What formal methods can (and can't) prove about a Gemma 4 + DFlash + DDTree blueprint?
Summary The G4-FlashTree stack — Gemma 4 31B as the target, a DFlash block drafter, and a DDTree tree verifier — is the kind of system where…
Gatsby Starter Blog RSS Feed
Amund Tveit's blog
Latest posts
Can Prompt Lookup Decoding Speed Up Gemma 4 31B 4-bit on Apple Silicon for RAG-style Workloads?
Summary This is an early investigation, not a victory lap. The headline finding from a controlled rerun (3-class bench, 800-token prompts,…
Exp minus Log is all you need for Deep Learning?
Note: This work builds on the 2026 discovery by Dr. Andrzej Odrzywołek (Institute of Theoretical Physics, Jagiellonian University, Kraków…
Mojo-tokenizer: Fastest AI Token Output → Readable Text on Apple Silicon?
mojo-tokenizer decodes tokens at 144M tokens/sec on Apple Silicon — 3.1x faster than tiktoken and 1.2x faster than rs-bpe. We benchmarked…
mojo-vec: Zero-Copy Buffers That Match Rust Performance
mojo-vec brings Rust-style zero-copy buffer performance to Mojo — inline storage, pointer-based slicing, and SIMD operations. Mojo's heap…
6.5 GB/s JSON Parsing in Mojo — Beating Rust and C++ on Apple Silicon
mojo-json parses JSON at 6.5 GB/s on Apple Silicon — 52% faster than the fastest C++ and Rust parsers. We benchmarked 6 JSON parsers across…
Building the Enterprise Foundation Mojo Needed
Mojo has speed. Now it has an enterprise toolkit. When Modular launched Mojo, they solved a fundamental problem: Python's beloved syntax…
4 bits is all you need (and 3.6 bit you have?) for resource-efficient LLMs?
A few months ago OpenAI published their open weights model(s) GPT-OSS (20B and 120B), and one of the eye-catching characteristics was that…
Towards milli-joules per token? AI on the Apple Watch
Can an Apple Watch run a (small) LLM (SLM)? Yes, it seems so. The Apple Watch Ultra has an S9 or S10 System on a Chip (SoC) 64 bit multicore…
Grokking implementations in Jax/Flax and Pytorch
https://github.com/atveit/torch_grokking https://github.com/atveit/jax_grokking What is Grokking in Deep Learning? Grokking is the phenomena…