Essential Reading for AI and the Singularity, From Someone Who Watched the Date Move

The book most literally about the singularity in 2026 is titled with a year, and that year is already slipping. AI 2027, the scenario forecast published in April 2025 by Daniel Kokotajlo, Eli Lifland and collaborators, put a superhuman coder around 2027. By this April the same team had moved Daniel’s median for that milestone to mid-2028 and Eli’s to mid-2030, with their AGI-equivalent forecast sliding about 18 months later than the version that made the headlines. Kokotajlo’s personal AGI median tells the fuller story: 2027 for most of 2022 through early 2025, then 2028, then “around 2030” by the end of 2025.

That is the right frame for a reading list on this subject. The arguments are worth your time. The dates attached to them are the first thing that ages. So here is a path from curious to informed, organized so each stage earns the next, with a note on how well each has held up.

Part 1: The frame before the machines

Start with history, because the singularity conversation is really an argument about what kind of thing humans are and whether that thing is about to be replaced. Yuval Noah Harari’s Sapiens (2011) makes the case that our species runs on shared fictions: money, nations, corporations, all of them stories we agree to believe. Its sequel Homo Deus (2015) extends the line forward into what Harari calls dataism, the idea that we are starting to trust algorithms over our own judgment. His 2024 book Nexus is the most useful of the three here, because it reframes AI as a new kind of information network rather than a mechanical brain, and asks how societies reorganize when the network starts making decisions.

Harari is a synthesizer rather than a technologist, and it shows. Take the sweep and leave the specifics. What you want from Part 1 is the instinct to ask what AI does to human institutions, beyond what it scores on a benchmark.

Part 2: How the thing actually works

Now open the box. The single best starting point is free: Stephen Wolfram’s essay What Is ChatGPT Doing … and Why Does It Work? It walks from “predict the next word” to attention and embeddings without hand-waving, and it is honest about how much of the result is still unexplained. Pair it with 3Blue1Brown’s neural-network series, where Grant Sanderson animates what a transformer is doing to a vector as it moves through the layers. Watching a matrix multiply turn into a prediction does more for intuition than any paragraph can.

When you want to build one, Andrej Karpathy is the teacher. His Zero to Hero lectures construct a language model from an empty file, and his nanochat project trains a small ChatGPT for about $100, tokenizer to web UI, in a single hackable codebase. The point of doing this is demystification. Once you have seen the whole pipeline, the machine stops being an oracle and becomes an engineering artifact with knobs.

Then read Brian Christian’s The Alignment Problem (2020), the bridge from “how it works” to “how it goes wrong.” Christian is careful and reportorial, tracing how systems learn from data, absorb its biases, and optimize the proxy you gave them instead of the goal you meant. That gap is not academic. It is the same failure that lets a coding agent three indirection steps from a prompt do something its operator never intended, which is alignment as a security problem you can watch happen today.

Part 3: The singularity arguments

With foundations in place, the speculative canon reads better, because you can tell an argument from a vibe.

Ray Kurzweil’s The Singularity Is Near (2005) and its sequel The Singularity Is Nearer (2024) are the origin of the popular thesis: exponential curves in computing carry us to a 2045 merger of human and machine intelligence. Read Kurzweil for the shape of the claim and the discipline of tracking trend lines. Read the critics for balance. Physicist Paul Davies’ line in Nature still lands, that the exhilarating speculation “needs to be taken with a huge dose of salt,” and the standing objection is that exponentials always look eternal right up until a resource wall bends them. Kurzweil’s 2045 has survived twenty years mostly by sitting far enough out to stay unfalsifiable.

Nick Bostrom’s Superintelligence (2014) is the canonical statement of the control problem: if you build something far smarter than yourself, getting its goals exactly right becomes the only thing that matters, and “exactly” is doing heavy lifting. It is dense and occasionally maddening, and it set the terms most of the field still argues in. Max Tegmark’s Life 3.0 (2017) is the friendlier companion, a tour of post-AI futures that takes both utopian and grim branches seriously. Mustafa Suleyman’s The Coming Wave (2023), from a DeepMind co-founder now running Microsoft AI, widens the lens to argue that AI and synthetic biology arrive together as one hard-to-contain wave, and that containment is the central political problem.

For the near-term forecasts, AI 2027 is the most rigorous scenario going, worth reading precisely because its authors document every revision in public; the 949-point Hacker News thread is a good map of where readers pushed back. Its louder cousin is Eliezer Yudkowsky and Nate Soares’ If Anyone Builds It, Everyone Dies, which hit the New York Times bestseller list last September and split reviewers straight down the middle. The Guardian called it “as clear as its conclusions are hard to swallow”; The Atlantic’s Adam Becker countered that the authors “are not grifters, they are just wrong” and fail to make an evidence-based case. Read it for the strongest version of the doom argument, and read the reviews to see where the argument thins out.

Part 4: Fiction that got there first

Novelists ran these experiments decades before the forecasters. Charles Stross’s Accelerando (2005) is the singularity from the inside, three generations of one family swept through an intelligence explosion and a solar system being disassembled into computronium. It is fast, weird, and unnervingly good on how economics mutates when cognition gets cheap. Isaac Asimov’s Robot stories supply the older foundation and the durable vocabulary: the Three Laws exist so Asimov could spend story after story showing them fail in ways their designers never anticipated, which is the alignment problem in a tweed jacket, published in 1950.

How to read the pile

Do not read for prophecy. Every dated prediction in these books will be checked against reality, and the forecasters closest to the models are the ones revising fastest, in public, with spreadsheets. Read instead for the arguments and the failure modes, the parts that stay true whether the milestone lands in 2028 or 2038.

A workable order: Wolfram and 3Blue1Brown first, so nothing downstream can bluff you on the mechanics. Then Christian, for how it breaks. Then one big-picture book (Harari’s Nexus or Tegmark’s Life 3.0) and one forecast (AI 2027). Save Bostrom for when you want the argument at full strength, and Stross for when you want to feel it. If you only have an afternoon, read the Wolfram essay and watch one 3Blue1Brown video. That alone puts you ahead of most of the people writing confident headlines about the machines.

Sources: AI 2027 · Clarifying how our AI timelines forecasts have changed since AI 2027 (LessWrong, 27 Jan 2026) · Q1 2026 Timelines Update (AI Futures, 2 Apr 2026) · What Is ChatGPT Doing … and Why Does It Work? (Stephen Wolfram) · 3Blue1Brown neural networks · Karpathy’s nanochat · The Alignment Problem · The Singularity Is Near (Wikipedia) · Superintelligence · Life 3.0 · The Coming Wave · Nexus · If Anyone Builds It, Everyone Dies (Wikipedia) · Accelerando · Coverage: AI 2027 on Hacker News · Related on this blog: Three Indirection Steps From a Reverse Shell · Twenty-Two Second Brains, and a Text File That Beats Most of Them