This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
This is a continuation of a topic brought up in one of the AAQCs for January. Hat tip to @birb_cromble
I value and believe what @birb_cromble wrote. I think AI is both over and under hyped (more on that below). I believe birb's report that a team of good devs are looking at it and saying "wtf ... this is ... ok ... maybe?" I think @RandomRanger had a similar comment that I am struggling to find (although, to be fair, it was pointed out that Ranger was using copilot which is a known dumpster fire).
On the other hand, I have direct, personal experience with AI (to be specific, as I kind of hate the blanket term, "AI", coding oriented LLMs) writing good code quickly and accurately. I've had past colleagues far more gifted than myself send 11pm "holy shit" texts based on their own projects. The head of Anthropic, has publicly stated that LLMs write 100% of the code at Antrhopic now. And the guy behind ClawdBot / MoltBook (or whatever its called now) has openly discussed how his own deployment of ClawdBot was thinking and executing ahead of him.
If it's all hype, it is the mother of all hype cycles and something that approaches a mass movement of hysteria. This would be outright falsehoods and lying on a level usually reserved for North Korean heads of state and Subsaharan cult leaders.
I don't think it's that. I am, however, developing the idea that both sides are actually right at the same time in different directions. To explain that, we're going to have to talk about software and software companies a little bit.
1. CRUD
Create, read, update, and delete or "CRUD" is what is at the core of almost every piece of software that is above the operating system level. CRUD is definitely at the core of almost every piece of software that is sold from one company to another (business-to-business or b2b) and most software sold to customers (business-to-customers or b2c). There are exceptions, of course, some of them quite large. But the fact remains that most software is about having data somewhere, storing it, asking it questions, modifying it (and unmodifying it), and, perhaps, deleting it (note, however, that with storage being fundamentally cheap now, deletion is a kind of philosophical state. Your e-mails for instance, are often not deleted until you double-for-serious-delete-them and then wait 30+ days).
A junior developer can build a CRUD app on their computer at home in less than a week. By hand, from scratch, zero LLM involved. Building a CRUD app is often a final assignment for mid-level undergraduate CompSci work. You, yes you, can build a CRUD app today with one good, long prompt to any of the big LLMs. It will be complete, with minimal to zero bugs.
Salesforce, at its core, is a CRUD app. Salesforce is worth almost $200 bn while the CRUD app you build is worth exactly nothing. Why is this?
2. Enterprise
The holy grail of all b2b software is their first enterprise customer. What defines "enterprise?" It's a bit of squishy term, but it means a big company. 1,000+ employees is more or less agreed upon as the minimum, though this may vary depending on the market niche you're in. Why are enterprises so prized? Because you're selling your product at scale (usually in terms of individual user licenses or "seats") to a customer who can pay a six, seven, or even eight figure annual bill without worrying about it and will not switch to one of your competitors quickly (....usually). This is where b2b software companies get their explosive valuations from and where founders get capital-F Fuck you money. Salesforce, our CRUD app supreme, has enterprise deals, probably, with every F500 company and thousands more very large companies. They recently announced a deal with the U.S. Army (lol, ELLE-OH-FUCKING-ELLE to that one). Salesforce has more enterprise than a Star Trek reboot.
But isn't an enterprise CRUD app still a CRUD app?
Yes, yes it is. But it's a CRUD app that;
To return to the CRUD app you just built at home, it works just fine on your laptop! Can it export seamlessly to Excel or Word? No. Can I log into it remotely from my laptop while I am in the Delta lounge at O'Hare? No. What if four people want to work on it together at the same time. Uh, no - you don't even have a login into it! You just start it and boom, you're CRUD-ing around.
So much of the value of "big" software is all of the non-core functionality that is bolted on top of it in overlapping layers. This is also the dirty secret of what a lot of FAANG engineers do - write integrations between one product or service and another. They are not thinking up the next killer app, but essentially acting as digital plumbers in the world's largest city.
In the startup world, core functionality is often complete within the first year or two. It kind of has to be to gain your first customers. Then, so much of "product development" is figuring out where you're going to spend your time building integrations and then balancing that against actual new feature requests. The smart product managers realize that they can unite those two things and integrate a new feature from a different product. Two birds, one stone, zero actual innovation. Give that man a promotion.
There was a unicorn that literally was an integration hub for different products and services.
3. New vs legacy software
This is where we start to get into "both sides may be right" territory. From my experience, it seems AI is now quite good at writing new software, even fairly complex systems. It can do this because it doesn't have to make any assumptions about how anything already works. If it makes assumptions based on the user's intent, it is usually decent at carrying those assumptions through development to the finished product. In cases where it is not, you, the human, have to debug. Debugging, in this case, however, is often no harder than saying "Hey, this part doesn't work, and I think it might be because of xyz..."
This is not the case when you deploy AI against a legacy codebase, which is exactly what @birb_cromble mentioned. This is because legacy codebases are evolutionary products of a system changing over time. Ideally, each major upgrade - and even the minor ones too - to a system are documented. What "documented" means, however, varies wildly across developer teams. For sometimes, it's nothing more than a quick changelog of bullet points. For other teams, they write about the decision making process that led to changes. Most documentation is incomplete or somewhat ambiguous. I would argue that, right now, almost all legacy documentation is in no way written for LLMs to use well in their context windows.
4. Documentation
Unless it is. That link is to a good blog post on the recent fracas at Tailwind labs. Tailwind labs makes software and gives its core functionality away for free. This is the same model as Red Hat linux. They make money by having developers realize that they, Tailwind, have already built premium features on top of the core and will sell those features and hosting to companies that want it. I actually really like this so called "open core" business model because I think it's philosophically more in line with OG software ideals. Linux and its various derivatives have been free - in some form - since the 1970s, and the world's infrastructure runs on it. If Linux had been locked down from the start, I am convinced computers would still be weirdo specialty scientific equipment.
Anyways, back to Tailwind. Tailwind had to lay off about 75% of its staff because AIs read their whole documentation - which was very, very good - and can, now, build all of the premium services on their own. This fucking sucks, it's bad, nobody likes it. OpenSource is a necessary part of the software ecosystem. Even the most evilest of the FAANGS pour millions of dollars into sponsoring open source projects every year - because they rely on lots of those projects in their own code bases. Now, however, LLMs that scrape the internet, potentially, pose an existential threat to opening up your documentation plus codebase. It's as if you've just created one million free forever expert devs. Furthermore, this also exposes a dark pattern. If you want to retain your IP, lock down your documentation, intentionally obfuscate it, or just don't post it and only support your product with bill-per-hour in-house tech support teams.
The good news, however, is that most documentation is such shit that this will not happen.
But let's return to the main thread: AI under and overhyped at the same time.
My suspicion with @birb_crombles code base is that it isn't completely documented. This is absolutely NOT a shot at birb. I say this because, for any legacy code base, it is essentially impossible to build and maintain complete documentation that describes not only how the system operations, but how it evolved over time. This is valuable and necessary context for an LLM. All of the assumptions it makes about various libraries and modules can be very, very wrong because it doesn't have the legacy "evolutionary" documentation to inform it of various design choices and modifications. Birb and his team have that context as tacit knowledge in their brains and shared collective intelligence. "Hey why does thing x do action y?" , "Oh, team A needed that special feature so they could do necessary report z" , "cool, got it." That 10 second exchange across the the aisle with another dev is worth approximately 1 million lines of well written context to an LLM (1 million may or may not be an exaggeration.)
Birb said as much in his post. He wrote:
Bravo, Birb! I mean this sincerely. Phrased differently, Birb is saying that once his team provided extra-context documentation, the LLM was performant. However, by doing so, his team pretty much arrived at a state where the fix was obvious and easy.
Very well done documentation does lead to this. However, documentation is literally endless if you want to cover not only the system now but how it evolved over time. Good technical writers at easily $100k+ and they are necessarily slower than writing new code. Most companies will not invest in this because, economically, they can't.
4. Ships and Planes
Existing legacy software is like a ship. It's big and slow, sure, but it's moving a lot of mass and is more or less steady and stable. One-shotted LLM applications - like Clawdbot - are like planes - fast, soaring, sexy, and, sometimes, they crash spectacularly. The thing to point out, however, is that planes cannot move, economically, the bulk that a ship can. What I mean here is that all of the evolutionary design choices, system revisions, and tacit knowledge that a legacy codebase reflects is a very bad payload to deploy an LLM against. There are too many unknown unknowns and relationships that are hidden so as to be very improbable. An LLM is a probabilistic machine, so it relies on what makes sense on average - not what is real in a specific circumstance.
But deploying an AI against the clear blue sky (like a plane) is its most advantageous arena because it can just assume the average and build the thing from scratch.
Big, legacy CRUD apps - and, absolutely, more specialized apps - aren't really in danger of being disrupted by AI in the immediate future. 5 to 7 years from now, ehhhh, I am not so sure. The folks who are absolutely totally fucked as in right now, today are any startups that have launched a CRUD app with the idea that they'll do all the dirty work of building it into an enterprise offering. The market for that is quickly evaporating. Instead, internal tool teams will just use LLMs to make their own CRUD app, wrap it in their existing security etc. stack and use it internally. This may equate out to as much as $250k of combined labor hours and API credits but, 1) that would be at the high end and 2) that would be a one time cost (besides internal maintenance) instead of the the recurring six, seven, eight figures of spend to a third party.
5. Conclusion
I hope I've done a reasonable job in showing how both sides are right. I believe @birb_cromble. I believe, because I see, that pretty big names in software, who were even AI skeptics (roon on twitter, for instance) are now admitting to 100% agentic coding. The difference is in the starting point and the legacy debt or bulk that a given party engages with.
I thought I was pretty far out on the 'singularity soon' wing of this website? In my experience AI is quite good for writing code, whether that's CRUD or more interesting code like pathfinding or O(n) tier operations or even writing out procedurally generated shaders and effects.
Not perfect, it does struggle and choke a bit right now on the more advanced or fiddly things... But what happens when it starts directing a 1000 subagents to attack your million line monstrosity of legacy code? What happens when it can error-test better?
More options
Context Copy link
With Tailwind it's less their specific documentation, it's more that it became an industry best practice for new projects. There is just so much Tailwind content on GitHub and in blogposts. It's also specifically targeted by the LLM teams developing models.
Claude actually knows Tailwind better than CSS and will sometimes try to use it in projects that don't have it installed.
I think that scraping public GitHub repos is actually more important to LLM performance than documentation about your specific project. That all gets baked into the core model. If you're doing something with a lot of public examples it will one shot it.
I have a specific example. I've been playing around with implementing a db compatible clone of themotte/rDrama in node / react to get around some of the issues the codebase has. Two slightly incompatible markdown renderers on the front and back, old school bootstrap modals, etc.
I mentioned themotte/rDrama in my instructions to Claude Code and it put in some very rDrama like features such as coloured indent level bars on comments.
So it was clearly aware.
As a result a lot of the benchmark projects people try to use to document model performance become useless. The model can look up public examples of the answer.
LLMs are very good at the sort of thing that really should have been automated by now anyways. eg converting nested json object from a POST request into rows in SQL tables.
When you're doing something less common it has a lot more trouble. It does seem a lot better at working on my minecraft mod project than it was six months ago. That's probably due in part to scraping the public repo of the mod itself. It has a rough image of the working endpoints without needing to look at any context.
I suspect that offline models will become good enough in the near future that large legacy projects will be able to fine tune a model against the codebase.
More options
Context Copy link
This is all true now, but your '5 to 7' year timeline seems long! LLMs were not anywhere near where they are now 2 years ago, and with simple extrapolation 2 years from now I think it's as likely as not they'll be able to handle legacy code just fine, just like humans can.
More options
Context Copy link
I wish I had the time to respond more deeply to both this thread and birbs. But I almost completely agree with you all.
I'm now of the opinion that I'll be able to amass enough capital to retire before AI takes my job.
Legacy codebases simply require too big of a context window for an AI to appreciably absorb, and they never have high enough quality documentation available even if they did. They also often deal with systems that simply aren't connected unless you provide them agentic capabilities that don't exist yet. That's assuming you can get your security guy, whose entire job is to say "no" endlessly without ever providing solutions to problems, to let a robot crawl through your network and source code.
I'm literally proposing a 7-8 figure project around this tomorrow. There's a 30+ year-old AS/400 working alongside thousands of Access databases, hundreds of SQL Servers, and dozens of client applications. Each of these abstraction layers has hidden business logic, various interchange patterns, and bugs that have become features over 3 decades.
Even once we've disentangled a tiny part of the elephant to improve the state of play and build modern REST APIs to serve various client applications and partners, then there's a cadre of humans who have worked with these systems for so long that they need to gently be brought into the light, which takes far more money and time than anyone would care to admit.
This is one company of thousands. They're a little behind the baseline, but not by much. I've got another decade of dealing with the hardest problems in software and ahead of me if I can keep stomaching how endlessly frustrating it can be.
Please provide a trigger warning before typing this out so that SysAdmins can make sure they have their therapy anime body pillow with them.
More options
Context Copy link
More options
Context Copy link
This is interesting and timely for me. We have a legendarily dysfunctional QA team at my company, and as the DevSecSREInfraPlatformOps manager, my biggest beef with them is that they have been killing our Lead Time to Release (time between dev having an idea and that idea getting released as a feature for customers). They manually test almost everything, do nonsensical "verifications" (mocking responses from external APIs when they could just... use the APIs), and don't know how to code at all. Dev and automated testing might finish in a few days, but QA takes 2-3 weeks(!) and they batch multiple changes together which confounds results. They have been given a half dozen opportunities to change and learn but they have always made excuses or refused. At one point they even had convinced a (former) director that they needed outsourced QA members to help with the workload -- and then promptly shifted 95% of their work to the outsourced QA!
Just this past week, I asked for a Claude Code account and started trying to vibe code a replacement for our QA team, partly out of curiosity, partly out of necessity (we have an OKR to reduce lead time), and just a little bit out of spite. I was not optimistic because this is a very poorly documented 10+ year old codebase cobbled together by devs who have all since resigned to escape the mess they've created.
First, I told Claude to pull down all the test suites described in Qase and cache them locally. Then I told it all of the paths to local copies of our frontend, backend, mobile, and infra repos. I asked it to analyze each one. Then, I asked it to begin writing tests for each Qase suite, starting with the simplest ones like "login." Sometimes it would get confused (it shows you its thinking) and I would interject a message (you can send messages while it's thinking, unlike other LLMs) to explain some important bit of knowledge. Eventually, after repeating the same info several times which had apparently gotten lost in the context window compaction process, I told it to create a file called TRIBAL.md to record all of these contextual bits I was telling it that were not evident from simply reading the code. I also had it write a CLAUDE.md that points to all the repos, instructs it to read and update TRIBAL.md, BUGS.md and TODO.md, and contains descriptions of other tools it has created for itself (helper functions, data seeding scripts, env vars and credentials, etc).
So far I don't think I've written a single line of code and it has automated 85% of web QA tests. I have had the test vetted by code rabbit and I plan to check them manually before release. I am quite impressed, though I'm still curious to see how it will try to handle mobile testing. Claude Code really is next level compared to Gemini, Copilot, or Grok.
All that said, I am very aware that this project is probably riddled with false assumptions and nonsense code. I am still not optimistic about the final result, although given how dire QA is, our director might try it out anyway just to see if we can reduce our dependence on them. Either way, it's been a good way to get familiar with a SOTA coding LLM.
More options
Context Copy link
As a related thing, I saw this article/video about vibe coding: https://atmoio.substack.com/p/after-two-years-of-vibecoding-im
I think I really agree with one of the core points, which is that the AI agent is really really good at making the diff it proposes look good, especially to the person who asked for it. But in the perspective of the entire project, or from the perspective of someone who didn't ask for those changes, the code is unbelievably retarded. In some sense, AI slop is a scissor statement cutting between the person who asked for it (who the AI is trying to please) and everyone else in the world.
I've had a lot of success vibe coding extremely self contained components for a larger application, such as a static web page that wraps an already existing api endpoint that I don't need to add more features to - the endpoint is stable and done. But on the other hand when AI makes changes deep into the internals of business logic, the code is absolutely dogshit garbage.
My assumption is that any large-ish vibe coded application that's not a bunch of self-contained parts is going to be completely unmaintanable, halfway broken, and just all around awful.
You can have the LLM describe the business logic as it's developing and create unit tests to confirm the output is correct and pair that with manual testing as well.
I've created an application with fairly involved business logic using Claude Code and it works really nicely - it simply required using good 'ol fashioned SLDC aka Design, Code, Test, Document.
Is it an enterprise level app with the ability to scale to 10000 concurrent users and SOC2 Authentication and privacy standards? No.... but those types of big applications require multiple teams. If I gave four or five competent software engineers Claude and divided up the work I think you could rebuild a complex application in a lot shorter time.
The idea that you can't use LLMs on business logic is like saying you can't use a car to drive across the country - sure it was true in 1920 but today with the interstate it's easy. Right now it's 1930 or so... but it won't remain 1930 forever.
The LLM will cheat and write test that don't test what they claim to, bypass tests when they fail, write docs that make zero sense, write code that doesn't follow the docs etc.
To be fair humans will do all of these things too. But the code that LLMs write is mind numbingly retarded. If anyone with a modicum of coding experience took a look at your slop app, he would only see an awful, unmaintainable mess. And every AI change you throw at it only increases the debt.
As I said, AI slop is a scissor statement cutting you, the slopmeister, off from everyone else in the world.
OK but that's why you do manual testing... also you can always ask other models to check on the work.
IDK I shared the code with my friend who's a Staff Engineer and he was pretty impressed especially since I've been really diligent about implementing good practices like separation of concerns, test driven development, DRY, determining architecture and reviewing it before implementing etc.
I'm not a moron - I've been a technical PM for 10 years now working on enterprise software and I've had plenty experience reviewing my dev's PRs and doing my own bug investigations as well as evaluating engineering solutions.
This dismissive attitude is only going to get less accurate as time goes on and models improve.
I wonder how long it'll be before "implementing" those practices is as simple as writing a good initial prompt for the coding agent to follow. And how long after that that "do it well" (or nothing at all) would sufficient for it to follow those practices by default.
Remember that LLM capabilities will only improve over time (barring severe government action, at least). Also remember that GPT-3 was released in 2020: Getting all the low-hanging fruit (never mind all the incremental improvements) from a novel technology in six years would be a fantastic achievement, so I don't think we're anywhere close to done.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Europe's strategy seems to be to bring down the U.S. tech sector by attempting to impose more onerous regulations like GDPR on it. See, for example, the "Online Safety Act" and how the UK's Ofcom is unsuccessfully enforcing it by emailing threats to American companies, notably ones not under UK jurisdiction.
No, the obvious answer is the true one here. Europe and the UK really really hate that the fundamental, society-altering technology that all of their citizens are using >5hrs a day is completely out of their control, as is the AI that they are hoping will become the new basis of their economy. And they are fundamentally incapable of conceiving that the answer might be less regulation rather than more. The closest American example is when America legislated the sale of TikTok (did that ever go through?).
I personally have mixed feelings about this. Having your public places under the control of another country is in some ways safer than having them under the control of your own country - broadly I like that Musk can tell Starmer to take a long walk off a short pier. But this cuts both ways, and I don't blame the various governments involved for being antsy around it.
This is in no way the obvious answer. The actual reason Europe and the UK hate the US tech companies, especially recently, is that the first amendment allows for freedom of speech which European governments absolutely cannot abide - exposure of the scandals of the elite and what they are doing is anathema to the corrupt and honestly evil governments that they have in place (see the recent disclosures about Peter Mandelson). Less regulation would in no way achieve their goals of censoring speech and keeping their population ignorant, which is why they are simply trying to use their existing powers to shut down foreign sources of uncensored communication services.
Yes, it did, and users are now abandoning the platform in droves due to the removal of pro-Palestinian content, the mandatory amplification of Trump/Zionist content and censorship which means private messages containing the word "Epstein" cannot be sent.
That is the same thing that I said, in much more polemical language, but it's only part of the story. Yes, various European and non-American (Aussie, UK, Canada) governments are very upset that, from their perspective, unfortunate dirty laundry is being aired in public. Some of them surely have things they would like to hide, others rightly or wrongly believe that the country would be better off and less febrile if matters weren't presented in a maximally inflammatory way and optimised for engagement.
But there are also lots of other things that people are concerned about. They really don't like the effect that addictive Instagram and TikTok etc. are having on the ability of young people to concentrate or socialise, they don't like Grok in general and the nudifying features in particular, etc.
Ultimately, both voters and governments generally prefer for regulation to be possible, even if they decide not to do it. Having a big part of life subject to the whims of Washington and Silicon Valley rubs people the wrong way.
More options
Context Copy link
We Euros love regulation and censorship both. It's not just one or the other.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Yeah, it is funny to see that underneath all that socialism and all that postmodern philosophical masturbation, Europe really still believes in feudalism and is furious that us new world peasants won't pay the King's Tax! Don't we know that they are our betters!
I don't know how to say this but you're the richest and most powerful people in the world. This kind of discussion always turns into a Bravery Debate but regulation like GDPR is more about clawing back some agency from America than it is trying to tax US industry.
As the Right discovered five years ago, and the Left discovered when Musk bought X, network effects and the overall stack just don't allow for 'make-your-own' social media.
(I don't actually like or agree with the vast majority of this regulation, though I think that GDPR specifically was a step in the right direction of forcing companies to give more than absolutely zero shits about the privacy of their customers).
Normally I wouldn't be quite so thin-skinned but the Greenland fiasco drove home for me just how worrying it is that half of the most powerful country in the world thinks of us as being essentially a pantomime villain from a Mel Gibson movie.
More options
Context Copy link
We do, though. In 2024, US tech companies paid more in fines alone (€3.8 billion) than the income tax revenue of the entire European tech sector (€3.2 billion).
https://atr.org/brussels-exploits-american-tech-companies-by-enforcing-heavy-fines-for-regulatory-non-compliance/
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Disclaimer: I am not a programmer, though I keep myself broadly aware of trends. I've only used LLMs for coding for toy problems or personal tooling (AutoHotkey macros, Rimworld mods, a mortar calculator, automating discharge paperwork at my dad's hospital)*. I've noted that they've been excellent at explaining production code to even a dilettante like me, which is by itself immensely useful. And for everything else, I'm so used to the utility they provide me personally that I can't imagine going back.
They being said, I am not in a position to judge the economic utility a professional programmer derives from using it for their day job, though it's abundantly clear that the demand for tokens is enormous, and that the capability of SOTA LLMs is a moving target, getting better every day on both benchmarks and actual projects. And look, I understand there's a position where you say "sure, but these things still aren't actually good" - but if you're claiming they haven't gotten better, then I'm going to gently suggest you might want to check yourself for early-onset dementia. The jump from GPT-3 barely coding a working React toy-example to current models is the kind of improvement curve that should at minimum make you sit up and notice.
In other words, even if you think they're not good enough today, you should at the very least notice that a large and ever-increasing fraction of US GDP is being invested in making them better, with consistent improvements.
However, here's a tweet from Andrej Karpathy which I will reproduce in full:
*Sadly they can't make the Rimworld mods I want. This is a combination of a skill-issue on my part (people have successfully made rather large and performant mods with AI), and because I wanted something niche as hell, in the form of compatibility with a very large overhaul mod called Combat Extended. Hey, at least Nano Banana Pro made the art assets with minimal human input, if you think my coding skills are trash, wait till you see my art.
I find it peculiar that Karpathy doesn't see a relationship between those two things. I've noticed a trend where the most glowing reviews of AI capabilities seem to be for people who are using it in areas where they, themselves, do not have enough skill to confidently perform the task. At its worst, it's a sort of tool-assisted Dunning Kruger effect that's actually breathtaking if you can decouple enough to look at it in the abstract.
"I clearly couldn't do this thing but I can clearly tell that Grok/Claude/ChatGPT/Gemini did it right" is a hell of a thing. It's already causing real stress on public software security databases. There's a continuous trend that looks something like this:
"I ran $LIBRARY through Claude and it says there was a potential denial of service attack. I asked Claude for a mitigation and it provided this code."
"Can you explain this piece of the code?"
"I asked Claude to explain this piece of code and it said the following."
Other than feeling like a scene from Office Space, it's effectively acting as a denial of service attack on its own. The amount of time necessary to submit something like that without a deep understanding of the problem is considerably lower than the time necessary for a genuine SME to comb through it and judge it on merits.
Hell, if I were a malicious actor I'd probably try to exploit that by shit-flooding the system to buy myself more time.
Hmm? That's not my takeaway from the tweet (xeet?). He's not denying a connection between AI capabilities and code quality decline, he's making a more subtle point about skill distribution.
The basic model goes like this: AI tools multiply your output at every skill level. Give a novice programmer access to ChatGPT, Claude or Copilot (maybe not Copilot, lol) , and they go from "can't write working code" to "can produce something that technically runs." Give an expert like Karpathy the same tools, and he goes from "excellent" to "somewhat more excellent." The multiplicative factor might even be similar! But that's the rub, there are way more novices than experts.
So you get a flood of new code. Most of it is mediocre-to-bad, because most people are mediocre-to-bad at programming, and multiplying "bad" by even a generous factor still gives you "not great." The experts are also producing more, and their output is better, but nobody writes news articles about the twentieth high-quality library that quietly does its job. We only notice when things break.
This maps onto basically every domain. Take medicine as a test case (yay, the one domain where I'm a quasi-expert) Any random person can feed their lab results into ChatGPT and get something interpretable back. This is genuinely useful! Going from "incomprehensible numbers" to "your kidneys are probably fine but your cholesterol needs work" is a huge upgrade for the average patient. They might miss nuances or accept hallucinated explanations, but they're still better off than before.
Meanwhile, as someone who actually understands medicine, I can extract even more value. I can write better prompts, catch inconsistencies, verify citations, and integrate the AI's suggestions into a broader clinical picture. The AI makes me more productive, but I was already productive, so the absolute gains are smaller relative to my baseline. And critically, I'm less likely to get fooled by confident-sounding nonsense (it's rare but happens at above negligible rates).
This is where I tentatively endorse a "skill issue" framing, where everyone's output getting multiplied, but bad times a multiplier is still usually bad, and there are simply more bad actors (in the neutral sense) than good ones. The denominator in "slop per good output" has gotten larger, but so has the numerator, and the numerator was already bigger to start with. From inside the system, if you're Andrej Karpathy, you mostly notice that you're faster. From outside, you notice that GitHub is full of garbage and the latest Windows update broke your system.
This isn't even a new pattern. Every productivity tool follows similar dynamics. When word processors became common, suddenly everyone could produce professional-looking documents. Did the average quality of written work improve? Well, the floor certainly rose (less illegible handwriting, if I continue to accurately insult my colleagues), but we also got an explosion of mediocre memos and reports that previously wouldn't have been written at all. The ceiling barely budged because good writers were already good. I get more use out of an LLM for writing advice than, say, Scott.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Or, like -- every western government except maybe Sweden, 4 years ago?
I'm not really kidding, but to engage with the meat of your argument -- translating natural language documentation to machine code is literally what programming is, and always has been.
If you have perfect documentation, the coding is trivial; so if LLMs can add another layer to this and become essentially a somewhat easier/more efficient programming language, that's great -- but it doesn't so far seem like they are particularly good at generating that documentation based on (complex, real-life) non-technical enduser requirements for broad problems. Which has been the Hard Problem of Programming at least since Fred Brooks.
If a programmer can say to an LLM "hey build me a Salesforce clone based on such-and-such requirements" and make it happen, that is a pretty big efficiency gain, but not really AI. Which would be a pointy-haired boss saying "hey build me this thing I thought of that doesn't currently exist, but is Salesforce scale" and making it happen; this would be kind of scary.
More options
Context Copy link
This feels like it's a less shitposty and thoroughly expanded version of my "Uber for artisanal cheeses, but on the blockchain" theory that I had.
Our flagship application has seen continuous development since the mid to late 2000s, and it's loosely based on a codebase and product that is considerably older than that. While it has CRUD elements (any application that functions as a long-running service must), it has some fairly extensive components that actually do things with that data in terms of business automation. Those are the areas where all the existing LLM solutions tend to fall apart. Given that they're statistical engines, going farther from CRUD is a very bad thing.
I'm not sure if I can fully buy into this. It wasn't that we were surfacing implicit context, so much as writing it for a very enthusiastic intern developer with absolutely no sense of self preservation. If we didn't break tasks down to an absurd level of guardrails and hand-holding, it would try to make enormous, system wide changes without any kind of midpoint validation. Sometimes we'd see the reasoning say things like "I have made a large number of changes. I should run unit tests to verify that I am correct", and then it just... wouldn't do it. Any of the server developers could have finished the full task in the time it took us to make the tickets that allowed the LLM to do the job without going off the rails.
Yep, I've seen this too. I have to ask, where you using any of the terminal based tools for code development (i.e. Claude Code). I know you said you were using Gemini, so I am doubting it was actually Claude Code (although you can run Gemini within CC).
There is a lot of guardrailing and handholding built into to these tools. If I pass a full system design doc to Claude Code and explicitly instruct it to do TDD with unit tests etc., it will.
LLMs aren't beings, people, or minds. If you think of it as having intention and character flaws, you're going to get frustrated quickly. If you think of it is a very imperfect and probabilistic tool that outputs into non-deterministic solution spaces, you'll get less frustrated and probably think differently on how you prompt it.
I am an unrepentant AI bull. I'll admit that and let people judge whatever I write with that bias in mind. I only request the same from the bears. When I see sentiment like this, which literally chastises a matrix of numbers, I have to assume a non-neutral bias.
I disagree with you here.
Setting aside the deep philosophical questions about personhood (which threaten to derail any productive discussion), I claim that LLMs are minds - albeit minds that are simultaneously startlingly human and deeply alien. Or at minimum, they can be usefully modeled as minds, which for practical purposes amounts to the same thing. (I should note: this position doesn't commit me to "AI welfare" concerns, or to thinking LLMs deserve legal rights or protections, or to losing sleep over potential machine suffering. You can believe something is a mind without believing it has moral weight. I do, I'm an unabashed transhumanist chauvinist.)
More importantly, I think there's nothing wrong at all with modeling them as having "intention or character flaws." if you use a variety of models on a regular basis, like I do, I think that becomes quite clear.
They have distinct personalities and flavors. o3 was a bright autist with a tendency to go into ADHD hyperfocus that I found charming. GPT-4o was a sycophantic retard. 5 Thinking is o3 with the edges sanded down. Claude Sonnets are personable and pleasant, being one of the few models that I very occasionally talk to for the sake of it. Gemini 2.5 Pro was clinically depressed, 3 Pro is a high-functioning paranoid schizophrenic who thinks anything that happens after 2025 is a simulation. Kimi K2 was @DaseindustriesLtd 's best friend, which I noted even before he sang its praises, being one of the weirdest models out there, being ridiculously prone to hallucinations while still being sharp and writing in a distinctly non-mode-collapsed style that makes other models seem lobotomized by comparison. If I close my eyes, I can easily see it as a depressed vodka swilling Russian intellectual, despite being of Chinese origin.
If these aren't character flaws, I don't know what is. Obviously they're not human, but they have traits that are well-described by terms that are cross-applicable to us. They're good at different things, Claude and Kimi (and sometimes Gemini) write at a level that makes the others seem broken. That being said, almost every model these days is good enough at a wide-spectrum of tasks. Hyperfocusing on benchmarks is increasingly unnecessary. Though I suppose, if you've got a bunch of Erdos problems to solve, GPT 5.2 Thinking at maximum reasoning effort is your go to.
nobody ever has any love for my best friend GPT-4.1
Hey, I'm fond of it, and I'll miss it when the imminent deprecation hits. I literally never used it for coding, but I found that it was excellent at rewriting text in arbitrary styles, better than any SOTA model at the time, and still better than many. Think "show me what this scifi story would be like if it was written by Peter Watts".
I have no idea why a trimmed down coding-focused LLM was so damn good at the job, but it was. RIP to a real one.
More options
Context Copy link
More options
Context Copy link
They're model weights. <-- This is a link.
That's literally, exactly, precisely what they are.
You can map your own preferred anthropomorphized traits to them all you want, but that's, at best, a metaphor or something. This is the same as when people say their car has a "personality." It's kind of fun, I'll grant you, but it's also plainly inaccurate.
This is correct. But it is correct because of training data, superparameters, and a whole host of very well defined ML concepts. It's not because of ... personalities.
They're model weights, and we're collections of atoms: bags of meat and miscellaneous chemicals. Both statements are technically correct. And yet... a tiger being made out of atoms doesn't make it any less capable of killing you. The problem with pure reductionism is that
it throws out exactly the information you need to make predictions at the level you actually care aboutcan be a cognitively and computationally intractable approach, even if it's more "technically correct". Too much of it can be as bad as too little.All models are false, some models are useful. That's a rationalist saw, but for good reason. What actually matters is whether a model constraints expectations, in other words, is it useful?
Gemini 2.5 Pro doesn't meet the DSM-5 or ICD-11 criteria for clinical depression. After all, it's hard for a model to demonstrate insomnia or reduced appetite. Yet the odd behaviors it regularly demonstrated are usefully described by that label.
If my friend let me drive his Lambo, and told me "be careful, she's fierce!", I'm going to drive more carefully than I would in a Fiat Pinto. That is still, to some degree, useful, but I think it's clear that anthromorphic analogies are more useful for LLMs, because they have more in common with us behavior-wise than any car (unless you're running Grok on your Tesla). They process language, they exhibit something that looks like reasoning, they have distinctive response patterns that persist across contexts.
This is true in the same way that human behavior is fully determined by neurotransmitter levels, synaptic weights, and neurological processes. But just as you can't predict whether someone will enjoy a particular movie by examining their brain with an electron microscope or a QCD-sim, you can't accurately predict an LLM's macroscopic behavior by staring at its training corpus and hyperparameters. No human can.
Nobody at Google intended for Gemini 2.5 Pro to be "neurotic" and "depressed" or to devolve into a spiral of self-flagellation when it fails at a task, nobody wanted Kimi to hallucinate as regularly as it does. These were emergent, macroscopic properties, there's no equivalent of a statistical scaling-law that lets you accurately predict log-loss for a given number of tokens in a corpus and a compute budget.
Training models is still as much an art as it is a science, particularly the post-training and personality tuning phrases (as explicitly done by Anthropic). You test your hypothesis iteratively, and adjust the dials as you go.
Anthropomorphism is a cognitive strategy. Like all cognitive strategies, it can be deployed appropriately or inappropriately. The question is not "is anthropomorphism ever valid?" but rather "when does anthropomorphic modeling produce accurate predictions?"
I maintain that, if applied judiciously, as I take pains to do, it's better than the alternative.
Your response is incoherent throughout.
Right from the jump;
As opposed to what? A tiger not made out of atoms? This isn't even strawman, it's just a weird thing to say presented as an argument.
You complete lost me here;
Regarding;
That something looks like, sounds like, and walks like a duck doesn't always make it a duck. For example, is Donald Duck a duck?. Well, we can yes and know that he's a representation of a conception of a duck with human like personality mapped onto him (see where I'm going ...) but it doesn't make him a duck made out of atoms - which seems to be, like, important or something.
We've only known that tigers are made out of atoms for a few hundred years. That is a fact of interest to biologists, I'm sure, but everyone else was and is well-served by a higher level description such as "angry yellow ball of fur that would love to eat me if it could." The point is that that the more reductionist framework doesn't obviate higher-level models. They are complimentary. Both models are useful, and differentially useful in practice.
(The tiger could be made out of 11-dimensional strings, or it, like us, could be instantiated in some kind of supercomputer simulating our universe, as opposed to atoms. This makes very little difference when the question is running zoos or how to behave when you see one lurking in your driveway.)
You think that LLMs being a "bunch of weights" makes ascribing a personality to them somehow incorrect. I don't see how that's the case, any more than someone arguing that humans (or tigers) being made of atoms precludes us from being conscious, being minds or having personalities, even if we don't know how those properties rise from atoms.
He's not Donald Goose, is he? Jokes aside, I'm not sure what the issue is here. Donald Duck on your TV is a collection of pixels, but his behavior can be better described by "short-tempered anthropomorphic duck with an exhibitionist streak" (he doesn't wear pants, probably OK principle).
If you sit down to play chess against Stockfish, you can say "this is just a matrix of evaluation functions and search trees." You would be correct. But if you actually want to win, you have to model it as a Grandmaster-level opponent. You have to ascribe it "intent" (it wants to capture my queen) and "foresight" (it is setting a trap), or you will lose.
My point is basically endorsing Daniel Dennett's "Intentionalist" stance. Quoting the relevant Wikipedia article:
As I've taken pains to explain, conceptualizing LLMs as a bunch of weights is correct, but not helpful in many contexts. Calling them "minds" or ascribing them personalities is simply another model of them, and one that's definitely more tractable for the end-user, and also useful to actual AI researchers and engineers, even if they're using both models.
Note that this conversation started off with a discussion about using LLMs for coding purposes. That is the level of abstraction that's relevant to the debate, and there noting the macroscopic properties I'm describing is more useful, or at least adds useful cognitive compression and produces better models than calling them a collection of weights.
I will even grant the main failure mode: anthropomorphism becomes actively harmful when it causes you to infer hidden integrity, stable goals, or moral patience, and then you stop doing the boring engineering checks. But that is an argument for using the heuristic carefully, not an argument that the heuristic is incoherent. As far as I'm aware, I don't make that kind of mistake.
More options
Context Copy link
More options
Context Copy link
I always find these arguments sort of annoying because it really conflates what is actually going on in ML/AI systems with this weird pseudo-science fiction mystification. Yes Tiger's are made of atoms, but no you can't use atomic physics to describe tiger-behavior. With AI models, you can describe behavior directly in terms of the underlying code. The model weights are deterministic parameters that literally decide how the system behaves.
Also you've gotten reductionism vs abstractions completely backwards. Abstractions "throw out information". High-level models compress details to make systems easier to reason about. Also not every useful abstraction corresponds to a mind, subject, or being.
Some Thought Experiments:
LLMs don't have minds and they aren't conscious. They are parameterized conditional probability functions, that are finite-order Markovian models over token sequences. Nothing exists outside their context window. They don't persist across interactions, there is no endogenous memory, and no self-updating parameters during inference. They have personality like programing languages or compilers have personality, as a biased function of how they were built, and what they were trained on.
That's what I get for arguing at 3 am. I do know the difference.
See my latest reply to Toll for more.
They are more "conscious" than a rock, since at . I do not know if they have qualia, but at least they contain conscious entities as sub-agents (humans).
Would you start objecting if someone were to say "China is becoming increasingly conscious of the risk posed by falling behind in the AI race against America"? Probably not. Are they actually conscious? Idk. The terminology is still helpful, and shorter than an exhaustive description of every person in China.
No, but the word "alive" is slightly more applicable here than it would be to a rock. Applying terms such as "alive" to a thermostat is a daft thing to do in practice, we have more useful frameworks: an engineer might use control theory, a home owner might only care about what the dials do in terms of the temperature in the toilet. Nobody gets anything useful out of arguing if it's living or dead.
Hold on there. You are claiming, in effect, to have solved the Hard Problem of consciousness. How exactly do you know that they're not conscious? Can you furnish a mechanistic model that demonstrates that humans made of atoms or meat are "conscious" in a way that an entity made of model weights can't be even in principle?
Entirely correct.
That is not mutually exclusive to anything I've said so far.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
We were using the Gemini cli for that series of tests. I can entertain the notion that Claude code is truly magical, but it's unlikely we'll get more funding to pilot it.
It's less that I think of it that way and more that I'm trying to describe it for an uninvolved observer. I made the statistical engine comparison just a few paragraphs further up.
For what it's worth, I've been using ChatGPT codex, Claude code, and Gemini CLI the last month
My ranking is codex>Claude code>>Gemini
Gemini is the worst, although not profoundly, but noticably
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link