I have to be honest about Fable
In a Nutshell
Fable (Claude Mythos with extra safeguards) produces unusually coherent, mistake-free code that lets non-experts one-shot complete games in a single pass, generating the striking "inside the model's mind" feeling users report. Yet these demos remain 80 % solutions; the final 20 % of bugs and polish still demands domain expertise the model cannot supply, which is why nothing ships. Mythos is also ruinously expensive—thousands of dollars per demo—and its best capabilities are gated behind export-controlled, strategically rationed compute, so human direction and economic reality will continue to limit what actually reaches production.
These notes were generated by AI and may contain inaccuracies.
An article on X described Fable with a sense of longing, noting that it had been removed and that people felt something meaningful when using the model. The speaker confirms experiencing this same feeling, describing Fable as a beautiful model. Fable refers to Claude Mythos with additional safeguards that prevent access to security and cybersecurity capabilities.
The speaker had Fable build multiple games. The first was Token Mania, a Factorio-style game where the objective is building an AI lab. Starting with resources on a grid including finite coal and infinite water, the game involves:
- Building a coal mine on coal deposits
- Constructing a coal plant
- Connecting via belts to transfer resources
- Building a data center that requires power
- Drawing wires from plant to data center
- Building a pump and pipes for water
- Purchasing GPUs and a medium cluster
- Training a model within the game
All assets were programmatically generated with no actual art.
A second game was created as an ethereal free-floating experience with a peaceful vibe. The player floats in infinite space collecting relics with no specific purpose beyond the experience and music. The game features:
- A map and directional arrow for navigation
- Relics to collect
- Shock waves that disorient the player when contacted, inverting controls
The speaker describes feeling like they were playing inside Fable's mind when experiencing the game. The animations and music created a sense of beauty and longing. The speaker notes this wasn't psychosis and clarifies they will explain why this doesn't represent the death of software engineering.
The speaker acknowledges uncertainty about whether Opus or GPT 5.5 could achieve similar results. They note that Fable appears to make no mistakes in these projects. However, the speaker hasn't examined the code, which is presumed to be blackboxy, neural network-style code that is jumbled and not amenable to human understanding. The games might exist frame-for-frame in training data. The speaker chose games specifically because they are not a game developer, noting that LLMs most impress non-experts in the field they're working on.
A third game was built as an open world crafting game using 3JS. The speaker instructed Fable to keep it simple and make it touring complete, allowing players to:
- Hit trees to gain wood
- Gather stone
- Build a stone axe
- Build electronics including wires, NOT gates, AND gates, and OR gates
- Build a computer within the game
The game was inspired by Rust. No visual guidance was provided to Fable.
A demo shared on X showed a Fable-built 3JS game with no external assets running at 18 frames per second. The game featured procedurally generated clouds, mountains, and trees. The speaker notes this was nicer than some games built with real game engines. The water in this demo appeared wonky when viewed from different angles.
The speaker explains why impressive demos don't translate to shipped products. LLMs can produce 80% of a working product, but the remaining 20% involves bugs that are difficult to fix without expertise. Describing visual bugs like water appearing blue and green from above but translucent from straight on would be challenging to communicate to Fable. Non-experts would struggle to navigate the code to fix such issues. This explains why people create crazy demos but no one ships products with these models.
The speaker had Fable build a piano tutorial application. This was not mind-blowing and could have been done by any model. The speaker had difficulty finding sheet music at their skill level. Fable generated original compositions, some of which the speaker liked. The application included:
- An algorithmic composer generating procedural music during inference based on seeds
- A practice mode accepting MIDI piano or microphone input that advances based on played notes
- An improv lab with a seven-step wizard teaching improvisation
The microphone feature wasn't perfect and would sometimes advance incorrectly.
The speaker attributes the beautiful feeling of Fable to its output, associating game music and visuals with the model itself. This creates anthropomorphization similar to attributing beauty in a game to its developer. The speaker notes that while the model was trained on collective human work, Fable's distinguishing characteristic is making no mistakes and not spinning in circles during development tasks.
The speaker references an article stating that Fable is special in ways that don't show up in benchmarks. The article emphasizes caring about the shape of a model's mind, how deeply it perceives users and infers intent, how far it can think and iterate, and how alive the model feels. The article notes that Fable makes users feel like they're back in 2023, and many reacted as if their wings had been torn off when it was disabled.
The speaker was only highly impressed with Fable's programming capabilities for one-shotting games. When tested for creative writing using previous video scripts, it felt like every other model with stereotypical LLM writing that wasn't particularly funny. The speaker notes that most progress is now in post-training reinforcement learning rather than base training data. Models excel at verifiable tasks like math and coding, but writing is harder to measure and shows less noticeable improvement.
The speaker references Anthropic's leaked Series C pitch from 2023, which predicted that companies training the best 2025-2026 models would be too far ahead for others to catch up in subsequent cycles. The speaker acknowledges Anthropic has been remarkably prescient about Mythos being unique and powerful, and that these capabilities continue scaling.
Mythos is described as ludicrously expensive and potentially economically unviable. A $200 subscription provides $8,000 worth of usage. The games demonstrated would have cost $2,000 to $3,000 in API credits. While spending $100,000 can build cool demos, the economic reality means most users cannot afford extensive usage. The bill will eventually come due for the impressive demos shared on X.
The speaker discusses how chip exports will be rationed, narrowed, politicized, and treated as instruments of national security. The Biden administration sketched export caps even for allies. Advanced compute will be treated like uranium, monitored, licensed, tracked, guarded, and kept within America's borders. Those controlling it will use all means necessary to prevent others from building their own stack.
The speaker maintains they don't buy the narrative that AI replaces humans. The common denominator in all demonstrations was the speaker prompting and directing the model at every step. Without human direction, the model produces nothing. The speaker references François Chollet, co-founder of the ARC Prize, stating that near-term AI is digital leverage and a force multiplier, but force without direction is noise, and it still requires a human in the loop at every level to be useful.
Keep Mo Bitar in your library
Save the videos and channels worth coming back to, and find them again in one place.
