The title of the paper itself is odd. "LLM" is an abstract computation that's performed on some substrate, could be GPUs, could be a piece of paper, could je Minecraft, could be AoE, or could be all of them together.
That's not a property of LLMs thought, that's a trivial property related to turing completeness. I don't get what LLMs have to do with this - if LLMs have anthropomorphic properties, so does AoEII, duh, it's called emergence
Anthropic, if you're reading this: I'm a developer trying to figure out LLMs, I use them 8h per day to do my work. Claude Desktop is a missing piece I don't have access to because I'm only on Linux! I'm happy with a simple AppImage.
What's with these journals all being so hostile? I remember years ago i tried to delete my Washington Post account - there was no button anywhere in the settings menu though, only: text if you want your data deleted. I texted them, they asked back, do you live in a GDPR region? I said no, they replied well tough luck. Insane
I understand that this is frustrating for people who mostly write thoughtful emails. But personally I use gmail for exactly the following things: account recovery, system notifications, and b2b email threads. For the latter, I really couldn't care less about form or shape. It's a tool to an end, to get a point across. I found the auto writing stuff pretty useless so far (suggestions change the intended tone or even meaning of the email) but summaries are very useful to get a grip what happened in a larger thread which I should only know the gist of anyway.
I might be in the minority but to me email is an annoying requirement to reach out to people, and that is not due to the AI tools, it's due to: thread management, the horrible noise of unasked for newsletter, and system messages and updates I theoretically do care about but that are just inconsistently formatted and badly listed. I welcome AI giving me a better overview over what's going on than what I myself have.
I agree, if I was the maintainer this would be an extremely tiring community feedback.
People coming in "I encountered a bug, I don't know what the bug is but I thought about it for a second and it's obviously your descision to do xyz".
As a maintainer, what are you supposed to do? It's not more useful than a ticket "somethings wrong idk what" which is useless enough to close without further action. But it puts the burden on the maintainer to a) figure out what's wrong based on basically no data whatsoever, then b) if they find it out figure out why then c), and that's the tiring part, review their process and create a defense for their approach, or admit that that thing that random user felt after trying out your software for 10 minutes is right, and that you were what? stupid to even think this would ever work? They never asked for any of this, and they're already doing so much work for free.
If the rsync maintainer reads this: You're doing incredible work and humanity appreciates your obviously incredibly competence in it, and not everyone feels the way these people do.
Moving to agentic workflows is obviously the right step and it already provides enough benefits to do it already. And mistakes are bound to happen (if the issue is even a mistake!) and there will always be people who cannot comprehend the power of agents and who will point the finger saying "I know it from the start! I've worked with these tools for 2 hours already and I can see they don't work! Idk why you think they do!". They're wrong. But mistakes will happen that otherwise wouldn't have - but that's the learning experience.
I don’t know the details of this exact instance, but saying that there are reliability issues is a valid feedback if reliability plummets.
As far as I know, nobody with data claims that vibe coding doesn’t affect reliability negatively.
People will connect these two things.
Many times, when reliability doesn’t plummet really. For example, there were huge negative news about a Samsung phone a few years back, that it easily causes fires. Sales were affected by this. Interestingly, next year, they released basically the same thing under different name, and complains were never that loud again. And as far as I know, when they were loud, there was nothing special about that particular model regarding this. So it’s possible that outrage is not validated at all.
They will also connect these, when reliability plummets, but it’s not because of vibe coding.
And they will connect, when it is the real culprit in general, but their problems are not affected by vibe coding.
And of course also when vibe coding really causes their problems.
In any case, the original statements will be true. Do we really want to make a product less reliable to implement features and bugs which we deemed not that important before? Especially with a stable product?
Of course, these on the maintainers, but it’s interesting that forcing AI and their consequences on us - like how Microsoft, Google, etc do - is the default, and not the other way around according to many in this thread and others.
Yes, reliability plummeting is valuable feedback, but it should be framed as such, and not attack the maintainers descision to use agentic tools to write code, and especially not in that high nose way with an undertone of: What you're doing is obviously wrong, did you even think?? Everyone here knows it, are you stupid?
And the maintainer can then choose to use that feedback to incorporate it into the workflow, on their own time. If they so choose (which I'm sure they will, unless they get burnt by the community right now).
If the maintainer used any other tool which is suspected to cause a number of recent problems, it'd be discussed. The tone is a problem but the reaction is equally problematic. It isn't even clear the maintainer hasn't been silently changed if agents are used, depending on the extent. That itself is worthy of discussion, and "maintainer decision" is not the right call in that situation. One comment basically insinuates that with instructions for AI, though it was written as a trollish joke.
It can be discussed but not like this. The tone is problematic and results in the reaction. You're basically saying: I found a few bugs and I saw that you use tool X, thus you're now not worthy of maintaining this software without my supervision. Which is tiring if the initial report doesn't even show what exactly was wrong, or if something was wrong. It's just a feeling of: something doesn't work for me; I don't like AI tools; I see you're using AI tools; therefore I now tell you that you're not capable of maintaining this package, and I have to intervene. Such a stark comment must be done on more than just vibes and feelings.
They're not negated, smarter is smarter, but you have to reach deeper in your pocket. I think this will happen more and more - the smartest models get more expensive. But it won't matter - the current models we have today will get cheaper and can still be used for what they're used today.
I would take all benchmarks with a grain of salt. I don't really use them. What's it supposed to tell me? "5% smarter", what does that mean? My experience will differ. Just try it!
I doubt Anthropic internally sets as a goal to improve this or that benchmark - it's just a way to visualize progress. They probably have much more complex metrics internally.
I can tell you my experience as a js package dev, last tried a few weeks ago. We're building an npm package that's supposed to run on both node.js, deno & bun & the web.
This is an annoying to do for exactly two platforms: node.js, and deno.
node.js bcs it requires a workaround whenever something networking comes in: fetch doesn't work the same. So you structure you're code around having a node.js workaround. Same story for some other APIs. But you can test if itn works!
Deno is more annoying, you just can't test your package with deno before publishing. Before we released to npm, we installed a tar file and sent those around for testing. Works in node, in vite (node, for browser), works in bun, like a charm. Doesn't work with deno unless you switch to package.json, and you use exactly the subset of the spec that deno supports. You can't "deno install xyz.tar", you have to use npm for that (inserts a single line into package.json), THEN you can use deno to execute. No docs, no hint, just trial & error.
Even more annoyingly, npm & bun both offer 'link': in package repo, call npm/bun link, in the test repo do npm/bun link @yourpackage, and that's it, it's installed. Creates a dyn link to the source's build dir so you can rebuild without packing or sending tars or anything like that, you just build in your package dir and the test project is immediately updated.
Deno doesn't have that. What's worse, they don't tell you they don't have that. Also basically no error messages. It just fails in weird ways. Spent hours trying to do it. Now I just publish without testing for deno and wait for bug reports.
So out of the three: bun just works. That's it. Better than any platform. It just works, and it has a nicer CLI & nicer error messages, and it's faster on startup. It has the web api and the node api (i think) and its own api that's very nice as well, nicer than e.g. node. And e.g. if you run bun link, it tells you exactly what happened: this is what just happened, this is what you have to do to use it elsewhere. Node doesn't have that!
I think deno recognized bun's strategy of using npm dev's backbone as being the better call - that's why they're now slowly introducing node.js features, even though that goes against their original USP.
This article assumes that AI only has an impact on the development phase which is certainly not true. It can speed up every part of the step. Including ideation, legal, documentation, development, and deployment.
Ideation: Throw ideas back & forth, cross reference with knowledge bases, generate design documents. Documentation: Generate large parts of docs. Development: Clear. Deployment: Generate deployment manifests, tooling around testing, knowledge around cloud platforms.
Every single step can be done better & faster with AI. Not all of them, but a lot.
Even development. Yes some part of your job involves understanding the problem better than anyone & making solutions. But some parts are also purely chore. If you know you keed a button doing X, then designing that button, placing it, figuring out edge cases with hover & press states, connecting to the backend etc - this is chore that can be skipped. Same principle applies to almost all steps.
A typical example of trying to add a new significant capability involves many meetings (days, weeks, months, etc. )with the business to understand how their work flows between systems X, Y and Z as well as all of the significant exceptions (e.g. we handle subset A this way and subset B that way, but for the final step we blend those groups together, except for subset C which requires special process 97).
Then with that understanding comes the system solutioning across multiple systems that can be a blend of internal system or vendor's system, each with different levels of ability to customize, which pushes the shape of the final solution in different directions.
There is certainly value in speeding up coding, but it's just one piece of the puzzle and today LLM's can't help with gathering the domain information and defining a solution.
What I've seen in an AI-forward looking environment is that it's much more common for PM/POs to be knocking up at least a UI prototype now, and experimentation is happening often even before writing the tickets. Similarly when devs are proposing something they often are coming with a couple of prototypes already implemented. Both of those mean decisions are coming a lot quicker.
I wouldn’t discount the value of moving small tasks away from developers, nor the value of fast cheap prototypes.
Product owners can very quickly get, for many problems, an interactive demo without coding. For lots of problems this can be somewhere from a static html page which shows the interactions to a hacked in feature that lets them actually test if it solves the customer need and try several variations before handing over much more concrete specs of what they want to happen. So much time is lost between getting an idea from someone’s head to code to use to then find out it wasn’t communicated well and then finally that the idea didn’t help anyway and we want it in a different way.
Yes yes I know someone is about to say that now there’s pressure to push the prototype out but that’s an organisational level problem that existed anyway.
And small problems can much faster to solve as well, or even move away from devs. Often people just need some text changed somewhere or html putting together, or some basic code for analysis. They could understand the logic, but the task of writing it from scratch and how to run things may be too much - now you don’t need to prioritise work for a dev to get some sql written and they can spend their time on the larger more software engineering level problems.
"that’s an organisational level problem that existed anyway"
That's very true to many organizations. One cannot just slap an AI tool on it when you are dealing with fundamental organizational problems in the first place.
"they can spend their time on the larger more software engineering level problems"
For sure, devs still needs to focus on the right type of work and maintain the balance. I built a tool to just do that: https://worktypefocus.com/
I've seen proposals for Product Managers to define those conditions themselves by speaking with the LLM. A continuing architectural diagram is constructed and graph is updated until all cases are covered and then the LLM writes the code, writes the validations, pushes to CI environments, runs tests, schedules prod deploy (by looking at company event schedule), gets CAB approval, deploys code, tests in prod, and fixes regressions.
I'm not saying this is the correct thing, but companies are implementing it and it is "working". I don't think keeping our head in the sand is helping.
> I've seen proposals for Product Managers to define those conditions themselves by speaking with the LLM.
But the LLM is not aware of how the business works and why, so someone needs to work with the business to extract the information. Typically it's not well documented.
> someone needs to work with the business to extract the information. Typically it's not well documented.
LLM extraction of the information from the Product Owner is becoming the way to overcome poorly-documented business context.
Non-technical folk are using things like `/grill-me` [0] to seed the LLM with the long-tail complexities that they didn't know they didn't know they needed to put out.
They can ask, they can do a back and forth and they can write documentation to be used from that point onwards and write it in a common style and structure.
These are language models, being able to talk through something with them and have them extract some information is what they excel at. Given that you’d probably get a halfway decent result with a literal fixed set of questions (an Eliza level docbot) gpt 5.5 is going to nail that as a task.
is it working though? The main outcome we've seen with companies that drink the AI Kool aid en masse is buggy unstable systems. clearly there's a level of rigor that's being missed for ship velocity
All of the above points align with our organization’s experience. But there is one more thing happening as well: we have more people in more roles able to create software solutions for issues that used to be brute forced via physical processes. (We are a small manufacturing business.) While these aren’t big giant enterprise projects that require deep swe experience, they are simple software tools that are improving process and productivity everywhere. It is pretty amazing what happens when your head of shipping can build a bespoke tool to solve a problem that previously they dealt with through burning through a lot of labor hours.
One of my beliefs about AI, for small / medium sized companies it allows them massive speed ups and generally increases their capability (I'm also in this space), existing employees of all types essentially get massive speed boosts / opens pathways not available before. For big companies, they are likely to have a bunch of problems due to size, communication pathways, management structures, decision making structures, etc.
I would be really interested in the details of these kind of tools that are improving processes and productivity.
Are they reasonably documented/audited/put into any sort of version control like a lot of internal tooling? Or are they the kind of the thing that gets whacked together on the fly in a "move spreadsheet data from A to B", "I want a list of people's schedules with custom highlighting" kind of things.
Not doubting your productivity increase, I'm just curious how people quantify that when they say it.
One of our BAs created a site that tests the effectiveness of copy / layout adjustments. I don't even know exactly what that's called but he's able to do statistical analysis much faster on what works and what doesn't. It's really cool to watch him thrive and I feel like some of the thinkers that were not devs are going to find themselves to be one but in their specific domain in a few years
Yes. In the same way that spreadsheets are the dev tools for non-devs, LLMs could step into that role, but with much more powerful end result. With the caveat that in the same way you can create a powerful foot-gun with a spreadsheet you can probably create a foot-cannon with an LLM.
yeah the Coinbase CEO gleefully pointed that out as well and now the market thinks they are totally incompetent every time some UX quirk is found
looks like orgs have to have engineers on for optics. like having a legal staff with no lawyers, or a cybersecurity staff with no IT or certified people. Software has famously not needed state licenses or industry certification, but maybe thats a direction to consider to give utility to company optics.
The article pretty much plays out whats happening in our place, heavy use of AI in software development but we dont see us shipping faster, about same or perhaps slower (for other reasons). Its a weird feeling as were waiting for this utopia to kick-in but its not and were cant fully put our fingers on it.
The article and the AI skepticism crowd on HN read like the blind leading the blind to me.
I'm at a FAANG. My org is moving much more quickly, maybe between 3-10x more quickly than we were pre-AI. We aren't seeing a spike in reliability issues. Things just get done faster. An org as large as mine has no right to move as fast as it does.
I’ve been back through your post history (not entirely) - you mention multiple times you work at a FAANG - so you work at one of 5 very public companies.
You have been asked multiple times by multiple commenters to provide a single example of something that reflects this incredible boost achieved by <massive tech org>, you have ignored every request for this, and I suspect will ignore this one as well. HN is going to die unless we all start calling these constant deceptive practices out. I’ll leave others to parse your history and make their own judgements.
Not going to break NDA and give up our competitive advantage for HN, sorry! I can tell you it's been useful for us, but thinking about how to use it is an exercise left for the reader.
Perfect excuse for avoiding any substance (except no one asked you to give up the advantage or secret sauce of using LLMs, only the end examples of what has been achieved with it). It is also funny how you always leave out the name of the company you allegedly work at. It is perfectly clear why you do that, though: no matter what company you name, its actual employees on HN will quickly disprove all the ridiculous claims you have made regarding LLMs and AI. Keeping the name ambiguous lets you get away with it.
It's highly team dependent. Shortly, the more "coding monkey" the work is, the more velocity you can get with AI. As soon as you need to interface with customers and extract requirements, that becomes the bottleneck.
The main problem is, it's not a one-size-fit-all tool, you need to understand what it speeds up to benefit from the speed up.
And if it is a chore, we already have some tools to speed it up, only if it is worth it though. Placing a button is actually easy if you get all the design system down usually with a component library, visual regression automation and testing automation.
If a team doesn't have tools and automation in place, AI might speed them up a little but it adds a layer of complexity, i.e. everyone have to manage their own workflow and tools. And when you try to align the team, you get the tools and automation that the team is supposed to have in the first place.
As for ideation, the problem isn't the speed of information ingestion but the ability to connect and understand different parts of the information, which require thinking. More information at times is just going to hinder the ability to think. For example, it is obvious to developers why there is a rate limit for the APIs but for PMs it might not be obvious. They might ask the AI whether or not a rate limit can be removed easily, how many days if you vibe code it and ignore the possibility that the rate limit might by abused by users just to improve a feature because it is too slow.
We are still doing alot of work with new tools but old methods though, it will be interesting to see how far can we go if we forget about the old rules and embrace the chaos entirely.
Indeed. I suspect most effective AI users are quietly making real progress toward their objectives.
Anecdotally, I see a lot of problems/solutions content about AI that doesn't reflect at all the challenges I face. But trying to tell people that there are other ways of doing things, especially when it conflicts with token-maxxing, is a lost cause
I know and I agree. It sounds incredibly arrogant but it's frankly is a bid sad to see how much HN is lagging behind AI adaption. It's been 90% noise over the last 3-6 months about problems that aren't truly problems if you really look hard at what AI is capable to do already today. It's mostly ppl & process problems. I could post a comment like the one above below almost every article on AI. But it is what it is. It's an opportunity for anyone who doesn't bite into the cynical tone here for sure.
The HN AI skeptics are just bizarre to me. They are insisting to us that, no, the productivity gains we're experiencing every day, simply don't exist!
It's not that they're using the tool wrong, it's that the tool just isn't capable of what we see before our own eyes! I guess our eyes and ears are simply lying to us?
And then they ask for how we are managing to make things move faster. When you refuse to breach NDA and give up your competitive advantage on HN, this somehow confirms their belief that AI is useless.
Precisely. People don't realize that it's all numbers. Given average IQ of people involved in a project is 140, an AI with an IQ of 150 can replicate each and every such individuals in the pipeline. People saying AI can't do this or AI can't do that should come to terms with the fact that this IQ gap is monotonously increasing.
1: When was the last time you worked on a project where you thought the average IQ was 140? I don’t even think I have worked on a project where the maximum IQ was 140.
2: Who thinks the IQ of people on the project determines its success? There’s so much more to it than just “high capability team members” (to give IQ a generous interpretation).
3: (math joke) A sequence like (AI IQ - Human IQ) can be negative and monotonicly increasing and still never reach 0.
Pattern matching against millions of IQ test questions from a training set in order to score 150 on an IQ test doesn't give you an intelligence equivalent to 150.
I agree. Inexperienced people (not necessarily "dumb") are likely to accept everything at face value, not apply critical thinking skills, and not even check the AI generated output.
What IQ "means" is separable from what it is. IQ is a measure of performance on IQ tests. That's literally what it is. If a computer system can complete IQ tests, it has an IQ.
The issue is that IQ means less than you want it to.
I don't want IQ to mean anything. pkoird clearly wants it to mean something.
IQ is a terribly flawed measure of human intelligence. But it measures nothing when you apply an LLM that contains multiple IQ tests in its corpus. IQ is deeply flawed, but the point is not to "measure performance on IQ tests". If someone cheats on an IQ test and scores 200, no reasonable person would say they have 200 IQ.
That's not a property of LLMs thought, that's a trivial property related to turing completeness. I don't get what LLMs have to do with this - if LLMs have anthropomorphic properties, so does AoEII, duh, it's called emergence