TLDR Dev 2026-11-19
Gemini 4 Neon π, CSS randomness π², better code reviews π
π§βπ»
Articles & Tutorials
The Magic of Partial Evaluation (12 minute read)
Partial evaluation executes the parts of a program whose inputs are already known and leaves specialized code for the remaining work. This post has examples that build from arithmetic and schema validation to the Futamura projections, showing how specializing an interpreter can produce compiled programs and even a compiler.
Understanding CSS random() (15 minute read)
CSS random() can give elements distinct values that stay consistent across breakpoints and animations. This worked example explains how cache keys, property positions, and explicit scopes determine whether random values are shared, regenerated, or fixed for debugging.
Sashiko: Lessons from Building an Agentic Code Review System for Linux (10 minute read)
Sashiko separates specialist bug discovery from verification, requiring code evidence for both findings and dismissals. Its review pipeline gathers surrounding context, investigates uncertain cases individually, and publishes its reasoning, with production data showing a correlation between clean reviews and fewer follow-up fixes.
We Stopped a Railway Migration Three Weeks Out (10 minute read)
A production migration was canceled after database import times and operational constraints made the cutover too difficult to predict. The retrospective argues for pricing recovery, observability, storage limits, and maintenance work alongside compute, while noting that its import runs were not a controlled within-tier benchmark.
The Codebase Is the Harness (2 minute read)
Agent tooling, instructions, review processes, and product code can evolve together inside one repository. One workflow uses session recordings to diagnose user-facing bugs and then updates instructions after fixes, with a person still setting the broader direction.
Traffic Mix Matters (2 minute read)
A load test can miss production failures when it reproduces request volume without reproducing workload distribution. Interactive simulations show key concentration reducing throughput in a lock-based system but improving it in a cached system, so the same skew can move performance in opposite directions.
Build software factories that work for humans with LaunchDarkly Release Autopilot (Sponsor)
Most AI software factories end at deploy, and when the automation stops, the risk begins. Release Autopilot helps eliminate that risk with new automation capabilities built to fit the factory you already have, automating releases within the rules engineers set so AI-written code reaches customers faster and safer.
Learn how
Mastra Factory (Website)
Mastra Factory is an open-source software factory that turns repository issues into plans, implementations, and reviewed pull requests. Teams configure the stages, skills, and approvals, with options to use Mastra's platform, run the server locally, or self-host the supporting infrastructure.
Search Browser (Website)
Search is a free, open-source Mac browser built on the WebKit engine already supplied by macOS. It combines spaces, pinned tabs, reading mode, and built-in blocking. The Chrome Web Store extension requires macOS 15.4 or later.
Become a founding member of TLDR's AI community (2 minute form)
TLDR is launching a curated community for AI builders to discuss major news and launches. The best links and comments will be featured in TLDR AI.
Apply here to join.
Introducing Gemini 4 Neon (10 minute read)
Gemini 4 Neon targets faster end-to-end coding, analytical, and agent workflows with a one-million-token context window and four thinking levels. Google prices it at $3 per million input tokens and $15 per million output tokens, with rollout through the Gemini API, enterprise products, Antigravity, and paid Gemini app plans.
Advancing Chip Design with OpenAI Models (5 minute read)
OpenAI reports model-written RTL improvements across its JalapeΓ±o chip and AMD designs, including up to 14% improvement in power, performance, and area relative to heavily optimized GPU baselines. Separate RISC-V evaluations found stronger results when models were designed from specifications, though omitted features and simplified timing checks limit how directly those scores transfer to production.
LLM-Assisted Reverse Engineering (7 minute read)
This reverse-engineering workflow combines agents with Ghidra, command-line disassemblers, runtime debugging, and persistent notes. When static analysis stalls, it shifts to one recorded hypothesis at a time, using probes or logs to prove or disprove explanations instead of letting the agent keep guessing.
Why Is Gemini 4 Carbon Taking So Long? (5 minute read)
Traces from a private bug-fixing evaluation attribute Carbon's long runtimes to slower inference, unnecessary investigation, and repeated file reads. About half its steps addressed work outside the tickets, while periodic check-ins shortened one trial from 37 to 28 minutes, suggesting that stopping behavior contributes to the gap.
The most important software engineering news in one daily email
Join 480,000 readers for
one daily email