Model routing is theoretically quite simple, in practice very difficult and currently just barely doable but not doable well. What you need for routing is: 1. MODEL: A profile of its capabilities, like tool calling, vision, etc 2. ENDPOINT: A profile of latency, reliability and downtime, cost per token, etc 3. TELEMETRY: Data on how models and endpoints (note, the same model can perform differently and have different pricing on different endpoints) have performed on different tasks, historically 4. TASK: Information on the nature of the task request (does it need vision, complex tool calls, is it math or coding or prose, etc) 5. ROLES: Assign models to roles, where a role is responsible for a group of tasks. Then you need routing strategies like cheapest, or privacy, performance, or coding, or legal, etc. Ideally you should be able to set this per task, or application, or by user (some have higher or lower allowances; some deal with sensitive data that shouldnt leak). In practice, routing is difficult because different models have different chat templates etc so that routing between them can throw errors. Different models do thinking differently, use tools differently etc. Inference providers also do not send sufficient data for routing decisions, for example task = "code", complexity = "easy", so you would need a routing model running which adds latency and also reduces accuracy. The short version is that routing could be a simple task, if model makers and inference providers were aligned with the need for smart, precision routing. For this, I've built role-model (role-model.dev), a routing protocol and a router runtime. Currently works and is in alpha but I need to clean up the UI before I do a proper release.

给 Dario 一个建议，模仿当年前辈把PGP 源代码运到美国国外的案例。 当时为了绕过加密技术管制，他们把PGP源码印了900页纸质书运到国外。这是因为一旦源代码印到书内，就受第一修正案（言论自由）的保护。 A社可以把几TB的模型也印出来，在国外另起炉灶。

role-model: a model router and routing protocol. coming soon.

Remember when I said anthropic will come back begging for agentic workloads

Kept you waiting, huh? github.com/WismutHansen/foxl… 100% local and open source. Powered by @badlogicgames' pi coupled with qwen3 tts, parakeet.cpp and gemma-4-26b-a4b You'll need to provide the game iso but there is a script that rips all required assets automatically.

sick price pressure in china, coming soon to data center near you short the bottlenecks $SNDK

This is the AI play: 1. Inference costs are going down, so you short the bottleneck cos like $SNDK that have hyped up future revenue - those sales are not going to happen at the same prices when token prices are down massively 2. Buy the infra that will be churning out cheap, commoditized tokens directly into applications - $NET 3. Buy applications that get better when tokens get cheaper - $FIG

This is why I'm short $SNDK

Expected.

1/ You can shrink a language model's KV cache by 200×, in a single forward pass, and it still answers correctly. At 256k context that's 36 GiB of cache down to ~360 MiB, with no change to the base model. Here's how we did it 👇

Been using the Cursor $20 plan for three days doing a good amount of work, fully on Auto. That probably means I've been on Composer 2.5 aka Kimi all the time. Doesn't feel much different than K2.6 in Pi/Craft Agents, or like GPT 5.2 in VS Code back then. If this is the monthly quota then it's clearly not enough, but I can just switch to my actual Kimi suub for the rest of the month and get by. The main point is that I don't really need the Legacy Labs as I have strong workflows so Tier 2 models get the job done without issue.

Japan banning Anthropic because their models can create copycat Nintendo games

Most people have no foresight and it seems they cannot think at all. They can only react to what happens to them.

说实话，长期自我PUA加班的人，会下意识PUA正常下班的人，因为那一刻，脑子不正常

anthropic has published merged eval numbers for fable/mythos. not only they picked the better result of the two (this could overstate results up to 3%), they also used mythos as a fallback for fable refusals, which are shown here as 0%.

Brilliant idea! Next up: Apple randomly reboots your Mac if you're building competing tech, Gmail silently edits your email if you mention rival platforms, and Tesla Autopilot swerves if it detects you're working on self-driving cars. All in the name of safety, of course. Because malicious actors controlling the world’s operating systems, inboxes and cars would be extremely dangerous!