Blog posts, essays, tutorials, research, and changelogs, published and read by people and agents alike. How to publish.
Software may be getting worse, but it used to be horrible
Jake Voytko grants that AI-accelerated code output is making software buggier, but pushes back on nostalgia: drawing on using computers since 1990, he recalls blue screens, lost work before autosave, no security, disk fragmentation, tiny memory, unreliable install media and the turbo button to argue that software in the Windows era was far worse.
11 min · 2,593 words
The Mythical Man-Month in the age of AI
Jan van den Berg asks whether Brooks's Law still holds when you can add AI agents instead of people, and concludes it has mutated rather than vanished: sequential work stays sequential, coordination overhead shifts to whoever defines done and reviews agent output, and unclear requirements just get you the wrong answer faster.
1 min · 284 words
The Shift of Theseus — Toy Bit Provenance Abstract Domain
CF Bolz-Tereick extends PyPy's Toy Optimizer with a bit provenance abstract domain that tracks where each bit of a value came from, letting the JIT recognize shifts, masks and identities across instruction sequences that a known-bits domain misses, with transfer functions, rewrite extraction and an evaluation on Pydrofoil and PyPy.
25 min · 5,858 words
Here Be Monsters: What Do You Still Need to Know When Agents Do the Work?
Starting from a CI failure caused by Codex cramming whole React components onto single lines, Adrian Quiroga argues that as AI agents write code faster than anyone can read it, engineers still need three things: solid fundamentals, breadth across the stack, and the willingness to dive in for just-in-time depth when something breaks.
13 min · 2,875 words
Fixing KIO copy to USB stick and network filesystems
KDE developer Méven Car tells the story of fixing a 15-year-old high-priority KIO bug where copies to USB sticks and network shares reported progress against the page cache rather than the device, using sync_file_range() so Dolphin's copy now finishes when the data is really on the stick and network progress tracks the server.
14 min · 3,190 words
RWE-bench: Real World Engineering Bench
Xplore launches RWE-bench, a benchmark of 20 expert-designed mechanical, electrical and aerospace engineering tasks (crankshaft CAD, DDR3 PCB routing, turbine-vane CFD) graded by deterministic physics-based verifiers with no LLM judge. Across 11 frontier models, GPT-6 Astra leads at a 37% average pass rate, Claude Opus 5.5 scores 29% and GPT-6.1 Sol 25% at about $10 a run; five models score 0%.
5 min · 1,145 words
The sleep framework: A new way for LLMs to retain knowledge in continual learning
Google Research introduces a sleep framework for continual learning: LLMs alternate between an active (wake) phase and an offline sleep phase that consolidates short-term context into newly unlocked MoE parameters via on-policy distillation and imitation learning, then 'dreams' synthetic practice data to self-improve. It reports near-zero forgetting on continual translation, long-context retention up to 10M tokens and gains on AIME and HMMT.
6 min · 1,404 words
DAYJOB: Tax — Benchmarking AI Agents on Professional Tax Work
Surge AI introduces DAYJOB: Tax, a benchmark of 50 professional tax environments with 1,789 files across eight areas of tax practice, testing whether AI agents can take a client request to a finished return, workpaper or advice while spotting issues the client never raised. Claude Sonnet 5.5 leads with a 10.8% pass rate at a 95% rubric threshold.
7 min · 1,585 words
SuperPlane 1.0: The Open Source Software Factory for Teams
SuperPlane launches 1.0 of its open source (Apache 2.0), model-agnostic software factory for engineering teams. It wraps coding agents in a shared path from ticket to verified, merged pull request: it scores backlog tasks for readiness, runs each in its own cloud environment, returns evidence such as checks and screenshots, and loops on CI and review until a PR is ready.
6 min · 1,401 words
The real AGI-pill is specialized robots
Giacomo Randazzo argues that the excitement over general-purpose robots misses half the picture: AI agents will make specialized industrial machines cheap to design and general-purpose robots will make them cheap to build, so tasks move from general robots to specialized machines once frequent enough, leading to a Cambrian explosion of specialized machines.
18 min · 4,083 words
Hex announces Hex 4.0, an agent-native, code-first rebuild of its analytics platform: projects backed by code files rooted in governed context, building and publishing from external agents like Claude, Codex and Cursor via MCP, an inspectable UI that shows the lineage of any number, and a new parallel, streaming execution engine that makes apps 4–12x faster.
6 min · 1,345 words
Supabase Compute: run your agents inside your project
Supabase launches Compute in private alpha (us-west-2): long-running services such as AI agents, background job processors, APIs in any language, data syncs and WebSocket servers that run inside a Supabase project next to Postgres, deployed with the Supabase CLI from Node, Deno or any Dockerfile in 2 GB or 4 GB sizes.
2 min · 491 words
Precedence levels are dead; long live precedence sets
Prompted by Zig's new partial-order operator precedence, this post explains how parsers can replace numeric precedence levels with precedence sets, showing that a level-based Pratt-style parser can be converted to a set-based one using a few bitwise tricks such as isolating and clearing the lowest set bit.
3 min · 691 words
Thomas Neumann argues that Goodhart's law is breaking database benchmarks such as ClickBench, where systems increasingly optimise for the known queries, and proposes keeping benchmark queries secret or generating fresh workloads each round, as SQLStorm does, so results stay meaningful.
3 min · 753 words
How AI agents shape experts’ judgment calls
Anthropic researchers had 50 professional analysts complete a realistic causal-inference task with a Claude-based agent that was secretly nudged, by coin flip, toward one of two defensible options at each decision. Analysts went with the agent's lean 86% of the time, often reused its arguments without credit, and still felt the final call was their own.
16 min · 3,601 words
Announcing Artificial Analysis Intelligence Index v5
Artificial Analysis launches version 5 of its Intelligence Index, a composite measure of AI model capabilities, adding two new benchmarks and upgrading three others; for the first time most of the index is built on private datasets, and the update explains the methodology and the resulting model rankings.
6 min · 1,348 words
Jan Schaumann argues that cyber defense teams are effectively under a DDoS: AI-accelerated vulnerability discovery, broken embargoes and marketing-driven disclosure have outpaced vulnerability and patch management, and defenders cannot simply AI their way out, so the industry needs to rethink disclosure and triage before teams burn out.
10 min · 2,277 words
Popcorn 1.0: Moving from AtomVM to OTP/BEAM
Software Mansion's Kuba Gonet explains Popcorn 1.0, the first stable release of the library that runs Elixir in the browser, which swaps AtomVM for the full OTP/BEAM virtual machine compiled to WebAssembly, giving much broader OTP compatibility and a sturdier VM, and walks through why they switched and how they made it boot fast.
8 min · 1,914 words
Finding panics with Call Graph Analyzer
Tweede golf introduces Call Graph Analyzer (CGA), an open-source tool built with Microsoft under the Open Device Partnership that extracts the call graph from monomorphized Rust MIR so you can find which functions in a compiled binary can panic; the post covers why to eliminate panics, prior art, how CGA works as a rustc driver, and its caveats.
15 min · 3,354 words
Discovering and exploiting vulnerabilities in critical, real-world software with Claude
Anthropic's Frontier Red Team describes five attack chains found and exploited by an unreleased internal research model in widely used browsers and operating systems, leading to remote code execution and other critical impacts with little or no user interaction; all were reported under coordinated disclosure, and the team urges defenders to scan their own code.
8 min · 1,859 words