BlogWilliam Liu
Essays, paper reviews, and notes on language models, coding agents, and the engineering practice around them.
Essay6 minInteractive
I labelled 168 clips of my Mandarin text-to-speech podcast by ear and ran 11 open speech-recognition models against them. Qwen3-ASR-1.7B was the most accurate, but the most accurate models sometimes silently fixed my AI voice's mispronunciations. What that means for choosing a quality checker, which second-pass model I am switching to, and what no transcript can catch.
Essay7 min
By the end of the release my agents answered consent prompts and released checkpoints. The three decisions I kept, what cross-vendor review was worth, and my real prompts graded.
Essay6 min
Thirteen launches, zero bugs in the release itself, one clean switch: the five mechanisms that made an agent-driven release fail safe, with code you can copy.
Essay6 min
An implementer model and a cross-vendor reviewer still let three real Upstash limits reach production. Why more agents add opinions but not ground truth, and the rehearsal that fixed it.
Essay7 min
The week before a careful migration, my agents broke production twice — once through a push I approved by clicking 'Recommended'. My real prompts, graded, and seven rules for organising agents.
Essay10 min
Part 6: finishing and launching the redesign brought the total to 6.26 billion tokens across Codex and Claude Code, $3,691 at API prices. 97% were cached context read again. The breakdown by kind, phase, role and detour, how I nearly undercounted it by 2 billion, and what the last four days added.
Essay11 min
Part 5: the redesign stayed off production until it was ready, by design. Getting it onto a staging URL took six failed deploys, a suspect CPU, an upload step that ate more than 50 GB of memory and a hidden dependency on private files. Then staging publishes filled a database production shares, and production writes stopped for about two hours.
Essay11 min
Part 4: I sent the agents 162 messages over fifteen days. Twenty were corrections, and they fall into two groups: five visual or product judgments, fifteen about how the agents were working. Neither group was caught by the checks in place.
Essay10 min
Part 3: one long-running Codex goal ran 333 agents with a three-role review loop and built the redesign in five days, then stalled at deployment. Claude Code took over as the orchestrator, sending fresh Codex runs to implement and review the fixes and making small corrections itself. What each mode was good at here, with numbers.
Essay10 min
Part 2: a design prototype carries taste, sample data and its own opinions about who wins a conflict. Here is how I turned mine into a spec agents could build against, and why every gate passing still left 20 places that didn't match.
Essay10 min
Part 1 of a six-part series on rebuilding williamliu.ai with Codex and Claude Code: 683 agent runs, 6.26 billion tokens, 20 human corrections, and the five detours that cost the most.
Essay8 min
The redesign puts the blog, podcasts and projects side by side, adds a Chinese twin of every page, lets you play episodes without leaving the list, and keeps working without JavaScript.
Essay5 min
Measurements of file reads and answer completeness show why token-saving hooks should follow workload evidence rather than a headline benchmark.
Essay9 min
Two coding-agent sessions responded differently to the same crash evidence, showing how diagnostic judgment and a clear human handoff can outweigh elaborate orchestration.
Essay7 min
The Autonomous Individual combines agents with initiative, judgment, and standards, turning independent execution into a hiring criterion beyond traditional role boundaries.
Essay4 min
fidx combines local keyword and embedding search without an LLM in the query path, reporting latency, retrieval quality, and limitations against QMD.
Essay3 min
AI coding tools compete on how well they encode engineering judgment, including evidence gathering, verification, boundaries, and review, rather than agent count.
Essay2 min
Claude Code’s dynamic workflows move orchestration into inspectable scripts, making repeatable processes more practical for audits, migrations, and other complex engineering work.
Essay7 min
Cloudflare’s layoffs prompt a distinction between creating value and routing information, with judgment and accountability determining which coordination work remains useful.
Essay7 min
Richer HTML artifacts can increase generation and revision costs, making measured token accounting essential before treating more compute as the solution.
Essay5 min
A controlled comparison of identical fixtures reports higher input-token counts for Opus 4.7, separating input accounting from output efficiency and task quality.
Essay7 min
Useful AI evaluation compares versions under consistent conditions, investigates failure patterns, and continually strengthens tasks, harnesses, and acceptance criteria.
Essay7 min
Tokenizer changes alter context capacity, serving costs, and model behavior, making Opus 4.7 a migration that requires measuring complete workflows.
Essay8 min
Vertical product teams and infrastructure teams can reduce handoffs when stable interfaces preserve independent ownership, allowing AI tools to amplify delivery.
Essay10 min
Agents that preserve human judgment complicate ownership and accountability, making permissions, escalation, auditability, and responsibility part of the design.
Essay4 min
Automating junior knowledge work may remove the apprenticeship tasks that develop expertise, forcing organizations to rethink how future experts gain judgment.
Essay6 min
Encoding expert judgment into repeatable agent workflows turns individual expertise into organizational infrastructure, while challenging how future experts learn.
Essay6 min
Agents that preserve memory, workflows, and preferences begin to carry delegated judgment, raising questions about representation, control, and human agency.
Essay2 min
Vibe coding lowers the cost of prototypes and experiments, but production reliability requires a more deliberate engineering approach than accepting generated code.