3 ms·

So I want to talk about the Claude Mythos mega-model that had a real marketing push in the last week. Marcus Hutchins (of IWannaCry fame) posted about the econo

by jmyeet 7mo ago

So I want to talk about the Claude Mythos mega-model that had a real marketing push in the last week. Marcus Hutchins (of IWannaCry fame) posted about the economics of this [1]. The upshot here is this was a longstanding BSD bug but mostly because nobody is paying BSD bug bounties and a null-pointer dereference may induce a crash but rarely (if ever) leads to privilege escalation as buffer overflows generally do so isn't as high-priority. The estimate on the cost is $20-50k of tokens. $50k gets you a lot of developer time, upwards of 2 months. As has been stated in a bunch of threads, much smaller models could also find this bug. The defense is "they didn't" and "once you know it's there, it's easier to find" but again, nobody has been paying BSD bug bounties. Put another way: far fewer people have been looking. My point is that the economics of these models are still highly debatable. What they will do is allow you to use 1 engineer with AI tools instead of 2-4 engineers. And we've arrived at the real goal here: to reduce head counts. Why? To suppress labor costs. To get people who didn't get laid off to do more free work. In addition, AI capex is used as an excuse to cut costs ie reduce labor, for exactly the same reasons. Lastly, the threat of this happening in the near to medium future is also used to scare labor and reduce costs. Sufficiently largecompanies can't grow anymore. Their only path forward is to raise prices and redcue costs. For tech companies, labor is a significant cost. That's what's going on here. [1]: https://www.tiktok.com/@itsmarcushutchins/video/7627740073534033183 https://www.tiktok.com/@itsmarcushutchins/video/762774007353...

xdennis 7mo ago
> $50k gets you a lot of developer time, upwards of 2 months. God, I wish I was American sometimes...
- yoyohello13 7mo ago
  Right up until you lose that job and get sick.
- jmyeet 7mo ago
  You have to consider all the costs of employing someone, not just the direct compensation. Benefits, health insurance, taxes (eg employer payroll taxes), office space, equipment, various insurances, utilities, costs related to just having that person (eg HR) and so on. So even a $100k salary, which is pretty modest for a US-based engineer, the additional costs may well add another $50-100k.
ianm218 7mo ago
You’re missing the cost curve of inference. What costs $20-50k now will cost 2-5k in a year or two at which point the math is a no brainer. It makes a lot of sense to build products that almost work now or are almost economical and ride the trend line.
- ehnto 7mo ago
  Why would it reduce ten fold? I think performance of the chips gets a little better but not that much better. The cost of the infra is probably going to go up if energy costs keep going up and presuming the US can keep getting cheap chips.
  - ianm218 7mo ago
    They have been exponentially decreasing so far [1] and the Vera Rubin generations chips will be going live that are 35X more efficient in terms of inference/ megawatt [2]. Even with rising prices 10x is possibly conservative. Maybe if demand is truly crazy the labs will take more margin 1. epoch.ai/data-insights/llm-inference-price-trends 2. http://hashrateindex.com/blog/nvidia-vera-rubin-nvl72-specs-breakdown/ http://hashrateindex.com/blog/nvidia-vera-rubin-nvl72-specs-...
jopsen 7mo ago
> What they will do is allow you to use 1 engineer with AI tools instead of 2-4 engineers. Stop there. The costs of today state-of-the-art hype-ware will be lower tomorrow. In a year, there will be fewer easy to find bug, and models like this will be cheaper. Doing math based on hardware deployed today is a bit naive. Assuming any of this is intentional about reducing devs is just dumb.
varispeed 7mo ago
Problem with these models is that they are not reliable. You never know when the provider decides to nerf the model and after good week of making progress you find the model is now running in circles and you are burning tokens like you are operating a steam locomotive and standing still.
- jmyeet 7mo ago
  Tbf people aren't reliable either. Not that I'm defending the AI models but remember these long-standing bugs were introduced by people and not found by people for years. Anytime anyone feels insecure that AI will genuinely do their job, I point them to creators like this [1]. Also, I find it so wild there are people, even here on HN, who don't admit that LLMs hallucinate, lie and forget. I would have a hard time trusting any critical code produced by an LLM. This was a frequent complaint with OpenClaw (being Vibe coded) and the general advice was to not run it on any computer you're logged onto as you'll get hit by prompt injection attacks and the like. Anyway, the models don't need to be reliable. They just need to be an excuse to fire people to suppress labor costs. That's the true use case. [1]: https://www.tiktok.com/@huskistaken https://www.tiktok.com/@huskistaken
  - varispeed 7mo ago
    > Tbf people aren't reliable either. I only seen one instance where developer came to work still under influence of drugs after the weekend. He was moonlighting as a DJ and as he said the afters was legendary. He was fired that day. Usually you get an odd Monday with younger devs that they call off sick or family/health leave with the older ones. But they never suddenly run in circles for weeks seemingly not knowing what they do after weeks of exemplary performance.
zeroonetwothree 7mo ago
> What they will do is allow you to use 1 engineer with AI tools instead of 2-4 engineers. In a competitive market this doesn’t work exactly that way because you will then lose to your competitor who has 2-4 engineer + AI.

by jmyeet 7mo ago

xdennis 7mo ago

> $50k gets you a lot of developer time, upwards of 2 months. God, I wish I was American sometimes...

yoyohello13 7mo ago
Right up until you lose that job and get sick.
jmyeet 7mo ago
You have to consider all the costs of employing someone, not just the direct compensation. Benefits, health insurance, taxes (eg employer payroll taxes), office space, equipment, various insurances, utilities, costs related to just having that person (eg HR) and so on. So even a $100k salary, which is pretty modest for a US-based engineer, the additional costs may well add another $50-100k.

ianm218 7mo ago

You’re missing the cost curve of inference. What costs $20-50k now will cost 2-5k in a year or two at which point the math is a no brainer. It makes a lot of sense to build products that almost work now or are almost economical and ride the trend line.

ehnto 7mo ago
Why would it reduce ten fold? I think performance of the chips gets a little better but not that much better. The cost of the infra is probably going to go up if energy costs keep going up and presuming the US can keep getting cheap chips.
- ianm218 7mo ago
  They have been exponentially decreasing so far [1] and the Vera Rubin generations chips will be going live that are 35X more efficient in terms of inference/ megawatt [2]. Even with rising prices 10x is possibly conservative. Maybe if demand is truly crazy the labs will take more margin 1. epoch.ai/data-insights/llm-inference-price-trends 2. http://hashrateindex.com/blog/nvidia-vera-rubin-nvl72-specs-breakdown/ http://hashrateindex.com/blog/nvidia-vera-rubin-nvl72-specs-...

jopsen 7mo ago

> What they will do is allow you to use 1 engineer with AI tools instead of 2-4 engineers. Stop there. The costs of today state-of-the-art hype-ware will be lower tomorrow. In a year, there will be fewer easy to find bug, and models like this will be cheaper. Doing math based on hardware deployed today is a bit naive. Assuming any of this is intentional about reducing devs is just dumb.

varispeed 7mo ago

Problem with these models is that they are not reliable. You never know when the provider decides to nerf the model and after good week of making progress you find the model is now running in circles and you are burning tokens like you are operating a steam locomotive and standing still.

jmyeet 7mo ago
Tbf people aren't reliable either. Not that I'm defending the AI models but remember these long-standing bugs were introduced by people and not found by people for years. Anytime anyone feels insecure that AI will genuinely do their job, I point them to creators like this [1]. Also, I find it so wild there are people, even here on HN, who don't admit that LLMs hallucinate, lie and forget. I would have a hard time trusting any critical code produced by an LLM. This was a frequent complaint with OpenClaw (being Vibe coded) and the general advice was to not run it on any computer you're logged onto as you'll get hit by prompt injection attacks and the like. Anyway, the models don't need to be reliable. They just need to be an excuse to fire people to suppress labor costs. That's the true use case. [1]: https://www.tiktok.com/@huskistaken https://www.tiktok.com/@huskistaken
- varispeed 7mo ago
  > Tbf people aren't reliable either. I only seen one instance where developer came to work still under influence of drugs after the weekend. He was moonlighting as a DJ and as he said the afters was legendary. He was fired that day. Usually you get an odd Monday with younger devs that they call off sick or family/health leave with the older ones. But they never suddenly run in circles for weeks seemingly not knowing what they do after weeks of exemplary performance.

zeroonetwothree 7mo ago

> What they will do is allow you to use 1 engineer with AI tools instead of 2-4 engineers. In a competitive market this doesn’t work exactly that way because you will then lose to your competitor who has 2-4 engineer + AI.