AI Friday — Saturday, May 23, 2026

Standards & Developer Experience

A New Standard: llms.txt for AI-Readable Documentation

This proposal introduces a simple standard — a llms.txt file that tells AI models how to best understand your project, API, or documentation. Think of it like robots.txt for LLMs. It's getting serious traction: popular on HN with thoughtful discussion about how docs should evolve. If you ship an API or SDK, this is worth considering.

Hacker News

Models.dev: An Open Directory of AI Model Specs and Pricing

A new open-source database of AI models with specs, pricing, and capabilities all in one place. Built for developers who need to compare models quickly without hunting across 10 different docs. The format is simple, machine-readable, and meant to be a community resource.

Hacker News

Market Shifts & Pricing

DeepSeek Cuts V4 Pro Pricing Permanently

DeepSeek just made its V4 Pro discount permanent, signaling a serious move toward aggressive pricing. Combined with the context we've been tracking all week — Anthropic's growth, OpenAI's scaling — the market is clearly entering a new pricing era where high performance and low cost can coexist.

Hacker News

The Current AI Pricing Model Was Always Going to End

A sharp analysis of why today's token-based pricing will eventually disappear. The piece digs into the economics: as inference gets cheaper and faster, the model that made sense 18 months ago stops working. Worth reading if you're planning product strategy around AI costs.

Hacker News

Skills & Practice

AI Has a Multiplying Effect on Existing Technical Skills

Josh Comeau's take on AI as a force multiplier is worth your time: it's not about replacing skill — it's about amplifying what you already know. The more you understand a problem domain, the better you can direct AI to solve it. A good counterbalance to the burnout stories we covered earlier this week.

Hacker News

AI-Assisted Engineers Are Burning Out (and What to Do About It)

From earlier this week: Evil Martians document the burnout pattern — writing more code faster, but without the breaks to think deeply. It's a real tension: AI can make you more productive, but it can also accelerate you into exhaustion. The piece offers practical suggestions for pacing.

Evil Martians

Industry & Infrastructure

All Model Labs Are Now Agent Labs

Latent Space ties together a quiet pattern: every major model lab is now positioning around agents, not just models. It's a shift in narrative from raw capability to practical orchestration — who can build the most useful system, not just the smartest weights.

Latent Space

Runtime (YC S26) Launches: Sandboxed Coding Agents for Teams

From yesterday: Runtime is Y Combinator's new bet on team-friendly AI coding. The angle: sandbox environments where agents can work alongside humans safely. Less about replacing engineers, more about giving teams a shared tool for faster iteration.

Y Combinator

Security & Trust

Lawyers Keep Citing Fake Court Cases Invented by AI

A sobering look at AI hallucinations in legal practice. It's not a new problem, but it's getting worse — LLMs are confident and fluent at making up citations that sound plausible. A reminder that for high-stakes domains, you need humans in the loop and verification as a first-class part of your workflow.

Scientific American

Domain-Camouflaged Injection Attacks in Multi-Agent LLM Systems

New research on prompt injection attacks that disguise malicious requests as legitimate domain-specific queries. Relevant if you're building multi-agent systems or anything that chains LLM calls together. The takeaway: agent coordination is powerful but needs robust input validation.

Hacker News

Also

Antigravity 2.0 Tops OpenSCAD 3D Architecture Benchmark — New model hitting new benchmarks — worth tracking if you care about spatial reasoning.
Microsoft Cancels Claude Code Licenses — Quick reversal: Microsoft is stepping back from pushing Claude integration after legal/partnership friction.
Don't Just Paste the AI at Me — A light-hearted manifesto for how to share AI-generated work responsibly. Funny and practical.
100K Lines of Rust Written with Claude: Lessons Learned — Real-world report from someone who let Claude write a LOT of Rust. The patterns that worked and didn't.
Bitcoin Trader Recovers $400K Wallet with Claude's Help — Claude brute-forced 3.5 trillion password combinations to recover an encrypted backup. Impressive and terrifying.
FTC Settles 'Active Listening' Deception Case for $1M — Companies claiming AI features that don't actually exist — FTC is starting to hold people accountable.
Steve Wozniak on AI and Actual Intelligence — Wozniak's take: AI is a tool, not intelligence. Refreshingly clear perspective from someone who's seen cycles before.
CODA: Rewriting Transformer Blocks as GEMM Operations — Optimization research that could make inference faster. Infrastructure-focused but interesting if you're building performance-critical systems.
Nemotron-Labs Diffusion Language Models: Speed-of-Light Generation — Nvidia's new approach to faster text generation via diffusion models. Early research but conceptually intriguing.
NTSB Pulls Docket After AI Reconstructs Dead Pilots' Voices — Thorny ethical moment: AI recreated voices from accident investigations. NTSB had to close the public docket. Shows the need for clearer guardrails.
Linux Audio Subsystem Fixes Driven by AI/LLMs — AI is helping fix low-level kernel code. Quiet victory in infrastructure that few will notice but everyone uses.

Standards & Developer Experience

Market Shifts & Pricing

Skills & Practice

Industry & Infrastructure

Security & Trust

Also

Today’s Sources