What surprises me is Hy3 on a second place, didn't know that! Is it really that good?
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Am I the only one who find it funny the X axis on that chart is in log scale while the Y axis is not only in linear scale but doesn't even start at 0, lol
In the real world what you actually need to measure is constant cost capability: how much can each model do for the same $10 in tokens.
On the same budget, its the difference between if you can develop one app with AI, or run an entire company on AI.
Damn, just tonight antrophic refuses to translate documents into simplified Chinese anymore... flagged it as illegal.. (its not, its a document that claude code itself generated and I merely ask it to translate to Chinese) Fucking Dario must have having a heart attack at the cheap deepseek api pricing and inferenced on Huawei in a nation with orders of magnitude electric power capability than the USAHe shouldn't be too worried. I'm pretty sure that eventually Western enterprise firms are going be walled off from Chinese AI, forcing them to use OpenAI and Anthropic. Same will probably be true of major compute providers, at least those who's infrastructure is based in the West.
The only way to get around it will be to either use your in-house solution/open source models or your own compute or both (like Google and Amazon).
Deepseek: After the discount event ends on May 31st, the V4-Pro price will be permanently adjusted to one-quarter of the original price.
The original price is time-limited; the discounted price is permanent
View attachment 175393
So Huawei's pod is online then? The one theyve been building for DS?
After Google got caught secretly downloaded a 4gb nano LLM into Chrome to offload some AI to consumer edge/gpu they then imposed per token limits and 5 hour limits just like Anthropic
Seems like American AI bubble is about to burst, the free lunch is over.... this is dot com energy all over again...
NetZero, Juno, AllAdvantage... History repeats itself.
Like I said, its up to China to be the savior of mankind
"""walled off"""
thats not what they were hoping for and its not what sustains their valuations
orginially US was supposed to lead in AI , China wasnt supposed to catch up at all
The rest of the world would have to pay the US a hefty premium via ai api , so America gets to charge an global "AI tax" and "AI" (intelligence as service) would start replacing oil/OPEC as the new reason countries had to use dollars thus extenting the dollar hegemony
That was supposed to be how the 40 trillion debt gets paid off, and why all the tech companies were building datacenters like no tomorrow
China not only catch up, but gave the weights for free, and also catch up with the gpu side with Huawei + Deepseek and SMIC/ etc beyond US calculations
And China is playing the rare earth card and the Iran War meant to chokehold China on energy also backfired
So now the house of cards is imploding, thats why you see Dario blocking Chinese translation for Claude, Google installing local models inside Chrome and gating Gemini usage by tokens and hours limits, and everything else going on....
"""walled off""" now means the only way these US AI tech companies can survive long term is to displace American white collar labor within US domestically (originally they wanted to do this to the rest of the world and so they own citizens can sit back and collect ubi off the backs of global AI Tax dividend) but now that China foreclosed that pipedream thats only going to lead to blackslash within domestic... see the recent protests about AI data centers or all the CEOs getting booed for talking about AI at graduation ceremonies....
Basically China rugpulled the US and killed America's AI dreams, no wonder Dario is pissed
For a company that names itself Antrophic the racists loser showed his true colors
I'm currently using a combination of Codex (GPT 5.5) and Claude Code (deepseek V4). Simple preliminary tasks like information gathering are handled by V4, while GPT 5.5 defining user needs and developing a plan.
Once the plan is finalized, I only need to use /goal to let deepseek V4 pro complete it. Because the plan is already defined, it's less likely to go astray. A task that would normally require using the full 5-hour CodexPlus limit can be completed with only 20% of the limit and V4 pro, in several times the time and at a less than a dollar's cost.
And these data will definitely benefit the training of V4.1. It will no longer be a free-roaming without Harness, nor will it be a distillation of GPT-5.5. It will be completed step by step under the guidance of the GPT-5.5's framework. There will be errors in the process, but it will not go astray.
Once the plan is finalized, I only need to use /goal to let deepseek V4 pro complete it. Because the plan is already defined, it's less likely to go astray. A task that would normally require using the full 5-hour CodexPlus limit can be completed with only 20% of the limit and V4 pro, in several times the time and at a less than a dollar's cost.
And these data will definitely benefit the training of V4.1. It will no longer be a free-roaming without Harness, nor will it be a distillation of GPT-5.5. It will be completed step by step under the guidance of the GPT-5.5's framework. There will be errors in the process, but it will not go astray.
Last edited:
Yes, I used it in Unreal Engine's blueprints, which are notoriously unfriendly to LLMs. Even the GPT 5.5 still can't fast and cheaply complete tasks that are simple for humans but time-consuming. So I had Deepseek V4 as its sub-agent to handle the cumbersome work using MCP, let GPT rewrite the original cumbersome blueprint assembly into a wrapped C++.
Oh, by the way, Codex particularly likes to write redundant code and tests to ensure security, while having Deepseek handle the actual coding can mitigate this to some extent.
Last edited:
Unreal Engine supports c++, why use blueprints at all?
Stick to c++ and even cuda or assembly language