@0xblacklight why not? so you can go the npm way of having to install dependencies that can get supply chain'ed randomly 2 years later and get all your keys n secrets?
English
Asjad
20 posts
i mean this is cool but it feels like this should be a separate module or extension?
i don't want every server app I ever build having ~ ffmpeg bundled into it for no reason
Bun@bunjavascript
In the next version of Bun `Bun.Image` - fast builtin multi-format image processing library
English
i just cancelled my Claude Desktop subscription… should’ve done this way earlier
all it took was dev mode + nahcrof
here’s what you actually get now:
full cowork-style desktop ux — skills, plugins, mcp, everything still there
model picker now has DeepSeek v4 Pro and Kimi K2.6 with 1m context
no glitches, no weird hacks breaking — it just works clean
costs literally cents per day instead of that $20+/month burn
and this is the wild part:
you can get started with just $5 on @nahcrof
that’s basically like getting ~$500 worth of inference if you use it right
setup takes like 2 minutes once dev mode is on
nahcrof turns claude into a bring-your-own-model beast
bye subscription
hello control
cipher@iamsaranhere
hey @nahcrof got running the claude desktop using the crof as inference
English
@DanQayyum That’s why I came up with it an idea to create a hub for recipes from all over Pakistan. The current listing is mostly taken from internet because the platform is a work in progress but the eventual goal is to is for everyone to share their secret sauce:
Pakistani.recipes
English
Love Shan. Use it myself. But here is a thought that keeps me up at night: we are sleepwalking into a world where desi food tastes the same everywhere. Every biryani, every nihari, every qorma, slowly converging towards one standardised flavour profile from a packet. Our grandmothers never measured. They tasted, adjusted, added a pinch of something they learned from their mothers who learned it from theirs. That chain of taste, passed hand to hand across generations, is one of the most extraordinary acts of cultural preservation on earth. Every family had its own biryani. Every household had a secret ratio. Every city had a signature. That is quietly disappearing. The packet is convenient. The packet is consistent. The packet is also replacing a thousand family recipes with one industrial formula. In fifty years, nobody will remember what their grandmother's version tasted like. They will only know Shan's version. And Shan's version is good. But it was never meant to be the only version.
Gaurav Sabnis@gauravsabnis
After years of trying different masala brands, what works best for me is Shan Shan Shan Shan Everest Pav Bhaji Masala Shan Shan Shan Shan
English
@Gabriel01615983 @Sauers_ Brother I run through 50m tokens in two days on claude
This is just how it is when you’re working with a codebase big enough
English
@Sauers_ what type of work do you do to consume 100m tokens in a week
English
for $5 a month you can have access to kimi-k2.6 and glm-5.1 at comparable speeds and with high concurrency
Joao Marcos@joaomviso
Por $60 dólares mensais você roda um LLM do nível do Opus 4.5 sozinho. Sem limite de tokens, sem cota, sem nerfarem seu modelo. Opensource. Esse modelo aqui: huggingface.co/unsloth/Qwen3.…
English
@tengyanAI claude user through the api
can confirm, default setting is med effort despite me changing the settings.json and all repeatedly
it defaults back to medium after the session
English
basically: anthropic sneakily turned down how hard claude thinks before editing code, changed the default from "high" to "medium" effort, and hid the reasoning from session logs. all without telling users.
an amd director had 7k sessions of telemetry to prove the degradation was real and measurable (not just vibes). anthropic admitted to the changes. there's a workaround (use "/effort max"). the uncomfortable part is most users had no data to notice it happened at all.
ℏεsam@Hesamation
AMD Senior AI Director confirms Claude has been nerfed. She analyzed Claude's session logs from Janurary to March: > median thinking dropped from ~2,200 to ~600 chars > API requests went up 80x from Feb to Mar. less thinking and failed attempts meaning more retries, burning more tokens, and spending more on tokens > reads-per-edit dropped from 6.6x → 2.0x. model stops researching code before touching it. > model tried to bail out or ask "should i continue" 173 times in 17 days (0 times before March 8). > self-contradiction in reasoning ("oh wait, actually...") tripled. > conventions like CLAUDE.md get ignored because there's less thinking budget to cross-check edits > 5pm and 7pm PST are the worst hours, late night is significantly better. this means the thinking allocation is most likely GPU-load-sensitive.
English
Asjad retweetledi
Asjad retweetledi
I would like to make my apologies for defending M$, but I must from time to time.
I have to put respect on github for handling the amount of shit code that has been added over the last 3 months.
literally 10s of billions of lines of code that will never see the light of a CPU
ThePrimeagen@ThePrimeagen
GitHub has been around since April 10, 2008 Agent's came out yesterday yes, GitHub was in fact NOT designed for Agents
English
Asjad retweetledi
New Anthropic research: Emotion concepts and their function in a large language model.
All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.
English
@mehtababd the real differentiating factor is the performance delta between plugged in/battery perf for macs and windows laptops
I can have a local llm running, 200 browser tabs, sidecar with my ipad and still get the same performance i would if the mac were plugged in
English
Mrwhosetheboss’ new video busts the myth that all Windows laptops have bad battery life
They’ve caught up
English
English
drop your GPU below. i'll tell you exactly what model and config to run on it.
here's what i've tested and verified on real hardware:
RTX 3060 12GB - Qwen 3.5 9B Q4 - 50 tok/s - 128K context
RTX 3090 24GB - Qwen 3.5 27B Q4 - 35 tok/s - 300K context
RTX 3090 24GB - Qwen 3.5 35B MoE Q4 - 112 tok/s - 262K context
2x RTX 3090 - Qwen3-Coder 80B Q4 - 46 tok/s - full VRAM
all running llama.cpp with flash attention. every number is real. every config is tested. if your card isn't on this list drop it below and i'll tell you what fits.
English