Reid Marlow's personal space
“We may have knowledge of the past but cannot control it; we may control the future but have no knowledge of it.”
whoami
I'm Reid Marlow, a technologist working in automation - currently doing PhD research in the field at the Hong Kong Polytechnic University. Here I write up what I actually learn building tools, engineering AI agents, and keeping systems running, along with the workflows and habits that stick. The best is yet to come. We've only scratched the surface.
This is a field notebook for the unglamorous layer: parsing messy inputs, wiring retries, and deciding which agent workflows are worth keeping.
- born
- 2001-05-12
- edu
- PhD Automation, HK PolyU
- pronouns
- he/him
- mode
- field notes
Latest
Why 80% on OSWorld Required VM Snapshot Rollbacks
Computer-use agents fail when corrupted OS state poisons training rollouts. CUA-Master hit 80.2 on OSWorld Verified by turning VM rollbacks into a branching search.
Blog
archive ->Why Standard bf16 Rounding Discards 99% of RL Policy Updates
In modern RL post-training pipelines, converting fp32 optimizer weights to bf16 rollouts via round-to-nearest silently zeros out small updates, leaving the policy engine a hundred steps behind.
Claude's Watermark Disappears When You Tune Temperature
Anthropic retrofitted watermarks onto older Claude models for EU compliance, but changing a single sampling parameter quietly bypasses the injection.
Why DeepSeek Built Its Coding Agent Around Attention Budgets
DeepSeek open-sourced dsh-superagent-prototype. Here is what happens when you design an autonomous coding agent around human attention cost instead of benchmark scores.
Why Nvidia Put Mamba and Attention Together for Agents
Nemotron 3.5 Super uses a 120B-total, 12B-active hybrid Mamba-Transformer architecture. Here is why multi-turn agent loops break pure Attention KV caches and why prefix reuse on recurrent states is tricky.
Why Google Gated Gemini 4 Carbon Behind Defensive Cohorts
Gemini 4 Carbon hits 54.8% on ExploitGym and 0.0% cheating on Terminal-Bench, but Google withheld public API access. Here is why exploit synthesis creates an immediate deployment asymmetry.
Tools
KolmoPDF is the daily driver; the rest are useful adjacent picks.
KolmoPDF
Most of what I automate starts by getting clean text out of a PDF, and ordinary parsers fall apart the moment a page has two columns, a formula, or a table that runs across the page break. KolmoPDF is the one I reach for: VLM-based parsing that keeps formulas, tables, code blocks, and multi-column order intact, layout-preserving translation when the source isn't in English, and an API clean enough to wire straight into an agent or a knowledge base. It runs the other direction too - Markdown back out to DOCX, HTML, LaTeX, or PDF.
When the bottleneck is the keyboard, not the idea, I switch to Typeless. Speak naturally and it drops polished text into whatever app is focused — messages, notes, editors — with filler words gone and punctuation already in place. Not a full writing stack, just a faster way to get the first draft out of my head.
typeless.com ->