4 ms·

Why do they need the best software engineers? I thought their product was supposed to replace such roles. Yet look at the positions they’re hiring for in market

Why do they need the best software engineers? I thought their product was supposed to replace such roles. Yet look at the positions they’re hiring for in marketing, finance etc.: https://www.anthropic.com/careers/jobs https://www.anthropic.com/careers/jobs Why aren’t they dogfooding their own products to replace such roles?

fredoliveira 5mo ago
They most certainly are. This is Jevons paradox.
abirch 5mo ago
AI can let you downsize the number of employees that you have and maintain the status quo or it can let you maintain the number of employees, reduce technical debt, improve products, and services.
- IncreasePosts 5mo ago
  Do the economics work out ? You can downsize the devs you have, but you need to maintain a smaller stable of very expensive devs, and then factor in the token usage. For example, a recent story about the openclaw creator using $1.3M of tokens/month. And let's assume he's getting paid $5M/yr which is probably a vast under estimate. Is he providing value that a traditional software development org with normal developers couldn't provide for $20M/yr?
  - ncphillips 5mo ago
    I feel like using $1.3M/year is a wild outlier. A $200/month Max sub is a pretty cheap way to get quite a lot of benefit.
    - IncreasePosts 5mo ago
      His usage is actually 12x that ...
  - fredophile 5mo ago
    The $1.3 million doesn't mean much. The article stated he could've switched to a significantly cheaper option and cut the bill to $300k. That's still a lot but since he worked for the company that sells the tokens it isn't as though they were paying the retail cost.
  - rft 5mo ago
    For those who, like me, had to do a double take on that number: https://www.businessinsider.com/openclaw-peter-steinberger-ai-token-bill-2026-5 https://www.businessinsider.com/openclaw-peter-steinberger-a... Yes, $1.3M in token cost in less than 30 days and some days were even off-peak, if you can call it that with that insane scale that likely hides quite a lot of tokens in the lower bars. HN thread: https://news.ycombinator.com/item?id=48159227 https://news.ycombinator.com/item?id=48159227
  - fnordpiglet 5mo ago
    The issue is there’s non linearities involved. Although I don’t know I would use the open claw guy, but let’s take Isaac Newton. You can’t sum up people and arrive at an Isaac Newton worth of talent. He’s singular, unique, and irreplaceable at what he did. There were others similarly outsized in their ability to change things, and there are today as well. But you can’t funge talent at some level with more people - in fact as we know there’s a rapidly diminishing return on people investment. Finally in some ways agentic workflows magnify the power of the individual who is adept at harnessing them, they don’t have to argue (much) with the agents to effect their ideas. I’ve found a lot of very bright engineers spend their days fighting to be heard by managers and peers who can’t / won’t understand them. By unshackling them from trying to debate down idiots, they deliver way way more, and of the right things, than they otherwise could have.
    - grogenaut 5mo ago
      | Newton... He's Singular, unique, and irreplaceable at what he did. Leibniz, literally parallel. Was Newton just a smart guy at the right place and the right time. These smart folks require other smart folks to understand and verify what they did. There are many who have amazing pedigrees in history.
      - fnordpiglet 5mo ago
        Both Newton and Leibniz intersected with calculus but had incredibly diverse careers spanning many subjects that were totally orthogonal. So, no, in fact the fact they discovered calculus in very different ways proves the point even further - they were so singular as to arrive at similar breathtaking insights from totally different perspectives totally independently, for the first time in all of mankind. You can’t substitute one for the other, and each was totally indispensable and irreplaceable.
    - tintor 5mo ago
      You can replace Newton with Leibniz.
      - 5mo ago
  - klooney 5mo ago
    I honestly believe that most Silicon Valley developers could have their pre-AI output replaced by 10 dollars/day in Gemini 3 pay as you go spend. I don't think existing companies will bite that bullet, but I think you'll see AI native companies in five years with like, a baffling small number of people.
    - bendmorris 5mo ago
      Why in five years? Where are these hyper productive small companies running laps around bigger ones right now?
      - arkh 5mo ago
        > Where are these hyper productive small companies running laps around bigger ones right now? Getting bought by the bigger companies as it is currently the goal of most start-up since no one cares about monopolies and anti competitive behavior nowadays.
- bandrami 5mo ago
  But Anthropic is adding employees
  - osigurdson 5mo ago
    Maybe smart to convince everyone else to fire and slow down while you hire and speed up. I don't think it is actually planned, but perhaps smart in hindsite.
joe_mamba 5mo ago
>Why do they need the best software engineers? I thought their product was supposed to replace such roles. Who claimed that? Their customers will be happy if their product replaces all the junior positions and midwit developers off the payroll. then that's already a huge saving to any company's bottom line. Even if it doesn't directly replace workers, reducing the bargaining power of those spoiled SW devs and not having to give them huge raises all the time or they leave, is still enough. That's the whole point of layoffs and offshoring anyway.
- bandrami 5mo ago
  > that's already a huge saving to any company's bottom line Possibly not if they are paying the full cost of inference
- resonious 5mo ago
  > Who claimed that? Dario Amodei
- JJOKOCHAA 5mo ago
  So you are happy with people losing jobs?
perplex 5mo ago
I've seen this at work. Giving Claude Code to a mediocre programmer gets you mediocre results. The really effective engineers with coding agent can accomplish a lot. Thousand monkeys...
- kaashif 5mo ago
  I've also seen this. But I'll extend it to saying that giving Claude to a bad programmer gives bad results. And seeing how people use it: good programmers review output and iterate to get better output. But bad programmers simply trust the output is good: they have no ability to review it themselves and often don't try.
  - Scoundreller 5mo ago
    As a mostly non-programmer it got me a lot done. With about 5-10h over the weekend using free tier Claude and ChatGPT I managed to put together a scraper for a particular thing on a website I’m interested, grab the item images, do an initial pass with local OCR, if it hit some keywords, run openCV to crop for better OCR and dump the hits for further investigation. Nothing particularly advanced but it would have taken me a horrendous amount of time without it to be half as good, like it did when I built a similar scraper 10 years ago. Neither were very good code quality i’m sure.
    - huflungdung 5mo ago
      [dead]
    - simplyluke 5mo ago
      This is broadly true of a bunch of jobs/fields with LLMs, but particularly true for programming. They raise the floor to a point where a generally capable person can put something like that together, or come up with a passably okay visual design, or decent-enough written language. I've been using them heavily to get some laughably basic CAD work done for small 3d printed projects. Stuff that absolutely makes my mechanical engineer friends roll their eyes at me. An expert can either use the tool more effectively, or see all the issues in a less experienced person's output. Both of these are good things, the mistake a ton of people are making is experiencing industrial scale Dunning-Kruger and thinking "Only my expertise is still valuable, every other white collar role is done!" The second-order mistake is thinking that raising the floor like that devalues expertise instead of increasing demand for it. The net-effect of me starting to play with CAD because it's a little easier now isn't that I don't hire my friends who are experts to make a tiny spacer I'm going to 3d print, I never would have hired them for that, it's that maybe I start learning the skills and decide to take on a more ambitious project where I do need to hire one of them for some help, or start ordering custom CNC'd parts -- scale that to the entire economy.
alexwwang 5mo ago
They need genius to promote the ceiling of LLM capability or dig out the potential of it. The best SEs often have better tastes in technique and business and they know why while knowing how. So there's a higher possibility for them to do master works quickly with the help and harness of LLM.
rienbdj 5mo ago
Effective AI use (in my experience) has human doing the load bearing parts by hand (schemas, api spec, overall architecture, domain types) then AI fills in the blanks.
- musebox35 5mo ago
  Could you briefly describe your workflow for doing that or give a pointer to a blog you wrote/like that aligns with the process? Thanks in any case, happy designing ;-)
  - lionkor 5mo ago
    I'm not who you asked, but my workflow is the same it always was, just that one tool in the process is an LLM. That's it. If I find a task to be annoying and repetetive, I use an LLM to do it, then validate it. The quality of my work only improves, and so does the speed, and I don't lose my skill over time.
    - musebox35 5mo ago
      I am exploring ways to document the design for the agent to read and update. What makes it difficult is the lack of structure. Spec writing is not my core skill. Schemas and APIs are easier, there are declarative ways to document them. Runtime concepts and workflows have less structure and writing prose seems so unstructured for my taste. Formal languages are too rigid. But I could not find a better way.
      - tssge 5mo ago
        For me writing manpages as a spec worked well. While the format is unintuitive and tedious (at least to me), I used AI to format the manpages from my descriptions and rendered to PDF to review. Though it's been mainly useful for CLI utilities. Once I got the manpages ready and the way I wanted, I just stored those in an empty project folder and told the agent to implement as specified in a language of my choosing. Maybe for other development scenarios like GUI there is some "native" way to spec the project in a common format like this.
troupo 5mo ago
> Why aren’t they dogfooding their own products to replace such roles? Oh, they do. That's why they buy actual engineers to fix the vibe-coded slop they produce en masse: https://x.com/jarredsumner/status/2026497606575398987 https://x.com/jarredsumner/status/2026497606575398987 Unfortunately those engineers slowly succumb to the same slopcoding vibes, so they go and buy a bunch of new engineers.

fredoliveira 5mo ago

They most certainly are. This is Jevons paradox.

abirch 5mo ago

AI can let you downsize the number of employees that you have and maintain the status quo or it can let you maintain the number of employees, reduce technical debt, improve products, and services.

IncreasePosts 5mo ago
Do the economics work out ? You can downsize the devs you have, but you need to maintain a smaller stable of very expensive devs, and then factor in the token usage. For example, a recent story about the openclaw creator using $1.3M of tokens/month. And let's assume he's getting paid $5M/yr which is probably a vast under estimate. Is he providing value that a traditional software development org with normal developers couldn't provide for $20M/yr?
- ncphillips 5mo ago
  I feel like using $1.3M/year is a wild outlier. A $200/month Max sub is a pretty cheap way to get quite a lot of benefit.
  - IncreasePosts 5mo ago
    His usage is actually 12x that ...
- fredophile 5mo ago
  The $1.3 million doesn't mean much. The article stated he could've switched to a significantly cheaper option and cut the bill to $300k. That's still a lot but since he worked for the company that sells the tokens it isn't as though they were paying the retail cost.
- rft 5mo ago
  For those who, like me, had to do a double take on that number: https://www.businessinsider.com/openclaw-peter-steinberger-ai-token-bill-2026-5 https://www.businessinsider.com/openclaw-peter-steinberger-a... Yes, $1.3M in token cost in less than 30 days and some days were even off-peak, if you can call it that with that insane scale that likely hides quite a lot of tokens in the lower bars. HN thread: https://news.ycombinator.com/item?id=48159227 https://news.ycombinator.com/item?id=48159227
- fnordpiglet 5mo ago
  The issue is there’s non linearities involved. Although I don’t know I would use the open claw guy, but let’s take Isaac Newton. You can’t sum up people and arrive at an Isaac Newton worth of talent. He’s singular, unique, and irreplaceable at what he did. There were others similarly outsized in their ability to change things, and there are today as well. But you can’t funge talent at some level with more people - in fact as we know there’s a rapidly diminishing return on people investment. Finally in some ways agentic workflows magnify the power of the individual who is adept at harnessing them, they don’t have to argue (much) with the agents to effect their ideas. I’ve found a lot of very bright engineers spend their days fighting to be heard by managers and peers who can’t / won’t understand them. By unshackling them from trying to debate down idiots, they deliver way way more, and of the right things, than they otherwise could have.
  - grogenaut 5mo ago
    | Newton... He's Singular, unique, and irreplaceable at what he did. Leibniz, literally parallel. Was Newton just a smart guy at the right place and the right time. These smart folks require other smart folks to understand and verify what they did. There are many who have amazing pedigrees in history.
    - fnordpiglet 5mo ago
      Both Newton and Leibniz intersected with calculus but had incredibly diverse careers spanning many subjects that were totally orthogonal. So, no, in fact the fact they discovered calculus in very different ways proves the point even further - they were so singular as to arrive at similar breathtaking insights from totally different perspectives totally independently, for the first time in all of mankind. You can’t substitute one for the other, and each was totally indispensable and irreplaceable.
  - tintor 5mo ago
    You can replace Newton with Leibniz.
    - 5mo ago
- klooney 5mo ago
  I honestly believe that most Silicon Valley developers could have their pre-AI output replaced by 10 dollars/day in Gemini 3 pay as you go spend. I don't think existing companies will bite that bullet, but I think you'll see AI native companies in five years with like, a baffling small number of people.
  - bendmorris 5mo ago
    Why in five years? Where are these hyper productive small companies running laps around bigger ones right now?
    - arkh 5mo ago
      > Where are these hyper productive small companies running laps around bigger ones right now? Getting bought by the bigger companies as it is currently the goal of most start-up since no one cares about monopolies and anti competitive behavior nowadays.
bandrami 5mo ago
But Anthropic is adding employees
- osigurdson 5mo ago
  Maybe smart to convince everyone else to fire and slow down while you hire and speed up. I don't think it is actually planned, but perhaps smart in hindsite.

joe_mamba 5mo ago

>Why do they need the best software engineers? I thought their product was supposed to replace such roles. Who claimed that? Their customers will be happy if their product replaces all the junior positions and midwit developers off the payroll. then that's already a huge saving to any company's bottom line. Even if it doesn't directly replace workers, reducing the bargaining power of those spoiled SW devs and not having to give them huge raises all the time or they leave, is still enough. That's the whole point of layoffs and offshoring anyway.

bandrami 5mo ago
> that's already a huge saving to any company's bottom line Possibly not if they are paying the full cost of inference
resonious 5mo ago
> Who claimed that? Dario Amodei
JJOKOCHAA 5mo ago
So you are happy with people losing jobs?

perplex 5mo ago

I've seen this at work. Giving Claude Code to a mediocre programmer gets you mediocre results. The really effective engineers with coding agent can accomplish a lot. Thousand monkeys...

kaashif 5mo ago
I've also seen this. But I'll extend it to saying that giving Claude to a bad programmer gives bad results. And seeing how people use it: good programmers review output and iterate to get better output. But bad programmers simply trust the output is good: they have no ability to review it themselves and often don't try.
- Scoundreller 5mo ago
  As a mostly non-programmer it got me a lot done. With about 5-10h over the weekend using free tier Claude and ChatGPT I managed to put together a scraper for a particular thing on a website I’m interested, grab the item images, do an initial pass with local OCR, if it hit some keywords, run openCV to crop for better OCR and dump the hits for further investigation. Nothing particularly advanced but it would have taken me a horrendous amount of time without it to be half as good, like it did when I built a similar scraper 10 years ago. Neither were very good code quality i’m sure.
  - huflungdung 5mo ago
    [dead]
  - simplyluke 5mo ago
    This is broadly true of a bunch of jobs/fields with LLMs, but particularly true for programming. They raise the floor to a point where a generally capable person can put something like that together, or come up with a passably okay visual design, or decent-enough written language. I've been using them heavily to get some laughably basic CAD work done for small 3d printed projects. Stuff that absolutely makes my mechanical engineer friends roll their eyes at me. An expert can either use the tool more effectively, or see all the issues in a less experienced person's output. Both of these are good things, the mistake a ton of people are making is experiencing industrial scale Dunning-Kruger and thinking "Only my expertise is still valuable, every other white collar role is done!" The second-order mistake is thinking that raising the floor like that devalues expertise instead of increasing demand for it. The net-effect of me starting to play with CAD because it's a little easier now isn't that I don't hire my friends who are experts to make a tiny spacer I'm going to 3d print, I never would have hired them for that, it's that maybe I start learning the skills and decide to take on a more ambitious project where I do need to hire one of them for some help, or start ordering custom CNC'd parts -- scale that to the entire economy.

alexwwang 5mo ago

They need genius to promote the ceiling of LLM capability or dig out the potential of it. The best SEs often have better tastes in technique and business and they know why while knowing how. So there's a higher possibility for them to do master works quickly with the help and harness of LLM.

rienbdj 5mo ago

Effective AI use (in my experience) has human doing the load bearing parts by hand (schemas, api spec, overall architecture, domain types) then AI fills in the blanks.

musebox35 5mo ago
Could you briefly describe your workflow for doing that or give a pointer to a blog you wrote/like that aligns with the process? Thanks in any case, happy designing ;-)
- lionkor 5mo ago
  I'm not who you asked, but my workflow is the same it always was, just that one tool in the process is an LLM. That's it. If I find a task to be annoying and repetetive, I use an LLM to do it, then validate it. The quality of my work only improves, and so does the speed, and I don't lose my skill over time.
  - musebox35 5mo ago
    I am exploring ways to document the design for the agent to read and update. What makes it difficult is the lack of structure. Spec writing is not my core skill. Schemas and APIs are easier, there are declarative ways to document them. Runtime concepts and workflows have less structure and writing prose seems so unstructured for my taste. Formal languages are too rigid. But I could not find a better way.
    - tssge 5mo ago
      For me writing manpages as a spec worked well. While the format is unintuitive and tedious (at least to me), I used AI to format the manpages from my descriptions and rendered to PDF to review. Though it's been mainly useful for CLI utilities. Once I got the manpages ready and the way I wanted, I just stored those in an empty project folder and told the agent to implement as specified in a language of my choosing. Maybe for other development scenarios like GUI there is some "native" way to spec the project in a common format like this.

troupo 5mo ago

> Why aren’t they dogfooding their own products to replace such roles? Oh, they do. That's why they buy actual engineers to fix the vibe-coded slop they produce en masse: https://x.com/jarredsumner/status/2026497606575398987 https://x.com/jarredsumner/status/2026497606575398987 Unfortunately those engineers slowly succumb to the same slopcoding vibes, so they go and buy a bunch of new engineers.