4 ms·

> This Rust rewrite would've taken a team of engineers with full-context on the codebase a year of work. With 1 engineer using Fable & closely monitoring Claude

by dabinat 3mo ago

> This Rust rewrite would've taken a team of engineers with full-context on the codebase a year of work. With 1 engineer using Fable & closely monitoring Claude Code, we went from start to 100% of the test suite passing on all platforms in 11 days. This is impressive from a technological standpoint, but it does gloss over the fact that it would have cost $165k in tokens were Bun not part of Anthropic. The comparison here isn’t completely fair - it would take a small team a year to port it if they spent $0 extra on it. I’d be interested to see a comparison between spending $165k in 11 days on Claude vs splitting that between 50 people over 11 days for a line-by-line rewrite of the Zig code. I suspect Claude might be faster and therefore cheaper, but maybe not by a lot.

tekacs 3mo ago
I feel like a core difference is that the AI implementor can get cheaper/faster (and indeed _uniformly_ better), whereas it would be very difficult for the same humans to do so. Even if this is not the right answer today, it can at the very least serve as a herald of a possible future, no?
Philpax 3mo ago
I think it'd take you at least eleven days to meaningfully coordinate 50 people!
Jenk 3mo ago
$165k won't get you far on salaried engineers. There's every chance that 1 engineer, assuming Anthropic employs them, is on $500k or more. Assuming average of $336k in that pool of 50 engineers, then for 11 days for 50 engineers you've spent $710k[0]. Salary info: https://www.levels.fyi/companies/anthropic/salaries/software-engineer?country=254 https://www.levels.fyi/companies/anthropic/salaries/software... [0]The maths I used (posting because I'm tired and prone to mistakes): $336,000 / 260 (working days of the year) = ~$1,292. $1,292 * 11 * 50 = ~$710,769
- bhaak 3mo ago
  You don't need top engineers to port a program from one language to the other. Outsource it to India. Of course, then you can also ask, could it have been done with a cheaper model. Probably yes. But then you wouldn't get free marketing.
  - minimaltom 3mo ago
    I agree you probably don't need top-dollar bay-area engineers for this, but hardcore outsourcing to a LCOL probably isnt going to work either due to novelty and generally being setup to do the more rote thing (generalizing a ton here obvi). This feels like something in the middle.
  - simonw 3mo ago
    Have you seen the "rewrite by outsourcing to India" thing work?
    - duhhhhh1212 3mo ago
      What a weird thing to say. The phrase “outsourcing to India” being used as shorthand for “you don’t need top engineers.” The nationality stereotypes are mean and degrading.
      - simonw 3mo ago
        That's fair, I shouldn't have commented that. I don't like the national stereotypes at all - I see "outsource to India" as being more about less expensive engineers than not needing "top engineers". That said, I don't think "rewrite from one language to another" with inexpensive engineers is a pattern that works. Happy to be proven wrong.
      - PeterHolzwarth 3mo ago
        And at a macro level, often found to be accurate due to how the businesses in India operate. Poll the west's engineers: you'll find that engineering from India is not currently viewed very favorably, in general. There are excellent engineers in India, but the system they operate in unfortunately doesn't allow them to shine.
jeremyloy_wt 3mo ago
They napkin math is fairly easy to do. One human works around 250 days per year, and if we assume Bay Area salaries we could assume ~300k/y conservatively for a fully loaded cost. $1200 per day. Your estimation is 50*11 days so $660,000. That’s 4x what Claude cost. That’s assuming that you actually get those 50 people to work without blockers, stepping on each other, or other coordination issues. The coordination complexity alone is astounding. I don’t like it necessarily, but Claude wins here, easily. It’s not close.
- piskov 3mo ago
  Unless you hire smart people from EU and what have you (especially ex-USSR) Which takes us to a point of future US dev salaries if this thing with agents gets better more and more
  - tclancy 3mo ago
    Sure, but can we not work out how to make humans more efficient for less money? There are obvious optimizations there that none of us would like to be part of.
    - cognitiveinline 3mo ago
      Humans can't become 100x efficient cost wise. That's how cheap LLMs are.
    - j16sdiz 3mo ago
      Less money? Cost of living is much lower in EU
- slopinthebag 3mo ago
  Why assume the upper level salary here? Using senior level developers making astronomical salaries for what is a mechanical line-by-line port would be a poor financial decision. What does the math look like with 25 devs making ~100k and doing it in 22 days? I’m sure you could find a reasonable combination which costs less. And if you’re already paying the devs the salary, it’s basically free (minus the opportunity cost of them not working on other things).
ignoramous 3mo ago
> I suspect Claude might be faster and therefore cheaper, but maybe not by a lot. While Jarred used Mythos-class model, some open weights, if they were as capable (certainly, GLM 5.2 looks the part), would have been way, way cheaper than professionals. Approx costs: DeepSeek v4 Pro & Mimo v2.5 Pro $3,426 ($2,567 / $600 / $259) Tencent HY3 $3,892 ($1,180 / $552 / $2,160) GLM 5.2 $30,016 ($8,260 / $3,036 / $18,720) Qwen 3.7 Max $37,925 ($14,750 / $5,175 / $18,000) Claude Opus 4.8 & GPT 5.5 xhigh $82,750 ($29,500 / $17,250 / $36,000) 5.9 billion uncached input tokens, 690 million output tokens, 72 billion cached input token reads.
- nsoonhui 3mo ago
  But not all models are equally capable, so I don't know your basis of comparison is even valid, let alone the numbers.
- KronisLV 3mo ago
  I did some more sizeable work with GLM 5.2 on Max reasoning (planning and implementing 8 features end to end) and it performed pretty well, but worse than Opus 4.8, with largely the same adversarial agent review loop. Opus 4.8 still found some real issues afterwards and spent about an hour fixing things, before the code was good enough to ship. Overall promising, wrote about it here: https://blog.kronis.dev/blog/z-ai-s-glm-5-2-is-a-great-model-but-is-it-good-value/ https://blog.kronis.dev/blog/z-ai-s-glm-5-2-is-a-great-model... The GLM Coding Plan seems to have lower token limits than the corresponding Anthropic Max subscriptions, but if you had to pay API rates for some LLM to do work somewhat reliably, it's a no brainer (unless you're swimming in money that you can give away and value your time more).
yomismoaqui 3mo ago
Your example of using 50 people for this reminded me of the classic “Nine women can’t make a baby in one month.” HINT: those 50 people must be coordinated...

by dabinat 3mo ago

tekacs 3mo ago

I feel like a core difference is that the AI implementor can get cheaper/faster (and indeed _uniformly_ better), whereas it would be very difficult for the same humans to do so. Even if this is not the right answer today, it can at the very least serve as a herald of a possible future, no?

Philpax 3mo ago

I think it'd take you at least eleven days to meaningfully coordinate 50 people!

Jenk 3mo ago

$165k won't get you far on salaried engineers. There's every chance that 1 engineer, assuming Anthropic employs them, is on $500k or more. Assuming average of $336k in that pool of 50 engineers, then for 11 days for 50 engineers you've spent $710k[0]. Salary info: https://www.levels.fyi/companies/anthropic/salaries/software-engineer?country=254 https://www.levels.fyi/companies/anthropic/salaries/software... [0]The maths I used (posting because I'm tired and prone to mistakes): $336,000 / 260 (working days of the year) = ~$1,292. $1,292 * 11 * 50 = ~$710,769

bhaak 3mo ago
You don't need top engineers to port a program from one language to the other. Outsource it to India. Of course, then you can also ask, could it have been done with a cheaper model. Probably yes. But then you wouldn't get free marketing.
- minimaltom 3mo ago
  I agree you probably don't need top-dollar bay-area engineers for this, but hardcore outsourcing to a LCOL probably isnt going to work either due to novelty and generally being setup to do the more rote thing (generalizing a ton here obvi). This feels like something in the middle.
- simonw 3mo ago
  Have you seen the "rewrite by outsourcing to India" thing work?
  - duhhhhh1212 3mo ago
    What a weird thing to say. The phrase “outsourcing to India” being used as shorthand for “you don’t need top engineers.” The nationality stereotypes are mean and degrading.
    - simonw 3mo ago
      That's fair, I shouldn't have commented that. I don't like the national stereotypes at all - I see "outsource to India" as being more about less expensive engineers than not needing "top engineers". That said, I don't think "rewrite from one language to another" with inexpensive engineers is a pattern that works. Happy to be proven wrong.
    - PeterHolzwarth 3mo ago
      And at a macro level, often found to be accurate due to how the businesses in India operate. Poll the west's engineers: you'll find that engineering from India is not currently viewed very favorably, in general. There are excellent engineers in India, but the system they operate in unfortunately doesn't allow them to shine.

jeremyloy_wt 3mo ago

They napkin math is fairly easy to do. One human works around 250 days per year, and if we assume Bay Area salaries we could assume ~300k/y conservatively for a fully loaded cost. $1200 per day. Your estimation is 50*11 days so $660,000. That’s 4x what Claude cost. That’s assuming that you actually get those 50 people to work without blockers, stepping on each other, or other coordination issues. The coordination complexity alone is astounding. I don’t like it necessarily, but Claude wins here, easily. It’s not close.

piskov 3mo ago
Unless you hire smart people from EU and what have you (especially ex-USSR) Which takes us to a point of future US dev salaries if this thing with agents gets better more and more
- tclancy 3mo ago
  Sure, but can we not work out how to make humans more efficient for less money? There are obvious optimizations there that none of us would like to be part of.
  - cognitiveinline 3mo ago
    Humans can't become 100x efficient cost wise. That's how cheap LLMs are.
  - j16sdiz 3mo ago
    Less money? Cost of living is much lower in EU
slopinthebag 3mo ago
Why assume the upper level salary here? Using senior level developers making astronomical salaries for what is a mechanical line-by-line port would be a poor financial decision. What does the math look like with 25 devs making ~100k and doing it in 22 days? I’m sure you could find a reasonable combination which costs less. And if you’re already paying the devs the salary, it’s basically free (minus the opportunity cost of them not working on other things).

ignoramous 3mo ago

> I suspect Claude might be faster and therefore cheaper, but maybe not by a lot. While Jarred used Mythos-class model, some open weights, if they were as capable (certainly, GLM 5.2 looks the part), would have been way, way cheaper than professionals. Approx costs: DeepSeek v4 Pro & Mimo v2.5 Pro $3,426 ($2,567 / $600 / $259) Tencent HY3 $3,892 ($1,180 / $552 / $2,160) GLM 5.2 $30,016 ($8,260 / $3,036 / $18,720) Qwen 3.7 Max $37,925 ($14,750 / $5,175 / $18,000) Claude Opus 4.8 & GPT 5.5 xhigh $82,750 ($29,500 / $17,250 / $36,000) 5.9 billion uncached input tokens, 690 million output tokens, 72 billion cached input token reads.

nsoonhui 3mo ago
But not all models are equally capable, so I don't know your basis of comparison is even valid, let alone the numbers.
KronisLV 3mo ago
I did some more sizeable work with GLM 5.2 on Max reasoning (planning and implementing 8 features end to end) and it performed pretty well, but worse than Opus 4.8, with largely the same adversarial agent review loop. Opus 4.8 still found some real issues afterwards and spent about an hour fixing things, before the code was good enough to ship. Overall promising, wrote about it here: https://blog.kronis.dev/blog/z-ai-s-glm-5-2-is-a-great-model-but-is-it-good-value/ https://blog.kronis.dev/blog/z-ai-s-glm-5-2-is-a-great-model... The GLM Coding Plan seems to have lower token limits than the corresponding Anthropic Max subscriptions, but if you had to pay API rates for some LLM to do work somewhat reliably, it's a no brainer (unless you're swimming in money that you can give away and value your time more).

yomismoaqui 3mo ago

Your example of using 50 people for this reminded me of the classic “Nine women can’t make a baby in one month.” HINT: those 50 people must be coordinated...