AI in April 2026 across Simon Willison, Martin Fowler, Interconnects, Zvi, and AI and Games

Because each of these five is a single named author with a recognizable point of view, not an SEO-driven roundup farm. Simon Willison ships the LLM CLI, so his release notes double as primary documentation. Martin Fowler's Fragments column is the closest thing to a working software architect's diary on AI. Nathan Lambert spends his day job on open-weight model post-training, so Interconnects is closer to the work than any press release. Zvi Mowshowitz is the only person publishing a weekly long-form digest with citations and counterevidence. Tommy Thompson is the only credible voice on AI inside game studios. Add them up and you have the closest thing to a real reader's digest for any one month in AI.

They are short interpretations written after reading each post, not extracted quotes. The single direct quotation in the Fowler April 14 entry ('LLMs lack the virtue of laziness') paraphrases the title and the body's framing. If you want the author's own framing in their own words, follow the link; the URL on each row is the canonical artifact, and the summary is only the lure to get you there.

Because Substack post URLs do not encode the date and the search results returned the slugs without timestamps. The three posts cited are the three April 2026 essays that the Interconnects archive surfaces for the topic; treating them as early, mid, and late April is honest given the available metadata. Click through if you need the exact publication day.

Because the original Google query (with five `site:` operators OR'd together) is asking for any of these axes from any of these sites. Games is one of the topical axes the user wanted covered. Tommy Thompson's April 22 piece on the AI-driven RAM crisis is also a constraint piece on the same hardware supply that bottlenecks training runs for the LLM side, so the overlap with the other four sites is genuine.

No. Fazm is a native macOS app that wraps Claude Code and Codex via the Agent Client Protocol. It ships an 856-line bundled skill called deep-research that, when invoked from the floating bar with a query like the one this page answers, runs an 8-phase pipeline on your machine to produce exactly this kind of multi-source curated digest. The page you are reading is the output shape; the skill is the engine. Both are public, both are verifiable, and the skill is in the repo at github.com/m13v/fazm.

It is at Desktop/Sources/BundledSkills/deep-research.skill.md inside the fazm checkout, 856 lines (verified by wc -l on 2026-05-24). The same directory holds seventeen sibling .skill.md files (ai-browser-profile, canvas-design, doc-coauthoring, docx, find-skills, frontend-design, google-workspace-setup, pdf, pptx, social-autoposter, social-autoposter-setup, telegram, travel-planner, video-edit, web-scraping, xlsx, composio-connect). SkillInstaller.swift hashes each file and seeds it into ~/.claude/skills/deep-research/SKILL.md on first launch, so the body shipping with the app is the body running on your machine.

Lines 116 to 122 of deep-research.skill.md spell it out: every factual claim must cite a specific source immediately; FACTS (from sources) are distinguished from SYNTHESIS (the agent's analysis); 'According to [N]' or '[N] reports' is the marker for source-grounded statements; inferences are marked 'This suggests' not 'Research shows'; if the agent is unsure whether a source actually says X it MUST NOT fabricate a citation; if the agent cannot find a source for a claim it must say 'No sources found for X' rather than invent references. The protocol is the difference between a digest like this one (every link is a real URL the agent retrieved) and a chatbot paragraph that confidently cites nothing.

Scope sets the boundaries: which five sites, which month, which topical axes. Plan picks the mode (Standard is the default for a trend query like this one: 5 to 10 minutes, 15 to 30 sources, 4,000+ words). Retrieve runs 5 to 10 WebSearches in a single message plus 3 to 5 parallel Task agents (the skill explicitly forbids sequential search and shows the wrong pattern as ❌). Triangulate requires 3+ sources per claim. Synthesize writes prose with the explicit cite marker. Critique (Deep mode and above) red-teams the synthesis for weak sourcing and missing counterevidence. Refine acts on the critique. Package writes the markdown, the HTML (McKinsey template), and the PDF into ~/Documents/[Topic]_Research_[YYYYMMDD]/. Verify scripts resolve DOIs and validate the report before delivery.

Close but not identical. The skill's outline-refinement phase (4.5) adapts the report structure to what the evidence says, not the structure the agent imagined at the start. You would get the same five-author spine, the same 8-phase pipeline running locally, and a similar per-author table of April 2026 posts with one-line summaries. The exact prose around each link would differ run to run because Synthesize writes prose rather than templating, and because the underlying search results shift as the archives change. The shape is reproducible; the wording is not.

The floating bar is the surface where you type the query. Cmd+Shift+Space (or hold-to-talk) opens it, the find-skills routing matches keywords in your prompt against each skill's frontmatter description, deep-research wins on 'comprehensive analysis' or 'compare X vs Y' or 'analyze trends', and the 8-phase pipeline starts immediately. You watch it run in a window that survives a Mac restart (Fazm chats are persistent on disk) and you can fork the run mid-flight if you want to branch the question without losing the original. The voice and persistence are the part that's specific to Fazm; the skill body would run inside any Claude Code session, which is why we ship it open in the github.com/m13v/fazm repo.