# Amund Tveit's blog

> Personal blog and technical articles by Amund Tveit (Principal AI Engineer, PhD in Computer Science).

- [Full text of all posts (llms-full.txt)](https://amund.blog/llms-full.txt)
- [RSS Feed](https://amund.blog/rss.xml)
- [Sitemap](https://amund.blog/sitemap-index.xml)

## Posts

- [What formal methods can (and can't) prove about a Gemma 4 + DFlash + DDTree blueprint?](https://amund.blog/g4-flashtree-formal-verification/) (2026-05-01): An audited-before-publish formal-methods write-up of a hypothetical G4-FlashTree speculative-decoding stack on Apple Silicon. Tier-classified evidence: 4 mechanically rechecked Lean theorems, 4 TLA+ properties under bounded TLC (with a load-bearing safe-vs-unsafe counter-example pair), 3 ABS behavioral models, and 9 deferred findings.
- [Can Prompt Lookup Decoding Speed Up Gemma 4 31B 4-bit on Apple Silicon for RAG-style Workloads?](https://amund.blog/accelerating-gemma4-31b-pld-mlx-lm/) (2026-05-01): An early investigation of Prompt Lookup Decoding on Gemma 4 31B 4-bit on M3 Ultra under mlx-lm 0.31.3, with a focus on RAG-style workloads (classic doc RAG, structured / JSON RAG, agentic RAG, code RAG). One class (structured / JSON RAG) shows a +54% decode lift with greedy parity holding at two PLD configurations. A second (passage / classic doc RAG) holds parity at one configuration and breaks at another — fragile evidence. A third (code RAG) breaks parity at every tested configuration. Numbers, RAG-applicability discussion, caveats, and an open implementation question included.
- [Exp minus Log is all you need for Deep Learning?](https://amund.blog/exp-minus-log-is-all-you-need/) (2026-04-21): From emulation to native representation. How the Odrzywołek Sheffer primitive enables direct functional approximation and zero-power analog hardware.
- [Mojo-tokenizer: Fastest AI Token Output → Readable Text on Apple Silicon?](https://amund.blog/fastest-ai-token-decoder-mojo-apple-silicon/) (2026-01-09): mojo-tokenizer converts AI output to readable text at 144 million tokens per second — 3.1x faster than the industry standard. 100% compatible, rigorously benchmarked.
- [mojo-vec: Zero-Copy Buffers That Match Rust Performance](https://amund.blog/mojo-vec-zero-copy-buffers/) (2026-01-08): InlineBuffer stores small data on the stack, provides zero-copy slice access via pointer arithmetic, and includes SIMD-accelerated operations. The missing piece for high-performance Mojo.
- [6.5 GB/s JSON Parsing in Mojo — Beating Rust and C++ on Apple Silicon](https://amund.blog/mojo-json-beats-rust-cpp/) (2026-01-05): mojo-json achieves 6.5 GB/s throughput, outperforming the fastest Rust and C++ parsers by 52% on standard benchmarks. Empirical results from Apple M3 Ultra.
- [Building the Enterprise Foundation Mojo Needed](https://amund.blog/mojo-enterprise-libraries/) (2026-01-02): Introducing mojo-contrib: 26 pure Mojo libraries for authentication, resilience, observability, and more. Enterprise-grade infrastructure for the high-performance language.
- [4 bits is all you need (and 3.6 bit you have?) for resource-efficient LLMs?](https://amund.blog/4-bits-is-all-you-need/) (2025-12-04): Exploring the lower bounds of LLM quantization - from OpenAIs 4-bit GPT-OSS to research suggesting 3.6 bits per parameter may be the theoretical minimum.
- [Towards milli-joules per token? AI on the Apple Watch](https://amund.blog/towards-millijoules-per-token/) (2025-11-26): Exploring ultra-efficient AI inference on wearable devices using MLX and Apple Silicon optimization.
- [Grokking implementations in Jax/Flax and Pytorch](https://amund.blog/pytorch_jax_grokking/) (2025-02-23): Grokking implementations in Jax/Flax and Pytorch