Top
Best
New

Posted by martinald 1 day ago

GLM 5.2 and the coming AI margin collapse(martinalderson.com)
653 points | 426 commentspage 4
dparkmit 6 hours ago|
I run a 2 Claude setup (one architect, one coder) and have been using it extensively. But I don't like the fact that I have to pay $200/month to use it. I'm going to try GLM with my current setup.
wg0 9 hours ago||
Everyone is declaring GLM 5.2 as something that's really a big deal.

I don't know about that but based on my own experience with Deepseek v4 Lite alone (with high effort) I have no doubt in my mind that anyone claiming such great things about GLM 5.2 must be true because Deepseek v4 already is really awesome.

jFriedensreich 11 hours ago||
I am confused by this article landing on front page, it does not seem to contain any new insights but besides mostly reading ok falls apart when mixing up model and harness comparison in an amateurish way. Why would the subpar search tool zai provides be relevant for comparing models? They did not even mention the >capability< to use said tools but talk about MCP/ search providers as if thats not an implementation detail.
moktonar 9 hours ago||
The problem is that the more AI eats labor, the more you can hike the costs until you pretty much can match the salary of the workers you replaced, with some margin, enough for the user to accept the cost. That’s what will happen in the next decade IMHO. price = base expense + what user accepts to pay
ilaksh 19 hours ago||
I think the profits depend on how well they manage their fleet purchases (or possible sub-leasing?) to get high utilization without overloading or idle racks.

Because accelerators like H200, B300 etc. are highly parallel and designed to run like 200 or maybe 300 sequences at once (depends on the model, just guessing). I assume they finance the hardware and that cost per device or rack is the same whether each unit is handling 10 requests or 150 requests (aside from electricity).

And probably international customers factor into it to get good utilization over more of the night time. And it likely is something that they look at quarterly more seriously than monthly. The biggest risk to profits might be a downturn in business that causes some portion of the financed AI accelerators to go idle or get low utilization for some weeks (that they can't sublease).

jaggederest 19 hours ago|
Someone on HN made a comment in one of these threads that we could bake the weights into something like Cerebras's wafer scale chips and serve essentially the entire world off a single wafer, which is a pretty wild thing to think about. You'd have to make new hardware any time you trained a model but that seems really worth it.
ilaksh 18 hours ago|||
Well, Taalas has that kind of technology, but the chip they demoed is probably 20-100 times smaller than necessary since it's only an 8b model.

But let's say they could someday scale that up to a much larger model, 72 large chips per wafer and each chip can do 1000 LLM requests at once (Vera Rubin?). So it's roughly the equivalent of an NVL72 rack.

You might be able to serve something like 50000-60000 requests at once. So I think it's more like handling a small city's worth of customers per wafer than the world if you had that.

I believe in less than 5 years we will get to that, but the model size and/or number of agents is going to keep going up also.

lelanthran 15 hours ago||||
I think the future will have to include specialised host boards for memory chips.

What I actually want is an FPGA board with a very large number of DDR3/DDR4 RAM slots arranged in banks (2, 4, 8 or even more banks). I want an FPGA board that can hold 1TB of DDR3/DDR4 RAM.

The throttling point right now is not RAM, it's bus speed. Having different busses for banks of RAM alleviates that.

Atotalnoob 18 hours ago|||
You’d never be able to update it’s knowledge.

LLMs need retraining to incorporate new knowledge.

Baking them into wafers means they will be out of date by the time they finish the first wafers.

jaggederest 17 hours ago|||
Yes, of course, but all the LLMs are already out of date, so that doesn't seem to me to be a hard limiting factor. Even if they had a knowledge basis ~3 months out of date additionally, being able to serve 100x the requests per watt seems totally reasonable to me.
Shorel 12 hours ago|||
So, what?

I don't see the C++ compiler standards or Newton's laws changing every day.

chrisss395 9 hours ago||
Where does the harness come in to play? I'd love to use GLM 5.2 for general chat, but I don't know of any harness that offers an experience close to ChatGPT or Claude (e.g., history, skills, projects) without requiring a PhD to set it up.
xienze 9 hours ago|
You can point Claude to another model if you want. See ANTHROPIC_BASE_URL: https://code.claude.com/docs/en/env-vars
scritty-dev 18 hours ago||
Braintrust which is a really solid eval tool/platform just compared it to Opus 4.8 to see if it could preserve exact long context retrieval under prod serving constraints and it did really well. I think 6-12 months before OSS has Fable-esque models
pier25 6 hours ago||
Inference margin is irrelevant.

That’s like a gas station saying they have 90% margin over pumps but still losing money.

sailfast 21 hours ago||
How long will that $4.40 rate persist? Until we know more about the real unit economics it will be damn near impossible to rely on steady inference costs or make them predictable at the enterprise level. Gonna be a wild ride for awhile.
markasoftware 21 hours ago|
Multiple providers (who need to make a profit) offer the same 4.40 rate for glm-5.2. It's not subsidized.

Deepseek's 0.86 or whatever is likely subsidized but alternate providers offer it for a price comparable to glm-5.2.

throwa356262 12 hours ago|||
According to deepseek themselves, their current rates are NOT subsidised.

They have published tons of articles dedicated to performance and efficiency engineering. Feel free to have a look...

markasoftware 5 hours ago||
Why is no other inference provider offering similar prices then?
throwa356262 3 hours ago||
How long did it take vLLM to implement deepseeks sparse attention from the r1 paper?

Does ananyone outside deepseek have a working code for the v4 compressed attention mechanism?

Has any other provider managed to bypass CUDA and program the compute engines in their native assembly language to get 10% more performance out of them?

There is your answer.

est31 21 hours ago|||
GPU/RAM/etc prices could continue to rise. If the world leaders decide it's time to build the robot armies, then that could price out the civilian uses for GPUs.
a_c 10 hours ago|
Regarding the lack of vision part, if you are using Claude or opencode, I've made a skill[1] that let's you talk with any models in Claude/opencode mid-session. You ask "Have claude opus to look at this PDF for a second opinion" during a session of claude with GLM5.2 or opencode with GLM5.2

It doesn't need to pass whole conversation history as context (unlike /model), you can ask follow up to that forked model (which sub agents in claude doesn't support AFAIK), and you can ask models from opencode while using claude.

[1] https://github.com/kmcheung12/second-opinion

msephton 9 hours ago|
FWIW I've seen subagents remain open for followups on latest Claude.
a_c 5 hours ago||
Will give it a try thanks for the heads up. Mind if I ask if you are referring to SendMessage? I was testing on Claude code 2.1.196. SendMessage was not available. Skimming through their change log didn’t seem to have anything related to SendMessage. There is “ Fixed SendMessage silently misrouting when a re-spawned agent reuses a previous agent’s name — the tool now detects the mismatch and asks the caller to retarget” from 2.1.199. Not sure if they are related
msephton 1 hour ago||
I have no idea what things are called but somebody showed me how they could start a bunch of subagents and steer each with additional messages; some closed after their task was completed, but one stayed open for followups and continued working after being prompted by the main agent who was acting as triage.
More comments...