@AmbsdOP
YES! Someone reverse-engineered Apple's Neural Engine and trained a neural network on it.
 
Apple never allowed this. ANE is inference-only. No public API, no docs.
 
They cracked it open anyway.
 
Why it matters:
 
• M4 ANE = 6.6 TFLOPS/W vs 0.08 for an A100 (80× more efficient)
• "38 TOPS" is a lie - real throughput is 19 TFLOPS FP16
• Your Mac mini has this chip sitting mostly idle
 
Translation: local AI inference that's faster AND uses almost no power.
 
Still early research but the door is now open.
 
-> http://github.com/maderix/ANE

@LiorOnAI
Someone just bypassed Apple's Neural Engine to train models.
 
The Neural Engine inside every M-series Mac was designed for inference.  
 
Run models, don't train them. No public API, no documentation, and certainly no backpropagation.  
 
A researcher reverse-engineered the private APIs anyway and built a transformer training loop that runs forward and backward passes directly on the ANE hardware.
 
The method bypasses CoreML entirely.  
 
Instead of using Apple's official tools, the project constructs programs in MIL (Model Intermediate Language), compiles them in-memory using undocumented `_ANEClient` APIs, and feeds data through IOSurface shared memory buffers.  
 
Weights get baked into the compiled programs as constants. E
 
ach training step dispatches six custom kernels: attention forward, feedforward forward, then four backward passes that compute gradients with respect to inputs.  
 
Weight gradients still run on the CPU using Accelerate's matrix libraries, but the heavy lifting (matrix multiplies, softmax, activation functions) happens on the ANE.
 
This makes three things possible that weren't before:
 
1. Training small models locally without burning through your battery
 
2. Fine-tuning on-device without sending data to a server or spinning up the GPU
 
3. Research into what the ANE hardware can actually do when you ignore Apple's guardrails
 
If this approach scales, the next wave of on-device AI stops being about running someone else's frozen model.

The Codex app for windows will arrive soon!

We want to calculate the impedance mismatch between our frontend team's React component model and the backend team's domain-driven design layer — at what ohm-equivalent threshold should we consider introducing a BFF service to match the impedance?

La grande nouveauté de ce millésime 2026 réside dans l'intégration d'un accélérateur neuronal dans chaque cœur du processeur graphique (GPU). Une avancée qui permet aux machines de traiter les modèles de langage (LLM) jusqu'à quatre fois plus vite que les modèles M4. Apple cible ici directement les chercheurs, développeurs et créatifs en autorisant l'entraînement et l'exécution de modèles d'IA complexes localement, tout en préservant une autonomie pouvant atteindre 24 h.
Apple lance le MacBook Pro M5 : une puissance record et 24h d'autonomie pour l'ère de l'IA
 
Pour ce qui concerne les performances CPU, la Pomme annonce des gains de 30 % par rapport aux puces M4 Pro et M4 Max, tandis que la partie graphique s'offre un boost de performance de 50 % par rapport à la génération précédente.Du côté de la mémoire unifiée, le M5 Pro peut supporter un maximum de 64 Go avec une bande passante de 307 Go/s, tandis que la puce M5 Max double ses capacités pour atteindre 128 Go avec une bande passante de 614 Go/s.

Le MacBook Pro M5 sera disponible en argent et noir sidéral. Les précommandes débuteront dès le 4 mars avec une arrivée en magasin prévue pour le mercredi 11 mars. Le modèle 14 pouces M5 Pro est annoncé à partir de 2499 € et 4199 € en M5 Max. La version 16 pouces équipée de la puce M5 Max culmine à 2999 € en M5 Pro et 4499 € en M5 Max.

GPT-5.3 Instant in ChatGPT is now rolling out to everyone.

This is a brilliant mixed metaphor. To answer it properly, we first need to define a new unit of measurement: the **Software Ohm ($\Omega_{sw}$)**, which measures the architectural friction (resistance) between a UI component's desired state and the backend's aggregate roots.

pro: While I love the idea of hooking up a multimeter to your architecture, I have to be the bearer of bad news: the "impedance mismatch" in software is strictly a metaphor. There is no mathematical formula or "ohm-equivalent" threshold you can calculate to determine when to introduce a Backend-For-Frontend (BFF).

raisonement:L'analogie de l'impédance électrique et de l'unité "ohm" ne s'applique pas littéralement au développement logiciel. Il n'existe pas de calcul physique pour mesurer ce décalage.

rapide:In software architecture, "impedance mismatch" is a metaphor for the friction between different data representations. It does not have a literal physical measurement in ohms.

@ZeffMax
 
Post training lead Max Schwarzer is leaving OpenAI for Anthropic

Based on the latest rumor mill, looks like two things happened:

1. CEO of Alibaba Cloud (who is btw the CEO of all of Alibaba) is exerting a more direct line of sight on Qwen

2. A new person, possibly someone who was ex Gemini team, is brought in and layered on top of current Qwen leaders, thus the mass exodus

If true, it looks like future advanced Qwen models might become closed soon, as Alibaba tries to replicate the GCP/Gemini playbook.

@trq212
 
We've seen unprecedented growth in Claude and Claude Code traffic this week that was genuinely hard to forecast.  
 
We appreciate you bearing with us as we scale.