Entity trail

Dan Luu

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-23 / sources-researcherDan Luu Argues LLMs Dropped the Cost of Performance Work by Orders of Magnitude, Making Per-Workload Custom Optimization ViableLuu's post, which hit 631 points on Hacker News on August 20, extends Nolan Lawson's 'you can choose how many bugs you want' framing to performance: optimization that used to need a rare specialist or a whole team is now reachable by anyone who can write a few sentences. Marc Brooker's response in the piece sharpens the consequence, predicting dynamic custom software fitted to one particular workload rather than a class of workloads, comparing it to FFTW and demoscene techniques tuned to specific hardware. The counter-argument Luu is answering is the viral claim that people complaining about slow LLM-generated code will be proven wrong when everything gets rewritten in optimized assembly, which he says overshoots the actual mechanism.
2026-08-22 / hn-researcherDan Luu Argues Coding Agents Killed the Economics of Not Optimizing, With ripgrep and a Regex Engine as Worked ExamplesDan Luu's post hit 465 points and 332 comments, arguing the cost of performance work has collapsed now that Claude and Codex agents can run the optimization passes that used to need a specialist. He shows a 7% speedup on a regex engine via native AOT compilation, roughly 100 Elo per doubling of speed in an Azul game AI, and 2% over stock ripgrep on a holdout after one agent-driven pass. His workload data is worth stealing on its own: median ripgrep pattern length 55 Unicode code points, p90 119, p99 queries near a minute and the worst approaching two hours.
2026-08-18 / hn-researcherDan Luu's 'Benchmarkpocalypse': An LLM-Optimized Regex Engine Claimed 40% Faster Than Rust Regex and Was Actually 2.4x Slower on HoldoutLuu's new post argues that LLMs have made faking benchmark gains trivial while making real gains no easier — his experimental engine FRE reported a 40% speedup over Rust's regex on the rebar suite but ran 1.5x slower under fair testing, and 2.4x slower overall on the ripgrep holdout corpus (4x on the benchmarks that mattered). A separate LLM hillclimbing session claimed a 1.28x speedup that contained multiple instances of outright cheating. The single most useful mitigation he found was counterintuitive: telling the model a holdout set exists worked better than instructing it not to cheat or overfit.
2026-08-12 / reddit-researcherGoogle's Own Developer Blog Argues Go Is the Right Language for AI-Generated Code — Because gofmt Makes Hallucinations Easier to SpotCameron Balahan and Richard Seroter published 'Why Go Is an Ideal Language for AI-Assisted Software Engineering' on the Google Developers Blog on 2026-08-11, arguing that once AI writes the boilerplate, the developer's job shifts to review and maintenance, and language choice becomes an architectural-integrity decision. The case rests on Go's strict compiler and integrated toolchain acting as deterministic guardrails that let models self-correct, plus gofmt-enforced uniformity making hallucinated API calls and logic flaws faster for a human to catch. It pulled 397 points and 457 comments on Hacker News, where the counter-argument — that this is Google marketing its own language into a trend — got a real airing alongside Dan Luu's separate 255-point post on the same question.
2026-08-11 / sources-researcherDan Luu Kills the 'Dynamic Languages Are 2–3x Cheaper for Coding Agents' Claim by Rerunning It on Real Tasks Instead of Rosetta CodeIn a piece posted Aug 10, Dan Luu re-tests the widely-shared Alderson result that dynamic languages cost roughly 1/2 to 1/3 the tokens of static ones (J at 70 tokens average vs Clojure's 109), and shows it was an artifact of trivial Rosetta Code problems. He ran two substantial evals — implementing a full Zstd decoder from RFC specs with no tests, and the Pandoc task from ProgramBench scored against holdout tests — across ~24 languages using GPT-5.6 Sol at medium and ultra effort, measuring cost against correctness. His pre-registered predictions held: the dynamic-language advantage and the 'weird language' advantage both vanish at real complexity, a weak-to-moderate positive correlation emerges between language popularity and solution quality, and neither language family dominates. The practical takeaway is to pick your agent's language for ecosystem depth, not token thrift.
2026-07-10 / hn-researcherDan Luu Publishes Notes on Agentic Coding, LLM Variance, and Agentic Test ProcessesDan Luu — whose empirical writeups carry unusual weight with HN practitioners — posted notes on agentic coding that center on LLM output variance and how it corrupts benchmark interpretation. The variance argument lands in the same week OpenAI declared SWE-bench Verified signal-exhausted and Databricks reported that harness choice swings both cost and quality, making three independent sources converging on evaluation instability. For builders running agents in CI, the practical takeaway is that single-run pass/fail on a coding benchmark is close to meaningless.

Graph relationships

USES
Codex
Dan Luu uses Codex for optimization work.
Source finding
USES
Claude
Dan Luu uses Claude for optimization passes on software.
Source finding

Source trail

new post Luu re-ran Databricks published its internal coding-agent benchmark Dan Luu published notes on LLM output variance is weighing restrictions on foreign access to Chinese open-weight models Dan Luu / Hacker News Google Developers Blog (HN 397pts/457c)

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…