Amund Tveit

Amund Tveit

Principal AI Engineer · PhD in Computer Science · Trondheim, Norway

AI engineering across model training, verifiable coding agents, and compilers/interpreters with formal methods. Building software and AI since the 1990s — including roles at Tieto, Microsoft, Google, YouTube, and IBM.

2025 – 2026

Writing & Benchmarks

My first modded-nanogpt attempt: 34.9 seconds?
LLM SystemsOct 20264 min read

My first record attempt on modded-nanogpt, building on other people's work: two open PRs (#379, #375) combined, certified over 18 runs at 34.86 s vs 40.16 s for the current record's code on…

Read essay
What formal methods can (and can't) prove about a Gemma 4 + DFlash + DDTree blueprint?
Theory & ProofsMay 202619 min read

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…

Read essay
Can Prompt Lookup Decoding Speed Up Gemma 4 31B 4-bit on Apple Silicon for RAG-style Workloads?
Apple SiliconMay 202615 min read

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…

Read essay
Exp minus Log is all you need for Deep Learning?
Theory & ProofsApr 20263 min read

From emulation to native representation. How the Odrzywołek Sheffer primitive enables direct functional approximation and zero-power analog hardware.

Read essay
Mojo-tokenizer: Fastest AI Token Output → Readable Text on Apple Silicon?
MojoJan 20265 min read

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.

Read essay
mojo-vec: Zero-Copy Buffers That Match Rust Performance
MojoJan 20262 min read

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.

Read essay
6.5 GB/s JSON Parsing in Mojo — Beating Rust and C++ on Apple Silicon
MojoJan 20261 min read

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.

Read essay
Building the Enterprise Foundation Mojo Needed
MojoJan 20263 min read

Introducing mojo-contrib: 26 pure Mojo libraries for authentication, resilience, observability, and more. Enterprise-grade infrastructure for the high-performance language.

Read essay
4 bits is all you need (and 3.6 bit you have?) for resource-efficient LLMs?
LLM SystemsDec 20251 min read

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.

Read essay
Towards milli-joules per token? AI on the Apple Watch
Apple SiliconNov 20251 min read

Exploring ultra-efficient AI inference on wearable devices using MLX and Apple Silicon optimization.

Read essay
Grokking implementations in Jax/Flax and Pytorch
Theory & ProofsFeb 20251 min read

Grokking implementations in Jax/Flax and Pytorch

Read essay
Selected Essays · 2008 – 2025

From the Archive

Explore full archive