May 2026 stack — understand before you build
This month's shape — a Foundations pillar that explains how AI works under the hood, paired with role-shaped paths through the library. The bet is that operators who understand the mechanism make better choices than ones who just pattern-match.
See How AI works pillarA note on what shipped in May, and the editorial thesis behind it.
The shape of the month
Three big shippings:
- A new "How AI works" pillar in
/learn/ai— 8 substantial guides covering the mechanism: how LLMs work, tokens, context windows, embeddings, transformers and attention, fine-tuning vs RAG vs prompting, how agents work, and how to evaluate LLM output. - Learning paths — sequenced routes through the existing library, shaped by role. Five paths today: operator, engineer, PM, founder, designer.
- 8 new tools filling category gaps (Grammarly, Sudowrite, Superhuman, OpusClip, PlayHT, Vista Social, AdCreative.ai, Textio), plus a guide on evaluating an indie AI tool before you commit.
The thesis
There's a reader segment growing fast: people who use AI every day, get real value from it, and have no mental model for how it works. They prompt by superstition. They get surprised by hallucinations. They overpay for tools that are wrappers. They under-invest in tools that are infrastructure.
The Foundations pillar exists for that reader. Not because everyone needs to know what attention is or how embeddings encode meaning — but because the people making real decisions about AI in their work usually do. The choices look different when you understand the mechanism.
What we'd reach for this month
If you're starting fresh today and want to learn AI seriously without falling for the listicle treadmill:
- For the mental model: How large language models work. Read it once. The "model is predicting the most plausible next token" framing explains 90% of what surprises people.
- For the operator job: the Zero → AI-augmented operator path. 8 guides over 4 weeks, roughly 15 min per week.
- For the engineering job: the Engineer → AI-coding fluency path. Foundation first, agent mechanism second, then the build-with-ai pillar.
- For tool selection: the evaluating-an-indie-ai-tool guide before paying for the next LTD.
What we deliberately didn't do
- No gamification. Considered progress tracking with checkboxes and completion badges. Pulled it back before shipping. AINews is editorial, not Duolingo. The path indicator stays (it's orientation). The achievement layer goes.
- No new pages. Foundations is a new pillar inside
/learn/ai, not a new top-level surface. Paths is a sub-route, not a separate product. The site count stayed flat. - No AppSumo-deals dump. The /deals concept is interesting and parked in roadmap, but auto-scraping AppSumo would create lifecycle risk without editorial value. When it ships, it ships manual-curated and with a "not worth it if" required per pick.
What's next
A few things on the roadmap, not yet shipped:
- A
/dealsroute, manually curated, AI tools only, weekly cadence — once the affiliate side is sorted - A pruning pass on
/bestand/digest— the editorial calculus may be that both should go, freeing space for higher-signal surfaces - More learning paths as new role-specific guides land
More spotlights: See the archive →