> Humans are very expensive, so the equation almost always falls against them.
You underestimate what these models cost. Uber's budget is $1,500/dev/month. I gather that was put in place because the dev's were going through $6,000/dev/month, which Uber decided could not be cost justified.
Fable costs at least twice as much, or $12,000/dev/month.
Fable can apparently work for hours without supervision, which means a skilled engineer can now have it working on many tasks concurrently. I would not be at all surprised if they can put a nought or two on that number. If you do that, you are well out of "what a human costs" territory.
Not to argue myself out of a job, but I cost around $20k/month, all costs considered(taxes, social fees, PTO, healthcare, benefits). If my efficiency is tripled(which it absolutely is, even before fable) for a mere 6k/month(in reality, 1k is more than enough though), that's ~10x ROI.
> You underestimate what these models cost. Uber's budget is $1,500/dev/month
$1,500/month needs to be contextualised against the fully-loaded cost of a software engineer. Uber's average TC for a US-based software engineer is around $350k, the fully-loaded cost is going to be in the $450k-$500k range. So we're talking around $38k/month for a software engineer.
$1,500/month isn't even a drop in the bucket. If LLM use lets them shave just one person off a team, that pays for tokens for the next 25 engineers.
These numbers don't mean anything without a denominator. You could burn $10 million/month of tokens if you want. We want to know how the cost per unit of useful output compares to a human. Does $6000 of usage buy you a man-month of work? Less? More?
Minor note, 2x $/tok is not 2x cost. Personally, I see Fable being significantly more token-efficient than Opus 4.8. Then, there's also the compounding costs of quality.
> I would not be at all surprised if they can put a nought or two on that number.
People keep saying this and it keeps not happening.
ChatGPT Pro was $200/mo when it launched in '23 for a ~100B class model with 8k context. Claude Max is now the same price for practically unlimited access to a ~1T class model with 1M context.
Moore's Law never died, it just switched architectures.
Not to mention, the cost/performance of the baseline keeps falling. The cost-effective Deepseek V4 Flash is better than frontier models from a year ago, at a fraction of the cost.
Good to know that LLMs will be removing all regulatory and legal risks, as well as creating a consumer economy that no longer employs or pays consumers.
I can't help thinking there might be some kind of strategic issue here.
One of the large (and enjoyable IMHO) challenges in this line of work is developing a de facto understanding of your process and the context it's in service to, and that's only possible if you're actually on your industry equivalent of a "shop floor" for each domain the project touches.
As far as I can tell this part of the job isn't really on anyone's radar anymore.
That's the beauty of these AI advancements. You, a human, will have to compete against a model for the same job.
If you get $100,000 per year as a SWE, and Anthropic offers a coding model for $100,000 per year (but working 24/7), then you'll have to give up all of those addons that make the fully burdened cost of the employee. Say goodbye to vacation, sick time, benefits, etc.
We know this model will be cheaper and faster with time.
And we have not even reached the timespan/timeframe were we have ASIC style models.
OpenAI has to do something which will beat Fable otherwise Anthropic won. China currently overtakes cars, pv, batteries and very soon silicon chip making, it has all the incentive to also take over AI.
> We know this model will be cheaper and faster with time
Why? Demand for AI compute seems to be increasing faster than new production is due to come online for the foreseeable future, particularly if more-intensive models induce demand.
Not OP, but for me, this model will get VERY expensive in 2 weeks. Now it is part of Pro plan, after 22nd it will get excluded and I will pay by token API usage (~10x more expensive).
The only thing they’ve overtaken is arguably batteries, and even that is questionable if the quality is as good as Korean manufacturers. I think it’s more likely that the Chinese chip industry overtaking competitors will remain like nuclear fusion, forever “just 5 years away”
They mostly have overtaken in cars too. Their EVs are just cheaper, and they have built the infrastructure around it, even in more rural provinces. Building infrastructure is something they excel at anyway.
The parent comment is describing a test they ran so they could assess their trust in the model for scenarios they don't have time to fully understand.
Do you not believe in running tests, evaluations, or experiments at all to better understand your environment?
The ROI in the case of a positive outcome is the reduced time needed to inspect the results in the future (the entire point of AI is to know what you can trust it on, so you can delegate everything at that level with less oversight). The ROI in the negative case is the tokens not wasted on tasks to ambitious for the model.
Humans are very expensive, so the equation almost always falls against them.
It's not just salary, but also safety/labor regulation, legal risk, vacations, sick time, personal conflicts, HR, benefits.
Even when automation is more expensive on paper, it's generally still cheaper
> Humans are very expensive, so the equation almost always falls against them.
You underestimate what these models cost. Uber's budget is $1,500/dev/month. I gather that was put in place because the dev's were going through $6,000/dev/month, which Uber decided could not be cost justified.
Fable costs at least twice as much, or $12,000/dev/month.
Fable can apparently work for hours without supervision, which means a skilled engineer can now have it working on many tasks concurrently. I would not be at all surprised if they can put a nought or two on that number. If you do that, you are well out of "what a human costs" territory.
Not to argue myself out of a job, but I cost around $20k/month, all costs considered(taxes, social fees, PTO, healthcare, benefits). If my efficiency is tripled(which it absolutely is, even before fable) for a mere 6k/month(in reality, 1k is more than enough though), that's ~10x ROI.
I kinda get why execs are excited
> You underestimate what these models cost. Uber's budget is $1,500/dev/month
$1,500/month needs to be contextualised against the fully-loaded cost of a software engineer. Uber's average TC for a US-based software engineer is around $350k, the fully-loaded cost is going to be in the $450k-$500k range. So we're talking around $38k/month for a software engineer.
$1,500/month isn't even a drop in the bucket. If LLM use lets them shave just one person off a team, that pays for tokens for the next 25 engineers.
These numbers don't mean anything without a denominator. You could burn $10 million/month of tokens if you want. We want to know how the cost per unit of useful output compares to a human. Does $6000 of usage buy you a man-month of work? Less? More?
Minor note, 2x $/tok is not 2x cost. Personally, I see Fable being significantly more token-efficient than Opus 4.8. Then, there's also the compounding costs of quality.
On top of which, as the article mentions, it delegates simpler tasks to cheaper models.
> I would not be at all surprised if they can put a nought or two on that number.
People keep saying this and it keeps not happening.
ChatGPT Pro was $200/mo when it launched in '23 for a ~100B class model with 8k context. Claude Max is now the same price for practically unlimited access to a ~1T class model with 1M context.
Moore's Law never died, it just switched architectures.
Not to mention, the cost/performance of the baseline keeps falling. The cost-effective Deepseek V4 Flash is better than frontier models from a year ago, at a fraction of the cost.
Good to know that LLMs will be removing all regulatory and legal risks, as well as creating a consumer economy that no longer employs or pays consumers.
I can't help thinking there might be some kind of strategic issue here.
Perhaps someone should ask Mythos about it.
By the point where we have work hours regulation for AI, all of our current debate about AI will be long irrelevant because we've clearly achieved AGI
One of the large (and enjoyable IMHO) challenges in this line of work is developing a de facto understanding of your process and the context it's in service to, and that's only possible if you're actually on your industry equivalent of a "shop floor" for each domain the project touches.
As far as I can tell this part of the job isn't really on anyone's radar anymore.
That's the beauty of these AI advancements. You, a human, will have to compete against a model for the same job.
If you get $100,000 per year as a SWE, and Anthropic offers a coding model for $100,000 per year (but working 24/7), then you'll have to give up all of those addons that make the fully burdened cost of the employee. Say goodbye to vacation, sick time, benefits, etc.
> "What have you got against machines?" said Buck.
> "They're slaves."
> "Well, what the heck," said Buck. "I mean, they aren't people. They don't suffer. They don't mind working."
> "No. But they compete with people."
> "That's a pretty good thing, isn't it--considering what a sloppy job most people do of anything?"
> "Anybody that competes with slaves becomes a slave," said Harrison thickly, and he left.
Kurt Vonnegut, Player Piano
They will do it for far less. Once manufacturing catches up and they have the data centers built out tokens are going to be dirt cheap.
It just got released, it shouldn't matter.
We know this model will be cheaper and faster with time.
And we have not even reached the timespan/timeframe were we have ASIC style models.
OpenAI has to do something which will beat Fable otherwise Anthropic won. China currently overtakes cars, pv, batteries and very soon silicon chip making, it has all the incentive to also take over AI.
> We know this model will be cheaper and faster with time
Why? Demand for AI compute seems to be increasing faster than new production is due to come online for the foreseeable future, particularly if more-intensive models induce demand.
LOL magical thinking
I'm happy to discuss arguments if you want to add any?
Not OP, but for me, this model will get VERY expensive in 2 weeks. Now it is part of Pro plan, after 22nd it will get excluded and I will pay by token API usage (~10x more expensive).
I find it good for code reviews.
Yeah my time scale is 'a handful of years' :)
The only thing they’ve overtaken is arguably batteries, and even that is questionable if the quality is as good as Korean manufacturers. I think it’s more likely that the Chinese chip industry overtaking competitors will remain like nuclear fusion, forever “just 5 years away”
The best batteries are currently from CATL. No one in the industry is doubting this.
Huawei just showed LogicFolding and have a roadmap for 1.4 nanometer by 2031; SMIC is going for 5nm.
And all of this WITHOUT EUV.
They mostly have overtaken in cars too. Their EVs are just cheaper, and they have built the infrastructure around it, even in more rural provinces. Building infrastructure is something they excel at anyway.
The parent comment is describing a test they ran so they could assess their trust in the model for scenarios they don't have time to fully understand.
Do you not believe in running tests, evaluations, or experiments at all to better understand your environment?
The ROI in the case of a positive outcome is the reduced time needed to inspect the results in the future (the entire point of AI is to know what you can trust it on, so you can delegate everything at that level with less oversight). The ROI in the negative case is the tokens not wasted on tasks to ambitious for the model.
It will be great when the price of compute/memory drops to normal level!
>Sam Altman has signed another Memoranda of Understanding: Buying all SDRAM till the heat death of the universe OR Musk relocates to mars.