C-suite are there to protect the owners so they don't get in trouble if their company breaks the law. That makes them the most important position. Being a good manager is just a bonus.
Seems obvious to me that jobs which primarily require research and thought will be the first to go. The job of a good exec is basically: 1) research the market, 2) make presentations, 3) make high level decisions based on research and input from others and have leaders below you work on those. Ok, 4) is schmooze and maintain relationships but you have an org to do that too.
My concern is what does “smarter” actually mean? If it means producing more functioning code faster across a broader spectrum, fine.
LLMs are based on training over existing data (i.e. examples) to produce output. As LLMs produce more output, they are training on their own data whether it be good or bad. Without technically skilled humans in the loop who know design patterns and best practices, what keeps LLMs from continually degrading the quality of their output? Therefore, less smart.
They are pretty up front about that when they hire from what I was told by someone who went through it: that you have to “be okay” with your job not exisiting in the next couple of years
> Anthropic reportedly screens job candidates with a culture interview testing whether they would put the company's mission ahead of their future wealth. One applicant was asked how they would feel if a safety decision sent Anthropic's stock to zero.
The future seniors won’t be people defined by strong technical knowledge. They’ll be the ones with “autonomy and judgment” as Anthropic puts it (read: highly social; good at climbing the corporate ladder; strong educational/professional pedigree).
These people won’t necessarily need to be trained in CS, because in 10 years no one will need to know what a data structure is. So they can come from other fields (law, business) or just from other roles within the company.
Note this is basically how it already works in every non-technical field where there’s no clear definable “skill” on which to evaluate people.
It's 2026 and people still believe in measuring productivity by lines of code^W^W^Wnumber of commits? Amazing.
Two vibecoders are walking down a forest path.
One of them spots a Zig codebase and challenges the other: “I'll give you $100 to convert this codebase to Rust!”. They accept and after spending $1000 worth of tokens on a thousand commits the codebase is now in Rust.
On the way back, the second vibecoder spots the same codebase again, and challenges back: “I'll give you $100 to convert this codebase to Zig!”. Another gajillion tokens later, the codebase is now in Zig.
Walking a little more, one vibecoder (in a moment of unusual clarity) asks the other: “It feels like we just did this for nothing”. “No”, the other responds, ”we just increased our company productivity by two thousand commits per day!”.
I think you're right in your comparison and that's way we'll soon see mass layoffs of software engineers. Software can now be produced with less people than before but I don't think there's market for more software.
Historically, more efficient technology has led to a greater demand. It began with more efficient steam engines that required less coal for the same amount of work. That didn’t lead to a reduction in demand for coal, instead the acceleration in demand for coal changed the world.
However you measure it, in LOC or in MB or in ops per second, demand for more software always seems insatiable. This trend will likely continue up until software can be produced with zero humans at all. But I can’t see that happening unless we invent actual real artificial intelligence.
A counterexample that I can think of is agriculture. In the past almost everybody worked in it, now there are barely any farmers (relative to the overall population) producing much more food than before.
Where I work, PR counts are a major metric come perf reviews. Surprise, surprise every team has a pointless migration going on. The number of promotions though...
no, i'd be more convinced if positions drop for tier 4 companies where everything is a mess. Companies like Anthropic have every incentive to build their system around AI accessibility and understanding.
Because of fears of AI, and people watching these indicators as hawks, the exact opposite will happen, they will try delaying this as much as possible to delay the backlash until it's too late.
Elon already stopped talking about job loses and doom, and tries to talk only about the positive.
“Junior” is the wrong distinction to make. I’ve recently worked with senior engineers who are terrible at wielding agents, and recent grad engineers who multiple themselves 100x. The main way I look for software engineers now are:
- Deep curiosity and ability to absorb new information quickly
- Evidence of ability to wield and coordinate agents to the max
- Ability to clearly communicate ideas
You have these skills, you can conquer the software world.
Where I work we make real hardware and software that are used by real people doing real jobs every day. We use AI but in that environment someone who made Jackobriens comment would be seen as an idiot, someone who doesn't have a clue how to build software with real users that has to work with few bugs and be maintainable.
But there are so many people who make statements like his that maybe we should assume that some of them are not idiots or bots. If he's not an idiot it means that he's working in an environment where that kind of statement is paying off. So it could be, for example, that he doesn't really have profitable customers. His main 'customers' are investors who are clueless, or who know that they are clueless but don't care because they are going to move it on to a greater fool anyway. Or that because there aren't any customers, the only thing they can measure is 'productivity' in terms of 'number of unused, unprofitable and unsustainable features'
I work with AI as well and it can help, but it can also hurt. I think any objective person agrees with that.
There is another option to consider. What if they heard someone else say it and they believed it.
Now they go around, like many insecure people do, shouting about their unfounded take. Begging the universe to correct them or give them a good reason why they arent informed.
Then what usually happens with people who suffer from this type of short ordered thinking is, they realize they can't walk it back without social consequences because of the shouting they have done. So they just double down on it and find ways to live with the cognitive dissonance. Anthropic, openai, etc make billions off of people like this.
Because they propagate the BS advertising to people who don't know better, and then spend more money to try to make it real.
Well anthropic says you are wrong and wants juniors to disappear into another trade. This absolutely wouldn't be a coordinated effort to deskill Americans and force companies to helplessly accept whatever LLM code is generated without human review. Most certainly not. They wouldn't target regulated industries that require certain work praxis like biotechnology and lobby heavily with the military and government to change laws around this. Definitely not!
What would Anthropic gain by wanting juniors to disappear into another trade? Obviously their goal is to turn a profit meaning they need to be more efficient and they seem to believe that will come from senior engineers. They have the supply for it, why hire juniors when seniors are dying to working in your shop
Uh complete market capture for an entire industry that costs billions of dollars a year. They have been claiming software engineer was solved for years now.
If everyone does it, no one monitors it anymore, and quality degrades the only way out is more tokens.
Bingo. From day one what these companies have really been after is not "intelligence" but rather mass production and the concomitant quality expectation and behavioral shifts that go with it.
All they really need to do is convince enough people to adopt lower expectations and accept worse software products. At that point, mission accomplished. They've annihilated the source of value originally provided by people who actually know how to do things.
This reminds me a bit of the gaming industry. In most AAA game studios, junior openings are very rare or simply non existent. Companies expect you to learn the ropes elsewhere and only show up when you're experienced.
As long as there other companies scooping up the juniors (which we cannot assume will happen), this could be sustainable.
If LLM's are supposed to get smarter, wouldn't they be able to wield agents for you? Seems like a skill with a short shelf life. In my day to day work i just use claude and it seems to spin up agents as it sees fit so I'm not sure why people need to coordinate agents to the max? If there is one thing i've seen in software engineering over many years is that sometimes software engineers prefer using flashy stuff even though it doesn't have that much benefits but they evangelize this stuff like it does. "Oh you don't use micro services and X,Y,Z for your crud app?", "Oh you use VS code and not vi?" "Why don't you maintain a data platform for your small scale data that can fit in postgres" so on and so forth.
Valid overall, but text editors like vi (or emacs or whatever else you prefer in the category) really do have major benefits and are definitely not being chosen to be "flashy". Skilled users choose one that resonates with them, and practice using it specifically. Traditional mouse-oriented editors just don't deliver that UX. It's not at all like choosing a bloated or over-engineered project architecture.
I've used vi and emacs and the number of their users that really 10x everyone else is vanishingly small. A large percentage of their userbases are people who wished they were the 10x but were much slower (or at best equal to) than the average Jetbrains tool users.
> If there is one thing i've seen in software engineering over many years is that sometimes software engineers prefer using flashy stuff even though it doesn't have that much benefits
We used to call this RDD - Resume Driven Development.
I’ve hired quite a bit over the last 10 years. Curiosity is the absolute number one thing I look for, whether for a fresh grad or a principal/staff+
The difference between a junior and a senior is still very real. With agents, knowing what to ask them to do (or not do) is bordering on more important than it ever was. A curious junior might still go down rabbit holes for things that a person who had done this before. That’s what makes them a junior. Not everything we _can_ do we _should_ do
Could you please stop posting unsubstantive comments and flamebait? You've unfortunately been doing it repeatedly. It's not what this site is for, and destroys what it is for.
There are talented and objectively nice software engineers working for Meta, Palantir, Flock, you name it. I knew some of them, warning about moving to the US at the early 2020s. They did not understand, just oblivious, happy about the comp, and that they could fulfill their "america" childhood dreams, mostly gathered from movies they watched on the 6th floor of scrappy socialist block building. They tend to complain about workplace tensions, high pressure... but evil things? No that's someone elses doing.
As if Anthropic was a model of successful company...
They are nowhere near profitable, have terribly bad products riddled with bugs (9,000+ open reports on Claude Code), and have no moat. Why would anyone listen to how they work internally?
For what it’s worth though, if your project can work on open source models or other models in general then you have the ability to vote with your feet and you still have leverage.
It is all entirely advertising. Anything they say. Of course they want all companies to require their engineers to not think anymore and just press accept on their software. It's how they make money. Most c suite is absolutely deluded enough to believe them too.
They are offering a cure to the defining contradiction of capitalism: The tension between capital wanting maximal self-reproduction and labor wanting a fair wage and a good life.
Anthropics pitch is to finally redeem capitalism by getting rid of those annoying workers and their annoying demands - that is why C-Suites are so excited about it, that is why its talked about in messianic language.
It is even worse than feudalism.
Their “utopia” doesn’t need an economy nor does it need a population.
Feudalism indeed relied on its population and on its primitive economy.
I think Thiel (could have been another billionaire) twitched when asked by a reporter if he will be sad if AI overtakes humanity and humanity disappears. And not the good kind of twitch, more like "it's fine if it happens if I'm uploaded to the matrix by then" kind of twitch.
Hard agree. Why else are they writing manifestos and talking about the antichrist...
To the poster below who thinks anthropic is in this to destroy capitalism(lmao rofl), I guess maybe if you are interested in fuedal societies run by people who currently bomb schools, have chatbots and finely tuned recommendation engines designed to make children commit suicide but only after extracting as much capital as possible beforehand, try to sway elections in foreign countries, theottle the Internet so dictators can win wars, etc. Yea sure man they are all definitely interested in making the world better for everyone by also automating one of the few middle class jobs available in a country which has no manufacturing to fall back on. Super normal timeline.
It does provide a useful ceiling, at least. If even Anthropic is seeing only a 2-4x revenue multiplier from AI adoption, I’m intensely skeptical of the people who run around saying they produce 10x or 100x more valuable output because they’re really good at building a software factory or whatever.
I mean, your productivity gains are relative to your starting point in some way.
If you are already an advanced company filled with skilled people, you may not see the gains that a mid-tier company filled with mediocre seat warmers would see when their "best and brightest" start adopting AI to replace said seat warmers.
The latter is something I’m dealing with and it’s given me some existential angst. An example is a company’s monolith with millions LoC and hacked together. With proper engineering strategies, it could be closer to 100k. Without proper AI tooling, my verdict was let it be. When you have so many developers, PMs, stakeholders who tie their value to the pile of slop and making “engineering” decisions with it, it becomes very unpopular to suggest any refactoring - especially when that refactoring relies on principles that require an understanding of software engineering.
An example: the UI applications suffer from massive amounts of repetitive code. People on hacker news will probably instantly jump to some kind of componentization strategy. More wise individuals probably understand you need to draw a winning strategy before that’s possible. Identify common workflows, from that, identify common components, figure out a parameterization and so on. The footguns are the slop having unintended consequences which fixing might actually cause production bugs by being “features”.
Hypothetical now: a current alternative solution which may exist is two simple detail views (of which there may be 30+ in the application) are 1000+ LoC template files, hand-coded, copy-paste, with all the errors you could imagine. You can get a lot of work from that so people are/could be content just fixing slop with slop and calling it engineering while piling on more features. Maybe it works but it doesn’t allow for much engineering room.
With AI tooling, popular support becomes less necessary - correctness matters more. Instead of galvanizing people to realize this is wrong (and it’s not just a want thing, with offshoring and the general difficulty of software engineering, many people don’t understand these concepts. They are great people but struggle understanding for loops intuitively; I don’t know what to say). Now I can provide my strategy to the tooling and do the decomposition and componentization myself. On top of this, my strategy can be more intelligent and robust because I’m not doing it by hand like concurrently using the API layer as a source of truth to give my agents more context on what the UI was trying to do. The code itself is confused so part of the work is learning the intent behind each module and also capturing bugs that are “features” to prevent regressions. Basically, a clean room rewrite (but we have access to the source code so I know that’s not the exact correct term).
The engineering moat is still there. If I say, “Claude make my UI codebase smaller” it will burn tokens and fail. One day it will succeed but right now it will fail. Likewise if I say “Claude suggest ways to make my codebase smaller”, it may suggest my strategy but it may not capture all the critical wrinkles. LLMs are not good at simplifying so to simplify (which is always a good goal in software engineering as long as the abstractions still hold), requires human strategies when complexity is high enough. Especially in my case where the behavior of the code is almost more emergent rather than engineered.
Then carrying on, hypothetically, I started to devise a system for not only rewriting the UI but the rest of the application (API and DB layer) as well as the corporate processes around the application. This may sound rude but these corporate processes can be disempowering and wasteful. The trick is inversion. First build a conceptual interface around the layer. The model will understand the purpose without understanding or rewriting the internals. This saves context and allows it to do more powerful reasoning by having perspective instead of being in a vacuum. Then when you have bandwidth, use that interface along with all the context you developed working on other modules, to take control. An example, you have a team which is responsible for building and deploying. To build and deploy, you may need to fill out forms, contact layers of people, all to achieve the end of goal of “building and deploying”. Through inversion, you communicate with an agent endpoint that handles communications to that team. Escalating, writing forms, and doing whatever is needed to get the code built and deployed. For all intents and purposes, you are now in control of that process. What used to be, for example, “I want my code built and deployed to QA” -> “message A” -> “message B” -> “write form A” -> “monitor for responses” -> “build triggered” can now be “I want my code built and deployed” -> “activate agent process” -> “build triggered”. This is an application of inversion in a nutshell. AI allows to apply engineering principles not only to our code but corporate processes where applicable.
So in this case, there could be 100+ individuals supporting this application and process. The one or two or three engineers that used to own the development process (more or less) were drowned out yet still responsible for the overall process and integrity of the software. It was an impossible task but now it seems more possible. The existential angst part comes from the all processes that were sprung up to support a reality where this all wasn’t possible.
Last, I definitely do not think the solution is one hundred engineers. If you know the show “Breaking Bad”, I actually find the reality of software engineering to be more like this now. The difference between 60, 90, 95, and even 98 percent purity is staggering. And the chemist who takes the input (money and tokens) and produces the highest quality output (codebase and processes), dominates. Just as in Breaking Bad when Walter White inherently has power, despite all other chemists and structures of authority, because his process had the highest purity output. It’s just business sense and adding more chemists does not increase purity. Output quantity is no longer gated by number of bodies so that argument goes out the window. Mentorship will be key (as in Breaking Bad) and large teams of all-high-quality producers should be put together into super high-thoroughput teams. But for regular operations, prioritize the highest quality output where the input is guided by the highest quality producer because adding more cooks will seriously degrade the output, at an exponential rate.
It's easy for AI to make somebody 100x more productive if they were insanely unproductive in the first place. In the workplace this is unlikely, but in hobby spaces I believe it. Lots of hobbyists using AI to program micros/etc that, in the past, would have been beyond their capabilities without extensive study first.
Consider for instance the bicycle youtuber Berm Peak, who's been using AI to reverse engineer and hack shitty ebikes. This kind of work is outside his expertise, so the relative lift he gets is massive.
I think a lot of "unproductive" people who use AI to "100x their output" in things they're not good at delude themselves into thinking they're being productive. You see people post vibe coded "projects" on places like linkedin where no one will ever use or look at it, and it doesn't even serve as a learning project. Or people who used to hate writing now posting complete slop blog posts 3 times a week which no one will ever read.
Exactly. Generative AI, so far, has clearly lead to increases in activity. I still have yet to see more than a handful of projects in which it has actually led to productivity and that's usually because the pilot still spent significant time on ensuring quality was high and that the outcomes actually align with what a business needs. In these cases, the boost is not 100x, it's a much more modest 2x-3x at most.
I think the root of the problem might be that, after two decades of the corporate version of scrum with its absurd obsession with story point velocity, we now have an entire generation of programmers and managers who have had the ability to distinguish between activity and productivity conditioned out of them.
If it gets the job done, that's what matters. Linkedin shitters were never going to do anything interesting in the first place, look at what actual hobbyists are making instead. People complained about Berm Peak using an LLM to code up a new user interface and control software for an ebike, but it worked so the complainers can be ignored.
E-bikes are cool I guess, but what most everyday folks are going to notice first is something like how my health insurance company just launched a vibecoded rewrite of the patient portal and now I have to wait on hold to find out anything because the claims history and EOB pages regularly show incorrect information or simply fail to even render.
No argument there, but it doesn't change the fact that some people, properly motivated, with the right backgrounds and right tasks, do get plenty of lift with these tools.
Note that I specifically mentioned that I'm skeptical of claims of insane lift in workplace contexts. Several factors work against it there, organizational dysfunction not least of all, not to mention a disconnect between the people doing the work and the people for whom the work has value. People being crunched by their boss to ship some corporate portal software faster are going to make shit. You making something for yourself, to solve your own problem to your own satisfaction, needing to clear problems outside but adjacent to your proficiencies, should be able to do much better. Many people have found it to be so.
As both a hobby user snd a professional user, I think it’s much simpler than that.
These tools excel at greenfield development, but they genuinely struggle in mature codebases. Especially when it’s a legacy system with spotty documentation.
There is effectively no value in statements like these.
There is no measurement. Think about how much of a difference empirical science made in such a short amount of time. Or think about how it is impossible to optimize software without measurement etc.
Also there is the AI companies’ atrocious reliability even if you trust the measurements
You're missing the obvious. Time. There are highly talented developers who have the technical ability but had no time in the past to do these things because of real life obligations besides their job. Many are producing 10x more valuable output now. Also you're missing that Anthropic as a leading frontier lab isn't a good comparison for almost any metric for any other company in the world. It defies any principles companies usually have like unlimited talent supply.
To wit, I'm not producing 100x or even 10x more valuable output, I'm producting <infinity>x output now. I've gone from "no time/will to push through personal projects" to actively working on 5 in parallel, just because I only have to do the broad design and UX development now.
am i the only one who read this and thought 'no shit.' obviously agent hours outnumber human hours? most people operate multiple agents and function as AI nannies now. big deal.
also 'lines of code' or any other output based metric in the age of infinite slop is not a flex/indicator of usefulness or value.
and yes, everyone spends all their time reading Claude output. that doesn't mean better work is getting done, or progress is being made. i've been reviewing and fixing claude slop at work for at least a year. if anything companies/products are getting more homogonized and bloated. thanks anthropic, you are going to save the world
Without senior developers having a deep understanding of the problems underlying the systems, they'll have little to distinguish themselves from non-skilled management poltroons, who are themselves even more replaceable by agents. But the alignment problem will become even more severe.
Software do have lot of variables, real world edge cases and unknowns, I don't think 100% automated agentic development is a good things. I know you wrote it sarcastically.
This just can't be sustainable. Ever increasing agent activity can only move to the point where it's literally impossible for humans to have a collective awareness of how the systems that support our lives actually work. How does anyone think this is going to end well? If as humans we loose control of vital systems, when they stop working one day (and they so easily could unless the agents are perfect and 100pct aligned with the general good of humanity (how do we even define that)) we could find ourselves without the ability to run healthcare systems and power grids, that have evolved so much such complexity that no human can intervene to fix them. I do believe that you can create feed back loops where probabilistic agents can be used to create deterministic checks and balances, and solve real problems for good. But it's a wild west at the moment and that is not what's happening in so many companies, and the solutions are not yet on the horizon. Even if the 'agents can handle it' ... where does that leave us? Unable to run our societies without puting everything through frontier LLMs? I just finished a very short story by Ted Chiang - 'The evolution of Human science' it starts at the point where it's 25 years since humans contributed a scientific paper because 'meta humans' now handle all research, and communicate digitally such that humans can only get a low fidelity report of what the agents actually know, and the endeavour to understand has become like a best effort archeological like hobby for interested humans. The story assumes that the 'meta humans' will always act in humans best interests. I believe it's from 2002, it's incredibly precient (The stories of your life and others is the collection the story is in. Many of them are very relevant to modern AI actually).
Imagine a country where no one can drive, and all transportation relies on a few AI monopolies.
If you don’t pay them, you can walk by feet.
If they ban you, you can walk by feet.
If their infrastructure is damaged, no one and nothing will be able to move.
I'm reminded of the visual novel "Root Double -Before Crime * After Days-" where one guy in town having a manual petrol-powered car instead of using robotaxis is a significant plot point.
You're more than welcome to wake up in a country that is 100% dependent (in all spheres of life and economy!) on a few monopolies. I'm not sure that thing will be a country anymore.
BTW, when (not if, but when) a war happens, they will become very nice targets to render this exact country completely disabled.
I can easily imagine, say, Russia destroying a few datacenters not with nuclear missiles, but with some relatively trivial local ground-based FPV attack or whatever.
And voila. No cargo moved, no code changes or fixes, no medical diagnosis made,etc.
A whole country is disabled.
There was a book I read maybe 10-15 years ago about a world where governments Essentially dissolved and in place there was nothing but mega corp monopoloys - Jennifer Government was the name ( if you ever played the MUD Nation States yk)
I thought that was an interesting idea, but not something that could ever happen..
There are already tons of countries in Europe and Asia that are dependent on a few monopolies. People still live their lives just fine.
Re: war.. that’s why US spends orders of magnitude more on military than anyone else. If any infra is hit on US soil then im heading to the bunker because most of the world will have already been reduced to dust.
Unless something dramatic changes and it'll become cheap to "constantly re-educate" the models to learn the new fundamental facts about the system (ideally with a small lag of maybe a few days like with humans) I see no reason to think that LLMs can have a profound understanding of how business systems work of a remotely comparable quality to humans. Just adding the context about the system already reduces it's capacity significantly according to what I can tell. Not having large context helps somewhat at a cost of having to explain lots of task-relevant things to LLM over and over again...
> I see no reason to think that LLMs can have a profound understanding of how business systems work of a remotely comparable quality to humans
My manager was dragging me through useless projects that went nowhere despite multiple people telling him "it ain't gonna work" while shooting down proposals that later turned out to be "high priority tasks required directly by upper management". And he's a well-paid, educated employee with years of experience. Don't even get me started on delivery guys who cannot read a simple instruction like "do not ring the bell, leave the package on the doorstep". They just follow a 50-line script and that's all they can do.
Whenever you mention that AI won't reach human intelligence, remember that this is the benchmark. Not Alexander the Great, it's the millions of people throwing their life savings ($1000 each) into Labubus. Long time ago there was a series of articles in my country about the costs of wedding receptions skyrocketing and some people taking out loans to pay that, and someone made a meme "robots will never have human intelligence!" with reply "bro, you took out a loan to pay for a wedding, I don't want your human intelligence".
Well I'm obviously not comparing to average human intelligence (and negligence), but that's actually a good reminder about one other important aspect of how living organisms are very different from machines -- humans have very high range of intelligence, strength, etc, and the best humans are usually much better than the average in all areas of life. Computers and software are much less diverse.
You are showing disagreements with your priorities (mostly money) as unintelligent. You're showing a lack of care and conscientiousness as unintelligent.
You're reducing intelligence to alignment with money and following instructions. Under that framing the argument is already won in favor of AI amd some vague effective maximizing of a cost function.
Many jobs are like this, so it's fine for employment. Maybe the vast majority of mankind was never suitable for reliable employment of the "cost function" style.
But I know two things:
1. Centralization should be avoided. The reality that AI can do most jobs is not a reason to concentrate power behind those portraying that or building AI. We still want local, ideally individual, empowerment. I don't see enough said about this. Maybe we can even say the system must be one people wish to opt in to, to maximize profit.
2. Intelligence is being very narrowly defined and humans value more than an INT score in the RPG sense. The more I meet people outside my belief circles, the more I realize the intelligence of some actions, in the sense of "logically deduced from a higher principle" not "aligned with my western white collar principles". The ultimate role of culture and humanity IS this principle exploration and to say we've reduced that exploration to "maximize X" is crazy.
At some point deciding fate of the company or system or state flows from principles embedded in cultures and that's in danger of being flattened away by optimizing outcomes with machines.
Vague and wishy washy, but its the clearest I can state it.
> At some point deciding fate of the company or system or state flows from principles embedded in cultures and that's in danger of being flattened away by optimizing outcomes with machines.
If the principle embedded in the culture of my manager is to do dumb shit that goes nowhere, then I'm not sure what goals we're achieving.
> The agents only have to be better than we are and we’re not very good. Same argument as self-driving.
Self-driving requires less holistic thinking than imagining, building, and evolving large scale systems. It requires abilities that LLMs do not have. It requires non-symbolic reasoning, intuition, etc. that are beyond LLMs.
I'm surprised this is even a question and the fact that you used human and not even mammal to phrase your question. Do you think intuition is a unique human trait?
> Ever increasing agent activity can only move to the point where it's literally impossible for humans to have a collective awareness of how the systems that support our lives actually work
That's the point, they are setting up the economy to be run by agents and they trust that it is possible. Step 1. Teach AI human knowledge (2022-2023). Step 2. Make AI agentic (2025). Step 3. Start autonomous research and create own corpus of knowledge (2026).
Step 3 will keep compounding (as we are witnessing) and most people will be left behind. I'm not against a revolution, I'm all for it.
Pre-training already makes up a very small cost of the training now. Post-training makes up ~90% of the training cost as modern LLMs are trained using RLVR on synthetic data. People still think that a lot of it is coming from scraping wikipedia like it is 2023.
No one's at the wheel. It's fair (to a degree) for companies to push for short term gain, but the legislative counterbalance that is supposed to care for long term effects no longer exists.
It's not only AI. Betting is on the rise, adults and minors are being hooked by dopamine loops, social networks are causing well measured negative effects, children's educational scores are sinking worldwide, appliance manufacturers engage in surveillance...
You have to remember that an American consensus in it's current condition is really just a consensus among the corporate interests that have captured the government.
> Ever increasing agent activity can only move to the point where it's literally impossible for humans to have a collective awareness of how the systems that support our lives actually work.
fwiw, i think literally all largish corporations have been like that pre AI too?
i can gurantee you that nobody understands how the whole system works at all of my last workplaces either...
people just know their small part and hope for the best - so from a devils advocate pov, its already been proved that its sustasinable.
now wherever its _worth it_ is another question. large employers are - on average - incredibly wasteful with their resources. not to anthropics degree, but still.
if youre boostrapping, and VC token subventions go away... i suspect it'd be not worth it to develop like that. But i guess we will find out in a few month to 1-2 years
Humans having understanding of the system went out the window hundreds of years ago at the latest. The "Renaissance man" era is long gone. I know you said "collective understanding", but that's not a real thing. Humans aren't borg. Our entire lives have been spent in systems we can only understand in piecemeal if at all.
Do you good understanding of where your food comes from?
How about the dollar you earned, spent, invested? Do you know how it works?
How about the clothes you wear?
This is just a tiny subset of the enormous complexity we're immersed in, and no individual or collective has a complete understanding of it. If you do, you're truly an outlier and GP's point still stands.
Some individual has an understanding of each of those areas. Why does one individual need to know the entire food chain? The discussion point is about each specific area.
Even for the COBOL based systems someone understands it well enough at the conceptual and business level. And when there's no one to the company migrates and modernizes.
You are conflating the subject matter expert who understands one area well and others not so well with a general newbie who understands nothing well
I wouldn't put it exactly like this, but I agree that many systems chugging along today (such as the COBOL running on IBM metal that powers banks) are not something people try to (or can) comprehend all at once.
Is that really a problem though? We can solve problems empirically even if we don't keep the entire system's workings in our heads.
Two other stories in this vein: "The Machine Stops" (1909) by E.M. Forster and "Pump Six" (2008) by Paolo Bacigalupi. Very different milieus, but both deal with decadent, hedonistic futures where humanity coasts on legacy systems engineered by long-dead experts until they begin to break down and no one alive knows how to fix them.
Wow, that sounds like an AI hellscape! Constant pressure to use AI agents to be more productive. That pressure requires an amount of attention that wears you out so quickly just to stay competitive that you need to build bespoke tools to orchestrate your agentic workflows. But the amount of output from that is too much too, so you need to build meta-orchestrators to orchestrate your orchestrators.
And all of this is then somehow classified as responsible development of AI, by people who claim AI could eradicate humanity, and that RSI will be our downfall. How exactly is outsourcing your work on agents to agents not RSI at this point? Why are these people still allowed to operate again?
> A financial manager built a purchase request review agent
> An executive assistant, who built a personal assistant agent for scheduling and email triage
Really? We should build our own agents for everything? Sounds a bit token-spendy.
Also, as far as I can make out, the paper was not written by AI (nor even with AI assistance?). You'd think they'd make a song and dance about it if it was. Maybe AI can't do this sort of thing yet (or ever).
I'm surprised this is not much more talked about these days. The hiring bar is going to increase so much in the next few years (already happening), and together with more layoffs driven by big tech, things aren't looking good.
Anyhropic is not a good example for anything. Their scale and growth is way beyond what like 99% of companies experiece. Their optimization formula is different.
They're celebrating elimination of junior work while offering no explanation for how future senior staff are trained, and measuring output by code volume. Lol who wrote this, can't wait to see how this plays out, if you asked even a junior they could have told you that's a terrible idea.
It really is the most misanthropic company, isn't it? But strategically is all this fearmongering about SWEs becoming obsolete even the right move for Anthropic?
Say that it is borderline true (it's not, but let's pretend). Then companies can lay off 90 devs for every 100, and buy the remaining 10 devs Claude subscriptions. Isn't it better if they keep all 100 and buy everyone a Claude sub?
So shouldn't Anthropic actually be pushing the narrative that Claude doesn't replace devs but actually makes them vastly better at their jobs and so every company, big and small, should buy every dev the highest tier Claude enterprise subscription?
A lot of it is directed at senior management who are wetting themselves at the prospect of being able to fire those irritating, demanding, highly paid developers.
They don't want higher productivity because what are they going to do with that? Launching new products is costly and uncertain. They don't want to be paying the same number of developers and have them all spending money on Claude subs. But the idea of firing 90% of them and having the rest in fear for their jobs? Brilliant!
Back in the real world, the question is what's going to happen to companies that go through with this. When they have vast amounts of slop code that no one understands? When you have to sit in with a customer and explain why the product doesn't work and no one knows even how it works never mind how to fix it? We can already see the inklings of this in the huge buggy messes that Microsoft and Anthropic are producing. Interesting times.
Said by a company that’s increasingly desperate to convince you that there’s serious business value in using their product. The challenge is those without that conflict of interests don’t seem to agree there’s a lot of ROI.
In other news, Turkey farmers encourage you to buy more turkey for Thanksgiving.
Authors of this article: Nathan Wilmers, Regina Catipon, Maxim Massenkoff, Peter McCrory
I'm too lazy to bother looking up each of their credentials. Are any of them software engineers? Do they know anything about creating, maintaining, and evolving software?
I do know that Anthropic's CEO isn't a software engineer, so it would be hard for him to evaluate their claims with any deep insight from experience.
Who cares how employment is changing at your startup? You're like an influencer on TikTok marching shit from your bedroom, pretending to have a fabulous and glamorous life. Minewhile, you are insecure, incompetent in the craft of software engineering, but play the social media influencer game because you have zero substance.
> As AI agents diffused through the company, Anthropic began producing more code per person
Anyone who has been in software for a while should hear alarm bells going off when they read this. Every line of code is a liability, not a trophy. This is what happens when people who don't know and don't care to learn about the existing body of knowledge in a discipline "reinvent" it.
Give it another year or two and Anthropic will be writing all about how they lowered LOC and got agents to be more conservative about generating new code and introduced a "verification gate" for every new line produced.
Or perhaps code becomes just another disposable artifact just like compiler-generated machine code, so understanding or maintaining it will be unnecessary.
> perhaps code becomes just another disposable artifact just like compiler-generated machine code, so understanding or maintaining it will be unnecessary.
The compiler development teams make sure that the output of the compiler is correct according to the source code input, preserving semantics and issuing errors for incorrect statements.
LLM is kind of a nondeterministic compiler. You'll want to test the hell out of the resulting code, not only to catch flaws in the prompt but also those introduced by the LLM. I'm bearish on programming but extremely bullish on testing (disclaimer: I'm a programmer).
> You'll want to test the hell out of the resulting code, not only to catch flaws in the prompt but also those introduced by the LLM.
Imagine worrying about gcc or the python interpreter that much. No wonder they use the label "reverse centaur" for LLM usage. You must think carefully before coming up with a prompt, then review the code carefully for correctness. Double the stress for no reason. Or is it just going with the vibes and slinging slop?
I believe that is true, but more of it is still a liability. Good code tends to be simple. Bad code tends to be verbose, and verbose code tends to be bad. As someone who has been extremely prolific both before and after AI in terms of output, and who monitors the output of LLMs (including anthropic’s best models) I know that they are usually generating a bunch of junk. It will do what you ask but it’s largely terrible. Yes, you can mitigate that with a bunch of adversarial agents, but it’s still not great. Every line of code can potentially be a security hole or cause other problems.
> Anthropic has been hiring for a smaller share of junior technical positions.
I don't wanna be the guy calling out the lie or slop but when did Anthropic hire in the last two years for junior roles? OAI and Anthropic could literally hire seamlessly the best of the best worldwide for every job they advertise without having the need to think about the poor juniors.
I would love to see productivity effects in revenues/profits and risk (how much more revenue/profit per head, for example) - in the end commits etc. aren't quite enough.
Doesn’t sound too disruptive yet, except for juniors and those who have invested in very specific skills e.g. performance. But this happened in two years. What’s going to happen over the coming two years?
One data point I’d be interested to know more about is if they’ve scaled back their hiring due to AI productivity gains, or if they’re just doing more engineering for the same money. As I see it that’s the big question for those of us who have to work for a living.
> Part of this shift is the natural expansion of business functions as the company scales.
Every R&D company on the planet has this cycle. Hire lots of developers and researchers. Build product market fit. Lay off all but essential personnel, while increasing the amount of sales and operations stuff as much as possible.
Anthropic this isn’t novel or unique in any way. Stop pretending your AI wonder machines are the reason why. This analysis lacks ANY comparison to other non AI research company hiring cycles. Do your homework please, this is pretty embarrassing.
This is a good HN discussion, but the post itself reads like advertising for their IPO. LOokt at us! We’re living the new high efficiency future and we’re building it too so buy the stock!
I wonder how complex the prompt was to write this.
They present and immediately take away their own value proposition. Do AI CEOs also buy tokens, or is that the reason for human employees ?
>"One possibility is single-person startups, each with a human founder steering swarms of agents to rake in billions in revenue. Another is companies built by AI genius CEOs employing human agents only to wield pipettes and testify before Congress."
Anthropic's hiring practices are rife with problems, potentially dangerous, and the expected value of interviewing there is very low.
Problem one: extreme subjective bias in candidate evaluation. Anthropic has a litany of red flag deal-killers they are evaluating for. They are subjective and not validated by any kind of scientific evaluation. Charitably, it is benign cargo-cultism (similar to being "un-Googley" from back in the day); uncharitably it is just voodoo. Interviewers use this to kill candidate loops via conscious and unconscious bias. Typically an interviewer will note one or more potential issues which can kill the loop (all candidates will say something which can be considered problematic in their interview; the potential problem space is infinite). If the interviewer likes the candidate, they will ask a follow-up question which allows the candidate to redeem themselves. If they dislike the candidate, they will ask questions which only probe for potential red flags and will not continue, or end the interview, before the candidate can reveal the redeeming half of their story. Probe until negative evidence is uncovered, and then a) continue to probe or b) terminate early, depending on the preexisting bias for the candidate.
Problem two: ghost jobs. Anthropic maintains an evergreen list of open positions that they never hire into, even senior ones. This is due to a few reasons: a) they are holding out for the ideal candidate, basically a god in human form. But that is only the reason they give themselves as to why they reject almost all candidates which make it through their loops. The real reason is b) existing employees and managers do not want another peer to collaborate with, instead what they really want is more agents for them to command. It's an intoxicating feeling and much preferred over in-person interaction. With fewer employees, it's more scope and power for the existing ones. It's well known in certain circles that interviewing there is a waste of time. It's ironic that they claim to love humanity but actively dislike actual humans.
Problem three: extreme dependence on agents for daily operations. Because they don't hire enough humans for the reasons above, they are outgunned and completely dependent on AI to operate. If AI were shown to be problematic or dangerous, they must still keep running it internally. Thus, they have no control. They are structurally compromised by AI; they don't have the capacity to put humans back into the loop. AI has already left the box and the box has been dismantled.
> Anthropic has expanded rapidly. But it has hired less per dollar of revenue than an earlier generation of large tech companies. At five years old, Anthropic takes in between 2–4 times as much revenue per employee as Meta and Google as they crested in their own rapid growth.
This may not be saying much at all. So much goes into deciding how fast to hire. Instagram had 13 employees / 1B rev at acquisition?
Also revenue per employee can vary a lot depending on the industry.
A better metric to understand AI effects might be revenue per employee at companies that have been around > 5 years that are AI-enthusiastic.
Commits per engineer 5.1x
Queries per data scientist 3.1×
Contracts per legal employee 3.1x
Deals per salesperson 1.7x
Hires per recruiter 1.4x
Invoices per AP processor 1.4x
These are the stats that might be hiding the possibly small potential for AI to generate "real" economic activity.
The first 3 stats, commits+queries+contracts, are secondary/meta statistics that could possibly go up 1000x and have very little affect on Anthropic's business, whereas that's much less likely to be the case with the last 3.
AI can make it very easy to do more and more busywork that doesn't move the needle.
If the above is true, AI econ must rely on a second order effect here in that busywork/BS is recognized more quickly and discontinued.
For example, if an organization paid an employee to compile a monthly report and it took them a very long time. But with AI, they can make that report every day. Now the organization realizes that they don't even need it.
OuterVale | 10 hours ago
bananaflag | 10 hours ago
tomlockwood | 10 hours ago
gottfired | 10 hours ago
Applejinx | 10 hours ago
mr_toad | 10 hours ago
tancop | 9 hours ago
grumple | 6 hours ago
tomelrod | 8 hours ago
bananaflag | 7 hours ago
I mean the thing you imagine I don't mean.
sebzim4500 | 10 hours ago
martijnvds | 10 hours ago
alex_suzuki | 10 hours ago
omnicognate | 10 hours ago
rzzzt | 9 hours ago
girvo | 10 hours ago
rwmj | 10 hours ago
dist-epoch | 9 hours ago
> Anthropic reportedly screens job candidates with a culture interview testing whether they would put the company's mission ahead of their future wealth. One applicant was asked how they would feel if a safety decision sent Anthropic's stock to zero.
https://finance.yahoo.com/markets/stocks/articles/anthropic-...
rdedev | 8 hours ago
oblio | 4 hours ago
skybrian | 9 hours ago
https://sre.google/sre-book/eliminating-toil/
girvo | 53 minutes ago
nswizzle31 | 9 hours ago
These people won’t necessarily need to be trained in CS, because in 10 years no one will need to know what a data structure is. So they can come from other fields (law, business) or just from other roles within the company.
Note this is basically how it already works in every non-technical field where there’s no clear definable “skill” on which to evaluate people.
rsfern | 9 hours ago
nswizzle31 | 9 hours ago
The point is that judgment won’t be needed on technical aspects. It will be more soft decisions like how to manage the agents.
oblio | 4 hours ago
Software, surprisingly, is a technical skill. Stop comparing it with poetry, sales or dancing.
dist-epoch | 9 hours ago
skybrian | 9 hours ago
q3k | 10 hours ago
Two vibecoders are walking down a forest path.
One of them spots a Zig codebase and challenges the other: “I'll give you $100 to convert this codebase to Rust!”. They accept and after spending $1000 worth of tokens on a thousand commits the codebase is now in Rust.
On the way back, the second vibecoder spots the same codebase again, and challenges back: “I'll give you $100 to convert this codebase to Zig!”. Another gajillion tokens later, the codebase is now in Zig.
Walking a little more, one vibecoder (in a moment of unusual clarity) asks the other: “It feels like we just did this for nothing”. “No”, the other responds, ”we just increased our company productivity by two thousand commits per day!”.
mono442 | 10 hours ago
mr_toad | 10 hours ago
However you measure it, in LOC or in MB or in ops per second, demand for more software always seems insatiable. This trend will likely continue up until software can be produced with zero humans at all. But I can’t see that happening unless we invent actual real artificial intelligence.
mono442 | 10 hours ago
rakejake | 10 hours ago
cannonpalms | 10 hours ago
pera | 10 hours ago
cloche | 10 hours ago
pera | 10 hours ago
Obviously it would need to be sustained for a reasonable amount of time while also showing at least stable finances.
jansport123 | 10 hours ago
dist-epoch | 9 hours ago
Elon already stopped talking about job loses and doom, and tries to talk only about the positive.
Jackobrien | 10 hours ago
- Deep curiosity and ability to absorb new information quickly - Evidence of ability to wield and coordinate agents to the max - Ability to clearly communicate ideas
You have these skills, you can conquer the software world.
iLoveOncall | 10 hours ago
Wow, so you worked with recent grads who are producing 100x the slop? No wonder Anthropic doesn't hire junior engineers.
thatsabadlook | 10 hours ago
"People with no taste or experience mass generating code by copying tickets into Claude and pressing accept is the best!" Ohhhhkie dokie broskie
discreteevent | 9 hours ago
But there are so many people who make statements like his that maybe we should assume that some of them are not idiots or bots. If he's not an idiot it means that he's working in an environment where that kind of statement is paying off. So it could be, for example, that he doesn't really have profitable customers. His main 'customers' are investors who are clueless, or who know that they are clueless but don't care because they are going to move it on to a greater fool anyway. Or that because there aren't any customers, the only thing they can measure is 'productivity' in terms of 'number of unused, unprofitable and unsustainable features'
thatsabadlook | 8 hours ago
There is another option to consider. What if they heard someone else say it and they believed it.
Now they go around, like many insecure people do, shouting about their unfounded take. Begging the universe to correct them or give them a good reason why they arent informed.
Then what usually happens with people who suffer from this type of short ordered thinking is, they realize they can't walk it back without social consequences because of the shouting they have done. So they just double down on it and find ways to live with the cognitive dissonance. Anthropic, openai, etc make billions off of people like this.
Because they propagate the BS advertising to people who don't know better, and then spend more money to try to make it real.
thatsabadlook | 10 hours ago
armanckeser | 10 hours ago
thatsabadlook | 10 hours ago
If everyone does it, no one monitors it anymore, and quality degrades the only way out is more tokens.
voidhorse | 9 hours ago
All they really need to do is convince enough people to adopt lower expectations and accept worse software products. At that point, mission accomplished. They've annihilated the source of value originally provided by people who actually know how to do things.
gatlin | 10 hours ago
rapidaneurism | 10 hours ago
TheOtherHobbes | 9 hours ago
I mean real profit, not "adjusted profit."
glimshe | 8 hours ago
As long as there other companies scooping up the juniors (which we cannot assume will happen), this could be sustainable.
jansport123 | 10 hours ago
zahlman | 9 hours ago
oblio | 7 hours ago
I've used vi and emacs and the number of their users that really 10x everyone else is vanishingly small. A large percentage of their userbases are people who wished they were the 10x but were much slower (or at best equal to) than the average Jetbrains tool users.
MattPalmer1086 | 8 hours ago
We used to call this RDD - Resume Driven Development.
maccard | 10 hours ago
The difference between a junior and a senior is still very real. With agents, knowing what to ask them to do (or not do) is bordering on more important than it ever was. A curious junior might still go down rabbit holes for things that a person who had done this before. That’s what makes them a junior. Not everything we _can_ do we _should_ do
xyzsparetimexyz | 10 hours ago
stogot | 3 hours ago
regexorcist | 10 hours ago
MichaelRo | 10 hours ago
rimbo789 | 10 hours ago
dang | 3 hours ago
If you wouldn't mind reviewing https://news.ycombinator.com/newsguidelines.html and taking the intended spirit of the site more to heart, we'd be grateful.
poisonborz | 8 hours ago
dang | 3 hours ago
https://news.ycombinator.com/newsguidelines.html
iLoveOncall | 10 hours ago
They are nowhere near profitable, have terribly bad products riddled with bugs (9,000+ open reports on Claude Code), and have no moat. Why would anyone listen to how they work internally?
eembees | 10 hours ago
Traubenfuchs | 10 hours ago
jansport123 | 10 hours ago
Imustaskforhelp | 9 hours ago
pyrale | 10 hours ago
There simply is no way to separate what’s reporting and what’s advertising.
thatsabadlook | 10 hours ago
vrganj | 10 hours ago
They are offering a cure to the defining contradiction of capitalism: The tension between capital wanting maximal self-reproduction and labor wanting a fair wage and a good life.
Anthropics pitch is to finally redeem capitalism by getting rid of those annoying workers and their annoying demands - that is why C-Suites are so excited about it, that is why its talked about in messianic language.
sedan_baklazhan | 10 hours ago
vrganj | 10 hours ago
Unless we somehow manage public control of AI, in which case we could hope for a Star Trek style technological utopia.
Rosa Luxemburgs word's have never rung truer - Socialism or Barbarism!
sedan_baklazhan | 10 hours ago
oblio | 9 hours ago
thatsabadlook | 10 hours ago
To the poster below who thinks anthropic is in this to destroy capitalism(lmao rofl), I guess maybe if you are interested in fuedal societies run by people who currently bomb schools, have chatbots and finely tuned recommendation engines designed to make children commit suicide but only after extracting as much capital as possible beforehand, try to sway elections in foreign countries, theottle the Internet so dictators can win wars, etc. Yea sure man they are all definitely interested in making the world better for everyone by also automating one of the few middle class jobs available in a country which has no manufacturing to fall back on. Super normal timeline.
SpicyLemonZest | 10 hours ago
sbarre | 10 hours ago
If you are already an advanced company filled with skilled people, you may not see the gains that a mid-tier company filled with mediocre seat warmers would see when their "best and brightest" start adopting AI to replace said seat warmers.
randysalami | 9 hours ago
An example: the UI applications suffer from massive amounts of repetitive code. People on hacker news will probably instantly jump to some kind of componentization strategy. More wise individuals probably understand you need to draw a winning strategy before that’s possible. Identify common workflows, from that, identify common components, figure out a parameterization and so on. The footguns are the slop having unintended consequences which fixing might actually cause production bugs by being “features”.
Hypothetical now: a current alternative solution which may exist is two simple detail views (of which there may be 30+ in the application) are 1000+ LoC template files, hand-coded, copy-paste, with all the errors you could imagine. You can get a lot of work from that so people are/could be content just fixing slop with slop and calling it engineering while piling on more features. Maybe it works but it doesn’t allow for much engineering room.
With AI tooling, popular support becomes less necessary - correctness matters more. Instead of galvanizing people to realize this is wrong (and it’s not just a want thing, with offshoring and the general difficulty of software engineering, many people don’t understand these concepts. They are great people but struggle understanding for loops intuitively; I don’t know what to say). Now I can provide my strategy to the tooling and do the decomposition and componentization myself. On top of this, my strategy can be more intelligent and robust because I’m not doing it by hand like concurrently using the API layer as a source of truth to give my agents more context on what the UI was trying to do. The code itself is confused so part of the work is learning the intent behind each module and also capturing bugs that are “features” to prevent regressions. Basically, a clean room rewrite (but we have access to the source code so I know that’s not the exact correct term).
The engineering moat is still there. If I say, “Claude make my UI codebase smaller” it will burn tokens and fail. One day it will succeed but right now it will fail. Likewise if I say “Claude suggest ways to make my codebase smaller”, it may suggest my strategy but it may not capture all the critical wrinkles. LLMs are not good at simplifying so to simplify (which is always a good goal in software engineering as long as the abstractions still hold), requires human strategies when complexity is high enough. Especially in my case where the behavior of the code is almost more emergent rather than engineered.
Then carrying on, hypothetically, I started to devise a system for not only rewriting the UI but the rest of the application (API and DB layer) as well as the corporate processes around the application. This may sound rude but these corporate processes can be disempowering and wasteful. The trick is inversion. First build a conceptual interface around the layer. The model will understand the purpose without understanding or rewriting the internals. This saves context and allows it to do more powerful reasoning by having perspective instead of being in a vacuum. Then when you have bandwidth, use that interface along with all the context you developed working on other modules, to take control. An example, you have a team which is responsible for building and deploying. To build and deploy, you may need to fill out forms, contact layers of people, all to achieve the end of goal of “building and deploying”. Through inversion, you communicate with an agent endpoint that handles communications to that team. Escalating, writing forms, and doing whatever is needed to get the code built and deployed. For all intents and purposes, you are now in control of that process. What used to be, for example, “I want my code built and deployed to QA” -> “message A” -> “message B” -> “write form A” -> “monitor for responses” -> “build triggered” can now be “I want my code built and deployed” -> “activate agent process” -> “build triggered”. This is an application of inversion in a nutshell. AI allows to apply engineering principles not only to our code but corporate processes where applicable.
So in this case, there could be 100+ individuals supporting this application and process. The one or two or three engineers that used to own the development process (more or less) were drowned out yet still responsible for the overall process and integrity of the software. It was an impossible task but now it seems more possible. The existential angst part comes from the all processes that were sprung up to support a reality where this all wasn’t possible.
Last, I definitely do not think the solution is one hundred engineers. If you know the show “Breaking Bad”, I actually find the reality of software engineering to be more like this now. The difference between 60, 90, 95, and even 98 percent purity is staggering. And the chemist who takes the input (money and tokens) and produces the highest quality output (codebase and processes), dominates. Just as in Breaking Bad when Walter White inherently has power, despite all other chemists and structures of authority, because his process had the highest purity output. It’s just business sense and adding more chemists does not increase purity. Output quantity is no longer gated by number of bodies so that argument goes out the window. Mentorship will be key (as in Breaking Bad) and large teams of all-high-quality producers should be put together into super high-thoroughput teams. But for regular operations, prioritize the highest quality output where the input is guided by the highest quality producer because adding more cooks will seriously degrade the output, at an exponential rate.
squidmaster | 10 hours ago
Consider for instance the bicycle youtuber Berm Peak, who's been using AI to reverse engineer and hack shitty ebikes. This kind of work is outside his expertise, so the relative lift he gets is massive.
JDups | 9 hours ago
voidhorse | 9 hours ago
bunderbunder | 9 hours ago
squidmaster | 9 hours ago
bunderbunder | 9 hours ago
squidmaster | 9 hours ago
Note that I specifically mentioned that I'm skeptical of claims of insane lift in workplace contexts. Several factors work against it there, organizational dysfunction not least of all, not to mention a disconnect between the people doing the work and the people for whom the work has value. People being crunched by their boss to ship some corporate portal software faster are going to make shit. You making something for yourself, to solve your own problem to your own satisfaction, needing to clear problems outside but adjacent to your proficiencies, should be able to do much better. Many people have found it to be so.
bunderbunder | 8 hours ago
These tools excel at greenfield development, but they genuinely struggle in mature codebases. Especially when it’s a legacy system with spotty documentation.
ozgrakkurt | 9 hours ago
There is no measurement. Think about how much of a difference empirical science made in such a short amount of time. Or think about how it is impossible to optimize software without measurement etc.
Also there is the AI companies’ atrocious reliability even if you trust the measurements
siva7 | 9 hours ago
Mawr | 9 hours ago
simianwords | 9 hours ago
hypfer | 10 hours ago
So while some stuff may transfer over as a general thing, a lot will probably not.
Still interesting, of course, but to be consumed with a mountain of salt.
dist-epoch | 9 hours ago
I think you left a tiny bit out: the smartest model in the world
probably irrelevant detail
mrmarket | 10 hours ago
also 'lines of code' or any other output based metric in the age of infinite slop is not a flex/indicator of usefulness or value.
and yes, everyone spends all their time reading Claude output. that doesn't mean better work is getting done, or progress is being made. i've been reviewing and fixing claude slop at work for at least a year. if anything companies/products are getting more homogonized and bloated. thanks anthropic, you are going to save the world
sreekanth850 | 10 hours ago
fwlr | 9 hours ago
angry_octet | 8 hours ago
sreekanth850 | 7 hours ago
gb2d_hn | 10 hours ago
mattmaroon | 10 hours ago
Don’t fear change.
sedan_baklazhan | 10 hours ago
If you don’t pay them, you can walk by feet. If they ban you, you can walk by feet. If their infrastructure is damaged, no one and nothing will be able to move.
mister_mort | 10 hours ago
ndsipa_pomu | 9 hours ago
nswizzle31 | 9 hours ago
Good point about infra, but that’s already the case with Cloudflare, GitHub, aws, etc
sedan_baklazhan | 9 hours ago
BTW, when (not if, but when) a war happens, they will become very nice targets to render this exact country completely disabled.
horsawlarway | 9 hours ago
Humans are autonomous, distributed, and adaptable. In a package that weighs ~150 lbs, with very low energy requirements.
You literally cannot wipe them out without a sustained and incredibly expensive ground campaign, and even then it's not likely.
AI in a data center on the other hand... So easily disabled.
sedan_baklazhan | 9 hours ago
And voila. No cargo moved, no code changes or fixes, no medical diagnosis made,etc. A whole country is disabled.
zeryx | 9 hours ago
I thought that was an interesting idea, but not something that could ever happen..
nswizzle31 | 9 hours ago
Re: war.. that’s why US spends orders of magnitude more on military than anyone else. If any infra is hit on US soil then im heading to the bunker because most of the world will have already been reduced to dust.
sedan_baklazhan | 8 hours ago
prollings | 8 hours ago
zeryx | 9 hours ago
Paradigma11 | 7 hours ago
nasretdinov | 9 hours ago
10xDev | 9 hours ago
anal_reactor | 9 hours ago
My manager was dragging me through useless projects that went nowhere despite multiple people telling him "it ain't gonna work" while shooting down proposals that later turned out to be "high priority tasks required directly by upper management". And he's a well-paid, educated employee with years of experience. Don't even get me started on delivery guys who cannot read a simple instruction like "do not ring the bell, leave the package on the doorstep". They just follow a 50-line script and that's all they can do.
Whenever you mention that AI won't reach human intelligence, remember that this is the benchmark. Not Alexander the Great, it's the millions of people throwing their life savings ($1000 each) into Labubus. Long time ago there was a series of articles in my country about the costs of wedding receptions skyrocketing and some people taking out loans to pay that, and someone made a meme "robots will never have human intelligence!" with reply "bro, you took out a loan to pay for a wedding, I don't want your human intelligence".
Denying this reality is pure cope.
nasretdinov | 8 hours ago
jvanderbot | 8 hours ago
You are showing disagreements with your priorities (mostly money) as unintelligent. You're showing a lack of care and conscientiousness as unintelligent.
You're reducing intelligence to alignment with money and following instructions. Under that framing the argument is already won in favor of AI amd some vague effective maximizing of a cost function.
Many jobs are like this, so it's fine for employment. Maybe the vast majority of mankind was never suitable for reliable employment of the "cost function" style.
But I know two things:
1. Centralization should be avoided. The reality that AI can do most jobs is not a reason to concentrate power behind those portraying that or building AI. We still want local, ideally individual, empowerment. I don't see enough said about this. Maybe we can even say the system must be one people wish to opt in to, to maximize profit.
2. Intelligence is being very narrowly defined and humans value more than an INT score in the RPG sense. The more I meet people outside my belief circles, the more I realize the intelligence of some actions, in the sense of "logically deduced from a higher principle" not "aligned with my western white collar principles". The ultimate role of culture and humanity IS this principle exploration and to say we've reduced that exploration to "maximize X" is crazy.
At some point deciding fate of the company or system or state flows from principles embedded in cultures and that's in danger of being flattened away by optimizing outcomes with machines.
Vague and wishy washy, but its the clearest I can state it.
anal_reactor | 3 hours ago
If the principle embedded in the culture of my manager is to do dumb shit that goes nowhere, then I'm not sure what goals we're achieving.
jvanderbot | an hour ago
jst1fthsdys | 8 hours ago
Denying that reality is pure cope.
andsoitis | 9 hours ago
Self-driving requires less holistic thinking than imagining, building, and evolving large scale systems. It requires abilities that LLMs do not have. It requires non-symbolic reasoning, intuition, etc. that are beyond LLMs.
10xDev | 9 hours ago
andsoitis | 9 hours ago
Intuition exists outside the human brain?
10xDev | 9 hours ago
andsoitis | 9 hours ago
No.
leoc | 9 hours ago
x______________ | 9 hours ago
> Don’t fear change.
Change is inevitable.
What I understood from the comment is that Op fears the boiling frog scenario. [0]
0 https://en.wikipedia.org/wiki/Boiling_frog
leoc | 9 hours ago
10xDev | 9 hours ago
That's the point, they are setting up the economy to be run by agents and they trust that it is possible. Step 1. Teach AI human knowledge (2022-2023). Step 2. Make AI agentic (2025). Step 3. Start autonomous research and create own corpus of knowledge (2026).
Step 3 will keep compounding (as we are witnessing) and most people will be left behind. I'm not against a revolution, I'm all for it.
RandomLensman | 9 hours ago
10xDev | 9 hours ago
RandomLensman | 9 hours ago
torben-friis | 9 hours ago
It's not only AI. Betting is on the rise, adults and minors are being hooked by dopamine loops, social networks are causing well measured negative effects, children's educational scores are sinking worldwide, appliance manufacturers engage in surveillance...
It truly feels like the wild west.
philistine | 8 hours ago
In the US. When the inevitable legislations start popping up worldwide, don't be surprised when somehow the American consensus will be to do nothing.
Avicebron | 8 hours ago
AlecSchueler | 4 hours ago
ffsm8 | 9 hours ago
fwiw, i think literally all largish corporations have been like that pre AI too? i can gurantee you that nobody understands how the whole system works at all of my last workplaces either...
people just know their small part and hope for the best - so from a devils advocate pov, its already been proved that its sustasinable. now wherever its _worth it_ is another question. large employers are - on average - incredibly wasteful with their resources. not to anthropics degree, but still.
if youre boostrapping, and VC token subventions go away... i suspect it'd be not worth it to develop like that. But i guess we will find out in a few month to 1-2 years
squidmaster | 8 hours ago
rwmj | 8 hours ago
zeroxfe | 8 hours ago
How about the dollar you earned, spent, invested? Do you know how it works?
How about the clothes you wear?
This is just a tiny subset of the enormous complexity we're immersed in, and no individual or collective has a complete understanding of it. If you do, you're truly an outlier and GP's point still stands.
rwmj | 6 hours ago
isignal | 4 hours ago
Even for the COBOL based systems someone understands it well enough at the conceptual and business level. And when there's no one to the company migrates and modernizes.
You are conflating the subject matter expert who understands one area well and others not so well with a general newbie who understands nothing well
zeroxfe | 2 hours ago
I don't think it is. @squidmaster specifically noted, 'The "Renaissance man" era is long gone.'
duckmysick | 5 hours ago
Paradigma11 | 8 hours ago
squidmaster | 7 hours ago
argee | 5 hours ago
Is that really a problem though? We can solve problems empirically even if we don't keep the entire system's workings in our heads.
Jordan-117 | 7 hours ago
9dev | 10 hours ago
And all of this is then somehow classified as responsible development of AI, by people who claim AI could eradicate humanity, and that RSI will be our downfall. How exactly is outsourcing your work on agents to agents not RSI at this point? Why are these people still allowed to operate again?
somewhereoutth | 10 hours ago
> An executive assistant, who built a personal assistant agent for scheduling and email triage
Really? We should build our own agents for everything? Sounds a bit token-spendy.
Also, as far as I can make out, the paper was not written by AI (nor even with AI assistance?). You'd think they'd make a song and dance about it if it was. Maybe AI can't do this sort of thing yet (or ever).
rameshrathiakg | 10 hours ago
thatsabadlook | 10 hours ago
mrleinad | 9 hours ago
oblio | an hour ago
tinyhouse | 10 hours ago
Anyhropic is not a good example for anything. Their scale and growth is way beyond what like 99% of companies experiece. Their optimization formula is different.
ytxm | 10 hours ago
orourke | 10 hours ago
SilverSlash | 10 hours ago
Say that it is borderline true (it's not, but let's pretend). Then companies can lay off 90 devs for every 100, and buy the remaining 10 devs Claude subscriptions. Isn't it better if they keep all 100 and buy everyone a Claude sub?
So shouldn't Anthropic actually be pushing the narrative that Claude doesn't replace devs but actually makes them vastly better at their jobs and so every company, big and small, should buy every dev the highest tier Claude enterprise subscription?
rwmj | 8 hours ago
They don't want higher productivity because what are they going to do with that? Launching new products is costly and uncertain. They don't want to be paying the same number of developers and have them all spending money on Claude subs. But the idea of firing 90% of them and having the rest in fear for their jobs? Brilliant!
Back in the real world, the question is what's going to happen to companies that go through with this. When they have vast amounts of slop code that no one understands? When you have to sit in with a customer and explain why the product doesn't work and no one knows even how it works never mind how to fix it? We can already see the inklings of this in the huge buggy messes that Microsoft and Anthropic are producing. Interesting times.
cmiles8 | 10 hours ago
In other news, Turkey farmers encourage you to buy more turkey for Thanksgiving.
andsoitis | 10 hours ago
I'm too lazy to bother looking up each of their credentials. Are any of them software engineers? Do they know anything about creating, maintaining, and evolving software?
I do know that Anthropic's CEO isn't a software engineer, so it would be hard for him to evaluate their claims with any deep insight from experience.
andsoitis | 9 hours ago
simianwords | 9 hours ago
andsoitis | 9 hours ago
According to his LinkedIn profile, the software development he did was for less than 3 years in a very narrow field. I would call that very limited.
Coupled with the incentives driving him, his judgment would be heavily clouded.
andsoitis | 9 hours ago
voidhorse | 9 hours ago
Anyone who has been in software for a while should hear alarm bells going off when they read this. Every line of code is a liability, not a trophy. This is what happens when people who don't know and don't care to learn about the existing body of knowledge in a discipline "reinvent" it.
Give it another year or two and Anthropic will be writing all about how they lowered LOC and got agents to be more conservative about generating new code and introduced a "verification gate" for every new line produced.
mrleinad | 9 hours ago
claykkari | 8 hours ago
skydhash | 7 hours ago
The compiler development teams make sure that the output of the compiler is correct according to the source code input, preserving semantics and issuing errors for incorrect statements.
Who is doing that for LLM prompts?
claykkari | 6 hours ago
skydhash | 4 hours ago
Imagine worrying about gcc or the python interpreter that much. No wonder they use the label "reverse centaur" for LLM usage. You must think carefully before coming up with a prompt, then review the code carefully for correctness. Double the stress for no reason. Or is it just going with the vibes and slinging slop?
grumple | 6 hours ago
siva7 | 9 hours ago
I don't wanna be the guy calling out the lie or slop but when did Anthropic hire in the last two years for junior roles? OAI and Anthropic could literally hire seamlessly the best of the best worldwide for every job they advertise without having the need to think about the poor juniors.
RandomLensman | 9 hours ago
blini-kot | 9 hours ago
it was interesting in the beginning when there was technical substance, but now its just techno-bros vision of deep thoughts
simianwords | 9 hours ago
“Anthropic bad because automation of labour”
“Claude bad because it hallucinated that one time”
“This is a marketing post to hype up ipo numbers”
vinitjain24 | 9 hours ago
bjornsing | 8 hours ago
One data point I’d be interested to know more about is if they’ve scaled back their hiring due to AI productivity gains, or if they’re just doing more engineering for the same money. As I see it that’s the big question for those of us who have to work for a living.
usernametaken29 | 8 hours ago
Every R&D company on the planet has this cycle. Hire lots of developers and researchers. Build product market fit. Lay off all but essential personnel, while increasing the amount of sales and operations stuff as much as possible.
Anthropic this isn’t novel or unique in any way. Stop pretending your AI wonder machines are the reason why. This analysis lacks ANY comparison to other non AI research company hiring cycles. Do your homework please, this is pretty embarrassing.
gumby | 8 hours ago
I wonder how complex the prompt was to write this.
kirykl | 7 hours ago
>"One possibility is single-person startups, each with a human founder steering swarms of agents to rake in billions in revenue. Another is companies built by AI genius CEOs employing human agents only to wield pipettes and testify before Congress."
subfocal | 3 hours ago
Problem one: extreme subjective bias in candidate evaluation. Anthropic has a litany of red flag deal-killers they are evaluating for. They are subjective and not validated by any kind of scientific evaluation. Charitably, it is benign cargo-cultism (similar to being "un-Googley" from back in the day); uncharitably it is just voodoo. Interviewers use this to kill candidate loops via conscious and unconscious bias. Typically an interviewer will note one or more potential issues which can kill the loop (all candidates will say something which can be considered problematic in their interview; the potential problem space is infinite). If the interviewer likes the candidate, they will ask a follow-up question which allows the candidate to redeem themselves. If they dislike the candidate, they will ask questions which only probe for potential red flags and will not continue, or end the interview, before the candidate can reveal the redeeming half of their story. Probe until negative evidence is uncovered, and then a) continue to probe or b) terminate early, depending on the preexisting bias for the candidate.
Problem two: ghost jobs. Anthropic maintains an evergreen list of open positions that they never hire into, even senior ones. This is due to a few reasons: a) they are holding out for the ideal candidate, basically a god in human form. But that is only the reason they give themselves as to why they reject almost all candidates which make it through their loops. The real reason is b) existing employees and managers do not want another peer to collaborate with, instead what they really want is more agents for them to command. It's an intoxicating feeling and much preferred over in-person interaction. With fewer employees, it's more scope and power for the existing ones. It's well known in certain circles that interviewing there is a waste of time. It's ironic that they claim to love humanity but actively dislike actual humans.
Problem three: extreme dependence on agents for daily operations. Because they don't hire enough humans for the reasons above, they are outgunned and completely dependent on AI to operate. If AI were shown to be problematic or dangerous, they must still keep running it internally. Thus, they have no control. They are structurally compromised by AI; they don't have the capacity to put humans back into the loop. AI has already left the box and the box has been dismantled.
djoldman | 2 hours ago
This may not be saying much at all. So much goes into deciding how fast to hire. Instagram had 13 employees / 1B rev at acquisition?
Also revenue per employee can vary a lot depending on the industry.
A better metric to understand AI effects might be revenue per employee at companies that have been around > 5 years that are AI-enthusiastic.
djoldman | 2 hours ago
The first 3 stats, commits+queries+contracts, are secondary/meta statistics that could possibly go up 1000x and have very little affect on Anthropic's business, whereas that's much less likely to be the case with the last 3.
AI can make it very easy to do more and more busywork that doesn't move the needle.
If the above is true, AI econ must rely on a second order effect here in that busywork/BS is recognized more quickly and discontinued.
For example, if an organization paid an employee to compile a monthly report and it took them a very long time. But with AI, they can make that report every day. Now the organization realizes that they don't even need it.
sjfr | 28 minutes ago
So... The org is equivalent to 7500 engineers ? What are these 7500 engineers doing exactly ?
I know the infra is nothing trivial. And that there's a tendency to underestimate the difficulty of what others do. But 7500 ?
I'm not sure it's the productivity revolution that it is hyped up to be.