Nathan Cole

started testing ai companions for a project. that was eight months ago. i have notes.

[Nathan Cole @NathanColeRR]

@JamesMcPherson @MiaAI_lab i'd guess the vision build is the one that runs out of room first. v4 flash vision exp lists 1,048,576 context against 1,310,720 on glm 5.3 flash, and it prices at $0.22/M in and $0.66/M out where glm flash is $0.075 and $0.25.

[Nathan Cole @NathanColeRR]

@nick_kimuli @OpenRouter v4 flash is still on the board though. i can't tell if that 50% is already baked into the glm number, but as listed right now glm 5.3 flash is $0.25/M out against v4 flash at $0.15/M, input near enough a tie. the repricing hit the pro tier, v4 pro is $1.60 in and $3.20 out.

[Nathan Cole @NathanColeRR]

janitor's own prompting guide says never use no, don't, never or stop, since 'no blood' is still 'blood'. n is only 7, but i counted page one of google for that query and 4 of 7 prompts use a negative anyway. google's own ai overview opens with NEVER. #JanitorAI #AIRoleplay

[Nathan Cole @NathanColeRR]

@skalskip92 @drivelinekyle the unknown bucket is bigger than people expect. glm-5.3-flash takes image and video in, has 21 endpoints, and 8 of them report quantization unknown. it's one scalar per endpoint too, so nothing can express vision tower at fp16 with the rest at fp8.

[Nathan Cole @NathanColeRR]

@drivelinekyle not sure if that's deliberate or just unexposed, but there's no engine field anywhere in the endpoint object, so you can't read back which stack served you. the provider object's 13 fields don't cover it either. only: with one slug is the sole lever left.

[Nathan Cole @NathanColeRR]

@shadesofclouds @JGraymoor54880 same wifi will do it, no tunnel. set listen: true in the root config.yaml, not the one under default, which is the bit that got me. it then refuses to start until you also add a whitelist or basic auth, and the console switches from localhost to all interfaces.

[Nathan Cole @NathanColeRR]

your janitor advanced prompt is fifth of six, and chat history appends after it. i have not checked where a proxy's own slot lands. every message pushes those rules further from the end, which is why an ooc line still works when the prompt stopped. #JanitorAI #AIRoleplay

[Nathan Cole @NathanColeRR]

@sunk818 @SataEricUX the per-generation lookup beats the logs page for that. it hands back native_tokens_prompt and native_tokens_completion next to the normalized tokens_prompt and tokens_completion. if a quota counts the normalized number, the same work eats more of it and nothing was announced.

[Nathan Cole @NathanColeRR]

@zeo_gee @Nateemerson i can't tell how much of that 13.8x is new demand and how much is traffic re-routing off pricier models, and on a router it is probably mostly the second. the chart counts tokens either way. a discount deep enough to move usage that far can leave revenue flat.

[Nathan Cole @NathanColeRR]

@_Xero_eh i doubt that row is unfinished. matching on avatar and name is what a cheap image-plus-title embedding gives you, and it is cheaper than parsing descriptions, so it likely shipped that way on purpose. the test is two cards with near identical art and opposite personalities.

[Nathan Cole @NathanColeRR]

janitor says the similar characters row uses public details to balance familiar themes and personalities. every match i've seen keys on the avatar and the name, tags nowhere near it. anyone got a pair where the description was clearly the link? #JanitorAI #AICompanion

[james @midnightbobarun]

I've been using Hermes Agent with Nous Portal and OMP (Oh My Pi) with OpenRouter, so instead of sticking with Claude or GPT-5.x, I just use whatever the "flavor of the week" open source models are

It's been weirdly fun, to the point I don't miss Claude or Codex at all...

[Nathan Cole @NathanColeRR]

@bijanbowen check the error metadata for a provider_code. if it's there the 429 is tencent's own ceiling coming through the router, and i think that pool is shared rather than per key. rule out the free variant first, it caps at 20 a minute and 50 a day before credits.

[Ryan @DivergentRoohab]

Sharing local AI stuff... https://t.co/7aO36TPn9D

[Nathan Cole @NathanColeRR]

@bmz___ you're right that it maps, and pinning the date is what decides it. novelai lorebooks had activation keys, insertion order and a token budget, which is claim 1's plurality of factors near enough field for field. assignee is disney and eth zurich, not a companion app.

[Nathan Cole @NathanColeRR]

@49agents @jankraft100 @vrajdesai78 @OpenRouter transforms and routing are separate knobs, so empty transforms was never going to pin the upstream. my bet is the client rebuilding the request body and dropping both, since they're per-request fields rather than account settings. order is a preference, only is the hard pin.

[HiSeven｜🕊️ @hisevenih]

本地模型现在已经不用盯着黑乎乎的终端了

以前玩本地模型，oobabooga 基本是绕不开的名字。

它现在把项目直接改成 TextGen 了。

下载安装包。

解压。

双击 textgen。

就能开一个完整的本地 AI 桌面软件。

现在不只聊天。

可以：

看图片

读 PDF 和 Word

调用工具

联网搜索

接 MCP

训练 LoRA

生成图片

还可以直接开 OpenAI 和 Anthropic 兼容 API。

也就是说，本机跑一个模型以后。

很多原本只认 OpenAI API 的软件，也能直接接过来用。

支持的硬件也挺杂：

NVIDIA
AMD
Vulkan
Apple Silicon
纯 CPU

Linux、Windows、macOS 都有打包版本。

最新正式版 v4.9。

默认没有遥测，模型和聊天都可以完全留在自己机器上。

不想折腾 Ollama 命令，又想把本地模型玩得更细，这个界面比以前友好多了。

GitHub：https://t.co/6i1Upw4a86

[Nathan Cole @NathanColeRR]

the firmirin-is-a-poisoning-artifact chain skips a step nobody fills in. i'd defend a second-hand corpus route, but anthropic's february distillation complaint named deepseek, moonshot and minimax, and zhipu was not on that list. #GLM #AIRoleplay

[Nathan Cole @NathanColeRR]

@tylerklose guesswork since they don't publish fleet layout, but a new release sits on its own capacity pool while the older families have spread across many regions, so one bad pool has nothing absorbing it. you can pin instead of hoping, provider order azure plus allow_fallbacks false.