THE COLLECTION

Library

Everything you’ve found. Nothing you need to save twice.

22 articles
Browse at your own pace.

Post 50: Good Research Takes are Not Sufficient for Good Strategic Takes — Neel Nanda

The author argues that being good at AI research is only weak evidence of being good at strategic thinking about AGI. Research offers clear feedback; strategy does not, so researchers can hold bad strategic views and never correct them. He lists skills good strategic takes need and asks readers to check what evidence supports a person's strategic claims.

MATS Applications Open (Due Aug 29) — Neel Nanda

This is a recruitment call for the MATS mechanistic interpretability program run by Neel Nanda. Applicants complete a 12-20 hour research task and write-up by August 29. About 32 candidates enter a 5 week paid online phase, then about 8 enter a 12 week paid in-person phase writing a paper. The author says past scholars often publish at top ML venues.

Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights

This article reports that a frontier LLM can read and write neural network weights without gradient descent. It identifies trained behaviors from raw weights, writes game-playing networks zero-shot, and edits a small share of a language model's weights to fix looping and raise math scores. Results are early; a model without looping defects could not be improved, showing limits of the method.

The era of personal software

The author describes building a custom health tracking web app to collect data from his devices. He uses automated exports, a hosted Postgres database, and AI agents to log meals and give insights. He argues that LLMs make personal software cheap to build, but not a full replacement for existing tools yet.

PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet

Wiz Research found memory-safety bugs in core PHP that let attackers run code or read worker memory without login. The bugs affect PHP versions back to 5.4, and for a decade exploit building was too complex for attackers. Frontier AI models made building exploit chains practical, and PHP released a patch on October 21, 2026.

Patch the Planet: global response coordination as agents uncover critical vulnerabilities

AI tools found two critical PHP vulnerabilities before attackers could use them. The authors propose a five-stage response framework. Maintainers confirm and fix the flaw, key groups prepare updates in private, and teams publish guidance after. The article also weighs sharing exploit details against leak risk and funds support for maintainers.

dlab Open Source Week: Frontier AI on Your Own Hardware — Tim Dettmers

Tim Dettmers argues that AI agents make research cheap, so small university labs can compete with big labs. His lab will release an agent harness, a local inference framework, and a compaction method that halves agent costs. He tells students to change how they work, keep their identity, and attack problems that need few resources.

How the Claude Code team does code review

This article describes the code review process of the Anthropic team behind Claude Code. Agents write most of the code, so machines do most of the reading and humans approve by risk tier. The tradeoff is less human review time, but mistakes still pass, so the team needs rollback, telemetry, and clear ownership.

How to keep enjoying programming in a world of LLMs

A Haskell developer, worried about LLM-driven burnout and skill atrophy, argues against full vibe coding: you should keep writing code yourself while delegating planning, research, review cycles, and low-risk cleanup to agents. The tradeoff is accepting modest productivity gains (roughly 2x by his estimate) instead of boasting numbers, in exchange for retaining enjoyment, codebase ownership, accountability, and independence from token limits and frontier-model vendors.

When AI writes almost all code, what happens to software engineering?

Arguing that late-2025 model releases (Opus 4.5, GPT-5.2, Gemini 3) made AI-generated code good enough to supplant hand-coding, the author blends anecdotes from prominent engineers with predictions about the profession. Specialized coding skills and hand-writing code lose value, while testing, architecture, product thinking, and verification become the engineer's core role. Tradeoffs include faster shipping at the cost of quality and accountability, though the piece rests on anecdotes and largely assumes its central claim holds.

Rebuilding AUTOMATIC1111 with Gradio Workflow

This article walks through Workflow1111, a rebuild of AUTOMATIC1111's stable-diffusion-webui as a 73-node Gradio Workflow canvas, arguing that node graphs can mix local Python functions with remote model calls. It shows that every output node becomes a REST endpoint and MCP tool, and that visitors can run pipelines under their own quota. Tradeoffs are acknowledged: there is no loop operator, and roughly two-thirds of nodes run in-process without a network call.

Your Agent Aced the Task. Will It Do It Again?

The article argues that average benchmark accuracy hides agent unreliability: a ReAct agent on GPT-4.1 scored 77.4% Mean@5 on AppWorld but only 53.0% Pass^5 (succeeding on every run). Their Consistency Analyzer resamples decision points from a single recorded trajectory to find flip-prone steps, and the resulting injected guidelines halved the gap (24.4pp to 12.0pp) without costing average accuracy. Tradeoffs: diagnosis needs one extra LLM call per decision step, and consistency is orthogonal to capability, so a bigger model won't fix it.

Should you read the code, is RAG dead, and did Skills kill MCP?

The article pushes back on five common AI hot takes: that you can skip reading AI-generated code, that not using AI hurts hiring, that Skills killed MCP, that RAG is dead, and that needing fine-tuning signals bad code. It argues each is oversimplified—review depth should match actual risk, AI fluency is about judgment, MCP and Skills solve different problems, and RAG still grounds answers. It closes by urging readers to test ideas by building rather than debating.

Your Agent Speaks MCP. Give It a Computer.

Fly.io pitches Sprites—instantly created, near-free-when-idle cloud computers with durable filesystems—as a better agent runtime than stateless sandboxes. The substantive argument: the MCP-vs-CLI debate conflates layers; progressive disclosure is about what enters the context window, while MCP handles transport, auth, and structured results, so plugins can layer skills over a hosted MCP server. It is vendor marketing, but the layering distinction, safety annotations, and guardrail defaults (org scoping, five-Sprite cap) are the meat.

My agent.md to improve LLM-assisted code quality

The author recounts moving from unusable LLM code in 2025 to productive agentic coding in 2026, then shares the agent.md file they built to encode recurring code-review corrections as prompt rules: concise comments, no magic numbers, early returns, private-by-default visibility, layered architecture, commit conventions, and test-first bug fixes. It reduces repetition and frees review time for architecture, but is no magic bullet — hallucination and verification remain, and 'context dilution' degrades adherence, mitigated by short sessions and reloading agent.md.

Laya — 33ms Multilingual System 1 Decision Engine

The author introduces Laya, an open-source family of bidirectional encoder models that answer typed questions (choice, ordinal score, boolean) with calibrated probabilities in roughly 33 ms, arguing generative LLMs are overkill for high-volume triage and routing. Self-reported benchmarks claim speed, calibration, and cost advantages over the proprietary Jev API. Acknowledged tradeoffs: choice questions degrade beyond ~20 options, base checkpoints are near-random zero-shot and require fine-tuning, and temperature calibration is needed.

How we used DSPy to turn AI evaluations into better responses in Dash chat

Dropbox built trajectory-based LLM-as-judge evaluations for its Dash chat agent, then used DSPy's GEPA and MIPROv2 optimizers in two stages: calibrating judges against a small human-labeled set, then optimizing the agent's system prompt via offline replay of historical chats. Reported gains include 26% fewer incomplete answers, 13% fewer missed key aspects, and 5.4% lower token usage. The tradeoffs: automation requires strict guardrails, and weak evaluation signals risk brittle improvements.

How well do agents use test/verification techniques?

An empirical eval of 26 prompt conditions (TDD, fuzzing, property-based testing, formal methods, and testing skills) given to coding agents implementing Zstd in Rust, ~80 runs each. Nothing beat the default no-instructions baseline; agents apply techniques superficially—vacuous proofs, trivial random tests—while TDD and popular testing skills underperformed. A brief hand-written skill nudging risky-area checks and structured randomization scored best, suggesting expert guidance matters more than naming techniques.

There's no point at which turning your brain off will work

The author argues that 'brain-off' LLM use—having agents write code and just relaying failures back (being a 'meat proxy')—still produces poor-quality software, failing on out-of-distribution problems and overfitting to tests or metrics. Even if it eventually yields good software, it won't benefit the employee, since a company could run the LLM loop without them and lay them off. He contends the present is the worst time to disengage, as companies are eager to cut low-contribution roles.

Have it both ways: stay discoverable in search while disallowing AI training

Cloudflare argues that mixed-use crawlers (Applebot, Bingbot, Googlebot) forced site owners to choose between search discoverability and refusing AI training. Its new Disallow AI Training setting publishes a robots.txt preference that 'Accountable' operators honor, keeping search while blocking training; Block now applies to mixed-use crawlers too, affecting search. Settings migrate automatically on September 15, with stricter presets for ad-funded sites. Cloudflare frames AI summaries, with only partial controls today, as the next unresolved tradeoff.

When scanners miss the attack: how Cloudflare Client-Side Security protects storefronts

Cloudflare details four malicious JavaScript campaigns found on storefronts—affiliate-commission hijacking, clickless affiliate theft, a remote-code backdoor, and a paid-traffic cloaker—that VirusTotal and URLScan largely missed, arguing behavioral ML analysis beats signature scanning for gated, conditionally executing scripts. The detection pipeline combines a graph neural network, a lightweight LLM on Workers AI, and a weighted frontier-model voting ensemble with human review. It is a vendor post promoting Cloudflare Client-Side Security, so evidence is detailed but conclusions favor the product.

Astra for Coding: Why Are We Doing This Again?

The author ran a fully autonomous 'software factory' with GPT 6 Astra for 35 hours (~$1200, 79 commits, 75k lines of code) and got nothing usable. His core complaint is that the model's token-efficient 'codegolf' Python for tool calls leaks into committed code, producing unreadable and bizarre output, and that it won't stop on oversized tasks. He argues training rewards local metrics over human readability, and doubts these costlier models still fit day-to-day software engineering versus other use cases.