on the other hand cyberespionage has changed drastically

I have to spend hours understanding how other humans have asked the computer to do something, and they are also bad at asking it to do so, so I could have gotten a better result if they just came to me with their desire...

Yup. So often with AI output, I want to know the prompt the human wrote instead.

It's the same price as regular processing. You get guarantees microsoft give you, which are ones OpenAI won't (or require dedicated spend,) and you can use azure identities for access.

We use it for access to gitlab, ado, github, datadog, rancher. I think the rancher one is the worst. We develop custom MCP servers for our internal stuff for agents to use, and it all gets accessed thru agentgateway.

I don’t really like having to use MCP but we don’t have a good solution for authorizing individual calls outbound from a sandbox without choosing to just not care about the sandbox.

not hard for secrets with explicit patterns and existing pipelines to detect them

Unless they're base64-encoded or compressed?

I would expect, although have no evidence, that any obviously high entropy crap like base64 and so on probably would get removed whether it's a secret or not.

because the products are great and I am a happy customer and stand by my opinions regardless of forum

someone should take these moves and build bots for each play style for Toribash.

this is insanely misleading, you can't run anything close to current chatgpt or gemini on local hardware

I am running GLM 5.3 across 2x DGX Sparks and was doing comparisons and it absolutely can beat Gemini. Yesterday it corrected a poor Fable 5 response even

I should have clarified, you can’t run it on their local hardware, which is a 16gb gpu. Glm-5.3 and k3 are of course near the frontier.

GLM 5.3 Flash? Qwen 3.8 Flash Next? I believe those both are as good as the best Gemini, competitive with Terra.

Yes they are quite good, but are not able to run on a 16GB RX 9070.

Quantized Qwen 3.8 Flash Next could maybe run eventually on that card with a highly optimized inference engine that dynamically caches the hottest layer experts. Even then you run into some hard limits.

many of them (kimi k3, glm-5.3) have license requirements to sell them with model-as-a-service.

do you have a source for this? it doesn't really make sense to me

In illegal output: https://asahilinux.org/slop/