WEIRD AI TOOLS
The unhinged side projects and apps that shouldn't exist but do.
Substack adds an AI detector to help spot blogs written by no one
Now you can finally confirm your favorite philosophical newsletter is just Claude in a trench coat.
Jack Dorsey is taking on Slack with Buzz, a group chat platform for teams and their AI agents
Jack Dorsey wants to put your silicon coworkers in the group chat.
DS V4 on single b300. only 770 tok/s batched in vLLM
POV: you bought a B300 and realize bottlenecks don't care about your budget.
Show HN: OSS Cross-Harness self hosted registry and analytics for AI Agents
Your agents are messy, now you can watch them fail in private.
Jack Dorsey launches Buzz to combine team chat, AI agents and Git hosting
Jack Dorsey is building a Slack-Git-Agent chimera and honestly, we're listening.
SenseNova-U1-8b-MoT-Infographic-V3 has been released (2 weeks after V2)
Finally, an AI that can fix a typo without turning your hand into a plate of spaghetti.
Google launches a cheaper alternative to large AI security models like Mythos
Google is now offering budget-friendly bug hunters for your codebase.
Nativ: Run AI models locally on your Mac
Your MacBook is finally earning its rent: LM Studio has a new MLX-native challenger.
I benchmarked Unsloth's Qwen3.6-27B NVFP4 on 1x/2x 5090s. MTP is great until it really isn't.
MTP is a speed demon until your context window decides to wake up and violence.
pi 0.81.0 adds support for llama.cpp
Local LLMs just got a VIP lane in the Pi dev stack.
Halliday’s latest smart glasses feature a much-improved display
New screens for your face, now with 50% less eye-strain-induced crying.
Grabette: an open system to record robot-manipulation data
Forget simulations, Hugging Face wants you to manually puppet your way to AGI.
There's a new PR for llamacpp claiming to boost prompt processing with rocm by around 15%, also fixes a bug which makes Q2_K 28x faster
Red team winning: AMD users just got a free 28x speedup while NVIDIA slept.
shot-scraper 1.11
Simon Willison just made your automated screenshots 30x less flaky.
Benchmarked every spec-decode method on Qwen3.6-27B across vLLM and SGLang (single RTX PRO 6000 Max-Q)
Local LLM math just dropped: DFlash is making 27B models run like they're on caffeine.
Deterministic Replay for AI Agent Systems
Infinite time loop for your agents, but for debugging instead of trauma.
Tri-Net v2: Open-source implementation of our Scientific Reports paper on unified skin lesion and symptom-based monkeypox detection [R]
training an ai to detect mpox so you don't have to google 'is this a rash'
MTP on MoE matters
Multi-Token Prediction is finally making local MoE models go brrr.
My learnings from optimizing training pipeline to go from 36 steps/minute to 47 steps/minute
POV: you're too broke for NVMe but too stubborn to stop training models.
DavidAU somehow managed to improve Qwen 3.6 27B
LocalLLaMA's favorite villain accidentally cooked a gourmet meal.
Those in the 1000+ prefill and 100+ decode range on Qwen3.6 35B at Q4, what hardware are you running?
The 'runs on a potato' arc is over; we're now in the 'ROCm or death' era.
nvfp4 kv-cache on 2x5060 ti, vllm
Blackwell optimization leaks to the masses: 4-bit KV-cache is officially here.
AI’s most important protocol is getting a little bit easier to use
Devs finally stop reinventing the wheel to let their LLM read a spreadsheet.
Have you built your own agent instead of using openclaw or Hermes, how’s it going for you?
When the abstraction layers get so thick you forget how the silicon actually thinks.
Reverse-engineering is cheap now
Your smart fridge has no secrets now that code is free.
Adobe camera app’s new feature will critique your photos using AI
Great, now your camera can tell you your photography sucks in real-time.
Would you trade speed for accuracy?
Your 4-bit model just got a brain transplant at the cost of your dopamine rush.
Launch HN: Bloomy (YC S26) – AI-powered mastery learning for K-12
RIP expensive private tutors, the AI is finally solving Bloom's 2-sigma problem.
Training a harness for model-agnostic and task-environment-agnostic capability improvements with PyTorch-like framework [P]
Why train the model when you can just train the vibe check around it?
Adobe’s ‘natural look’ camera app embraces generative AI
Adobe pivoted from 'natural' to 'generative' fast enough to give you whiplash.
Good ASR and TTS models?
Whisper is the new 'hello world' and your ears deserve an upgrade.
1-Bit LLM in the Browser
Your browser is now a 1-bit beast thanks to WebGPU.
Introducing Scylla's Band, a new TTS model + inference framework with Android sample!
Your phone's internal monologue just got a lot more emotional and way faster.
AnovaX: A Local, Multi-Agent Voice Assistant with LLM Planning, Typed Executors, and Adaptive Recovery
Jarvis just moved out of the cloud and into your RAM.
Gemma 4 is still lazy
POV: You spent 4 hours on a config just for the AI to hit the 'nah' button.
I built Ghost: 60+ verified Mac actions, local document search, use your LocalLM model privately all from your menu bar notch
Your MacBook notch just grew a brain that doesn't report back to the mothership.
Are there some textbooks that take a primarily engineering approach to machine learning (as opposed to a "scientific" approach)? [D]
POV: You realized building a model is 5% of the work and the rest is just fancy plumbing.
BeeLlama.cpp v0.4.0: KVarN, KV precision tail, q2_0-q3_1 KV cache, upstream rebase
Your 8GB VRAM is screaming for mercy, and BeeLlama finally answered.
Most AI agents have no concept of opportunity cost [D]
Your agent is basically a toddler with your credit card and zero sense of time.
poor man's way to local inference on the go
Local LLM fans will literally carry a GPU in a briefcase instead of paying for an API.
Qwen 3.6 27B + Opencode: what am i doing wrong?
When your 27B model decides that 'meaningful compression' means 'selective amnesia'.
Follow up: GPT-2's vocabulary as a hyperbolic tree — 32,070 tokens in a Poincaré ball you can fly through [P]
POV: you're 30 minutes into a trip inside a Poincaré ball wondering why 'apple' is near 'jobs'.
Claude Code uses Bun written in Rust now
The stack is getting so deep even the robots are writing Rust now.
LLM-Integrated Multivariable Calculus Course
Calc 3: Now with a personalized math wizard that doesn't judge your basic arithmetic errors.
i couldnt find anyone making full movies locally on a mac, so heres what that looks like right now
Pixar in a briefcase: making entire movies while you sleep with the WiFi off.
What’s your favorite underrated local model?
POV: You're tired of giving Gemini your data and want to run a giga-brain on a microwave.
Deepseek v4 Flash on 80 GB VRAM and 128 GB DDR4 RAM
POV: you're conducting a symphony of GPUs just to run a 'flash' model at home.
Claude Code uses Bun written in Rust now
Anthropics ninja-swapped Bun for Rust and literally nobody noticed until now.
I made a very detailed guide on how to run Qwen 27B through GHCP harness at highest performance and best quality
Local models are coming for your subscription fees.
Interactive map of GPT-2's token embedding space - tap any token and explore [P]
finally, a way to see what your LLM is actually smoking inside its brain
Setting up your spare Mac for Claude Code to control, a step-by-step guide
Watching your computer build itself while you drink coffee just reached peaked levels.
CMP 170HX 8gb — Perf + Memory + PCIe Gen2 Unlock - NVIDIA Driver 610.43.03 (patched open kernel modules)
NVIDIA is punching the air right now while you unlock 64GB of VRAM for pennies.
Sharing MiniBot v2, this is what I'm currently using I gave it a major update so I thought I'd share. I make things that do work for me, always have... and this is the latest.
A 20k-line PowerShell script is the 'run-in-memory' AI minion you didn't know you needed.
SQLite Query Explainer
When you're too lazy to read docs so you have an AI build a debugger for you.
This open-source AI canvas keeps the entire image-to-video workflow in one reusable graph
Why use one model when you can chain ten together on a spatial whiteboard?
How are y’all stomaching the “AI Boom” prices?
POV: you’re trying to build an AGI on a budget while NVIDIA gaslights your wallet.
TabFM Studio: point-and-click predictions on spreadsheets with tabular foundation models, fully local [P]
Finally, a way to use AI on your messy spreadsheets without a computer science degree.
[Model] catmind-1.2b
Reasoning models are peaked: this one replaces logic with elaborate cat stories.
CrimeNER Demo: Named-Entity Recognition in the Crime Domain
The 'Law & Order' unit for your unstructured text data is finally here.
Qwen and Gemma providers
POV: you're discovering that not all 4-bit quants are created equal.
GET THE DAILY CHAOS
The only newsletter for people who read AI news at 3am and feel things. One email a day.