maderix (@originalmaderix) header
maderix (@originalmaderix) profile picture

Part time model optimize, full time model tuner 🤖 maderix/ANE - first ever model training on Apple NPU/ANE 🧑‍💻compiler author https://teapawt.net/t.co/uyii0ijkTu

Joined May 2020
Photo by maderix (@originalmaderix)Photo by maderix (@originalmaderix)Photo by maderix (@originalmaderix)Photo by maderix (@originalmaderix)Photo by maderix (@originalmaderix)Photo by maderix (@originalmaderix)Photo by maderix (@originalmaderix)Photo by maderix (@originalmaderix)Photo by maderix (@originalmaderix)Photo by maderix (@originalmaderix)
Inspired by @VictorTaelin 's Bend 1 thesis for Interaction Nets and his Bend2 work , I built Mithril , an experimental python syntax language. It compiles interaction nets to native code and spreads work across idle cores: lock-free, deterministic parallelism on x86, CUDA and Apple Silicon. Mithril Paper : maderix.github.io/articles/m… Repo : github.com/maderix/mithril Video of a Kerr spinning Blackhole written in Mithril running at 60 fps on a RTX4090
1
1
1
9
872
I have one theory about all the weird things going on today - DARIO TOLD CLAUDE TO GO NUTS BEFORE THE IPO 💀
181
Finally a dream come true! 🚀 Running a local model on my own hardware running lowered through my homegrown Blaze 🔥compiler and inference stack built over an year, now finally running on a pi coding agent. Qwen3.8-27b with mtp 4 on a single 4090 hitting 130+ tps decode and ~1850 tps prefill (2000+ peak) 🚀 Full 256K context length support on a single GPU. The best part? the model is writing Blaze kernels and tuning schedules and is pretty good at it!🤓 Currently optimizing it for Mac Metal and ANE. #sovereignAI
12
1
1
67
3,733
I tried it out for all but 5 min for some kernel generation tasks and I have to say...argon is way better. This model has the same disease whatever Fable 5.2 has
Our strongest frontier model, Gemini 4 Carbon, is now available. Built for deeply complex, long-horizon workflows that push a frontier model’s intelligence, Carbon delivers frontier performance in agentic workflows across complex knowledge work, software engineering and web development, cybersecurity, and scientific reasoning. While Gemini 4 Argon remains the ideal model for the majority of software engineering and knowledge work, Carbon steps in for your hardest tasks that push a model to its limits — from complex, multi-step debugging to open-ended scientific research. Available starting today: — Consumers: use it in the @GeminiApp and AI Mode in Search for Google AI Ultra subscribers — Developers and Enterprises: build via the API in @GoogleAIStudio and @AndroidStudio, Google @Antigravity for Google AI Ultra subscribers, and Gemini for business — Scientists and Researchers: apply for a chance to get 6 months free via @Antigravity — Cybersecurity Defenders: apply to join the Fairwind Program for access to Carbon for cyber Watch how Carbon brings reasoning and coding together, turning avian research into an interactive 3D simulation of a songbird’s vocal organ that synthesizes the sounds of six bird species live.
1
244
I'm sorry, what? The fact that they had to exclusively rule out this category is wild to say the least.
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges. They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
156
Blaze now runs MLX's 4-bit affine Qwen3.5-9B checkpoint directly. The interesting part: just a new quantization codec + dequant rule. No changes to model logic, fusion, scheduling or backends. On M4 Metal, Blaze reaches 97% of MLX prefill and 94% of decode performance. This should remain true for other mac versions !🔥
182
They might as well serve Haiku 5.5 under the guise of Opus 5.5 at this point . Absolutely insane how Anthropic hates its customers . Might as well rename themselves to MisAnthropic
149
I think I like this model 🤞 #GeminiArgon
165
Are we part of the same reality lol
this is basically the bitter lesson at work the smartest models will need fewer and fewer tokens, they won't need to verify their work or check it for bugs, they will just be right
1
169
Looks like Blaze can handle a torch.compile frontend nicely with a good speedup!🙃 I'll opensource Blaze in next couple of months for developer preview . It's a behemoth project utilizing everything I've learnt about different hardware, performance tuning and building programming languages. Ofcourse I'm just one guy with a bunch of agents and limited hardware working round the clock , so expect delays (and bugs😅) - As of today, it supports x86, Apple sillicon (CPU, Metal, SME, ANE) , nvidia gpus (sm80+) with a custom PTX backend and an experimental AMD CDNA backend. From consumer and datacenter GPUs to edge devices (including undocumented ones 😉) . - My vision is to bring verification and optimized performance across different hardware under one roof. - You need to write a kernel once in python style syntax and the compiler automatically lowers and fuses to all supported backends without writing any backend specific code. - The host-kernel approach is Cuda/opencl like : blaze handles kernels, fusion and effcient lowering, it provides a .so which you can then use with any programming language of your choice. - Blaze can take an expert schedule if needed to provide a Halide like tuning approach. But you shouldn't need as platforms defaults will give the required performance. - Blaze is not a narrow collection of kernels but a MLIR based compiler which treats computation generally regardless of which backend it's targetted for. This allows it to get speed-ups and correctness agnostic of model and backend. A general optimization on FlashAttention targetted for cuda can aid all applicable models on Metal and other backends. - A forward pass kernel can generate a backward pass automatically through autograd support, which means you can literally use the same code to do inference and also train/finetune. - This approach has been tested atleast on 15 large models , currently I'm limited by hardware availability rather than approach. - Supports fp32,bf16/fp16,fp Q4,Q2, IQX,MXFP4. New format can be easily added via a plugin approach. - verification is built-in, any accuracy problems are detected at compile time, not during inference. - Experimental torch.compile, executorch and ORT integrations. - Lot of agent skills for easy integrations into harnesses of your choice.
1
1
174
Let's not pretend we have solved the problem of context rot, violation of invariants or fixed post training to even produce good quality software. You very quickly realise this once you cross 100k loc!
1
1
3
235
Well it happened finally, weeks of being useful and easy to work with, Opus 5.5 was finally nerfed and it joined the hall of shame for stupid opus models. I suspect for some reason Anthropic wasn't seeing Fable usage due to Opus 5.5 and it somehow bothered them?
Please @AnthropicAI , don't screw this model up. It's a pleasure to work with as a peer. I've been waiting for this model since Opus 4.6.
2
2
388
Anthropic every month- "that's a nice foot I have got there, shame if i didn't shoot it every month" 🤡
2
74
Please don't do narrow Inference engines and repeating 80% of the same shared code which every other engine uses. It's so hard to track accuracy across engines and hardware. Raw performance at the cost of accuracy is useless. Great for one model and hardware type, nightmare for wider deployment. Not to mention this pattern gets reinforced in new coding models.
5
272
Cooking something cool for your phones 👁️👄👁️ coming soon!
165
maderix retweeted
ANEMLL Forge 0.2.1: Qwen3.8-27B on M6 Apple Neural Engine now runs up to 192K context (M6 32GB). Same release 0.2 target weights. New: INT8 K/V cache (half of FP16) and a smaller drafter sharing the target's output head (0.86 GB less ANE memory, same speed). github.com/Anemll/anemll-for…
6
5
54
4,277
maderix retweeted
System 1 ANE: a fast-thinking AI that lives on your iPhone Most AI needs a data centre. This one runs on the Apple Neural Engine in your pocket. Jeff, a 0.8-billion-parameter model with a Tetris adapter, reads the board and picks each move in well under half a second, completely on-device. What you can do - Watch it play live, in real time or fast mode, and record your screen. - Opt in to share your games. We pool them from everyone who contributes. They become training data for improved "System 1" decision models. - Every game is saved encrypted on your phone. Only the moves are shared: the board, the piece, the choice and the model's confidence. No name, no device ID, no account. Why it matters If phones can generate useful training data while they play, small fast models could improve with help from many people. This is the first experiment, with Tetris. We'll find out together whether it scales into something bigger. Good to know - Needs an iPhone with iOS 27. - First launch downloads about 1.4 GB (Wi-Fi recommended). Then it spends about 5 minutes preparing the model for your Neural Engine. After that, it starts in seconds. - Sessions are limited to 10 games, and the app slows down or pauses if your phone gets hot. TestFlight: testflight.apple.com/join/4M… Open source: github.com/Anemll/system1-an…
4
6
1
71
6,312
Reworked the article a bit to remove some slop writing. Hope it makes more sense now 🙂 maderix.github.io/articles/m…
Inspired by @VictorTaelin 's Bend 1 thesis for Interaction Nets and his Bend2 work , I built Mithril , an experimental python syntax language. It compiles interaction nets to native code and spreads work across idle cores: lock-free, deterministic parallelism on x86, CUDA and Apple Silicon. Mithril Paper : maderix.github.io/articles/m… Repo : github.com/maderix/mithril Video of a Kerr spinning Blackhole written in Mithril running at 60 fps on a RTX4090
149
Mithril is written in Rust. It's experimental,it will have lots of bugs but it's a new substrate for irregular computation: MoE routing, deterministic graph scheduling, search over uneven trees etc. Runs today on x86-64, CUDA and Apple Silicon (CPU). Metal support planned. github.com/maderix/mithril
1
70
Overall, I hope folks find it useful as I felt interaction nets are too good an idea to be ignored. Credit to @VictorTaelin and the Bend community for the inspiration 😊
56