Basically what you’re saying is “another person made a wrong prediction in the past, so your prediction about something entirely different is wrong too”.
People used that same quote when talking about NFTs, and VR, and 3D TVs, and literally every grift and failed “big idea” ever since.
That quote is not an argument, and using it is a tired technique. The only thing it signals is a blind belief in whatever the quoter thinks is good at the time.
Yes, I know that quote is used to make fun of Internet nay-sayers, but I think younger people underestimate just how important fax machines were even as late as the early 2000s (and reportedly, they are still popular in Japan, where handwriting a note and faxing it is still easier than typing the characters).
I'm not sure that "scam" is the right word to use in this case, unless I am misconstruing what they are calling a scam.
However...
Did you train your own model? Fine tune it? Or did you just use some model that you had free/cheap access to?
AI/LLMs can/will be very useful things. But there is every indication that spending huge money to create a model is flushing money down the drain - you'll never be able to generate the revenue to justify that spend. And nevermind the cost to create - there are some models that cost more to operate than they can generate in revenue.
If there is money to be made with AI/LLMs (and there probably is), it is going to be made by some entity that is unburdened by the cost of creating the model.
What has it helped you automate?
Copilot was cool to use for a bit but I ultimately stopped paying for it. ChatGPT was exciting to play around with but it never became a core part of my workflow. We are 2+ years from GPT 3.5 but AI still hasn't made its way into my daily toolkit, so would be cool to hear how others are using it effectively.
Helped automate digging through ad ridden google searches to string together a series of articles explaining automation of mind mapping visualizations pointing to a sharepoint database.
A 4090 can do tens of GH/s on SHA256D (BTC) for a substantial capital investment and 300W - other GPUs substantially less.
A $10 ASIC can do >500 GH/s for 5W.
If you possess these GPUs, at any energy price it becomes orders of magnitude more cost-effective to mine something else, rent for LLM uses, etc. and then just buy Bitcoin.
I thought ASICS only worked on certain protocols like BTC, whereas ETH was implemented to intentionally limit their utility and require more general purpose hardware. Is that false?
BTC ASICs only work on BTC. ETH no longer operates using proof-of-work, all of this equipment moved to other coins (many of them using the same algorithm.)
Because $40k GPUs with three year warranties renting for $3.50/hr or more is a literal money printing machine. It's a no brainer (if you can get the GPUs or already have them) for companies that already have data center infra in low cost energy markets. However, 700w for a single h100 is honestly trivial compared to how profitable selling the compute is (for now). Maybe I'm missing something but this seems like an (temporary) infinate money glitch and if I had 8 figures to throw at I would be doing the same.
Well they usually locate in industrial buildings which means rent per square foot can be under a dollar a month. And energy where some miners are at is 7 cents per kilowatt hour. So the things you are talking about might add $.25/hr we can go with $1.75. If you are renting this for $3.5 you are doubling your money. Any wise investor would take this action all day long. DCs full of non mission critical workloads that require low bandwidth aren’t hard to setup and run.
Source: I live in a town with dozens of industrial sized mining ops. We have the cheapest power in the country.
It would be really interesting if this could fundamentally damage bitcoin, but given the existing ASICs are specific to bitcoin, it may be a while. It's a happy and not so surprising thought if there is just a better and more profitable way to use power to compute than bitcoin.
Thinking out loud here:
Obviously because that's how predatory investments work.
Step 1: Notice high potential market that could, but doesn't need to, disrupt established practices and hierarchies by federating stuff.
Step 2: Suck all the leaking behavioral and private data out of it.
Stup 3: Gain as much control over the supply chains & the market as possible, eliminating federation.
Also: AGI hasn't found its optimal hardware-software combo yet, so "porting" it all to networked mining rigs is worth trying and progressing. And neural nets and blockchains have quite a bit to give & take to & from each other.
32 comments
[ 3.0 ms ] story [ 87.9 ms ] threadInternet's Effect on the World Economy Would Be No Greater Than the Fax Machine's.
--Paul Krugman
People used that same quote when talking about NFTs, and VR, and 3D TVs, and literally every grift and failed “big idea” ever since.
That quote is not an argument, and using it is a tired technique. The only thing it signals is a blind belief in whatever the quoter thinks is good at the time.
However...
Did you train your own model? Fine tune it? Or did you just use some model that you had free/cheap access to?
AI/LLMs can/will be very useful things. But there is every indication that spending huge money to create a model is flushing money down the drain - you'll never be able to generate the revenue to justify that spend. And nevermind the cost to create - there are some models that cost more to operate than they can generate in revenue.
If there is money to be made with AI/LLMs (and there probably is), it is going to be made by some entity that is unburdened by the cost of creating the model.
A $10 ASIC can do >500 GH/s for 5W.
If you possess these GPUs, at any energy price it becomes orders of magnitude more cost-effective to mine something else, rent for LLM uses, etc. and then just buy Bitcoin.
Source: I live in a town with dozens of industrial sized mining ops. We have the cheapest power in the country.
It would be really interesting if this could fundamentally damage bitcoin, but given the existing ASICs are specific to bitcoin, it may be a while. It's a happy and not so surprising thought if there is just a better and more profitable way to use power to compute than bitcoin.
Step 1: Notice high potential market that could, but doesn't need to, disrupt established practices and hierarchies by federating stuff. Step 2: Suck all the leaking behavioral and private data out of it. Stup 3: Gain as much control over the supply chains & the market as possible, eliminating federation.
Also: AGI hasn't found its optimal hardware-software combo yet, so "porting" it all to networked mining rigs is worth trying and progressing. And neural nets and blockchains have quite a bit to give & take to & from each other.
*Thinking quietly again ...