Google/Rulesofthumb(nonint.com) |
Google/Rulesofthumb(nonint.com) |
Do I buy expensive tool X or get buy with cheap ones that make the work take longer or be less precise? Do I buy fancy machine Y or pay 10 people to do it manually?
The only new development with AI is that this was traditionally limited to relatively manual or repetitive processes, but is now expanding to knowledge work as well.
In the medium term the question will be do I just pay for people, or do I spend resources on collecting data and training a model?
Unions are stronger than the last 40 years especially at auto/medical/trucking. Likely, deglobalization is going to bring some production home, but really the question is, why dismiss or even relish worker insecurity and inefficiency?
You know, for other people you are the other people that things happen to.
Minor quibble unrelated to the main content of the post: the measures are not fixed costs of assets, but a blend of depreciation and the operating cost of power usage for those assets. Sort of a regular snapshot of an average daily accounting cost, so to speak (which is reasonable). And this was cost to Google, not taking into account what Google could make by charging Cloud customers for the use of those assets (opportunity cost).
My understanding is that the main use of this tool was actually for engineers to give reasonable-ish impact statements for their performance work. I hadn't heard of anyone using it to make serious trade-offs in project planning, since at the level where that matters, the capacity planning teams had more precise costs related to their actual budgets, as well as short term goals like "RAM has a supply chain shock so we can't get any more than X for the next quarter."
Also a pet peeve of mine, people constantly screwed up the units. "10 SWEs" is rate (cost per time), same as "10 TB of RAM", but "SWE-years" is cost (ie dollars). Many design documents use these inconsistently.
But the exact numbers are important. An H100 is ~$2/hour [1], so 1000 is $16M/year (24/7). Even if Google gets a massive internal discount that's still way more than 3 people's total cost. If you have to choose between 16 (highly paid, senior) people or 1000 H100s would you have to think about the choice?
Then again when you revisit this comment in a few years' time the original comparison may be correct.
[1] https://gpus.llm-utils.org/h100-gpu-cloud-availability-and-p...
And if you really insist on 24/7 comparison, you would come to need about 12 people, as the expected productivity per person is of about 6 hrs/day. Factor in the fact that people need vacations, sick days, weekends off… it looks like 1000 H100 might actually be a good trade-off in the very near future.
Very unlikely that Google pays $16M for a single H100. Amazon has it at $44k: https://www.amazon.com/Tesla-NVIDIA-Learning-Compute-Graphic...
Unfortunately 1 Jeff Dean SWE != SWE
If anything, software based solutions (LLMs or perhaps something more domain specific and accurate, eventually) will just broaden the labor pool for software. So you'll see more people enter the field which will depress wages, but there will also probably consequently be more jobs.
I'm not sure if you are talking about a company or a person or what kind of entity really here.
I'll assume a private person.
So you think people who clean bathrooms do it because they don't have "different priorities for their time"?
That sounds backwards to me.
Edit: and it is, of course, logically.
What I mean is: to have different priorities for your time, you either need to be able to earn more money in the same time, or not need to earn money.
That brings to mind for me another fact: much of art, literature, philosophy, science and other things that we historically have built on was done by people who could afford other priorities in their life than cleaning their bathroom, preparing food, earning money to obtain food...
Of course this will stir up the question of meritocracy and how capitalistic society really functions for some.
But the way you put this question seems naive to me.
I don't think that most people who clean bathrooms for a living do that because they set this priority for their (life-)time.
That doesn't preclude being a janitor from being a potentially satisfying and certainly valuable job.
I know I'm mixing up the terms janitor and cleaning bathrooms here, but I felt that it was already unclear from your comment what kind of job you are talking about. Might be the language barrier.
A janitor hase more responsibility of course than just cleaning bathrooms.
I personally don’t know what impact AI will have on the job market, but is is not going to be an overnight revolution.
But I'm still not convinced that it would result in more jobs. If you saved money by automating away labor, why spend the savings on labor?
The proponents argue the savings would be spent on laborers doing other (possibly new) jobs that are not yet automated and hired by different companies.
EDIT reply to: >I can't see what new jobs would be created, especially high value ones.
Yes, I agree about not being able to see new jobs. However, history has shown we (society) have always failed to imagine what the new jobs would be that takes the place of old jobs automated away.
Farmers and field workers replaced by tractors and harvesting combines. Human telephone switchboard operators replaced by automatic digital relay circuits. Travel agents replaced by website airline ticket bookings. [Thousands of other examples...] And yet the total # of people employed still keep going up instead of society dealing with massive 95% unemployment.
The paradox happens because the farmers and switchboard operators that were disrupted can't possibly envision new types of jobs that exist in the future. Understandably, the displaced farmer can't imagine there would be a job where somebody ... presses keys all day on a "typewriter like mechanism" that sends instructions to an "interactive tv". That basically describes what today's programmer does at his computer the entire day.
So, the new AI could usher a new wave of jobs. (Something more unpredictable than the immediate memes of ChatGPT "prompt engineers").
As we're living in the moment, we're also "blind to the future possibilities" like those farmers. If one could go back in time and tell that farmer in 1920s driving a Model T that there would be new and different jobs replacing farm work, you wouldn't be able to convince him. In the world that he's familiar with, our attempted descriptions would all just be speculative science fiction. Likewise for those of us today that's pessimistic about AI affecting jobs, is there really anything anyone could say that would change our minds? It's just predictions that we'd just dismiss.
The counterpoint to AI optimism is that _this_ new type of automation with powerful AI is unlike the tractors/microchips/websites in the past that replaced people and much more disruptive.
The total number of employed people goes up because the population goes up. The labor participation rate has been steadily declining for a while, only really increasing since 2020 as a correction to return back to pre-pandemic levels, which were still on a decline.
There is a myth that displaced workers will result in a higher unemployment rate. The official unemployment rate (U-3) shows people who are looking for jobs and cannot find them. In actuality, displaced workers often leave the workforce entirely and manage to survive in ways not registered by the Bureau of Labor statistics (often through government entitlement programs and welfare). Men, in particular, are more likely to drop out of the labor force when they've been made redundant. A whole generation of young men are declining the enter the labor force at all, opting instead to live with their parents long term or use school as a way to avoid getting a job.
People who worked in blue collar jobs and agriculture and were displaced over the course of the 20th century didn't all get "better jobs", many of them languished in broken communities and died deaths of despair. The coal miners didn't learn javascript