So, more #genAI pain.
-
So those of you who are burning millions of tokens on Claude Code should probably be aware that you’re one of the few remaining forces inflating the bubble.
Whatever, it’ll pop anyhow, the #genAI space will become a tiny, mostly open-source, fraction of what it is today, and at that point. we devs can use it without feeling guilty.
@timbray weirdly I think contract law might also succeed for the same reason. Small market also.
-
So those of you who are burning millions of tokens on Claude Code should probably be aware that you’re one of the few remaining forces inflating the bubble.
Whatever, it’ll pop anyhow, the #genAI space will become a tiny, mostly open-source, fraction of what it is today, and at that point. we devs can use it without feeling guilty.
@timbray I'm not feeling guilty.
My budget is US$10/month. I get a lot for that.
-
@timbray Not to mention what happens when they finally decide to raise the token price to reflect all-in costs, including capital depreciation.
Claude Code itself works really well against local models if you have the hardware to run something in the 30b parameter range. And hey it's already been open sourced so when Anthropic ceases to exist the agent will still be around in some form.
-
The software-dev community is many orders of magnitude too small to keep the bubble inflated.
And in fact #genAI works way better on code than on actual human language. Because the vocabulary is so much smaller and “doing the same as everyone else already has” is actually a good engineering practice.
So…
@timbray disclaimer: not necessarily if a genuine innovation may be about to happen, or in a brand new language dialect, in which case LLMs are mostly useless anyway due to their mean reversion effects.
-
The software-dev community is many orders of magnitude too small to keep the bubble inflated.
And in fact #genAI works way better on code than on actual human language. Because the vocabulary is so much smaller and “doing the same as everyone else already has” is actually a good engineering practice.
So…
@timbray programming tools were barely profitable back when people paid serious money for compilers, IDEs, and related tools.
-
The software-dev community is many orders of magnitude too small to keep the bubble inflated.
And in fact #genAI works way better on code than on actual human language. Because the vocabulary is so much smaller and “doing the same as everyone else already has” is actually a good engineering practice.
So…
@timbray I think that another major reason why this works for software is that there are tools to help determine if it is correct. Compilers, linters, static analysis and tests. Harder to apply to domains that don’t have these kind of checks.
-
So those of you who are burning millions of tokens on Claude Code should probably be aware that you’re one of the few remaining forces inflating the bubble.
Whatever, it’ll pop anyhow, the #genAI space will become a tiny, mostly open-source, fraction of what it is today, and at that point. we devs can use it without feeling guilty.
@timbray but what if my employer is the one spending $400 on my tokens a month?
-
Claude Code itself works really well against local models if you have the hardware to run something in the 30b parameter range. And hey it's already been open sourced so when Anthropic ceases to exist the agent will still be around in some form.
@bflipp @virtuous_sloth I'm curious: What kind of hardware is that?
-
So those of you who are burning millions of tokens on Claude Code should probably be aware that you’re one of the few remaining forces inflating the bubble.
Whatever, it’ll pop anyhow, the #genAI space will become a tiny, mostly open-source, fraction of what it is today, and at that point. we devs can use it without feeling guilty.
@timbray I do think there is a bubble that will pop, but I also think that useful AI is already escaping from the dev community via tooling like OpenClaw that is taking off and using lots of tokens automating tasks for people.
-
So those of you who are burning millions of tokens on Claude Code should probably be aware that you’re one of the few remaining forces inflating the bubble.
Whatever, it’ll pop anyhow, the #genAI space will become a tiny, mostly open-source, fraction of what it is today, and at that point. we devs can use it without feeling guilty.
@timbray So, if it's real value that people are willing to pay for, is it a bubble? Or a business?
I guess if software developers are paying enough that investors think it's interesting, but not enough to pay the expenses, and not by a large margin, that's probably a bubble.
-
@timbray So, if it's real value that people are willing to pay for, is it a bubble? Or a business?
I guess if software developers are paying enough that investors think it's interesting, but not enough to pay the expenses, and not by a large margin, that's probably a bubble.
@evan depends how many people are willing to pay, I guess.
-
@evan depends how many people are willing to pay, I guess.
@timbray Right. If revenue is only 0.5% of expenses, or even 5%, there's no way that unit economics can be tuned to make a profit. But if it's 50%, say, I think the traditional VC idea is to keep funding it, and then make it pay well in the future. I don't know if those percentages are anywhere near accurate, btw.
-
So those of you who are burning millions of tokens on Claude Code should probably be aware that you’re one of the few remaining forces inflating the bubble.
Whatever, it’ll pop anyhow, the #genAI space will become a tiny, mostly open-source, fraction of what it is today, and at that point. we devs can use it without feeling guilty.
@timbray
I do see some other clearly useful, clearly worth paying for use cases at work. We're a multilingual environment, and translation is if anything more of a killer app than coding. A large fraction of the entire workplace already uses it, and there's a clear business case for paying for it. -
@timbray Right. If revenue is only 0.5% of expenses, or even 5%, there's no way that unit economics can be tuned to make a profit. But if it's 50%, say, I think the traditional VC idea is to keep funding it, and then make it pay well in the future. I don't know if those percentages are anywhere near accurate, btw.
-
So those of you who are burning millions of tokens on Claude Code should probably be aware that you’re one of the few remaining forces inflating the bubble.
Whatever, it’ll pop anyhow, the #genAI space will become a tiny, mostly open-source, fraction of what it is today, and at that point. we devs can use it without feeling guilty.
@timbray The bubble is all business. Investors will shutdown successful companies that make a nice 3 million a year because they decided that is the market and there is no growth.
Even in software the rate of adoption is deceptive since there are mandates on using it coming from business leaders, many told my the same investors to create demand sort of like return to office
-
So those of you who are burning millions of tokens on Claude Code should probably be aware that you’re one of the few remaining forces inflating the bubble.
Whatever, it’ll pop anyhow, the #genAI space will become a tiny, mostly open-source, fraction of what it is today, and at that point. we devs can use it without feeling guilty.
@timbray Getting away from guilt will be hard given the training data and the lack of regard for copyright and licenses. Anyway, free software depends on copyright being honored for its licenses to work. Bubble or not, there are some vicious legal gotchas in there. And even small-scale use makes the climate crisis worse. It was already bad before this craze.
I'm not seeing a bright future for the stuff.
Happily, I love coding, so a lack of chatbot aid won't phase me. Not using it now, won't miss it tomorrow.
-
@bflipp @virtuous_sloth I'm curious: What kind of hardware is that?
I bought a used radeon 7900 off of ebay. I have a very old dual xeon supermicro motherboard with 256GB of DDR3 that it’s installed in. I can run much larger models than 30b but that’s the cutoff of for what I consider a close parity for my use cases between open source models (I use Qwen) and commercial ones like Opus, Sonnet, or GPT.
If you have a GPU with 16GB of VRAM and 32GB of system RAM you can have a decent experience for some use cases. I use it primarily as my Joplin interface for notes and task planning with OpenClaw, and for scripting/prototyping with Claude Code.
-
So those of you who are burning millions of tokens on Claude Code should probably be aware that you’re one of the few remaining forces inflating the bubble.
Whatever, it’ll pop anyhow, the #genAI space will become a tiny, mostly open-source, fraction of what it is today, and at that point. we devs can use it without feeling guilty.
@timbray Let’s wait for this to happen. No need to have FOMO. If it comes, fine. If it doesn’t, finer.
-
I bought a used radeon 7900 off of ebay. I have a very old dual xeon supermicro motherboard with 256GB of DDR3 that it’s installed in. I can run much larger models than 30b but that’s the cutoff of for what I consider a close parity for my use cases between open source models (I use Qwen) and commercial ones like Opus, Sonnet, or GPT.
If you have a GPU with 16GB of VRAM and 32GB of system RAM you can have a decent experience for some use cases. I use it primarily as my Joplin interface for notes and task planning with OpenClaw, and for scripting/prototyping with Claude Code.
@bflipp @virtuous_sloth Linux?
-
So those of you who are burning millions of tokens on Claude Code should probably be aware that you’re one of the few remaining forces inflating the bubble.
Whatever, it’ll pop anyhow, the #genAI space will become a tiny, mostly open-source, fraction of what it is today, and at that point. we devs can use it without feeling guilty.
@timbray the open source models are still built on theft and exploitation. They still reinforce social biases. Sure, much better than paying Anthropic.
