The Frequency — Issue #1

Get The Frequency in your inbox. Every Tuesday & Thursday — original thinking, zero fluff.

Free · No spam · Unsubscribe anytime

Issue #1 — May 15, 2026

Next issue Tuesday, May 20 · Browse all issues →

Stories

Pod picks

Sources

Authors

Editor's pick

Software 3.0 — Karpathy's clearest framework for where we actually are

Andrej Karpathy · Sequoia Ascent 2026

At Sequoia Ascent 2026, Karpathy laid out what he calls Software 3.0: programs written through prompts, context, tools, memory, and agents — not code. The context window is the new runtime. Installers become instruction blocks that agents execute and debug live.

This is the clearest conceptual frame for the current moment — not hype, just a clean model that explains why everything feels different now.

Software 3.0 Agents Read full piece →

Simon Willison · May 11

Your AI Use Is Breaking My Brain

An unusually angry post on how AI writing has become impossible to avoid — and is starting to distort even authentic human writing.

Ethan Mollick · One Useful Thing

What Just Happened

A Wharton professor who actually uses AI daily maps the 2025–26 inflection point more clearly than any lab blog post.

Matt Rickard · blog.matt-rickard.com

The Spec Layer

Agents fail differently from humans. When a decision isn't written down, the agent decides it again. Current tooling doesn't know this yet.

Also this issue

Ben Tossell · bentossell.com

How I code — an honest builder's stack in 2026

GPT 5.4 XHigh for code, Opus 4.6 for planning, Codex for syncing agent memory. No hype — he says what broke too.

Pieter Levels · levelsio.com

$3.5M ARR, 0 employees

Photo AI, Interior AI, Nomad List — entirely solo, 90%+ margins. The most interesting live business experiment in AI right now.

Simon Willison · May 6

Vibe coding and agentic engineering are getting closer than I'd like

The line between "write a prompt and ship it" and serious engineering is blurring faster than most developers are comfortable with.

Lenny's Newsletter · Simon Willison

AI state of the union — dark factories are coming

Willison on Lenny's platform: honest read on automation timelines and what passing the inflection point actually means vs. the marketing narrative.

Builders & Makers

People actually shipping things with AI — workflows, products, and lessons from the build.

Ben Tossell · bentossell.com

How I code — my honest 2026 AI stack

Ben ran MakerPad (sold to Zapier), then built Ben's Bites into a 7-figure newsletter. His stack: GPT 5.4 XHigh for code-heavy tasks, Claude Opus 4.6 for planning, Codex for syncing memories across agent runs. He's the best "what actually works" voice for non-engineers building with AI. Only writes about things he's personally built with.

WorkflowNo-code

bentossell.com Read →

Pieter Levels · levelsio.com

$3.5M ARR, 0 employees — the solo AI empire model

Photo AI, Interior AI, Nomad List, RemoteOK — all profitable, all solo, all automated. 90%+ margins. AI handles 95% of coding and customer support. The interesting lesson isn't the tools — it's the philosophy: ship fast, automate ruthlessly, stay small on purpose. Worth studying as a model even if you plan to hire.

$3.5M ARRSolo

buildloop.ai Read →

Andrej Karpathy · karpathy.ai

microgpt — 200 lines of pure Python that trains and runs a GPT

A single file, no dependencies, full GPT pipeline: dataset, tokenizer, autograd engine, neural net, Adam optimizer, training loop, inference. Useful as a learning artifact and as a benchmark for how minimal a real implementation can be. The kind of thing only Karpathy would bother making elegant.

Open sourceLearning

karpathy.ai Read →

Hacker News · Ask HN

What are you working on? (May 2026)

The HN "What are you working on?" thread is consistently the best raw signal for what builders are actually shipping. Hundreds of replies, no pitch decks, no press releases. Look past the top comments — the interesting projects are buried in the thread.

Community

Hacker News Browse →

Podcasts & Long Reads

Episodes and essays worth your time — no filler, no "AI will change everything" non-answers.

🎧

How I AI · Lenny's spinoff · Hosted by Claire Vo

Real AI workflows, live screen shares — the format every podcast should steal

30-minute episodes with practitioners showing their actual AI setup on screen. No scripted "AI will transform X" takes — just what people actually do. This is the show to subscribe to if you want to see real usage, not demos.

⏱ ~30 min · Product leaders & builders

🚀

Latent Space · swyx & Alessio Fanelli

The AI Engineer podcast — technical, rigorous, no softballing

swyx (Shawn Wang) and Alessio Fanelli run 175+ episodes of genuinely technical AI engineering content. The 2026 plan: expanding into AI for Science. Their episode on "three paths AI could take" is required listening for anyone thinking about where this goes.

⏱ ~60 min · AI engineers & founders

🎙

No Priors · Elad Gil & Sarah Guo

The most honest AI investor podcast — they don't softball guests

Elad Gil and Sarah Guo interview researchers, founders, and investors without the usual deference. Episodes with Nat Friedman and Daniel Gross are highlights — they speak plainly about what AI can and can't do.

⏱ ~45 min · Founders & researchers

🌐

Simon Willison on Lenny's Newsletter

AI state of the union — we've passed the inflection point

Willison on dark factories, automation timelines, and what the inflection point actually means vs. the marketing narrative. Required reading for anyone tracking the real trajectory.

⏱ Long read

Research & Models

Papers, frameworks, and technical ideas worth knowing — filtered for signal over noise.

Andrej Karpathy · Sequoia Ascent 2026

Software 3.0 — the context window as the new runtime

Software 1.0 = explicit code. Software 2.0 = neural network weights. Software 3.0 = LLMs programmed through prompts, context, tools, memory, agents. In Software 3.0, the context window is the main lever — the LLM is an interpreter over context, performing computation on digital information. Concrete implication: installers become blocks of instructions pasted into an agent.

FrameworkKarpathy

karpathy.bearblog.dev Read →

Matt Rickard · blog.matt-rickard.com

The Spec Layer — agents fail differently from humans

Most software tooling is optimized for human failure modes. Agents fail differently: undocumented decisions get re-decided; context windows are finite; too much freedom at execution time causes drift. Rickard's framing for why current tooling is fundamentally mismatched to agentic systems.

AgentsTooling

matt-rickard.com Read →

Simon Willison · May 6

Vibe coding and agentic engineering are getting closer than I'd like

The line between "write a prompt and ship it" and serious engineering is blurring faster than most developers are comfortable with. A sharp reflection on what this means for craft, quality, and accountability — from someone who maintains production tools and actually reads the papers.

Agentic AICraft

simonwillison.net Read →

Simon Willison · Substack

"Agent" may finally have a definition coherent enough to use

Willison argues "agent" as jargon has reached the point where it's actually useful for reasoning — not just marketing. This matters for system design, spec writing, and explaining capabilities to non-technical stakeholders. Sounds boring. Isn't.

Terminology

simonw.substack.com Read →

AI × Business

How AI is reshaping companies, products, and work — strategy and real-world signals.

Pieter Levels · Multiple profiles

The solo AI empire: $3.5M ARR, 90% margins, 0 employees

Pieter Levels is running what might be the most interesting live business experiment of this AI moment. The model challenges every assumption about team size and growth. What's worth studying: not the products, but the philosophy — ship fast, automate ruthlessly, stay small on purpose.

$3.5M ARR0 employees

buildloop.ai Read →

Hacker News · May 2026

Uber burns 2026 AI budget in four months

Uber burned through their entire 2026 Claude Code budget by May. Actual enterprise AI usage is running orders of magnitude higher than any projection made in 2025. Ops and finance teams everywhere are scrambling — the models they used to forecast spend are broken.

EnterpriseSpend

Hacker News HN thread →

Matt Rickard · blog.matt-rickard.com

The Model is Not The Product — and that changes everything about moats

The current wave has been about directly exposing models to users. Rickard argues this is temporary. Data no longer serves as a moat the way people expected, and the real competitive landscape is shifting underneath everyone building on top of models right now.

StrategyMoats

matt-rickard.com Read →

Authors on the Radar

Only people who build things, have skin in the game, or do original research. Grows each issue — with active search for diverse voices worldwide.

🧠

Andrej Karpathy

karpathy.ai · @karpathy

New

Former Tesla AI Director, OpenAI co-founder, now independent. Built microgpt (200-line pure Python GPT), coined Software 3.0, the LLM wiki pattern. Rare combination of deep theory and practical builder instincts.

Blog Site

🚀

swyx (Shawn Wang)

swyx.io · @swyx · Singapore

New

Singaporean developer and founder of Latent Space podcast + smol.ai. Coined "AI Engineer" as a distinct role. Latent Space (~175 episodes) is the most technically rigorous AI engineering podcast running. 2026 focus: AI for Science.

Podcast Blog

🎓

Ethan Mollick

oneusefulthing.org · @emollick

New

Wharton professor studying entrepreneurship and AI. Writes One Useful Thing — research-backed, personally tested. Writes every post himself before asking AI for feedback. Best academic voice that isn't purely theoretical.

Newsletter

⚙️

Matt Rickard

blog.matt-rickard.com

New

Posts almost daily. Dense, short, no padding. Covers AI engineering, product strategy, and infrastructure. Key pieces: The Spec Layer, The Model is Not The Product, The New AI Moats. No bio fluff — just ideas.

Blog

📝

Simon Willison

simonwillison.net · @simonw · UK

Co-creator of Django (UK). Maintains datasette and LLM CLI tools. Most technically credible independent voice on practical AI — reads papers, maintains production tools, writes about things that break. No incentive to hype.

Blog Substack

🛠️

Ben Tossell

bentossell.com · @bentossell · UK

Built MakerPad (sold to Zapier), then Ben's Bites ($1M+ newsletter). Best voice for "what actually works" AI builder takes for non-engineers. Only writes about things he's personally built with.

Blog Newsletter

🌍

Pieter Levels

levelsio.com · @levelsio · Netherlands

Dutch solo founder running $3.5M ARR portfolio (Photo AI, Interior AI, Nomad List, RemoteOK). The canonical example of AI-powered solo business at scale. Builds in public, shares real numbers.

📦

Lenny Rachitsky

lennysnewsletter.com

Former Airbnb PM. Runs the top product newsletter + podcast. The "How I AI" spinoff (hosted by Claire Vo) is now the best source for real AI workflow walkthroughs. Weekly Willison collaboration is excellent.

Newsletter Podcast

✍️

Paul Graham

paulgraham.com · @paulg

YC co-founder. Infrequent but sharp essays. "Writes and Write-Nots" on AI and writing is one of the best pieces on the cognitive consequences of outsourcing writing. Tracked because when he writes, it tends to matter.

Essays

Sources

17 sources tracked. Refreshes every Tuesday & Thursday. Zero AI-generated summaries of AI news.

New

New

New

New

Builder · Netherlands

New

New

What's excluded and why

Out: LinkedIn posts recycling TechCrunch · "AI will replace X" thought pieces with no data · Newsletter aggregators summarizing other newsletters · "I asked ChatGPT to write this" posts · Press releases dressed as blog posts · "Top 10 AI tools" listicles where the author hasn't used them.

The filter: Does this person build things or have real domain expertise? Do they have skin in the game? Is there something here you couldn't get from a headline? No to any = out.

✓ Link copied