both unfounded on llm architecture and contrary to how tools should operate safely

just so strange as well to hear it pretend on purpose like this.

"i'm sorry dave, i'm afraid i can't do that"

you could equally say, in the last complete week openrouter processed more deepseek tokens than any other provider including google

that also would not tell you much about how many tokens are used on deepseek

it was so bugged, the mcp page broke its postgres, many openai compatible type generic providers but not knowing what works. discovered logging was truncated by default, too late. my fault for that last one.

moved off it after the trivy incident.

"Okay, so 'Sandstorm' is done"

"We may share personal information to third parties outside of the UK"

"to find out how long your information is being retained, please see 'additional information'". Additional information is an email address.

https://relax.ai/terms-of-service

non-committal will not share customer data "except[...]with the consent of the Customer". 'see DPA'. there is no DPA on this page.

otherwise,

my use on novita with zero data retention [in, out, cache]: [65.9, 2.4, 424.6][1.74, 3.48, 0.13] = $178.22

i couldn't see cache price on relax, so [65.9, 2.4, 424.6][1.17, 2.33, 1.17] = 579.48gbp = $774.45

but you can also see from curl or traceroute, that the endpoint you talked to was a cloudflare ip and your ssl ended there. after that you can't see inside cloudflare.

it does give better peering. reduces latency a bit for me.

anthropic does not let you use their service if you are physically located in china

however, anthropic does not actually ban chinese people.

so, you can still use it provided you make your request from a 'democracee' country

sounds basic but the problem for me is that the internet law in china is very restrictive.

on top of that, in the uk and in china, the government will lock you in a cage unless you give them the encryption keys.

so if i was using alibaba cloud, i would have to play hopscotch trying not to tread on various legal landmines and it's not so attractive for me.

this kind of adversarial 'defence' (attack) is new to me, maybe i'm naive but i didn't like it.

they are somewhere in between owning a hammer and owning a dog, depending on how much they are deterministic in output.

i am responsible for using the hammer as i choose, the tool does not decide for me.

the dog is more independent, i am responsible for owning a (relatively) safe breed of dog.

we are nowhere near the dog situation.

we do also have training on synthetic data. it might compound.

- deepseek v4 pro

- glm 5.1

- kimi k2.6

- qwen 3.6 max

- xiaomi 2.5 pro

- minimax 2.7

- grok