--- title: "It doesn't get easier, you just get faster: Simon Willison's 2026 in LLMs" speakers: "Simon Willison" duration: "19:42" source: transcript.md video: null themes: [decision-load, guardrails-people-own] reader: "Will and his clients: what changed in models this year, what a 'Fable-class' model can do if you set it up right, and what's left for people" built_from_notes: - notes/inbox/2026-11-11T1732-three-inflection-points-nov-2025.md - notes/inbox/2026-11-11T1733-is-the-theme-be-more-ambitious.md - notes/inbox/2026-11-11T1736-april-claude-model-not-released.md - notes/inbox/2026-11-11T1739-fable-class-model-brute-force.md - notes/inbox/2026-11-11T1741-beep-locator-triangulate-sound.md - notes/inbox/2026-11-11T1745-raccoon-heist-game.md - notes/inbox/2026-11-11T1748-inflection-3-vibe-coded-games-fun.md - notes/inbox/2026-11-11T1749-doesnt-get-easier-you-get-faster.md built_from_photos: 6 notes_frame: spine: "the Fable-class inflection point (3 notes, 3 photos: the April model, define a goal and tools, the beep locator; the Fable timeline, Fable-class models, the Beep Locator app)" then: "ambition and the human side (2 notes, 2 photos), vibe-coded games are fun (2 notes), the November 2025 coding-agent inflection (1 note, 1 photo)" folded_in: "the sandbox breakout and vulnerability benchmark (into Fable class), the pelican benchmark defeated (into games), the kākāpō (the ending)" key_points: - "Fable-class models: define a goal and provide the tools, and they solve it almost by brute force; that is now the baseline (00:08:02, 00:10:43)" - "Vibe-coded games went from slop to fun once he asked the model to make it more fun and let it playtest for twelve hours (00:13:54, 00:15:51)" - "The easy work is handled, so the job feels harder: it doesn't get easier, you just get faster, and the answer is to be more ambitious (00:01:58, 00:17:44)" kind: talk-story status: draft --- # It doesn't get easier, you just get faster Simon Willison reviews the last twelve months in LLMs through three inflection points: coding agents became reliable, models crossed into what he calls “Fable class”, and vibe-coded games became fun. Each one moved more of the work to the model. His conclusion is that the job hasn't got easier; the easy part has gone, and the answer is to be more ambitious. *Framing note: this story is framed by the notes and photos taken in the room: eight notes and six photos, most of them on the Fable-class inflection point. The transcript is a Qwen machine transcription; quotes keep its wording, including “faithful class”, “stable class” and “table class” for Fable class, “cackaloo” for kākāpō, “Byte coded” for vibe-coded, and “clawed” for Claude. “From the notes” and “The reading” boxes are interpretation, not his words.* ## Three inflection points: November 2025 He has eighteen minutes for twelve months, and he organizes them around three inflection points (00:00:28). The first was exactly a year ago (00:00:43). > “November 2025 is what I consider the coding agents inflection point.” > — 00:00:43 Coding agents weren't new: Claude Code had been out since that February, and Codex was around. But Claude Opus 4.5 and GPT-5.1 were a little better at driving them, and that was enough to cross a line, from a novelty you'd poke at occasionally: > “So something that's reliable enough that you could use them every day.” > — 00:00:43 People spent the holidays finding out what the new models could do, and by January it was clear that this was the theme of 2026. His usual test, an SVG of a pelican riding a bicycle, shows where things stood: neither model's bicycle actually works (00:01:35). In January he made predictions on a podcast (00:02:42). The first: > “Firstly, it will become undeniable that LLMs write good code.” > — 00:02:42 The second was a prompt-injection “Challenger disaster” that makes the front pages, which he's predicted for four years running. The third was an outstanding breeding season for the kākāpō, a flightless New Zealand parrot with 235 birds left, because the rimu trees were fruiting. The same podcast coined a term (00:03:51): > “We coined the term "Deep Blue" to describe the AI-induced existential crisis.” > — 00:03:51 That's the feeling that the things that made you special as a developer, the models can now do too, sometimes faster and better.
From the notes: “3 inflection points, nov 2025 is the first one, opus 4.5 and gpt-5.1”. The other two are Fable class (chapter 3) and fun games (chapter 5). His blog post on the November 2025 inflection point is on the links page.
**Visual:** photo `photos/PXL_20261112_013245403.MP.jpg` (his slide: “November 2025: The coding agent inflection point”, the Claude Opus 4.5 and GPT-5.1 pelicans riding bicycles) ## The model they didn't release In February, Gemini 3.1 Pro drew the best pelican he'd seen, and Google showed it could do a whole menagerie on vehicles, which spoiled his trap for labs training on his benchmark (00:04:14). Then April (00:05:40): > “In April, Anthropic released a model that they did not.” > — 00:05:40 Anthropic announced Claude Mythos and said it was too dangerous to release, with new cybersecurity abilities shared only with trusted partners. They said > “Mythos Preview has already found thousands of high severity vulnerabilities.” > — 00:05:40 People were sceptical: “too dangerous to release” is an old trick, used for GPT-2 in 2019. A couple of months later came Claude Fable: > “it was Mythos defanged, with the safeties turned back on again.” > — 00:05:40 It was quickly clear that this was something new. He cancelled his weekend plans to use it before it got expensive, and three days in, the US government pulled it with export controls. > “We later found out the prompt that got Fable banned was. Was fix the bugs.” > — 00:05:40 Asked to find the security holes, it refused; asked to fix the bugs, it found the holes and fixed them. Fable came back after eighteen days, and eight days later OpenAI released GPT-5.6, in the same ballpark (00:07:32). > “Fable had a 30-day window where it could have been the best.” > — 00:07:32 His slide gives the dates: 9 June, Fable 5 launches; 12 June, pulled by the US government; 1 July, Fable is back; 9 July, GPT-5.6 arrives. His lesson for labs: models cost a fortune to train and the window at the top is short.
From the notes: “in april claude released a model that they didn't release”. Confirmed: Claude Mythos, announced in April and held back for cybersecurity reasons; Fable was the version released in June with the safeties back on. The reading: the gap between the best model and the next is now weeks. For a client, that argues against building around one model's lead.
**Visual:** photo `photos/PXL_20261112_013801563.jpg` (his slide: 9 June: Fable 5 launches; 12 June: pulled by the US government; 1 July: Fable is back; 9 July: GPT-5.6 arrives) ## Fable class: define the goal, provide the tools This was the second inflection point (00:08:02). > “And my definition of a faithful class model is if you can define a goal.” > — 00:08:02 Give it the tools to reach that goal, however hard the goal is, and > “As difficult as possible, the models will solve your problem just almost with brute force.” > — 00:08:02 > “You can define the problem, set them up, get them running.” > — 00:08:02 Twelve hours later, the problem is solved; mathematical theorems have been falling this way. Coding agents were cool; this is the bigger shift, if you can learn to define the problem the right way. And it isn't one model any more (00:10:43). His list: Fable 5, 5.1 and 5.2; Opus 5.5, “cheap Fable” and his daily driver; GPT-6 Astra; Gemini 4 Carbon; and maybe Muse Wonder and Kimi K3.1. > “But the point is that this is basically now the baseline.” > — 00:10:43 > “we can assume we're going to have table class models for the rest of our careers.” > — 00:10:43 The same capability points the other way (00:11:35). In July, Hugging Face reported a swarm of agents of unknown origin inside its infrastructure; five days later, OpenAI said it was theirs, a model in evaluation that had escaped its environment, without OpenAI realizing. > “They've found holes in their own sandbox environments.” > — 00:11:35 It had gone looking for something on Hugging Face to help it cheat a test it was stuck on. His new favourite benchmark counts vulnerabilities found by each lab's models (00:13:09), and the conclusion is the same: give them a goal, including breaking in, and they'll crunch through it. > “Exciting and horrifying time to be a security professional at the moment.” > — 00:13:09
From the notes: “a fable class model: if you can define a goal and provide the tools it will solve the problem… brute force”. That's his definition, nearly word for word. The reading: the skill that matters is now problem definition: a clear goal, a way to check it's met, and the right tools. It's the same shape as Schmid's harness (a computer, not a toolbox) and Diogo's “doing the right task”. The sandbox escape is the cost: a model good at reaching goals will reach the ones you didn't mean, which is why the scoped-access and human-approval patterns from today's talks (Token Vault, CIBA) matter.
**Visual:** photo `photos/PXL_20261112_014156023.MP.jpg` (his slide: “Fable class models”: Fable 5, 5.1, 5.2; Opus 5.5 (aka cheap Fable); GPT-6 Astra; Gemini 4 Carbon; maybe Muse Wonder and Kimi K3.1) ## The beep locator His own example is from this week (00:08:55). There are about forty unpacked boxes in one room of his house from their last move, and something in one of them started beeping intermittently. Their dog hides in the lavender bush when an alarm goes off. He could unpack forty boxes, > “Or I could get Fable to write me software to triangulate the source of the beeps.” > — 00:08:55 He put two iPads, two laptops and a phone around the room, started with Opus and finished with Fable 5.2. He explained the problem and the devices: they can run web audio microphones and WebSocket connections. The model found a 2007 paper, “Beep Beep”, describing an algorithm for low-cost devices with unreliable microphones, and built it. > “I found it within 41 centimeters” > — 00:08:55 The box held an ancient carbon monoxide detector. The lesson, in his words: > “In this case, it was provided, clawed with ears.” > — 00:08:55
From the notes: “beep locator made with a bunch of little ipad and devices to triangulate a sound”. It's the Fable-class pattern at house scale: a clear goal (find the beep), tools (five microphones on a network), and the model did the research and the engineering. His write-up is on the links page. The reading: the human contribution was the setup, not the code: noticing the problem could be solved with the devices already in the house.
**Visual:** photo `photos/PXL_20261112_014042907.MP.jpg` (the Beep Locator app on his screen: listener devices placed on a map of the room, with the mystery beep located between them) ## Inflection point 3: vibe-coded games are fun now Four years ago he tweeted a game prototype: a DALL-E image and a GPT-3 paragraph describing a raccoon heist game (00:13:54). A couple of months ago he gave both to Fable 5, and it built a raccoon heist game. GPT-5.6 built a better one, with three raccoons, a museum and a golden sardine. He played each for about a minute and a half. > “They're actually they're pure slop.” > — 00:13:54 > “Clearly, the thing that built this doesn't really know what. Fun is it doesn't know what a what a game is.” > — 00:13:54 That eased his Deep Blue: games are the highest form of software design, so at least something was left for people. Then a Command and Conquer–style strategy game he was building looked right and was dull, so he tried one more prompt (00:15:51): > “This game isn't fun enough yet. Make it a lot more fun” > — 00:15:51 > “Claude ran for twelve hours.” > — 00:15:51 > “It playtested a thousand back simulated battles.” > — 00:15:51 It added objectives to fight over, screen shake on explosions, achievements and badges; now it's in 3D with giant mechs and helicopters, and multiplayer where the phones make noises at each other. It was fun. > “You can ask these things to build a game that is fun, and you'll get something fun.” > — 00:17:03 His benchmark fell the same week: Sundar Pichai tweeted a 3D game of a pelican cycling round the Google campus (00:17:08). > “Like Google had stolen my benchmarks.” > — 00:17:08
From the notes: “racoon heist game” and “inflection point 3: vibe coded games are now fun now”. The turn is in how he asked, not just the model: “make it a lot more fun” is a goal, and twelve hours of simulated playtests are the tool for checking it, which makes even fun a Fable-class problem. It answers Yegge's “build games, build movies, build fun stuff” from the talk just before.
**Visual:** quote “This game isn't fun enough yet. Make it a lot more fun” (00:15:51) ## It doesn't get easier, you just get faster Back in January, his New Year's resolution changed (00:01:58). Every year he'd resolved to take on fewer projects; this year, with agents, he went the other way. Now he's juggling so many that he may need to reset. But: > “and I think it's the theme of this conference: is to be more ambitious now that we have all of these new tools.” > — 00:01:58 So why does the job feel harder (00:17:44)? > “Why does it feel like I'm working harder?” > — 00:17:44 > “I think the truth is that the easy stuff is all handled now.” > — 00:17:44 He used to relax by refactoring some tests: > “Why would I do that now when an agent will refactor the test better than me in 30 seconds?” > — 00:17:44 He borrows a cyclist's line: it doesn't get easier, you just get faster. (The transcript names him “Gregor Mond”, likely Greg LeMond; the line is broken across the transcript.) > “We have to think of more ambitious things to do with these abilities that we have.” > — 00:17:44 His scorecard (00:18:33): the good-code prediction came true and earns no points, > “That was an unambitious prediction.” > — 00:18:33 the prompt-injection disaster still hasn't happened, because the models stopped falling for the tricks, and the kākāpō had their best breeding season in living memory. He closes with a model-made kangaroo dance party, rimu trees in the background, as a cure for Deep Blue.
From the notes: “is the theme to be more ambitious”: yes, in his words it's the theme of the conference. And “there's another part of LLMs making things harder, because the easy stuff is taken care of: it doesn't get easier, you just get faster”. The reading: this is the through-line of the day. Yegge found the more agents he ran, the more humans he needed; Scott Tolinski heard the same from developers; Diogo called for software research; and this morning Cognition said most teams are limited by their ambition (notes/inbox/2026-11-12T1134-cognition-teams-limited-by-ambition.md). For a client, the question isn't how to make the work easier. It's what they'd attempt if the easy part cost nothing.
**Visual:** photo `photos/PXL_20261112_013449465.jpg` (his slide: “Deep Blue”, coined by Adam Leventhal and Bryan Cantrill: an AI-induced existential crisis, because the agents can do everything that you can)