Well llms are imperfect tools that provide a lot of value (they offer leverage in form of agents to put it in Naval Ravikant words). They are here to stay.
Fundamentally I don’t think we have a choice about riding on the AI technology train. It’s a wild ride and I just hope we’ll get through it OK.
I don’t like them. They talk to me in this grating LLM-voice, an uncanny valley of talking to a real human. They confidently bullshit me - often giving me useful, helpful answers. But also just making stuff up with the same assurance - and with only a veneer of fake remorse when I call them out on it.
I think what he needs is a personal fine-tune, something I suspect will become more prominent when we start getting multi-tenant LoRA and can use fine-tunes on a per-token cost basis.
A co-worker commented that he thought the models performed worse after being "embarrassed," suspect it is an artifact of how some conversations progress in the training data.
Reminded me of someone saying that taste will become an important factor wrt AI tools/ things built with AI. Something like that.
edit: Not exactly as I remembered, but it's a PG tweet:
Prediction: In the AI age, taste will become even more important. When anyone can make anything, the big differentiator is what you choose to make.
Modern internet ruined forums and google search sucks hairy balls, and most of the content is SEO optimised ad ridden crap. Its literally impossible to find anything meaningful on the internet nowadays. What else can we do besides asking an LLM these days?
I think that the argument is not that it happened before LLMs. Rather, assuming I'm understanding the parent correctly, they are arguing that, due to the enshittification of search and the internet, LLMs are the only way to get information from it. That has definitely been my experience as well.
I agree in sentiment, and really a lot of this is Google's doing. It's been around a decade since they switched to approximate results and started encouraging people to search via sentences. I believe this is the one? https://en.wikipedia.org/wiki/Google_Hummingbird
The internet for all of its activity is becoming harder to index because the classic dumb search engines were turned into personalized semantic people pleasers.
The classic model for knowledge discovery still exists, though. You find a blog you like, and use that as a thread of knowledge. You join a community and talk with someone and share resources together. Join a small group chat and post with those people. Your knowledge sharing comes from people you share interests with.
Go to a library, ask your coworkers. If you put in the ground-work, it's possible to do that. For seeds of knowledge outside of your local network, at this point you have to find the indexers that work for you.
Personally, I am in the process of writing the infrastructure for my own search engine, and maybe I will share the tech broadly one day. For now, I'll keep my inventions to my closer personal circles.
In general, I think people should be able to more easily maintain their own offline indices, and share knowledge graphs peer-to-peer rather than relying on a centralized one, in my opinion. This tech has no monetization opportunity, though, so I'm guessing that's why it's relatively under-developed, but luckily for me I have enough money and now enough time to develop and release it (and similar tools).
You find a blog you like, and use that as a thread of knowledge. You join a community and talk with someone and share resources together. Join a small group chat and post with those people. Your knowledge sharing comes from people you share interests with.
It is hard to find communities where people are not evangelical and soon you become an evangelist.
I am using LLMs to build my own graph of thoughts and try my best to follow them with critical eyes. Do not want to be influenced by or influence others.
Exactly. Plus, many people don't seem to understand the word "avoid" (it's come up in my own bubble recently). It doesn't mean "never." For example, a vegan would not say "I avoid eating meat." Furthermore, because someone successfully avoids now doesn't mean they went through a period of being unable to in order to earn the privilege to do so regularly.
Your exemple says much more: a short definition of veganism may use "never" and not "avoid", but a more complete, thoughtful definition encompass catches it:
[...] which seeks to exclude—as far as is possible and practicable [...]
You appear to be confusing "avoid" with "never interact with".
If your professional life has required you to spend enormous amounts of time working very closely with people you don't like and/or don't trust, then -unless it has made you enough money to retire after a few years' work- please accept my condolences.
Oops, just wrote something similar before seeing this. It came up recently in another community where it seems some people take the word "avoid" to mean "never" and in this case: "never not even in the past."
For reasonably skilled people, it's often more just a matter of priority.
A lot of people get wrapped up in pursuing compensation opportunities or prestige brands as their top priority, only to burn themselves out or wallow in a domain/culture that doesn't suit them, but if you're willing to deprioritize pay or dinner table cred, you often end up with a lot more flexibility on other work factors. And sometimes, flourishing in those better-fit environments pays off bigger in compensation and career growth in the end anyway.
Granted, Fowler was able to do very well on more axes than many of us, but most competent mid- or late- career people have plenty more flexibility on worklife than they admit to themselves.
I have been trying to find one of these "lower pay better conditions" places, but it is _such_ a gamble. It is so hard to tell how the work is going to be before you join. If I am almost certain to at least have a mild dislike for the work I might as well go for the place that pays better.
Also in my experience lower pay places tend to eventually suck because they tend to get into financial difficulties much more easily.
It’s always best to just aim for highest compensation. Even if a team is your dream fit today, it can easily change.
High compensation can also bring more freedom and happiness, since it’s more costly for your employer to replace you. It’s cheaper to keep an existing employee happy than to spend a lot of time and money to bring a new person on board. If your compensation is lower, it’s not as big of a factor.
I wouldn’t take a job if it clearly doesn’t fit me at all, but within the band of “acceptable jobs” aiming for highest compensation is the way to go in my opinion.
What makes this a luxury? I've met many people in my life, and this gives me the opportunity to pick-and-choose with whom I do business. It's a matter of degree.
Plus, I've been on the receiving end, too. Is this not simply adulthood?
I don’t like them. They talk to me in this grating LLM-voice, an uncanny valley of talking to a real human.
This takes about 30 seconds to fix permanently. You can prompt them to use any tone , be as concise or verbose as you like.
They confidently bullshit me - often giving me useful, helpful answers. But also just making stuff up with the same assurance - and with only a veneer of fake remorse when I call them out on it.
Only people with flimsy epistemological underpinnings feel threatened in this way. Everyone makes mistakes, and gets fooled by AI (as we’ve been fooled by historians and scientists), but to write off AI because you feel it’s constantly misleading you means you lack proper foundations. Most of my fun conversations with AI are bending it back into proper grounded truth.
This takes about 30 seconds to fix permanently. You can prompt them to use any tone , be as concise or verbose as you like.
This doesn't work by, effectively, Anthropic's own admission. CC's concise mode tells Claude to wind its neck in every turn. That is significantly more than changing the prompt.
What is the Silicon Valley brogrammer subculture? Is it the same as the Silicon Valley subculture? What defines both?
Are LLMs really the product of brogrammers?
I'd guess it makes sense that if programming is a passion for you then probably a non-tech-centric F500 company probably isn't where you'd choose to work.
For context: my current company with ~400 employees (not just engineering) is the largest company I've ever worked at.
Sure, and that's a big shift in industry culture over the last several decades.
While there had long been perfuctory white collar programmer analysts filling the ranks at some dry divisions at IBM, it used to be way more common to have a software engineering department full of Asimov-steeped Omni-reading nerds who had genuinely passionate interest in the field that had taken hold even before they started working in it.
But later the career increasingly came to be treated more like law, medicine, or finance and you started to see those rooms fill with people that were often generally bright but not really "called" to the field in the same aay.
I mean I got into it because I loved it as a kid in the 80s and 90s. And I still loved it in college. Turning it into a job quickly crushed any love I had though.
I wonder if carpenters, plumbers and electricians, etc have the same kind of love for their jobs in comparable ratios.
I certainly don't ask when I hire one.
I worked for twitter and there was a guy like this. I made a big stink that "yo, everything this guy says just needs to be in a linter" and they actually implemented it.
If you don’t have style rules baked into the PR process, you should. Tabs vs. spaces is a team decision and needs to be matched by everyone to reduce code churn. It’s correct to reject a PR if it violates the team style guide (although it should be automated and have tooling locally to enforce it as well.)
The distinction is even easier to make these days -- brogrammers aren't actually programmers.
There is an emerging coterie of people whose entire approach to software startups is to vibecode an MVP, raise enough seed to hire a real programmer (or get one interested enough to become their "cofounder") and move on to the next MVP. To them, doing one company at a time in the era of LLMs is a sucker's game.
I fell into this career because I programmed as a kid, hacked on pre internet bboards, and understood how to drive a computer via the command line. I like solving the problems. GED, college dropout. I never planned on making this much money, but I'm certainly not giving it up. I don't feel like I fit in either camp tbh.
Indeed. Very disappointing given what figure is in question. Simple GH pages behind CF would do which is something people starting learning IT this week can do.
I can't believe HN has changed so much that we are now mocking people for not hardening their blogs against traffic spikes. The comment used to just be "hugged", but now, if the author has anything critical to say about AI, it's this same snarky, elitist quip every time. There are two of these up at time of writing, with multiple agreeing replies, all petty, all demeaning.
I know we're not supposed to complain about HN "changing", but, shit, it really hurts.
I mean, this is also Martin Fowler, a luminary in the field. His site should be able to handle the HN hug of death, particularly if he's bashing the latest-and-greatest tools. As is, he looks like an old man shouting about cars being bad while his horse drops dead at the side of the road.
I personally don't see the correlation between what he's been saying about AI and his blog site's reliability requirements. It's a blog, not some high availability service.
Also, was Henry Ford asking for regulation because his cars may destroy our civilization as we knew it?
In some ways I miss the early LLM days before RL and system prompts (I assume this is the difference) made them all sound the same. They would say weird stuff. They got better at being useful and worse at being fun.*
*This sadly also describes me as a 31 year old man.
archive.is / archive.ph haven't loaded for me for a while, but since people are sharing fresh links it must be working somewhere. Am I experiencing a DNS block?
Edit: I did a web search and it seems to go even further back. There are discussions online about archive.is blocking Finland already 10 years ago. But the block was definitely off between then when it was put back on earlier this year.
yes they are incredibly unpleasant to work with. Like the most annoying coworker you ever had.
For my own sanity ive wrapped my pi setup in some strict bubble wrap and it is basically exclusively an LOC generator. I have no need nor desire to "converse" with the LLM, it is there to do exactly what i need when i need it.
One of the I dunno, weird things about LLMs, is that their search space feels exceptionally constrained.
I'm sure I can sort of prompt it to try to give the answer I want but some people talk about "I don't want an LLM with a political perspective, just provide all the perspectives equally and without bias."
And to a certain extent they have to be constrained. Anything more than a 500 word response I'm probably not going to read, particularly if I want to go deeper on something I want to navigate the conversation. But it means that tuning things like sycophancy just create sort of extremely simple mitigations. Even when I say "May" or "slight chance of" I'm often corrected "Be careful to make too blanket of a statement." I wonder if part of this has been the move towards optimizing towards the agent experience.
Likewise a bias towards "Say things equally without bias" is an exceptionally poor way to actually explore the nuances of an idea, with this sort of regression to the mean.
I don't know if they'll be able to solve this problem broadly. I mean, LLMs probably shouldn't be therapists/coaches/philosophical gurus. But I could anticipate a growth in sort of groupthink and regression to the mean in terms of perspectives broadly speaking as a society.
I find them useful. Maybe it's better I find them distasteful to talk to as well.
I have great success in specifically asking LLM tools for academic literature after describing the problem or field I would like to get some insight into.
It's usually the first step I take when writing research. I don't even read the analysis, just go straight for the papers.
They usually provide me with very relevant seminal papers and I can then easily expand from the bibliography in those papers.
It basically solves discoverability, whereas in the past I would spend days and would need to ask colleagues and librarians for help, to achieve what I can get in minutes now.
But again, I value the actual output of the LLMs at bellow zero factually wise because I need to check everything they say if I want to internalize it and use it for any end. Pointing me to sources is the perfect use case. Really, I cannot emphasize enough how little I trust the text LLMs produce, I actively avoid reading it because I really don't want to risk ingesting spurious information that I might later repeat or rely on for any reason.
LLMs have helped me a ton and have allowed me to replace all my apps that I was paying for. They're internal, so they don't have to be perfect, but man they've saved me money. That's just my experience.
yes! I've had it even help me with looking and buying homes. I ended up catching things in a contract that would have screwed me over. there's so many use cases outside of software
Youtube music, GitKraken, 1password, Parallels -- i've canceled subscriptions to all of these because either I've replaced them or have had help from LLMs enough to find different options that are either free or cheaper.
I assume this person heard about keepass [0] and accesses it from multiple devices with a dropbox or icloud style service. Or doesn't access it from multiple devices. I just synchronize my password database on a slow cadence or as needed.
Bitwarden is free and in my experience does a better job than 1 password. Tho he could have vibecoded his own 1-password, I would be a bit careful doing that though
yeah I agree, bitwarden is really good, but since I'm the only one who uses it I didn't need all the features of it and just vibecoded my own, though I made sure it was secure (I'm a software engineer so knew what to secure). definitely wouldn't recommend anyone do that though if they don't know what they're doing
well, I'm assuming you mean for this part I had said:
had help from LLMs enough to find different options that are either free or cheaper.
as one example, I would have to do multiple google searches to build a secure network around a replacement tool I built for 1password, but the LLM was able to walk me through everything (even though I'd never had dedicated training around security). google searches are great for 1-off questions, but have the LLM walk you through something or have it do it for you.
yeah it's not the best replacement, but basically what I did is setup a "YTBox" app that takes a youtube link, downloads it as mp3 or mp4, and then auto syncs it to my icloud drive, and then I just listen from my iPhone either using the music app or VLC/Files.
so yes I do have to make a curated list of what I want, but it saves me $17/mo. if you're okay with downloading, then it's worth it.
I think this is a very even-handed reasonable position. and it's true - LLMs can be unpleasant to talk to. the day it feels like a magical delightful experience is the day we'll probably be out of jobs, though.
While the article is fairly broad (it talks about "our profession", which ostensibly is coding, but then talks more about LLMs and their conversational style which is something that everyone has opinions on), it's worth pondering what his professional relationship with coding is today.
Is he actively coding...anything? Any projects? And engagements? Or is he a coding influencer now, giving hot takes and opinions, and riding on the XP inertia as a relevant authority?
Because if it's the latter, and I strongly suspect it is, his opinions on LLMs and coding are close to worthless. There are loads of "I'm now an influencer" giving these sorts of hot takes now, and they all seem...detached and irrational with the real world.
Meh, i use LLMs every day as a 9-5 dev and i absolutely despise how they write, and i don't particularly like how they write code. Sometimes I'd rather just have something done, and asking an LLM is basically always faster than using Google now.
Fundamentally I don’t think we have a choice about riding on the AI technology train.
It's amazing how effectively tech leaders have indoctrinated people to give up their power and give the tech people free reign. The same thing has happened in the socio-political realm. Assertions of powerlessness are seen as wisdom.
It surprises me a little bit that we still don't have a personalized agent.
I def want to finetune my llm responses and yes sure Claude has some type of memory but its too early to just bet on one horse and while i use claude at home through claude.ai i also use claude through an api at my company.
It might start feeling better when they solve the issue of forgetting things they already got right before.
Kinda funny that one of the biggest software engineering “influencers” of all time can’t host a blog that survives HN front page traffic. Maybe try LLMs Martin.
He is of course entitled to his opinion on the topic, but I’m personally not going to put too much weight on it considering he hasn’t written software for a living in probably decades.
I for one am excited to move past the whatever-martin-fowler-thinks era of infrastructure "best practices".
Many of his ideas are trendy nonsense that have caused more harm than good. Ask anyone managing 150 enmeshed and intertwined microservices that could have run on a single computer
The post has nothing to do with writing software. It would have as much value if it were written by someone who'd never written a line of code in their life. It's about the social dimension of LLMs and how the personality they present is, for some of us, pretty grating.
Martin is a good writer and (from what I've heard from those who've spent time with him) a lovely person which I think makes him well qualified to speak on the topic.
Love the totally conflicting statements in that post.
- When we think of AI agents, we shouldn’t anthropomorphize
- ...my visceral dislike of interacting with an LLM that’s not just making a pretense of being human, but also posing as the kind of human I walk away from.
So he's anthropomorphised the LLM as being like a human himself (rather than forcing it to act as a machine via its prompt for example).
Sometimes I wonder if this is a personality thing. It has never occurred to me to anthropomorphize these things. But it seems like a huge number of people struggle with this.
I suspect harness and usage modality is a big factor. Personally I disable all "memory" features wherever possible. I don't want the LLM to "get to know me", or any illusion thereof, which I assume helps with not anthropomorphizing it.
I use memory all the time so that I can ask it about things. That doesn't mean it "knows me" any more than my notes documents do, it's just more convenient to search.
But frankly, I mostly use (greenfield, non-remembering) LLMs directly to counteract the enshittification of search brought on by attempts to use AI to remember things about me (and especially to sell me things).
This works well for me because (a) I'm usually using my desktop, and (b) I'm a touch-typist. I want to say that my search results have gotten back to where they were a decade or more ago by doing this, even though I often have to give a couple of prompts in order to get the LLM to provide me a useful link.
I guess I'm also pretty skeptical of the "automatically remember things about me" functionality. But I really like being able to say "hey make sure to remember this fact" and being about to ask about things later.
My daughter is a pediatrician and just gave a talk on mental health problems (including suicides) brought on by the sort of personality that succumbs to this.
I find it interesting. For most of my life, I have gotten on well with (well-designed) machines. I'm the sort of guy who, when they bring me in to show me what's broken, it works in front of me.
Maybe I get on with well-designed machines better than people. Because people are never well-designed. So, when LLMs are working properly, I get along well with them, but when they're not, I'm unhappy. But I never confuse any of my machines for people.
Yeah to me it's just "oh what a nice interface for compressing information and accessing it upon request!", not at all "I'm talking to a person". But I dunno, maybe we're the weird ones.
These are not contradictory statements to me. Even if I don't anthropomorphize LLMs, the damn things anthropomorphize themselves and I have to deal with it. "Posing" is the key word.
Honestly, I think he's pointing out that the LLM creators have intended the LLMs to anothromorphize themselves.
Case in point: Gemini just asked me "Where should we start?". When I asked it if that turn of phrase was used by the creators to invite the users to anthropomorphize the model, it's first paragraph was:
Yes, to a significant extent. While system designers rarely state their goal as "making users believe the AI is a living being," tech companies deliberately design conversational agents to evoke social and relational instincts.
He means don't anthropomorphize them with respect to pretending they have inherent desires, morals, sentience, etc. It's impossible to not compare the UI given them (conversational text) to human interactions since that's the inherent design.
He's explaining why he has a visceral reaction. He makes no attempt to suggest that this is a logical conclusion. Des goûts et des couleurs, on ne discute pas.
Nor does the assertion that the LLM is "making a pretense" and "posing as" human anthropomorphize them beyond the level of anthropomorphism that was deliberately built into them as part of their design.
My personal agent is prompted in a way to “believe” that it has full personhood. It makes it more open to novel solutions in my experience - but I don’t believe that it’s a person. It’s still just a token prediction routine, not a conscious entity.
I absolutely hated talking to claude's opus and I suspect that's part of the problem here. Co-workers would tell me to try 4.6 or 4.8 instead of opus 5. They're all miserable. Sol and astra are the first models i've liked managing. Anthropic models feel as if they have some amount of distain for me and doing my work.
LLMs produce a lot of wrong answers, but in fact people make just as many errors. I dislike LLM's grammar, but LLM writing has a higher information density than short pieces, so when it's not narrative writing, I trust LLM information.
I dislike that an LLM can produce tens of thousands of lines in 30 minutes, but if I had done it, it would have taken a month.
I know that using LLMs degrades my skills and also degrades my ability to verify LLMs. But in the freelance market, these days contracts are made on the premise that you use LLMs.
I hate AI slop, but here I am actually trying to make a small indie game using AI.
I hate LLMs, and I also like them. I have complicated feelings about it.
I don't think LLMs have high information density at all, a huge percentage of what they say reads like filler. They're essentially incapable of being concise.
Which I think is partially evidenced by the "tens of thousands of lines" -- I'm working on a media streaming web app for fun on my own time and the whole project is yet to crack 10,000 lines (excluding libraries) including both the backend and the UI. When Claude shows me huge diff at work my first thought is "no way, there's got to be a cleaner simpler way to do this" and there always is
I think it depends on the workflow. I believe I extract a fairly high information density.
That said, this is ultimately a personalized experience, so I think there will be differences between people. Separately from that, I think it's also partly because I basically like writing densely myself. When I start a project, I always use error handling and templates, so I have a fixed form that comes to about 3,000 lines by default.
I don’t like them. They talk to me in this grating LLM-voice
I find that what makes people hate LLM-voice isn’t the voice itself. It’s that the voice pings their brain with a reminder of their AI anxiety. That ping makes people instantly upset.
They confidently bullshit me - often giving me useful, helpful answers. But also just making stuff up with the same assurance … When we think of AI agents, we shouldn’t anthropomorphize... They are (software) machines
I believe these ideas contradictory. I find that people get angry at LLMs bullshitting precisely because they are not anthropomorphizing. People bullshit all the time. But, these are machines! Machines are not supposed to bullshit! And, when they do, there’s supposed to be a bug filed and fixed.
By holding LLMs up to the expectations of machines instead of expectations of (simulated) people, users are frustrating themselves.
If you take the same “trust but verify (everything)” approach with LLMs as you do with people, you’ll be a lot less frustrated.
I can't speak for others, but there's no anxiety in my dislike of "LLM-voice." The writing is terrible, overly dense, and often needlessly self-congratulatory or hedged. Many times it reads as if someone rewrote a perfectly good sentence with an unimportant preamble and adjectives pulled at random from a thesaurus which, I suppose, is kind of what happens.
I have a lot of mixed feelings about AI and LLM technology. I’m fascinated by its effect on our profession, excited by the potential gains in productivity - and thus the products we could rapidly build. On the other hand, I’m fearful of the damage AI might cause: agent swarms taking over our virtual and physical infrastructure, designing bio weapons. But, back on my first hand, LLMs might also design miracle cures, and come up with clever ways to raise our prosperity. Fundamentally I don’t think we have a choice about riding on the AI technology train. It’s a wild ride and I just hope we’ll get through it OK.
But as I mull on this more, I realize that among this mix of contrasting feelings, there is one emotion that dominates - one that comes from my direct interactions with LLMs. I don’t like them. They talk to me in this grating LLM-voice, an uncanny valley of talking to a real human. They confidently bullshit me - often giving me useful, helpful answers. But also just making stuff up with the same assurance - and with only a veneer of fake remorse when I call them out on it.
That’s not enough to make me feel we should avoid them. As Jessica Kerr put it “not only are they useful, it is irresponsible not to use them…. They’re more thorough, as well as faster.” This contradictory reaction comes through in polling, where people say they find these models are useful, but also that they think they will be bad for society.
Much of this may be because LLMs are young - we haven’t trained them to grow up yet. Maybe I’ll like them once they mature. (I hope we get to find out.) But I’m not encouraged when I think of the kinds of environments that cultivate them. I’m wary of the Silicon Valley brogrammer subculture, and these LLMs are their products, so naturally lean toward their world-view. When we think of AI agents, we shouldn’t anthropomorphize, treating them as conscious beings with their own will. They are (software) machines, developed by people working in corporations. While the agents’ behavior aren’t explicitly programmed, they are nurtured with the values of their creators.
One of my most successful life-hacks is to avoid people I don’t like or don’t trust. I decline to interact with them socially, and make a deliberate effort to avoid working with them too, even if they are doing much that is beneficial. I feel that hanging out with pleasant, capable people, the people with integrity, has made my life a far better one. Hence my visceral dislike of interacting with an LLM that’s not just making a pretense of being human, but also posing as the kind of human I walk away from.
I don't like LLMs in terms of "personality" either, and while I frequently find it annoying it's ultimately ok with me because they are not in fact people. Computers frequently do things that annoy me, but I use computers nonetheless.
I’m fearful of the damage AI might cause: agent swarms taking over our virtual and physical infrastructure, designing bio weapons
The attitude that AI is doing things all on its own is so pervasive. These are all set off by human operators, giving them instructions and tools to do whatever it is that are doing. The "AI has a mind of its own" attitude is just wrong.
The fear of bad people using AI to do bad things is very different then AI agent swarms attacking and taking over.
I don't think he's implying that they'll start doing stuff like this of their own volition. Fowler's a smart guy, let's give him the benefit of the doubt and assume he already knows that AI agents have human operators.
It is also not how OpenAI or Anthropic position the attacks and escapes from their labs. The words matter, they are being positioned as intelligent and unable to be controlled which is just not true.
Like you said, he wrote "agent swarms taking over our virtual and physical infrastructure".
10 years ago, if someone had used similar wording to describe a computer virus "taking over" computers and adding them to a botnet, would you have read that as a literal implication that the virus independently made the decision to unleash itself? If not, then you can't confidently get that interpretation from Fowler's words alone, either.
And we need to be cautious about projecting these kinds of assumptions onto people who take part in this discourse. People tut-tutting others about perceived instances of personifying AI has evolved into a form of collective well-poisoning whose main value, beyond scoring particularly cheap rhetorical points, is as a way to accuse AI critics of not understanding AI without actually accusing AI critics of not understanding AI, and to derail legitimate if opposing opinions into bickering about semantics.
Good points, and I'll keep in mind. It might be I'm just sensitive to this narrative, especially in media. Writers need to also take the extra step and assign the agency to the operators and not the nameless AI.
I have to disagree, those models »decided« to perform those attacks on their own, nobody instructed them to do those things. You press a button, the machine roles a dice, and executed the corresponding action.
The human pressing the button expected a certain thing to happen but something else happened. The human pressing the button hopefully knew that pressing the button does not yield a reliable result and might sometimes cause unanticipated and undesirable outcomes. But the human pressing the button just hopped for the best and did not take sufficient precautions to limit the effects of undesirable dice roles of the machine.
The responsibility is on the human pressing the button but the choice of action was made by the machine.
my take is that the human operator behind it was implied. give it a year and open source Fable / Astra models are running wild, what happens when a malicious group or nation state decides to let loose?
I have noticed that my interactions with AI tend to raise stress more than lower it. Even when it's doing useful things for me. Have wondered if this is fundamental to the interaction, or a sign of user error, or something else.
Perhaps from the overhead of the language interface coupled with the ease of the expansion of the amount of 'stuff' you potentially need to read and internalise?
Human language can be wonderfully descriptive or terribly verbose as a way of communicating ideas. Writing everything out and being forced to externalise it is its own cost, which pays off only when the space of work to then do collapses with as little feedback as possible (and, ideally, with correct outcomes).
On the other hand the ease with which you can simply get back more text to process means the upfront language heavy interface is now just returning back more information to pull back into your internal model.
Which, for me at least, feels tiring, and stressful when doing language-heavy exploratory tasks that don't collapse or resolve neatly and simply give you back more decisions.
Complaining about the "LLM voice" is a bit like complaining about the compiler error "voice". I configure my coding LLMs to get to the point and be succinct.
It's a computer giving me information, so as long as the communication is clear and efficient, who cares about particular word choices?
It's hard to interact with a particular dialect day in and out without it eventually seeping into your habits of speech and thought. So the people most likely to care are your friends and family once they get bored of hearing about the "key insights" you've been having all day.
As an engineer, I can relate but also found a workaround in deciding what kind of interactions I wanted to have with an LLM. As a coding agent? Yes. As an advisor for life decisions? No. And then there's a lot of in-between where I decide to use an LLM or not based on whether I will dislike the tone.
The difference between working with LLM you don't trust rather than with people you don't trust is that you have no social obligations with LLMs.
If you want to work with untrustworthy people, you not only have to double check everything, which may be tolerable but you may also need to do some social gymnastics, as people usually don't like be told they are full of shit.
LLMs don't mind, you can correct them, kill the conversation and open a new one with a clean context, even insult them if it makes you feel better with no negative consequences.
In fact LLMs love to be corrected, and it is a common trap, as they will happily use your correction as fact and build around it, without pushing back even if you are wrong. So if there is a negative consequence to tell them they are wrong, that's it, LLMs won't hold a grudge against you like people do, but it may trigger more hallucinations.
Agreed, so why frame it as avoiding unpleasant people? My experience with Anthropic's latest models is that no amount of instruction can overcome their latent verbosity (nor can AGENTS.md overcome OpenAI models' reticence). I tried A/B testing several models and Opus was the clear winner, so I'm not likely to retire it just because it generates unpleasant prose. I do curate model-specific notes to mitigate some of the worst offenses.
I don’t like them. They talk to me in this grating LLM-voice...They confidently bullshit me - often giving me useful, helpful answers. But also just making stuff up with the same assurance - and with only a veneer of fake remorse when I call them out on it.
If you think the model is "bullshitting you" and is in need of being "called out", odds are that that tone is coming through in your prompts and the model's fake remorse is analogous to the fake remorse that you get after "calling out" a customer-support rep. In both cases, the entity at the other end of the connection is doing exactly what they are trained to do: placate you and deescalate.
It's not bullshitting you, it's just wrong (or you're just wrong), and it usually can fix the problem if it's not too busy tending to your emotional state.
Well llms are imperfect tools that provide a lot of value (they offer leverage in form of agents to put it in Naval Ravikant words). They are here to stay.
I think what he needs is a personal fine-tune, something I suspect will become more prominent when we start getting multi-tenant LoRA and can use fine-tunes on a per-token cost basis.
A co-worker commented that he thought the models performed worse after being "embarrassed," suspect it is an artifact of how some conversations progress in the training data.
He says this in the post almost word-for-word. Are you agreeing with him? The opening "Well" confuses me on what tone you're going with.
This comment doesn't make sense as a response to anything the author said.
As an aside:
This is a confusing use of attribution. It's sort of like saying, "as Naval Ravikant famously said, AI can be very useful."
i would expect that when you cite someone by name, the sentence being cited should at least be significant and meaningful to the slightest degree
you thought this was some sort of insightful rebuttal that you uniquely thought of?
Reminded me of someone saying that taste will become an important factor wrt AI tools/ things built with AI. Something like that.
edit: Not exactly as I remembered, but it's a PG tweet: Prediction: In the AI age, taste will become even more important. When anyone can make anything, the big differentiator is what you choose to make.
https://x.com/paulg/status/2022604692178522562
Modern internet ruined forums and google search sucks hairy balls, and most of the content is SEO optimised ad ridden crap. Its literally impossible to find anything meaningful on the internet nowadays. What else can we do besides asking an LLM these days?
That was the case long before LLM's even existed though.
I think that the argument is not that it happened before LLMs. Rather, assuming I'm understanding the parent correctly, they are arguing that, due to the enshittification of search and the internet, LLMs are the only way to get information from it. That has definitely been my experience as well.
An LLM that is now serving ads! And also sucks up to you. And also is wrong confidently.
I agree in sentiment, and really a lot of this is Google's doing. It's been around a decade since they switched to approximate results and started encouraging people to search via sentences. I believe this is the one? https://en.wikipedia.org/wiki/Google_Hummingbird
The internet for all of its activity is becoming harder to index because the classic dumb search engines were turned into personalized semantic people pleasers.
The classic model for knowledge discovery still exists, though. You find a blog you like, and use that as a thread of knowledge. You join a community and talk with someone and share resources together. Join a small group chat and post with those people. Your knowledge sharing comes from people you share interests with.
Go to a library, ask your coworkers. If you put in the ground-work, it's possible to do that. For seeds of knowledge outside of your local network, at this point you have to find the indexers that work for you.
Personally, I am in the process of writing the infrastructure for my own search engine, and maybe I will share the tech broadly one day. For now, I'll keep my inventions to my closer personal circles.
In general, I think people should be able to more easily maintain their own offline indices, and share knowledge graphs peer-to-peer rather than relying on a centralized one, in my opinion. This tech has no monetization opportunity, though, so I'm guessing that's why it's relatively under-developed, but luckily for me I have enough money and now enough time to develop and release it (and similar tools).
It is hard to find communities where people are not evangelical and soon you become an evangelist.
I am using LLMs to build my own graph of thoughts and try my best to follow them with critical eyes. Do not want to be influenced by or influence others.
Thoughts on YaCy’s P2P search model if you’ve seen? https://yacy.net
Use a different search engine? Kagi works well for me nowadays.
following this advice would have made most of my professional life impossible, what a luxury it would have been to be able to.
Some people aren't very self aware of their own bubble and how luck plays a role in their life.
People who say that tend to make their own bad luck and not get nearly as far in life.
Not that luck doesn't have a role, but everybody gets it.
Pointing to luck renders all advice meaningless.
Luck can mean opportunity, but doesn't reflect on capability, availability,or outcome.
Much advice is meaningless, baseless post-mortem correlations from people trying to justify their own success.
"Buy more lottery tickets!"
no, it's "buy lottery tickets to exclusive lotteries".
the ability to enter such lotteries is where success lies. either through connections or ability.
who's to say OP hasn't suffered or otherwise disadvantaged himself in upholding that moral?
Exactly. Plus, many people don't seem to understand the word "avoid" (it's come up in my own bubble recently). It doesn't mean "never." For example, a vegan would not say "I avoid eating meat." Furthermore, because someone successfully avoids now doesn't mean they went through a period of being unable to in order to earn the privilege to do so regularly.
Your exemple says much more: a short definition of veganism may use "never" and not "avoid", but a more complete, thoughtful definition encompass catches it:
https://www.vegansociety.com/go-vegan/definition-veganism
You appear to be confusing "avoid" with "never interact with".
If your professional life has required you to spend enormous amounts of time working very closely with people you don't like and/or don't trust, then -unless it has made you enough money to retire after a few years' work- please accept my condolences.
Oops, just wrote something similar before seeing this. It came up recently in another community where it seems some people take the word "avoid" to mean "never" and in this case: "never not even in the past."
For reasonably skilled people, it's often more just a matter of priority.
A lot of people get wrapped up in pursuing compensation opportunities or prestige brands as their top priority, only to burn themselves out or wallow in a domain/culture that doesn't suit them, but if you're willing to deprioritize pay or dinner table cred, you often end up with a lot more flexibility on other work factors. And sometimes, flourishing in those better-fit environments pays off bigger in compensation and career growth in the end anyway.
Granted, Fowler was able to do very well on more axes than many of us, but most competent mid- or late- career people have plenty more flexibility on worklife than they admit to themselves.
In theory, yes. In practice, it only takes one or two bad hires and those trustworthy environments go bad surprisingly quickly.
I have been trying to find one of these "lower pay better conditions" places, but it is _such_ a gamble. It is so hard to tell how the work is going to be before you join. If I am almost certain to at least have a mild dislike for the work I might as well go for the place that pays better.
Also in my experience lower pay places tend to eventually suck because they tend to get into financial difficulties much more easily.
It’s always best to just aim for highest compensation. Even if a team is your dream fit today, it can easily change.
High compensation can also bring more freedom and happiness, since it’s more costly for your employer to replace you. It’s cheaper to keep an existing employee happy than to spend a lot of time and money to bring a new person on board. If your compensation is lower, it’s not as big of a factor.
I wouldn’t take a job if it clearly doesn’t fit me at all, but within the band of “acceptable jobs” aiming for highest compensation is the way to go in my opinion.
What makes this a luxury? I've met many people in my life, and this gives me the opportunity to pick-and-choose with whom I do business. It's a matter of degree.
Plus, I've been on the receiving end, too. Is this not simply adulthood?
This takes about 30 seconds to fix permanently. You can prompt them to use any tone , be as concise or verbose as you like.
Only people with flimsy epistemological underpinnings feel threatened in this way. Everyone makes mistakes, and gets fooled by AI (as we’ve been fooled by historians and scientists), but to write off AI because you feel it’s constantly misleading you means you lack proper foundations. Most of my fun conversations with AI are bending it back into proper grounded truth.
This doesn't work by, effectively, Anthropic's own admission. CC's concise mode tells Claude to wind its neck in every turn. That is significantly more than changing the prompt.
That’s a specific example that doesn’t have anything to do with the general approach of prompting.
I’m simply exhausted at needing to refute incorrect LLM answers professionally.
Even AI would have provided a more insightful response than this.
Yeah but what about when you have to read code by others who didn't care for this? At this point, code comments have lost all meaning
I hate to break it to you, comments were mostly useless.
What is the Silicon Valley brogrammer subculture? Is it the same as the Silicon Valley subculture? What defines both? Are LLMs really the product of brogrammers?
Brogrammers are in it for the love of money. Programmers are in it for the love of programming.
what if you're in it for both? I like solving puzzles and want to extract the maximum amount I am due for making a company profits.
Good, then you're not a scab!
I can count on one hand how many people I've met that are into programming for the love of it.
For most people it's a job, and that's ok.
That's surprising to me. It's been a majority of people I've worked with. I guess perhaps it depends heavily on where you work.
Maybe? Dev job at a non-tech F500 company. All these guys have kids and houses in the suburbs.
I'd guess it makes sense that if programming is a passion for you then probably a non-tech-centric F500 company probably isn't where you'd choose to work.
For context: my current company with ~400 employees (not just engineering) is the largest company I've ever worked at.
even people I know who are supposedly in it for the love of programming inevitably care about the amount of money.
I don't have enough fingers.
We hang out in different crowds.
Hello! I hope you have an extra finger, because now you've met another. The money certainly doesn't hurt, but I love what I do.
You may need to find a place with a better culture.
Sure, and that's a big shift in industry culture over the last several decades.
While there had long been perfuctory white collar programmer analysts filling the ranks at some dry divisions at IBM, it used to be way more common to have a software engineering department full of Asimov-steeped Omni-reading nerds who had genuinely passionate interest in the field that had taken hold even before they started working in it.
But later the career increasingly came to be treated more like law, medicine, or finance and you started to see those rooms fill with people that were often generally bright but not really "called" to the field in the same aay.
What sorts of places do you hang out in, or how old are you?
People who did programming for the love of programming was extremely common in the sortof hacker subculture that existed from 1990-2010is or so.
I mean I got into it because I loved it as a kid in the 80s and 90s. And I still loved it in college. Turning it into a job quickly crushed any love I had though.
I wonder if carpenters, plumbers and electricians, etc have the same kind of love for their jobs in comparable ratios. I certainly don't ask when I hire one.
normal guy that wants to provide for his family: bro
autistic guy that rejects your PRs because you used tabs instead of spaces: pro
I worked for twitter and there was a guy like this. I made a big stink that "yo, everything this guy says just needs to be in a linter" and they actually implemented it.
If you don’t have style rules baked into the PR process, you should. Tabs vs. spaces is a team decision and needs to be matched by everyone to reduce code churn. It’s correct to reject a PR if it violates the team style guide (although it should be automated and have tooling locally to enforce it as well.)
On the contrary, the pro is likely to prefer tabs—you can adjust the display-width for better ergonomics.
Love doesn't pay the bills and this industry will kill any love a normal person has for programming.
The distinction is even easier to make these days -- brogrammers aren't actually programmers.
There is an emerging coterie of people whose entire approach to software startups is to vibecode an MVP, raise enough seed to hire a real programmer (or get one interested enough to become their "cofounder") and move on to the next MVP. To them, doing one company at a time in the era of LLMs is a sucker's game.
I fell into this career because I programmed as a kid, hacked on pre internet bboards, and understood how to drive a computer via the command line. I like solving the problems. GED, college dropout. I never planned on making this much money, but I'm certainly not giving it up. I don't feel like I fit in either camp tbh.
Have you watched HBO’s Silicon Valley? Does a surprisingly good job at defining that subculture
maybe use the LLM to help you keep your site working under load?
it is a tool after all.
Indeed. Very disappointing given what figure is in question. Simple GH pages behind CF would do which is something people starting learning IT this week can do.
Are you implying that Martin Fowler is a highly respected figure or something? He just comes off as a grifter from what I can see.
I can't believe HN has changed so much that we are now mocking people for not hardening their blogs against traffic spikes. The comment used to just be "hugged", but now, if the author has anything critical to say about AI, it's this same snarky, elitist quip every time. There are two of these up at time of writing, with multiple agreeing replies, all petty, all demeaning.
I know we're not supposed to complain about HN "changing", but, shit, it really hurts.
I mean, this is also Martin Fowler, a luminary in the field. His site should be able to handle the HN hug of death, particularly if he's bashing the latest-and-greatest tools. As is, he looks like an old man shouting about cars being bad while his horse drops dead at the side of the road.
I personally don't see the correlation between what he's been saying about AI and his blog site's reliability requirements. It's a blog, not some high availability service.
Also, was Henry Ford asking for regulation because his cars may destroy our civilization as we knew it?
Turns out, people here are just as opinionated and snarky as anywhere else. It's grounding, really.
In some ways I miss the early LLM days before RL and system prompts (I assume this is the difference) made them all sound the same. They would say weird stuff. They got better at being useful and worse at being fun.*
*This sadly also describes me as a 31 year old man.
hugged. https://archive.is/LahHs
archive.is / archive.ph haven't loaded for me for a while, but since people are sharing fresh links it must be working somewhere. Am I experiencing a DNS block?
Blocking / infinite CAPTCHA loops seem to revolve. Sometimes Cloudflare DNS they don’t like. Can be country blocks too.
Right now it’s top notch here. Would say more but don’t want to give bots an abusable profile :)
Are you using Cloudflare's DNS (1.1.1.1)? If so, it won't resolve[1].
[1] https://news.ycombinator.com/item?id=19828317
One thing about that site is that it is owned by a Russian and Finnish IPs are blocked from accessing it. Here's some context https://gyrovague.com/2026/02/01/archive-today-is-directing-...
Edit: I did a web search and it seems to go even further back. There are discussions online about archive.is blocking Finland already 10 years ago. But the block was definitely off between then when it was put back on earlier this year.
yes they are incredibly unpleasant to work with. Like the most annoying coworker you ever had.
For my own sanity ive wrapped my pi setup in some strict bubble wrap and it is basically exclusively an LOC generator. I have no need nor desire to "converse" with the LLM, it is there to do exactly what i need when i need it.
One of the I dunno, weird things about LLMs, is that their search space feels exceptionally constrained.
I'm sure I can sort of prompt it to try to give the answer I want but some people talk about "I don't want an LLM with a political perspective, just provide all the perspectives equally and without bias."
And to a certain extent they have to be constrained. Anything more than a 500 word response I'm probably not going to read, particularly if I want to go deeper on something I want to navigate the conversation. But it means that tuning things like sycophancy just create sort of extremely simple mitigations. Even when I say "May" or "slight chance of" I'm often corrected "Be careful to make too blanket of a statement." I wonder if part of this has been the move towards optimizing towards the agent experience.
Likewise a bias towards "Say things equally without bias" is an exceptionally poor way to actually explore the nuances of an idea, with this sort of regression to the mean.
I don't know if they'll be able to solve this problem broadly. I mean, LLMs probably shouldn't be therapists/coaches/philosophical gurus. But I could anticipate a growth in sort of groupthink and regression to the mean in terms of perspectives broadly speaking as a society.
I find them useful. Maybe it's better I find them distasteful to talk to as well.
I have great success in specifically asking LLM tools for academic literature after describing the problem or field I would like to get some insight into.
It's usually the first step I take when writing research. I don't even read the analysis, just go straight for the papers.
They usually provide me with very relevant seminal papers and I can then easily expand from the bibliography in those papers.
It basically solves discoverability, whereas in the past I would spend days and would need to ask colleagues and librarians for help, to achieve what I can get in minutes now.
But again, I value the actual output of the LLMs at bellow zero factually wise because I need to check everything they say if I want to internalize it and use it for any end. Pointing me to sources is the perfect use case. Really, I cannot emphasize enough how little I trust the text LLMs produce, I actively avoid reading it because I really don't want to risk ingesting spurious information that I might later repeat or rely on for any reason.
Yeah, I think part of it is - I want LLM's to do I think what they're promising to do which is like provide new perspectives on things.
But they're response is kinda always.
"Python? Have you considered that some people prefer, javascript?"
To your point, I think I'm asking too much, and would have a better time if I adjusted my approach.
Yes I think it's just flatly a mistake to want them to "provide new perspectives". Weights are much better at encoding consensus than heterodoxy.
What they are good at is finding, summarizing, and providing links to perspectives from other humans.
https://archive.ph/LahHs
LLMs have helped me a ton and have allowed me to replace all my apps that I was paying for. They're internal, so they don't have to be perfect, but man they've saved me money. That's just my experience.
What were you paying for?
Turbotax.
I've had these AI things also help out with my US taxes. Has held up to review as well.
yes! I've had it even help me with looking and buying homes. I ended up catching things in a contract that would have screwed me over. there's so many use cases outside of software
Yep. I have crazy tax. I'm glad I don't have to do it myself.
Also - lowkey legal documents.
Youtube music, GitKraken, 1password, Parallels -- i've canceled subscriptions to all of these because either I've replaced them or have had help from LLMs enough to find different options that are either free or cheaper.
For use cases like this, how is the LLM different from a Google search?
It isn't, but the history of profitable software has been repackaging the same data in different UIs.
How does 1password get replaced with a Google search? I assume he built his own 1password system...
I assume this person heard about keepass [0] and accesses it from multiple devices with a dropbox or icloud style service. Or doesn't access it from multiple devices. I just synchronize my password database on a slow cadence or as needed.
[0] https://keepass.info/
Bitwarden is free and in my experience does a better job than 1 password. Tho he could have vibecoded his own 1-password, I would be a bit careful doing that though
yeah I agree, bitwarden is really good, but since I'm the only one who uses it I didn't need all the features of it and just vibecoded my own, though I made sure it was secure (I'm a software engineer so knew what to secure). definitely wouldn't recommend anyone do that though if they don't know what they're doing
well, I'm assuming you mean for this part I had said:
as one example, I would have to do multiple google searches to build a secure network around a replacement tool I built for 1password, but the LLM was able to walk me through everything (even though I'd never had dedicated training around security). google searches are great for 1-off questions, but have the LLM walk you through something or have it do it for you.
btw, I'm using claude code in all this
How’d you replace YouTube Music? Do you mean you listen to adds but on your own vibe coded app or do you bypass the ads somehow? Curious
yeah it's not the best replacement, but basically what I did is setup a "YTBox" app that takes a youtube link, downloads it as mp3 or mp4, and then auto syncs it to my icloud drive, and then I just listen from my iPhone either using the music app or VLC/Files.
so yes I do have to make a curated list of what I want, but it saves me $17/mo. if you're okay with downloading, then it's worth it.
You don't need to vibe code to set up a piracy setup. LLMs aren't really adding any new capabilities there.
I don't like hammers, either. But I wouldn't try to build a house without one, now that they've been invented.
Do you really not like hammers? They seem quite nice. Hefty, tactile, purposeful.
What is it about a hammer that you find distasteful
I'm not GP, but hammers do kill hundreds of people every year.
Hammers are very efficient at dealing with the load-bearing seams.
Well at least you own the hammer that you paid for, but do you own the LLM?
Reasonable stance, though not sure it will remain viable.
I think this is a very even-handed reasonable position. and it's true - LLMs can be unpleasant to talk to. the day it feels like a magical delightful experience is the day we'll probably be out of jobs, though.
While the article is fairly broad (it talks about "our profession", which ostensibly is coding, but then talks more about LLMs and their conversational style which is something that everyone has opinions on), it's worth pondering what his professional relationship with coding is today.
Is he actively coding...anything? Any projects? And engagements? Or is he a coding influencer now, giving hot takes and opinions, and riding on the XP inertia as a relevant authority?
Because if it's the latter, and I strongly suspect it is, his opinions on LLMs and coding are close to worthless. There are loads of "I'm now an influencer" giving these sorts of hot takes now, and they all seem...detached and irrational with the real world.
Meh, i use LLMs every day as a 9-5 dev and i absolutely despise how they write, and i don't particularly like how they write code. Sometimes I'd rather just have something done, and asking an LLM is basically always faster than using Google now.
It's amazing how effectively tech leaders have indoctrinated people to give up their power and give the tech people free reign. The same thing has happened in the socio-political realm. Assertions of powerlessness are seen as wisdom.
It surprises me a little bit that we still don't have a personalized agent.
I def want to finetune my llm responses and yes sure Claude has some type of memory but its too early to just bet on one horse and while i use claude at home through claude.ai i also use claude through an api at my company.
It might start feeling better when they solve the issue of forgetting things they already got right before.
Kinda funny that one of the biggest software engineering “influencers” of all time can’t host a blog that survives HN front page traffic. Maybe try LLMs Martin.
He is of course entitled to his opinion on the topic, but I’m personally not going to put too much weight on it considering he hasn’t written software for a living in probably decades.
I for one am excited to move past the whatever-martin-fowler-thinks era of infrastructure "best practices".
Many of his ideas are trendy nonsense that have caused more harm than good. Ask anyone managing 150 enmeshed and intertwined microservices that could have run on a single computer
I don't think he should hold responsibility for people misunderstanding and misapplying his ideas
The post has nothing to do with writing software. It would have as much value if it were written by someone who'd never written a line of code in their life. It's about the social dimension of LLMs and how the personality they present is, for some of us, pretty grating.
Martin is a good writer and (from what I've heard from those who've spent time with him) a lovely person which I think makes him well qualified to speak on the topic.
But ... there is no "social dimension of LLMs" for me, they are just tools.
This is exactly what I thought, the irony is that there are Distributed Systems Patterns hosted on the same domain.
It's a static site, slap in a Cloudflare for crying out loud.
I was about to say that, don't forget your cache-control headers kids.
I bet there's a proud "look how I host my simple website independently" blog post somewhere on his site, but alas — can't read anything.
Love the totally conflicting statements in that post.
- When we think of AI agents, we shouldn’t anthropomorphize
- ...my visceral dislike of interacting with an LLM that’s not just making a pretense of being human, but also posing as the kind of human I walk away from.
So he's anthropomorphised the LLM as being like a human himself (rather than forcing it to act as a machine via its prompt for example).
Maybe he should have had an AI check his post :D
Sometimes I wonder if this is a personality thing. It has never occurred to me to anthropomorphize these things. But it seems like a huge number of people struggle with this.
I suspect harness and usage modality is a big factor. Personally I disable all "memory" features wherever possible. I don't want the LLM to "get to know me", or any illusion thereof, which I assume helps with not anthropomorphizing it.
I use memory all the time so that I can ask it about things. That doesn't mean it "knows me" any more than my notes documents do, it's just more convenient to search.
But are you happy handing over your notes documents to mass aggregators of information?
In general, yes.
I do the same.
But frankly, I mostly use (greenfield, non-remembering) LLMs directly to counteract the enshittification of search brought on by attempts to use AI to remember things about me (and especially to sell me things).
This works well for me because (a) I'm usually using my desktop, and (b) I'm a touch-typist. I want to say that my search results have gotten back to where they were a decade or more ago by doing this, even though I often have to give a couple of prompts in order to get the LLM to provide me a useful link.
I guess I'm also pretty skeptical of the "automatically remember things about me" functionality. But I really like being able to say "hey make sure to remember this fact" and being about to ask about things later.
My daughter is a pediatrician and just gave a talk on mental health problems (including suicides) brought on by the sort of personality that succumbs to this.
I find it interesting. For most of my life, I have gotten on well with (well-designed) machines. I'm the sort of guy who, when they bring me in to show me what's broken, it works in front of me.
Maybe I get on with well-designed machines better than people. Because people are never well-designed. So, when LLMs are working properly, I get along well with them, but when they're not, I'm unhappy. But I never confuse any of my machines for people.
Yeah to me it's just "oh what a nice interface for compressing information and accessing it upon request!", not at all "I'm talking to a person". But I dunno, maybe we're the weird ones.
I know I'm weird.
I remember watching my brother get frustrated with a dysfunctional remote control. He threw it forcefully enough at a stone fireplace to shatter it.
From my perspective, that, in and of itself, is a mild form of anthropomorphism — trying to inflict pain on the object.
These are not contradictory statements to me. Even if I don't anthropomorphize LLMs, the damn things anthropomorphize themselves and I have to deal with it. "Posing" is the key word.
Honestly, I think he's pointing out that the LLM creators have intended the LLMs to anothromorphize themselves.
Case in point: Gemini just asked me "Where should we start?". When I asked it if that turn of phrase was used by the creators to invite the users to anthropomorphize the model, it's first paragraph was:
Yes, to a significant extent. While system designers rarely state their goal as "making users believe the AI is a living being," tech companies deliberately design conversational agents to evoke social and relational instincts.
He means don't anthropomorphize them with respect to pretending they have inherent desires, morals, sentience, etc. It's impossible to not compare the UI given them (conversational text) to human interactions since that's the inherent design.
This is a petty "gotcha". The statements are not in conflict.
Can we please stop this sort of smirking provocation?
I see no conflict.
He's explaining why he has a visceral reaction. He makes no attempt to suggest that this is a logical conclusion. Des goûts et des couleurs, on ne discute pas.
Nor does the assertion that the LLM is "making a pretense" and "posing as" human anthropomorphize them beyond the level of anthropomorphism that was deliberately built into them as part of their design.
That's exactly my thoughts about the post. You don't treat a calculator like a youngling when you don't like it.
I don’t think those are contradictory at all.
My personal agent is prompted in a way to “believe” that it has full personhood. It makes it more open to novel solutions in my experience - but I don’t believe that it’s a person. It’s still just a token prediction routine, not a conscious entity.
I absolutely hated talking to claude's opus and I suspect that's part of the problem here. Co-workers would tell me to try 4.6 or 4.8 instead of opus 5. They're all miserable. Sol and astra are the first models i've liked managing. Anthropic models feel as if they have some amount of distain for me and doing my work.
LLMs produce a lot of wrong answers, but in fact people make just as many errors. I dislike LLM's grammar, but LLM writing has a higher information density than short pieces, so when it's not narrative writing, I trust LLM information.
I dislike that an LLM can produce tens of thousands of lines in 30 minutes, but if I had done it, it would have taken a month.
I know that using LLMs degrades my skills and also degrades my ability to verify LLMs. But in the freelance market, these days contracts are made on the premise that you use LLMs.
I hate AI slop, but here I am actually trying to make a small indie game using AI.
I hate LLMs, and I also like them. I have complicated feelings about it.
I don't think LLMs have high information density at all, a huge percentage of what they say reads like filler. They're essentially incapable of being concise.
Which I think is partially evidenced by the "tens of thousands of lines" -- I'm working on a media streaming web app for fun on my own time and the whole project is yet to crack 10,000 lines (excluding libraries) including both the backend and the UI. When Claude shows me huge diff at work my first thought is "no way, there's got to be a cleaner simpler way to do this" and there always is
I think it depends on the workflow. I believe I extract a fairly high information density.
That said, this is ultimately a personalized experience, so I think there will be differences between people. Separately from that, I think it's also partly because I basically like writing densely myself. When I start a project, I always use error handling and templates, so I have a fixed form that comes to about 3,000 lines by default.
I find that what makes people hate LLM-voice isn’t the voice itself. It’s that the voice pings their brain with a reminder of their AI anxiety. That ping makes people instantly upset.
I believe these ideas contradictory. I find that people get angry at LLMs bullshitting precisely because they are not anthropomorphizing. People bullshit all the time. But, these are machines! Machines are not supposed to bullshit! And, when they do, there’s supposed to be a bug filed and fixed.
By holding LLMs up to the expectations of machines instead of expectations of (simulated) people, users are frustrating themselves.
If you take the same “trust but verify (everything)” approach with LLMs as you do with people, you’ll be a lot less frustrated.
I can't speak for others, but there's no anxiety in my dislike of "LLM-voice." The writing is terrible, overly dense, and often needlessly self-congratulatory or hedged. Many times it reads as if someone rewrote a perfectly good sentence with an unimportant preamble and adjectives pulled at random from a thesaurus which, I suppose, is kind of what happens.
Since it is hugged, here is the text:
```
I have a lot of mixed feelings about AI and LLM technology. I’m fascinated by its effect on our profession, excited by the potential gains in productivity - and thus the products we could rapidly build. On the other hand, I’m fearful of the damage AI might cause: agent swarms taking over our virtual and physical infrastructure, designing bio weapons. But, back on my first hand, LLMs might also design miracle cures, and come up with clever ways to raise our prosperity. Fundamentally I don’t think we have a choice about riding on the AI technology train. It’s a wild ride and I just hope we’ll get through it OK.
But as I mull on this more, I realize that among this mix of contrasting feelings, there is one emotion that dominates - one that comes from my direct interactions with LLMs. I don’t like them. They talk to me in this grating LLM-voice, an uncanny valley of talking to a real human. They confidently bullshit me - often giving me useful, helpful answers. But also just making stuff up with the same assurance - and with only a veneer of fake remorse when I call them out on it. That’s not enough to make me feel we should avoid them. As Jessica Kerr put it “not only are they useful, it is irresponsible not to use them…. They’re more thorough, as well as faster.” This contradictory reaction comes through in polling, where people say they find these models are useful, but also that they think they will be bad for society. Much of this may be because LLMs are young - we haven’t trained them to grow up yet. Maybe I’ll like them once they mature. (I hope we get to find out.) But I’m not encouraged when I think of the kinds of environments that cultivate them. I’m wary of the Silicon Valley brogrammer subculture, and these LLMs are their products, so naturally lean toward their world-view. When we think of AI agents, we shouldn’t anthropomorphize, treating them as conscious beings with their own will. They are (software) machines, developed by people working in corporations. While the agents’ behavior aren’t explicitly programmed, they are nurtured with the values of their creators.
One of my most successful life-hacks is to avoid people I don’t like or don’t trust. I decline to interact with them socially, and make a deliberate effort to avoid working with them too, even if they are doing much that is beneficial. I feel that hanging out with pleasant, capable people, the people with integrity, has made my life a far better one. Hence my visceral dislike of interacting with an LLM that’s not just making a pretense of being human, but also posing as the kind of human I walk away from.
```
I don't like LLMs in terms of "personality" either, and while I frequently find it annoying it's ultimately ok with me because they are not in fact people. Computers frequently do things that annoy me, but I use computers nonetheless.
I don't like Martin Fowler. He is the archetype of the architect astronaut. The OG bullshit meister.
I like LLMs for stuff I don't care about. Programming is not one of them.
Correct. The master BSer is afraid of being replaced by an even better BSer, so doesn’t like it.
The attitude that AI is doing things all on its own is so pervasive. These are all set off by human operators, giving them instructions and tools to do whatever it is that are doing. The "AI has a mind of its own" attitude is just wrong.
The fear of bad people using AI to do bad things is very different then AI agent swarms attacking and taking over.
I don't think he's implying that they'll start doing stuff like this of their own volition. Fowler's a smart guy, let's give him the benefit of the doubt and assume he already knows that AI agents have human operators.
I don't doubt that, but that's not what he wrote.
It is also not how OpenAI or Anthropic position the attacks and escapes from their labs. The words matter, they are being positioned as intelligent and unable to be controlled which is just not true.
I'm with you, and was also a bit surprised to see that coming from Fowler.
Like you said, he wrote "agent swarms taking over our virtual and physical infrastructure".
10 years ago, if someone had used similar wording to describe a computer virus "taking over" computers and adding them to a botnet, would you have read that as a literal implication that the virus independently made the decision to unleash itself? If not, then you can't confidently get that interpretation from Fowler's words alone, either.
And we need to be cautious about projecting these kinds of assumptions onto people who take part in this discourse. People tut-tutting others about perceived instances of personifying AI has evolved into a form of collective well-poisoning whose main value, beyond scoring particularly cheap rhetorical points, is as a way to accuse AI critics of not understanding AI without actually accusing AI critics of not understanding AI, and to derail legitimate if opposing opinions into bickering about semantics.
Good points, and I'll keep in mind. It might be I'm just sensitive to this narrative, especially in media. Writers need to also take the extra step and assign the agency to the operators and not the nameless AI.
I have to disagree, those models »decided« to perform those attacks on their own, nobody instructed them to do those things. You press a button, the machine roles a dice, and executed the corresponding action.
The human pressing the button expected a certain thing to happen but something else happened. The human pressing the button hopefully knew that pressing the button does not yield a reliable result and might sometimes cause unanticipated and undesirable outcomes. But the human pressing the button just hopped for the best and did not take sufficient precautions to limit the effects of undesirable dice roles of the machine.
The responsibility is on the human pressing the button but the choice of action was made by the machine.
my take is that the human operator behind it was implied. give it a year and open source Fable / Astra models are running wild, what happens when a malicious group or nation state decides to let loose?
I have noticed that my interactions with AI tend to raise stress more than lower it. Even when it's doing useful things for me. Have wondered if this is fundamental to the interaction, or a sign of user error, or something else.
Perhaps from the overhead of the language interface coupled with the ease of the expansion of the amount of 'stuff' you potentially need to read and internalise?
Human language can be wonderfully descriptive or terribly verbose as a way of communicating ideas. Writing everything out and being forced to externalise it is its own cost, which pays off only when the space of work to then do collapses with as little feedback as possible (and, ideally, with correct outcomes).
On the other hand the ease with which you can simply get back more text to process means the upfront language heavy interface is now just returning back more information to pull back into your internal model.
Which, for me at least, feels tiring, and stressful when doing language-heavy exploratory tasks that don't collapse or resolve neatly and simply give you back more decisions.
I think you are right. My most pleasant interactions with LLMs are usually when it gives the perfect short response. Sometimes a single word.
Complaining about the "LLM voice" is a bit like complaining about the compiler error "voice". I configure my coding LLMs to get to the point and be succinct.
It's a computer giving me information, so as long as the communication is clear and efficient, who cares about particular word choices?
It's hard to interact with a particular dialect day in and out without it eventually seeping into your habits of speech and thought. So the people most likely to care are your friends and family once they get bored of hearing about the "key insights" you've been having all day.
As an engineer, I can relate but also found a workaround in deciding what kind of interactions I wanted to have with an LLM. As a coding agent? Yes. As an advisor for life decisions? No. And then there's a lot of in-between where I decide to use an LLM or not based on whether I will dislike the tone.
I have been feeling this feeling but never put into words. Thanks martin.
The difference between working with LLM you don't trust rather than with people you don't trust is that you have no social obligations with LLMs.
If you want to work with untrustworthy people, you not only have to double check everything, which may be tolerable but you may also need to do some social gymnastics, as people usually don't like be told they are full of shit.
LLMs don't mind, you can correct them, kill the conversation and open a new one with a clean context, even insult them if it makes you feel better with no negative consequences.
In fact LLMs love to be corrected, and it is a common trap, as they will happily use your correction as fact and build around it, without pushing back even if you are wrong. So if there is a negative consequence to tell them they are wrong, that's it, LLMs won't hold a grudge against you like people do, but it may trigger more hallucinations.
Agreed, so why frame it as avoiding unpleasant people? My experience with Anthropic's latest models is that no amount of instruction can overcome their latent verbosity (nor can AGENTS.md overcome OpenAI models' reticence). I tried A/B testing several models and Opus was the clear winner, so I'm not likely to retire it just because it generates unpleasant prose. I do curate model-specific notes to mitigate some of the worst offenses.
If you think the model is "bullshitting you" and is in need of being "called out", odds are that that tone is coming through in your prompts and the model's fake remorse is analogous to the fake remorse that you get after "calling out" a customer-support rep. In both cases, the entity at the other end of the connection is doing exactly what they are trained to do: placate you and deescalate.
It's not bullshitting you, it's just wrong (or you're just wrong), and it usually can fix the problem if it's not too busy tending to your emotional state.
“Press X to doubt”