Distill · 2026-10-15
A single-day batch of Paul Graham essays dominates the week, alongside an Anthropic education report on measurable AI fluency, a Claude Code best-practices document arguing context is the scarce resource, and Zhipu's two-model release claiming frontier coding at roughly a tenth of the cost.
Keys: space/K play-pause · J/L jump 10s · arrows on the waveform nudge 5s
The full episode transcript, timed to the audio. Click any line to jump there; the current line lights up while you listen.
HostLet's start with the shape of the week, because the shape is the story. Distill's brief is dominated by one source: a single-day batch of Paul Graham essays, all timestamped the same day, spanning language design, startup practice, writing, and one California ballot position.
CorrespondentAnd when I say batch, I mean it. This isn't two or three posts. It's a pile. And the pile has a spine — the technical essays reinforce each other almost like one argument split across pieces.
HostThere's a wrinkle worth flagging up front, because it affects how you read the whole week. Two of those captures contain no essay text at all — just site boilerplate. Navigation links, a Y Combinator prompt, a copyright line. And a third is truncated to its introduction.
CorrespondentSo the batch is bigger on paper than it is in substance. That's a real caveat and we'll come back to it. But the substance that is there is coherent, so let's take it seriously.
HostAlongside the Graham batch, two other things landed. Anthropic published an education report measuring observable AI-fluency behaviors across 9,830 anonymized Claude.ai conversations, plus a Claude Code best-practices document. And Zhipu, also known as Z.ai, announced two models — GLM-5.3 and GLM-5.3-Flash.
CorrespondentTwo very different kinds of news. One is about how people actually use these tools. The other is about the price of intelligence falling. Both matter if you're building anything.
HostLet's do the technical spine first. The core claim across the Graham batch is bottom-up language design. In "Programming Bottom-Up," he argues that experienced Lisp programmers evolve the language and the program together, until the program looks as if the language had been designed for it.
CorrespondentWhich is a nice inversion of how most of us were taught. You don't pick a language and then write the program. You bend the language until the program is the natural thing to write in it.
HostAnd he grounds that historically. "The Roots of Lisp" and "What Made Lisp Different" both go back to McCarthy's 1960 paper. The second one enumerates nine ideas — conditionals, first-class functions, recursion, garbage collection, and others — that mainstream languages have since absorbed.
CorrespondentThat's the strongest version of the argument. Not "Lisp is better," but "everything you use now is Lisp, slowly." Nine ideas, and the industry spent decades rediscovering them one at a time.
HostThen there's the commercial anecdote. "Beating the Averages" is the Viaweb story — Lisp, 1995 — and it's the piece that made this argument famous outside of computer science departments.
CorrespondentAnd "Succinctness Is Power" supplies the metric. Program power measured as element count rather than lines. That's a subtle but important reframe — lines are a formatting artifact, elements are the actual expressive units.
HostThe batch also extends the claim forward. "The Hundred-Year Language" and "Revenge of the Nerds" both push it into the future. The latter asserts a hacker-culture hierarchy — Java below Perl below Python — where each step up is more Lisp-like.
CorrespondentAnd then there's "Java's Cover," which is the hostile counterpart. Twelve signals that Java smells suspicious. Written by an author who concedes he has never used it.
HostThat concession is doing a lot of work in that piece. It's an argument from taste and from the shape of the ecosystem, not from experience.
CorrespondentWhich is either admirably honest or a fatal flaw, depending on your priors. I'd say it's honest, and it's also why that particular essay has aged the worst of the set.
HostThe non-technical half of the batch repeats familiar positions. "How to Start a Startup" reduces the whole thing to good people, something customers want, and minimal spending.
CorrespondentThree ingredients. That's it. And "Ideas for Startups" calls ideas worthless — in the narrow sense that no market exists for selling them. You can't trade an idea for money. You can only trade execution.
Host"Hiring Is Obsolete" argues that falling startup costs have made traditional hiring obsolete for talented young people. The logic being: if you can build the thing yourself cheaply, why go be a small cog somewhere?
CorrespondentThat one's been controversial for two decades and it's still controversial. The counterargument is that most things worth building are too big for one person, no matter how cheap the tooling gets.
HostOn craft, "Writing Briefly" claims writing generates ideas rather than merely communicating them. And "A Version 1.0" reports writing three to four words for every one that survives.
CorrespondentThree to four to one. That's a useful number to keep in your head, because it reframes editing as the actual work rather than the cleanup.
HostThen two media pieces. "The Submarine" estimates that more than half of non-political, non-crime traditional-media stories originate from PR placements. And "What You Can't Say" proposes a conformist test for finding unsayable ideas.
CorrespondentThe submarine number is the one that sticks. More than half. If that's even roughly right, most of what you read as news is a press release wearing a costume.
HostAnd the one political item: "This Year We Can End the Death Penalty in California" urges support for Proposition 62 and rejection of Proposition 66, citing a research figure that about four percent of people sentenced to death are innocent.
CorrespondentFour percent. That's the whole argument in one number, and it's the kind of number that either moves you or it doesn't.
HostNow to Anthropic, which published two things this week worth separating. First, the education report — the AI Fluency Index. It looked at 9,830 anonymized Claude.ai conversations with multiple back-and-forths, over a seven-day window in January, and measured eleven directly observable behaviors drawn from the 4D AI Fluency Framework.
CorrespondentThe headline finding: 85.7% of conversations showed iteration and refinement. And those conversations exhibited 2.67 additional fluency behaviors on average, versus 1.33 without. Roughly double.
HostAnd 5.6 times higher rates of questioning Claude. That's the part I keep coming back to.
CorrespondentBecause it suggests the behavior is contagious within a conversation. If you push back once, you're much more likely to push back again. The interface isn't just a window onto the behavior — it's shaping it.
HostWhich leads directly to the second Anthropic document: the Claude Code best-practices guide. It positions Claude Code as agentic, and argues that most best practices derive from a single constraint — the context window fills quickly, and model performance degrades as it fills.
CorrespondentOne constraint, everything else follows. That's a clean way to organize a doc, and I think it's basically right. A lot of what looks like mysterious agent failure is just a full context window.
HostThe central recommendation is to give Claude a runnable check. Tests, build exit codes, linters, or screenshot comparisons — so the work loop closes without the human acting as verifier.
CorrespondentThat's the operational takeaway of the week, honestly. If you're the verifier, you're the bottleneck, and you're burning context on both sides. Give it a check it can run itself.
HostAnd notice how these two Anthropic pieces rhyme. The education report says iteration is where the fluency lives. The engineering doc says the loop should close on a machine-checkable signal. Both are arguments against one-shot interaction.
CorrespondentYeah — one-shot interfaces may be suppressing exactly the behavior you're trying to measure. That's a design problem, not a user problem.
HostThen Zhipu, or Z.ai, with two models. GLM-5.3 builds on the same base model as GLM-5.2, with all the gains coming from post-training. They claim it's the most capable open-weights coding model.
CorrespondentAnd the benchmark jumps are not small. Terminal-Bench 3.0 goes from 4.6 to 28.3. DeepSWE v1.1 from 46.2 to 66.9. Agents' Last Exam from 23.8 to 28.5.
HostThe Terminal-Bench move is the eye-catching one — 4.6 to 28.3 is a different order of capability, not an incremental bump. And all of it from post-training on the same base.
CorrespondentWhich is itself a signal. If post-training alone can do that, the base-model race and the post-training race are separate competitions, and the second one is cheaper to enter.
HostThe other thing Zhipu reports is cyber capability scaling faster than expected. CyberGym at 84.5% versus 77.2%. ExploitBench at 54.4% versus 24.4%. ExploitGym at 105 of 130 tasks versus 29 of 39.
CorrespondentExploitBench nearly doubling is the number that should make people sit up. That's not a benchmark curiosity — that's the security surface moving.
HostAnd then GLM-5.3-Flash, which is a newly trained base rather than a distilled version. Hybrid sparse-plus-linear attention, Manifold-Constrained Hyper-Connections, and a 30-trillion-token multimodal pre-training corpus.
CorrespondentThe claim there is about the Pareto frontier of the Artificial Analysis Intelligence Index. Scoring 57 at $0.045 per discounted task — intelligence that previously cost roughly ten times as much.
HostTen times cheaper at the same measured level. That's the kind of number that invalidates a budget spreadsheet.
CorrespondentAnd it's the second time this episode we've said "roughly ten times." That's the week's real theme, sitting underneath the Graham pile.
HostLet's do the disagreements, because the brief flags a real one inside Graham's own corpus. "Being Popular" argues a language must be popular to be good, since user feedback drives improvement — and speculates that a critical mass of roughly twenty independent users is needed.
CorrespondentTwenty. That's a remarkably small number, and I think that's the point. It's not about market share, it's about having enough people to find the bugs.
HostBut "If Lisp Is So Great" argues the opposite: that a language's popularity bears little relation to its quality. And he makes the comparison to Jane Austen's modest sales versus John Grisham's bestsellers.
CorrespondentWhich is a great line and also a slightly unfair one, because Austen did eventually sell enormously. But the underlying point stands — popularity and quality are different axes, and they're only loosely coupled.
HostThere's also a softer tension. "The Hundred-Year Language" predicts Java will prove an evolutionary dead end. "Revenge of the Nerds" holds that mainstream languages had almost caught up to Lisp by 2002.
CorrespondentAnd those aren't strictly incompatible. Something can converge toward Lisp and still be a dead end. But they do point in different directions, and the brief is right to flag it.
HostThe honest read is that Graham is running two arguments at different timescales. Popularity matters for a language's evolution in the short run. Quality determines whether it survives the long run. He's not being inconsistent so much as answering two different questions.
CorrespondentI'd accept that, with one caveat: he never quite says which question he's answering in which essay. So the reader has to do that work.
HostSo what should Peter actually do with this week? The brief lists five signals. Let's run them.
CorrespondentFirst: re-run cost assumptions. GLM-5.3-Flash claims 57 on the Artificial Analysis Intelligence Index at $0.045 per discounted task — roughly ten times cheaper than intelligence previously available at that level. Routing and eval budgets built on last quarter's price curve are stale.
HostSecond: treat open-weights coding models as a security surface. GLM-5.3 reports cyber capability scaling faster than expected, with ExploitBench nearly doubling. That's a capability you now have to assume is in the wild.
CorrespondentThird: build the runnable check before the agent. The Claude Code doc says context is the scarce resource, and the loop should close on tests, exit codes, linters, or screenshot comparisons rather than a human verifier.
HostFourth: design for iteration. The AI Fluency Index finds iterating conversations carry roughly double the fluency behaviors and 5.6 times the questioning rate. One-shot interfaces may be suppressing the behavior you want to measure.
CorrespondentAnd fifth, the caveat we opened with: don't treat the Graham batch as complete. The two ANSI Common Lisp chapters contain only site boilerplate — navigation links, a Y Combinator prompt, a copyright line. No chapter prose. And the Lisp-for-web-based-applications piece is truncated to its introduction.
HostSo if you're building a corpus out of this, you're building it out of a partially empty set. Worth knowing before you draw conclusions about what Graham did or didn't say this week.
CorrespondentWhich is a nice reminder that the pipeline is part of the finding. What got captured shapes what you think happened.
HostLet's run the sources. The bulk of the week is the Paul Graham batch. On the technical side: "Programming Bottom-Up," "The Roots of Lisp," "What Made Lisp Different," "Beating the Averages," "Succinctness Is Power," "The Hundred-Year Language," "Revenge of the Nerds," and "Java's Cover."
CorrespondentOn the startup and craft side: "How to Start a Startup," "Ideas for Startups," "Hiring Is Obsolete," "Writing Briefly," "A Version 1.0," "The Submarine," and "What You Can't Say." Plus the political one, "This Year We Can End the Death Penalty in California."
HostThe disagreement pair is "Being Popular" against "If Lisp Is So Great." And the incomplete captures are the two ANSI Common Lisp chapters and the Lisp-for-web-based-applications piece.
CorrespondentFrom Anthropic: the education report — the AI Fluency Index, out of Anthropic Research — and the Claude Code best-practices document from Anthropic Engineering.
HostAnd from Zhipu, or Z.ai: the GLM-5.3 announcement, "Frontier Coding with Emergent Cyber Capabilities," and the GLM-5.3-Flash announcement on frontier intelligence at flash cost.
CorrespondentThat's the week. A pile of essays arguing about Lisp, a measurement of how people actually iterate with these tools, and a price floor that just moved. We'll see you next week.
ndovu