess from not using the command line with LLMs, I'm not doing much more than generating documents and web apps.

skybrian | 21 hours ago

Whenever I ask an LLM to review its own work (for example, a design doc), it finds issues too. After I fix them, it will find more issues, but with repeated reviews, they become less serious and then just nitpicky or inapplicable.

If you generated something with Claude and asked ChatGPT to review it, I expect that it would also find issues. They try to give you what you asked for.

updawg | 13 hours ago

Yes, I'm aware. Even just asking "are you sure?" a few times will make it bounce back and forth like an insane person. That's not what I'm talking about.

A month ago, when I asked all four big American frontier models to generate the same brief for a (public information, non-work-related) presentation my boss asked me to give, using the same prompt with all of them (Would you please do a deep dive into every state's laws regarding x AND y in a report? Ideally it would provide the state law references and excerpts plus links and other helpful information that can help [my coworkers]? I will later convert it to a more presentation/user-friendly format, but I want to gather all of the information first.), Claude spent like 45 minutes putting almost everything in an artifact with just one prompt, with just a handful of extra refining prompts needed (it can only quote 125 characters so summarize the law instead, remove extraneous info that isn't needed for a brief, reminding it if AI is going to destroy us all then surely it can access a copy of Alaska's laws, etc.).

Every other one gave me some bare bones response. They were several paragraphs long, but they weren't a report and they didn't cover every state.

When I handed it off to Gemini to check the information (Please analyze this for accuracy.), Gemini basically gave a summary of places where the laws were contradictory and a few places where the wording was too strong ("Claude said the state requires x but the website says it may require x).

This is consistent every time I try the same prompts with multiple models or have one check another. As another example, I was emailing my landlord about a structural issue and Gemini talked about how Claude's summary of my rights was wrong because the law stated xyz and when I passed it back to Claude, it was like "well, that's a good summary of the 2001 law that has been obsolete since 2013.”

Every single time, the others are just so much worse.

Gemini is better at giving recommendations in an area, though. It has a good sense of the physical world and is better at pulling info from Google Maps.

arqalite | 21 hours ago

I do see the general opinion is that Fable 5.2 and Opus 5.5 are a league above everything else, but personally I fail to understand why. (to be fair I also do not use them enough to really dive into their strengths and weaknesses)

I'm doing just fine with Haiku 5.5 and GPT-6.1 Terra and I rarely find them stupid enough to warrant that reaction.

That said I am very hands-on with the work I do with them, and I assume you're talking more of a vibe-coding approach where you're going from prompt to PR?

For that use-case, I can see how Fable/Opus models would outperform the rest, but personally I wouldn't be comfortable with that workflow at all in the first place, so my opinion might not be worth that much.

teaearlgraycold | 21 hours ago

It sounds like you're mostly using LLMs for synthesis, but I find them also very useful for analysis. Try asking Haiku to find security issues, or UI/code design inconsistencies, and then run that same prompt through Fable or Opus and compare. I recommend using sub-agents for this so it's a comprehensive search rather than a random walk (you can say "ultracode" in your prompt to get parallel LLMs but unless you're on the $100/month plan or greater you'll definitely run out of usage credits when using Fable and ultracode).

arqalite | 11 hours ago

Oh yeah, for analysis I do not trust Haiku in the slightest. Terra is decent but it does sometimes miss stuff or focus on the wrong findings, but I've managed with it.

For the few instances where I needed a full codebase analysis, I've been generally using GPT-6.1 Sol or Astra and I haven't felt disappointed by the results but I am absolutely convinced Fable would do a much better job.

karsaroth | 22 hours ago

I distrust the big AI companies for various reasons, especially because it's clear their business case is that they want to get rid of me (a software engineer).

However, it's clear that LLMs have a place in my industry, and in the world. What I think we as a global society should aim for (for once) is to use them to make human work easier. Give them the difficult manual pattern matching tasks (e.g. in software: live checks for runtime failures, security holes, inefficiency etc), and leave the key creative work to creatives.

vingtcinqunvingtcinq | 19 hours ago

If I could go back in time and strangle Baby Hitler GPT in the cradle, and it would give me another 10 years on my career, I 10000% would. That being said:

For the good stuff:

  1. As someone without much CS academic background, I begrudgingly am trying to like the democratizing effect it has on someone like me who cannot sit through 100 hours of Kubernetes / AWS dry-ass videos for a cert to figure out how to do something perfect for a hyperscaled system design. Instead, it feels like there's been a shift towards prototyping, which is more fun if I can accept that the industry has no heroes in programming quality and culture to emulate right now. It feels like there's more value in surface-level understanding versus acing every single detail?
  2. I like it for looking into old code bases and getting file / line references for its assertions, explaining code snippets when the docstrings are too canned (or, in some libraries, the public docstring just says something like "wrapper around (internal method)", thanks guys), quick unblocking on tasks that might create a context switch for me (e.g. "what is the syntax", "how do I call this library", "what are libraries I should look at for X criteria"), and anything where the alternative was probably never doing this task or where the outcome is easily verifiable. For example, you can make visual e2e testing by giving even just Qwen two images of the same page, which is not something that can always be meaningfully caught by Playwright.

Back to the bad stuff:
3. I do not like it for general coding. I think the best effects come from naturally small contexts, and are extremely poor in ill-defined areas. This sounds obvious, but it's aggravating to be gas lit into thinking I have a 'skill issue' because I don't have all the requirements in front of me. Dude, my job is to work around the lack of requirements by doing surgically small changes and understanding the external abstractions without delving into details, it's so hard to do without rolling in muddy code first. The canned stuff the LLM copy-and-pasted I could do just as fast or faster lol. This is the use case that's made me the laziest, and I became exponentially happier not doing this anymore.
a. for this same reason, It gives decent feedback on code ("judge this snippet") but terrible feedback for PRs ("is merging this a net benefit with no meaningful additional problems?"): "Look for issues Claude, max effort." "Well, on a full moon, when mercury is in retrograde, the two writes may cause a data race where one value is written before the other but the end result looks the same." While stuff like this was always undesirable, to be honest, when you have to dig through the shit yourself, you develop a better sense of... "I'm not in this code base to save the world today." You optimize time because you can intuit the effort and complexity better. I find this challenging when using an LLM.

Neutral:
4. I think I would like it more if a) we didn't call it AI b) if people weren't clearly clamoring to replace people, say what you will about previous era programming but I generally didn't see it as effectively replacing people and c) if I wasn't mandated to use it and in a specific way to please people.
5. I think there's a weird approach to both black boxing it and engineering it that I have seen. People treat context like they're using black magic to manipulate black boxes? There's a weird level of flirting with deterministic behavior but keeping other stuff nondeterministic. It feels confused and inefficient. Like... some things are better copy and pasted, but other times, wouldn't a library be better? I made "skills," but some of those skills were just shit my IDE did for free on save lol (formatting, linting, etc.). I think there's a lot to be unearthed on the UX side here for "power users," as I feel like a lot of it is geared toward smaller projects or nontechnical users where one-shotting things works out better.

I pay for Claude Pro - their cheapest plan - but I don't use it for "useful" things. I don't make it do work for me unless that work is painfully mundane or something I don't care about. Examples might be something like finding good coffeeshop options in an area I'm not familiar with, or customizing a resume to give to specific employers.

The most important job I've offloaded to Claude so far has been to analyze Terms of Use documents when they let me know things have changed. I've written about how much I hate how companies immorally force you into arbitration agreements (it really should be illegal for companies to use their leverage to force you to give up your civil rights!), so I came up with a project with extensive instructions to examine the document, compare it with the previous versions, find language about arbitration clauses, and for those with opt-out clauses, to draft a long indignant letter that lightly curses at them while fulfilling the requirements to opt out.

That being said, the thing I primarily use it for is to explore concepts and learn about things. I know hallucinations are problems for any AI, so I always make sure to dig in and ask for it to research and provide sources. I've written before about how bad the curriculum was for my university classes, and Claude has an excellent learn skill that will automatically load when I tell it to teach me something. I can provide it the text and the information from the curriculum so that it can give me the information in a way I can understand it and then quiz me with questions that I can answer, and if I don't understand or give the wrong answer, it usually tries to give me hints to lead me to the answer instead of just giving it to me. It's super useful for this.

There is a notable downside, and I'm not sure how common it is among LLM applications and models. I often use incogneto chats for things that are inconsequential or I simply don't want to have added to my history later; those conversations are not used for training and they disappear when I'm done. But I've noticed that when I ask it for information, especially on social issues, it goes hard on false equivalence. I can't recall the specifics, but the most noteworthy one was when I asked it to fill me in on the happenings with the last Brazilian presidential election. It was giving me pros and cons for the two of them, and one of them were that Lula might be corrupt because of some scandals. It only told me about Bolsonaro's own corruption scandals after I asked for more details about his father and asked it to confirm that the younger was not also implicated in them. The need for equivalence somewhat often ends up giving me neoliberal lassez-faire capitalist responses about how some things that cause personal harm are good for the economy. I can't report them as being bad responses because they're incognito, and it doesn't seem to happen when I'm doing a regular saved chat.

What do we expect? False-equivalence is the default state of reporting these days.

JCPhoenix | 15 hours ago

I actually pay for my own Claude subscription for work. And so do several of my coworkers for themselves, for work. Even though we're not supposed to use external AIs. But it's just that useful. All the projects and tasks I worked on in the last year would've taken me far, far longer to figure out, and plan, and complete -- if they got completed at all -- if I didn't have AI.

Which has led to conversations with coworkers, friends, and family about AI. I work in IT. I'm a "knowledge worker." But who's knowledge, who's brain, is doing the work here? Sure, I need to know what to ask. I need to know what's important and what's not. I'm still the one coming up with test plans and such. Claude doesn't have a full understanding of our infrastructure and systems and processes. But clearly AI is bringing something to the table. A lot of something.

Idk. Guess we'll see in like 5yrs if I'm still in IT or if I've hung it all up to become a goat farmer or something.

VerdantNinja | 12 hours ago

I'm fine with them, and I will use Claude for Linux troubleshooting and coding help.

That being said, I've had a few freelance contracts to do AI training, and I know how woefully behind my limited knowledge base they are. So I know that they're usually less good at things if I'm moderately well-versed in the thing, and better than me if I'm not well-versed.

Would I ever use raw LLM output in my areas of expertise? Hell no. It's never good enough, but it can be a starting point that saves me a lot of time.

pete_the_paper_boat | 11 hours ago

Make it do the work while you do the thinking.

At least in code, the most successful workflow seems to be to turn yourself into a maintainer. But only once a solid foundation has been set.

Paul85 | 6 hours ago

I use it sparingly, usually when I’m stuck at work and need answers quickly instead of visiting several web pages littered with ads until I finally find something ok.

I’ll use it at work to fill gaps, where I don’t have time to spend on more studious research, especially if it’s a one-off task versus a skill I’d like to learn properly.

At home I don’t use it much at all. When I do, it is same as above: don’t want to surf a handful of pages or dig through comment threads on Reddit. I just want a straight answer.

I add instructions in the settings to reduce all the things that make it friendlier and that offer follow up questions for engagement.

It’s basically what a good Google search used to look like back in the day, when the top 1-2 results were the actual answer and solved the problem or answered the question.

Markpelly | 6 hours ago

I have been using different LLMs for years now but more so in my current role. I am a data architect who does solutions engineering for financial companies. We have a small team and a lot of customers. The work I do is very detailed and I can be a bit forgetful at times. My agents and tools I have been using keep me honest and use my knowledge of banks to produce projects very quickly and accurately. There are times that I need to jump in and make improvements, but honestly it does a great job to maybe 80% . We have a basic chatGPT enterprise account and it does so well.
My friends and family tell me I am way too efficient for my own good, but this tool has allowed me to be even more efficient and help our business grow.

crulife | 11 hours ago

Stream of thought:

I've been using LLMs since GPT2, so it's completely normalized for me at this point. It's crystal clear obvious to me that LLMs are the biggest step forward we have had in centuries and they are eventually going to change everything in a way comparable to industrialization. They could eventually pave the way to post scarcity, not alone but with other inventions they will enable. The problem is, humanity seems to cope with 1% of the people being idle rich, but what will happen when it's 99%?

Currently, I have Codex and Claude via work, and I'm paying for Mistral and Melious subscriptions from my own wallet. If I don't manage to use my weekly Codex/Claude tokens, I pump the excess to my own projects. Other times, I 24/7 use deepseek v4.1 flash from mistral/melious for coding, GLM5.3 for reviewing.

On my own projects, I usually do the fundamental work with a hands-on approach but once the cement has been poured in, I mostly just command the agents to grind away for 4 hours, 8 hours or more with the tasks (stored in beads) until they're done -- working in a review/fix loop without my input. Usually they come back roughly once a day with questions that need judgment calls. I answer those and put them back to work.

I've been able to create things that I have dreamed about for decades. I sometimes worry that I might be losing some edge I might have had because of this, but then again, I was never that good of a coder. The fundamental design and the question of what to build, are still interesting. It's like you're an alchemist, you pour in a the ingredients and a drop of your own soul and let it brew.

Another nice thing is how good they are at configuring Linux. An LLM will take a 30 second read-only look at your system (if you dare allow it) and will likely suggest a dozen things that would improve things. They are all things that I could've figured out by myself with my 20 years of Linux experience surely, but also things that I would never have had the perseverance or deep interest to detect.

I've built a Fedora Atomic system for myself that gives me roughly the benefits of NixOS without being that weird.