Artificial Intelligence thread

[iewgnem]

That's why I said it's just a metric derived from token count. $600k might be the nominal value on the open market, but not what we are paying.

Also most developers are nowhere close to using it up lol. I myself only use about 10%, despite automating most of my daily workflow.

Thats what I mean by people dont actually use it, the company pay bulk price for everyone and people who use less average out ones who use more, which means the system breaks down if everyone actually start using it at high rates, which is probably what MS, Uber and Coinbase found out.

I'm sitting on around 1B GLM5.2 tokens on my dashboard for last 30 days and I have my team sharing my $100/mo Kimi plan, on GLM I only came close to 5h limit once and I dont even think about rate on Kimi. So Im sitting at a team of 4 people doing enterprise dev without even checking tokens on $250/mo.

So as I said, the only reason anyone would use Claude is if they dont know the bill

Ps. These are only for AI at around Opus level that I need for coding, for general purpose, task automation, doc summery, etc, Deepseek is for all practical purposes free, I still havent worked through my initial $50

 

[iewgnem]

I will say again with respect for Anthropic and Claude Code. If you personally use Claude Code, then don't waste your time calling it out here, because you are a giant hypocrite. I specifically asked Zai person I interviewed on what I can use to try GLM models out without using Claude Code. I only migrated to GLM after ZCode came out. And frankly, it's a really good program. They update it so frequently.

GLM works great on OpenCode. Zcode is great too but I prefer CLI, if Zai come out with a CLI I might switch.

I wouldnt touch Claude Code with a 10ft pole and that was true even when I was using Claude

 

[temporary1]

GLM works great on OpenCode. Zcode is great too but I prefer CLI, if Zai come out with a CLI I might switch.

I wouldnt touch Claude Code with a 10ft pole and that was true even when I was using Claude

Any harness that isn't opensourced/available on GitHub, I don't install, trust or use.

Codex biggest success was actually OpenAI open sourcing it. Anyone can go and inspect the code, download and modify on their own, developers from all over all the world can scrutinize it etc

 

[Wrought]

Thats what I mean by people dont actually use it, the company pay bulk price for everyone and people who use less average out ones who use more, which means the system breaks down if everyone actually start using it at high rates, which is probably what MS, Uber and Coinbase found out.

I'm sitting on around 1B GLM5.2 tokens on my dashboard for last 30 days and I have my team sharing my $100/mo Kimi plan, on GLM I only came close to 5h limit once and I dont even think about rate on Kimi. So Im sitting at a team of 4 people doing enterprise dev without even checking tokens on $250/mo.

So as I said, the only reason anyone would use Claude is if they dont know the bill

Ps. These are only for AI at around Opus level that I need for coding, for general purpose, task automation, doc summery, etc, Deepseek is for all practical purposes free, I still havent worked through my initial $50

People literally can't use it fast enough because we are all producing way more code than we can review and approve without blindly trusting Claude to not take down our systems, which would cost the company 1000x more than what we get from shipping code faster. Not like it's bad at reviewing, mind you, we just can't afford to make mistakes on production. The limiting factor is human attention, not AI budgets.

Just big corporate things.

 

[iewgnem]

People literally can't use it fast enough because we are all producing way more code than we can review and approve without blindly trusting Claude to not take down our systems, which would cost the company 1000x more than what we get from shipping code faster. The limiting factor is human attention, not AI budgets.

Just big corporate things.

I think at some point one just has to acknowledge if you're going to use agentic coding to full capacity. you just have to accept agentic reviews. Different companies have different cultures and processes so I cant comment on everyone, but I will say I probably burn majority of my tokens on requesting full reviews, automated tests and writing new customized test harnesses at every step.

 

[Wrought]

I think at some point one just has to acknowledge if you're going to use agentic coding to full capacity. you just have to accept agentic reviews. Different companies have different cultures and processes so I cant comment on everyone, but I will say I probably burn majority of my tokens on requesting full reviews, automated tests and writing new customized test harnesses at every step.

Yes, there is considerable effort devoted to revamping our cumbersome test suite and deployment pipeline to try and bring its coverage up to the level where we can trust Claude to handle the rest, but at some level the risk tolerance is a financial decision that gets made by finance types, not developers. There's also talk about beefing up the safeguards and fallbacks so they can handle a lot more post-deployment churn than they have historically, but again, that kind of infra isn't really my wheelhouse.

 

[temporary1]

IMO the one I am waiting for is Deepseek V4 release and if it's GLM 5.2-tier. If they manage to do it, and keep the costs reasonably low along with their cache pricing, it's all over.

Imagine having semi-frontier level coding intelligence for 0.2 cents or whatever, per million cached input tokens lol

 

[gpt]

IMO people get attached to Claude because Anthropic put a lot more effort into giving it "personality" than Chinese models, but the world would not feel a single iota of difference if Claude disapeared tomorrow

There is a fair amount of circumstantial evidence to suggest that their secret sauce is running very aggressive RLVF loops over curated datasets, much of it obtained by having the necessary '''creativity''' to not get ceased and desisted into oblivion from all the copyrighted material you trained your model on.
Moves like this:

Please, Log in or Register to view URLs content!

are just one facet of a much broader ingestion strategy.

 

[Michael90]

It's legit really good, but not worth the money unless your company is paying the subscription (like mine lol). It can still make the occasional dumb mistake, but it reliably gets you ~80% of the way there even on complex tasks. Also benefits hugely from MCP connectors to internal documentation, messaging, codebase, etc, which lets the agents go wild.

And speaking of not worth it:



View attachment 177880

Please, Log in or Register to view URLs content!

Who are those others/unknown?

 

[Michael90]

Thats what I mean by people dont actually use it, the company pay bulk price for everyone and people who use less average out ones who use more, which means the system breaks down if everyone actually start using it at high rates, which is probably what MS, Uber and Coinbase found out.

I'm sitting on around 1B GLM5.2 tokens on my dashboard for last 30 days and I have my team sharing my $100/mo Kimi plan, on GLM I only came close to 5h limit once and I dont even think about rate on Kimi. So Im sitting at a team of 4 people doing enterprise dev without even checking tokens on $250/mo.

So as I said, the only reason anyone would use Claude is if they dont know the bill

Ps. These are only for AI at around Opus level that I need for coding, for general purpose, task automation, doc summery, etc, Deepseek is for all practical purposes free, I still havent worked through my initial $50

How do they make money at the end of the day? Since being so cheap(almost free) gotta incur some costs no? Else all Chinese models will be as cheap as deepseek.