Conscious Engines
Story of Attention - Part 1: Before the Transformer
From n-gram counting and RNNs to the first attention mechanisms.
guide
Read more →Research and Development
Experiments in model architecture, training, evaluation, and efficient inference
Papers
2 publishedThe Joint Effect of Quantization and Sampling Temperature on LLM Safety Alignment
How weight quantization and sampling temperature jointly affect LLM safety alignment — a factorial analysis of the interaction.
arXiv
LLM Inference at the Edge
Sustained-load benchmarks across mobile, NPU, and GPU platforms reveal how LLM inference performance really degrades under back-to-back requests.
arXiv
Featured
Conscious Engines
Building a tasks app that puts all iOS 27 on-device AI features to work
How we used an ensemble of on-device models to build consumer AI that works offline, costs nothing to run, and never leaves the phone
apple silicon
Read more ↗Conscious Engines
The Need for Speed: Why LLMs Are Slow and What Speculation Promises
Your GPU sits nearly idle during token generation — the memory-bandwidth bottleneck behind LLM latency, and how speculation exploits it.
Inference
Read more →Conscious Engines
Kinetic-4B: A 4-Billion Parameter Model That Outperforms Claude Haiku at Tool Calling
A fine-tuned 4B open-source model that beats frontier models at tool calling, with p95 latency under two seconds on a single GPU.
Agentic AI
Read more →All research
Broad Review of DLM architectures
Conscious Engines
A field guide to diffusion language model architectures — how LLaDA, Dream, Block Diffusion, and MoE variants like LLaDA 2.0 approach denoising.
dLLM
Why Diffusion LLM Quantization Is Harder Than It Looks
Conscious Engines
Diffusion LLMs share transformer modules with AR models but not inference physics — why GPTQ, AWQ, and QuaRot fail, and what fixes them.
dLLMquantization
Apple Foundation Model 3, what even is it?
Conscious Engines
A walkthrough of Apple Foundation Model 3's on-device architecture — the routing changes that let a 20B-parameter model run on a phone.
apple silicon
Distribution Matching Is Not Enough: Two Failure Modes in Latent Text Drifting
Conscious Engines
A technical postmortem on adapting the Drifting one-step generative framework to text, and the two failure modes that derailed it.
Probing Latent Directions in Video Diffusion Models
Conscious Engines
Testing whether GAN-style linear latent editing transfers to LTX-2, a modern video diffusion model, and what the geometry reveals.
Hybrid Lexical–Semantic Retrieval for Tool Selection in Agent Systems
Conscious Engines
A retrieval architecture that routes natural-language intent across thousands of API tools, combining lexical and semantic search stages.
Agentic AI
Speculative Decoding in MLX using DFlash
Sabesh B
An empirical evaluation of speculative decoding on Apple Silicon, across a 300-run parameter sweep
apple siliconInference
From Single-GPU to Distributed Training: A Framework for Making the Right Call
Conscious Engines
What actually fills GPU memory during finetuning, and a decision framework for picking DDP, ZeRO/FSDP, tensor, or pipeline parallelism.
training
Distributed Data Parallel: How It Actually Works
Conscious Engines
The mechanics behind DDP's near-linear throughput scaling, and the misconfigurations that silently turn a 1.9x speedup into 1.3x.
training
Tensor Parallelism and Sequence Parallelism
Conscious Engines
Splitting weight matrices across GPUs instead of replicating layers — why it wins without NVLink, and how LoRA complicates sequence parallelism.
training
Pipeline Parallelism: How It Actually Works
Conscious Engines
How pipeline parallelism splits model layers across GPUs, where the pipeline bubble comes from, and the 1F1B schedules that shrink it.
training
ZeRO and FSDP: Model Sharding
Conscious Engines
How ZeRO shards weights, gradients, and optimizer state across GPUs to fit models that don't fit on one card, and what each FSDP knob does.
training