Stop Reading the Code, Thariq's Field Guide to Fable & Sholto's Classifier Pile-On
The "Stop Reading the Code" Fight
Theo spent the night of the 4th deliberately starting a fire: "How much better do the models have to get before you'll stop reading the code?" (370K views, 653 replies), followed by "At this point I'm genuinely convinced most of you would have kept reading the assembly code after C got popular" and the confession that "I'll be honest, I barely even read the code back when I wrote it by hand..."
The replies are the best part:
- Tim Sweeney (Epic Games) gave the historical anchor: "Moving from assembly language to compilers, there was a 24 month window where it mattered." Theo: "This checks out."
- Glauber Costa (Turso) took the other side: "for database or OS code, I will expect that never. A lot of us still routinely read assembly code..."
- David Cramer (Sentry) set the bar at "two orders of magnitude with actual real verification capabilities" — to which Theo shot back "Bold coming from someone whose code is gpt-3.5 level."
- A recurring pragmatist position from the replies: read only what breaks — "i stopped reading most of it two models ago. now i read the parts that broke. its the best way to learn what the model can and can't do yet."
Theo says a video is coming on the "you should still read your code" debate that will "piss both sides off" — a direct sequel to Geoffrey Litt's "understanding is the new bottleneck" thread from AIE that dominated Thursday's discourse.
Field Guide to Fable & the Orchestration Meta
Three days into the Fable re-release, the "how to actually drive this thing" genre produced its biggest hits yet.
Thariq's Field Guide. Anthropic's Thariq (Claude Code team) published A Field Guide to Fable: Finding Your Unknowns (1.2M views, ~5K likes) — an X article arguing "the map is not the territory": Fable is the first model where the quality of your output is bottlenecked by your own unknowns, so the job is discovering what you don't know before prompting. Techniques include having Fable quiz you before merging and using HTML artifacts to surface unknowns (examples thread). He credits Geoffrey Litt's AIE discussion for the quiz idea — this is the practitioner's companion to Litt's "understanding is the new bottleneck" thesis. Notable pushback in the replies: multiple people asking Anthropic to publish these on an agent-accessible site instead of X articles, since you can't feed an X article to Claude.
Simon's "use your judgement" tip. Simon Willison shared the most interesting Fable tip he's heard (227K views): tell it "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent" — and let Fable be the router. More notes on his blog. The replies became a mini-survey of delegation patterns:
- Dave Thomas Jr: "Fable dispatched Sonnet 5 but then didn't like the result so sent it off again with a bit of a slap on the wrist."
- Morgan Linton builds the model and effort level selection into the plan itself, with pre-populated model-switch commands.
- The counter-experience from Dominik Lukes: Fable let Sonnet 5 loose for 6 hours and "produced garbage" — it worked only after explicit instructions (plan better first, Opus for code, Fable for judgement).
- The inverse pattern also has fans: use Opus/Sonnet for everything and spin up a Fable subagent only for architectural specs.
40K LOC on a $20 plan. LLMJunky posted the most concrete orchestration receipt of the week: Fable proposed ten improvements to his StarSwap app, he picked seven, had Fable write independent plans, delegate each to GPT-5.5 High via Codex Exec in its own worktree, review/fix the results, and merge — 40,000 LOC written using 96% of a single 5-hour session on the Claude Pro $20 plan. Best reply, on the "working out of the box" claim: "The metric I'd watch the next morning: how many files you can still explain... Parallel worktrees are magic until merge debt shows up." LLMJunky: "whats regression? /s"
Theo's effort-level guide. Theo published a one-image guide to maximizing Fable usage whose effort-tier taxonomy tops out at "mental illness" (xhigh/max users happily self-identified in the replies). Consensus in the thread: medium/high is the sweet spot, xhigh+ is for benchmarks, and the pros pair Fable-as-planner with Codex as the workhorse — "Plan and review with Fable, give it tools to drive Codex. 1% used per hour." One cautionary tale: "fable setting off 19 fables in a swarm ate a full 5 hr max 5x limit in under 10 minutes."
Matt Pocock's taxonomy crisis. Related: Matt Pocock asked how anyone categorizes models now — "I used to put models in the bucket of Opus-like, Sonnet-like, or Haiku-like. But now we have Fable. Now Sonnet 5 behaves like Opus." Best answers: "Fable is Opus-like, Opus is Sonnet-like, Sonnet is Sonnet-like", and the pragmatic "I treat different effort levels as different models". Matt's own follow-up: effort levels make everything harder, and on models vs harness he calls it 50/50.
Sholto's Feedback Thread & the Classifier Pile-On
Anthropic's Sholto Douglas asked for feedback post-Fable-jump: "what axes do you feel like we've made great progress on, and what do we need to improve on?" The thread turned into a remarkably consistent bug report, and the dominant axis wasn't capability — it was the safety classifiers:
- xjdr: "the risk classifier is making it impossible to use... even basic 'center a div in this react codebase' [fails] on safety grounds... any indication as to why so I can adapt my prompt" would help.
- A parade of scientists: near-100% refusals on neuroscience, computational neuroscience, aging research ("What good is a model that can't do biology?"), a genealogist blocked from helping an adoptee find her biological parents, and one user whose résumé cascades failures down to Sonnet because "biosecurity" is in his job title.
- The other big axes: keep-it-on-subscriptions pleas (a proposed $500 tier came up twice), writing quality ("incredible at math, physics, finance, coding... terrible at writing"), vision as the verification bottleneck, and needle-in-haystack web research where GPT-5.5 Pro still wins.
Buried in the replies, the best data point of the day — Shrivu Shankar's controlled /goal+ultracode experiment, one hard task (replicate an image as a 3D scene) across three stacks:
| Stack | Time | Cost | Result |
|---|---|---|---|
| Codex, GPT-5.5 xhigh | 19m 31s | $11.09 | OK-ish |
| Claude Code, Opus 4.8 ultracode | 1h 48m | $62.64 | broken |
| Claude Code, Fable 5 ultracode | 4h 23m | $268.18 | "not bad, but still far from the goal" |
His conclusion: Fable is still a step function above the other two, but it's "very eager to fan-out to more fables and $$$" — which reinforces the value argument for Codex and explains why the orchestration meta above exists at all.
Honorable mention: Danielle Fong is building "DJ Claude" — Opus 4.8 fast-mode execution plus Fable with variable thinking, in tmux, with the faders on her DJ deck controlling the thinking level. "Perfect kind of 4th of July weekend project."
Harness Corner: Phantom Tool Params & Skill Evals
Armin's phantom parameters. Armin Ronacher has a genuine mystery on his hands: a handful of pi users report Anthropic models injecting extra tool parameters into edit calls, and he can't reproduce it — except that a user-supplied session, when continued, does show it. Replies point at context-prefix effects in resumed sessions; it's an open issue if you have ideas. This follows his AIE-prompted post on pi's intentionally-strict edit tool and the related observation that "Fable really likes its comments. Damn."
Skill evals are unbelievably hard. Matt Pocock, replying to the Effect team: "Evals on skills are hard" is the understatement of the year. His escalation ladder: "Evals on a classifier is trivial. Evals on a coding agent is extremely hard. Evals on a harness-agnostic skill is unbelievably hard" — and it's a data problem, not a token problem. David Cramer pushed back (no harder than any high-value eval). Two useful concretes from the thread: David K's baseline strategy ("does the skill produce better results vs no skill at all, per rubric?") and Will Hampson actually benchmarking the 57K-star ponytail plugin on DeepSWE: double the baseline's solved problems, 26% fewer LOC, 12% fewer tokens, 10% cheaper, 7% faster.
Agents Get Their Own Computers
Peter Steinberger: "Give your agent its own computer to REALLY end to end test stuff" (136K views) — a macOS VM via crabbox.sh (which has a Parallels backend), with the killer feature in all caps: "IT WILL CLICK ON ALL THE ANNOYING MACOS ALERTS FOR YOU." For mobile, he says Codex + computer use "works amazing." The replies are full of people converging on the same pattern — VMs as ephemeral, security-isolated E2E environments, and one user's "minions": one pulls CI, the rest test it with computer-use end-to-end.
In the same spirit, LLMJunky moved his Codex Plugin Marketplace off Vercel onto his Hetzner box by just asking Codex: "Migration, DNS, auto-deploy, SSL, traefik, etc. Codex deftly handled it all via SSH and Chrome in 12 minutes with minimal guidance."
The Creative Writing Gap
Lee Robinson — now doing ML at Cursor — asked the weekend's best non-coding question: Are current LLMs incompatible with great creative writing? Even council-of-models grading ten drafts yields "lowest common denominator slop," because coding has verifiable rewards and writing doesn't. His darker follow-up: "You can't help but wonder if we already have the best writer of all time sitting in the weights, but it just gets fried with RL and inference optimizations."
The standout reply came from Karina Nguyen (ex-OpenAI/Anthropic): non-fiction is fixable — the gap comes from conflicting RL objectives, since chat-pleasant traits fight the precision/density/restraint good technical writing needs (they fixed this for a GPT-4o version). True creative writing is a different problem entirely: coherent worlds and emotional arcs over hundreds of pages, with no training data capturing the process ("hard to simulate Pixar's braintrust process for AI") and no lab investing because it's unmeasurable and low-value economically.
Jerry Liu offered the practical 70% solution: a writing-style skill hill-climbed against your real writing samples — which fixes style but not "the inherent ability to articulate clear, differentiated insights, which is probably an inherent posttraining problem."
Also Worth a Look
- pxpipe — ~60% Fable cost cut by "transparently turning the code into an image and having the model OCR it." Steipete, retweeting: "WILD idea. also hilarious."
- swyx on tools for thought — "'tools for thought' people spent like a decade making cool pretty demos with canvases and then got completely mogged by low contrast poorly designed CLIs just winning because they do commodity thinking for you". Sharpest reply: "places to think vs things that think."
- AIE World's Fair session index — individual talk videos aren't out yet, but there's a timestamped index of all three days of main-stage recordings (via), so you can jump straight to any session.
- Video: Fable 5 UltraCode game playthrough — full playthrough of a game where every asset was generated by Fable, from the reviewer who also made his own one-image Fable usage guide.
- The roast meme — Theo had Fable roast him based on an old codebase ("It kinda cooked me"), and Steinberger escalated: "I fed Fable 80,000 of my tweets so it could roast me even more. 💀"
- Open-source realtime voice — Thomas Wolf (Hugging Face): "Most people should probably update their priors on the state of open-source speech-to-speech" — a fully open-source realtime voice demo built with Cerebras.
- Robert Miles' bar for software — "If your software product doesn't have a text box where I can write a feature request and have it show up in a few minutes, I'm going to constantly be tempted to get Claude to replace you" (RT'd by steipete).