The AI Friday Brief
Good Sunday, NOLA. July 12th is a quiet day after an absolutely wild week — Latent Space called it exactly. But we've got some genuinely interesting reads on the state of AI tooling, a clever font that dodges AI vision models, and some solid infrastructure moves. Let's dig in.
Interesting Reads & Takes
Ghost Font: A font that humans can read but AI cannot
A designer created a typeface optimized to be readable by humans but invisible to AI vision models. It's a clever adversarial approach to the watermarking problem — if your text needs to stay hidden from AI scraping or training, this might actually work. Worth trying if you're protecting content.
Hacker News
Stop Telling Me to Ask an LLM
A sharp piece on the reflexive "just ask ChatGPT" advice that's become ubiquitous. The author argues that dumping every problem into an LLM is lazy thinking, not progress. If you've been feeling weird about the automation creep, this validates it.
Hacker News
AI Can't Recreate the Thrust Game (But It Can Help You Understand It)
A fascinating deep-dive on using AI to reverse-engineer a 1984 Commodore arcade game. The author shows where AI shines (generating plausible code, spotting patterns) and where it completely fails (the actual physics of the game). Great example of AI as a thinking partner, not a replacement.
Hacker News
AI 2040 and the cult of intelligence
A contrarian take on where AI is actually headed and what we're getting wrong about the hype. Definitely opinionated, but worth reading if you want to push back on some of the consensus narratives.
Hacker News
Tools & Infrastructure
Mesh LLM: Distributed AI computing on iroh
A new approach to running LLMs across a distributed network without a central server. If you're thinking about edge inference or peer-to-peer AI, this is the kind of architecture that could matter. Early-stage but promising.
Hacker News
Sqlsure: Deterministic semantic checks for AI-generated SQL
Open-source tool that validates AI-generated SQL queries before they hit your database. This is the kind of practical safety layer that matters when you're using Claude or GPT to write database code. Worth pinning if you're building agents.
Hacker News
sqlite-utils 4.1
sqlite-utils got a dot release with some nice quality-of-life improvements for the insert and upsert commands. If you're building scripts that touch SQLite, this keeps getting better.Business & Product Moves
OpenAI bets on families as ChatGPT goes deeper into households
OpenAI is hiring a dedicated product manager to build family-focused features for ChatGPT — caregiving tools, age-gated content, family accounts. A signal that they're thinking beyond individual power users to mainstream adoption.
TechCrunch
Microsoft's 25% emissions jump tied to AI data centers
Microsoft just disclosed that building out AI infrastructure for Azure and Copilot drove a significant spike in their carbon footprint. Not doom — just reality. If sustainability matters for your stack decisions, this is worth tracking.
Hacker News
Also
- Reverse centaurs are the answer to the AI paradox — Cory Doctorow on humans and AI working together effectively — a good counterpoint to both 'AI will replace us' and 'AI is useless' narratives.
- Rowboat: Open-source local-first Claude alternative — If you want Claude on your machine without cloud, this is the self-hosted option. Still early but worth watching.
- Geosql: Claude agents for geospatial data — AI agents that understand geographic queries and databases. Niche but cool if you work with maps or location data.
- FableCut: Browser video editor AI agents can drive — An open-source video editor designed so AI can actually use it. The vision is agents editing video like humans do.
- Kastor: Terraform for AI agents — Infrastructure-as-code but for agent deployments. Early project but interesting angle on agent ops.
- Microsoft Flint: Visualizing AI agent behavior — Open-source tool for understanding what your AI agent is actually doing step-by-step. Super useful for debugging.
- AMD Ryzen AI Halo hands-on review — New AMD chips with built-in AI acceleration. Laptop inference is getting real.
- Brown University AI cheating scandal: 50% drop in in-person exam scores — When students have ChatGPT for take-homes but can't use it on exams, the gap is brutal. An honest look at the ed-tech reckoning.
Pass it on