The Sweet Side of the Bitter Lesson

Unlearning fear

Why Bash, React, and the Shoggoth didn’t trap us

After a year away from Anthropic’s Claude Code team, Shihipar begins with three predictions he got wrong. First, because early agents were trained heavily on Unix pipes, builders assumed every capability would collapse into CLI tools:

“There is just going to be bash and CLI's, and that's going to like take over the world. That's actually kind of not how it turned out. Like I actually prefer MCP to CLI's now.”00:01:40

Second, he worried React would become the last UI language because models wrote it best; instead, models write fluently across stacks—HTML, React, Swift, and even proprietary formats:

“They can write in HTML. They can write in React. They can write in Swift. They can write in the Figma proprietary JSON format.”00:01:40

Third, diffusion models did not swallow human creative control; LLMs drive Blender, Photoshop, and Figma through code and computer use. All three fears misread the Bitter Lesson as submission to an alien mind:

“These were all sort of like fears I had about like I think. The bitter lesson, you know, our alien intelligence.”00:01:40

“And so I think the like the question I come back to a lot is like okay like how can we be AGI built and human centered.”00:04:28

From the notes: “Bash is not all you need; React is not all you need.” The reading: a direct counterpoint to Philipp Schmid’s single-bash-tool harness—simplifying scaffolding makes sense, yet model breadth frees us from forcing every human workflow through a terminal.

Evidence: transcript 00:00:00, 00:01:40, 00:04:28

The thesis

Both an Industrial Revolution and a Renaissance

Software factories dominated the conference conversation—turning unreliable craft into dependable industrial output. Shihipar argues that two opposing shifts are happening simultaneously:

“What I like to think about it as AI is both at the same time an industrial revolution. And the Renaissance, right?”00:05:09

When software execution gets industrialized, commodity software gets cheaper while expectations shoot upward—making taste, delight, and intentionality worth more per marginal hour, not less:

“there's also kind of a Renaissance happening, right? Where people want and higher and higher quality software, and every marginal. Bit of like creativity, utility, like spirit. Like soul that you put into it is like is useful and like economically valuable, right?”00:05:09

From the notes: “Comparing it to the Industrial Revolution vs the Renaissance.” The reading: placed at the end of a week full of factory blueprints, this reframes Dexter Horthy’s judgment warning—factories supply the reliability, while Renaissance craft supplies what makes people actually care.

Evidence: transcript 00:05:09

The capability shape

Forget the jagged frontier—look at the smooth interior

Benchmarks chase the jagged frontier: can a model clone Mario Kart from scratch or port it to the Wii? Inside that outer edge, however, sits a rapidly filling smooth interior where capabilities stop being spiky:

“And what you see is that like it's not jagged below the frontier, right? Like if. It can write React. It can write Figma. It can write JSON. It can do sort of like all of these capabilities all together.”00:06:19

To show how broad pretraining recall has grown underneath narrow coding benchmarks, he displays twelve “blind” Claude models asked simply whether 16,200 latitude/longitude points are land or water—climbing from 60.5% on Sonnet 4.5 to 99.0% on Opus 5.5. As that interior smooths out, heavy QA scaffolding fades and builders can focus intelligence on how they work:

“And what's exciting, I think, about this is that we can use. Intelligence to understand and shape how we use intelligence, right?”00:09:42

From the notes: “Jagged edge vs the smooth interior.” The reading: smoothness is the “sweet side” of the Bitter Lesson—when models handle every format equally well, humans stop bending to the model’s quirks and choose representations that fit human thought.

Evidence: transcript 00:06:19, 00:09:42

The wall

Where vibe coding breaks: losing touch with the work

Building a 2D isometric fighting game inspired by Avatar and Smash Bros. exposed where pure prompt-and-wait vibe coding falls apart. Tell an agent to make a character throw a rock, and dozens of unasked game-feel decisions disappear under the hood—speed, charge scaling, whether rocks pierce summoned walls:

“I find that like the ball just makes a bunch of choices for me, and then I try it out. I'm like, wow, I have no idea what that actually does.”00:11:06

Sending a prompt and drifting off to another tab kills creative flow. Referencing Dexter Horthy’s point on messy judgment, Shihipar stresses that great work comes from tight control over what gets left out:

“It's more about the choices you made and what wasn't involved. Or was it what wasn't included versus like what was included”00:11:06

From the notes: bridges the Renaissance thesis into his live demo. The reading: even a former Claude Code builder hits the comprehension ceiling when chat prompts scatter product choices across opaque files.

Evidence: transcript 00:11:06

The architecture

Splitting `engine/` from `human/`—and comment-then-tab

To keep his hands on the steering wheel without writing camera math manually, Shihipar split his repository into two layers:

“I've got like this like engine part of the codebase, and then I have a folder. Like I think I call it spec, but I think I'm just calling it human or something. And that folder is the code that I write.”00:13:27

Claude builds and maintains the engine/ primitives underneath, while strict linter rules keep the human/ spec folder concise, written in real meters instead of raw screen coordinates, and immediately readable:

“It's supposed to feel like English. It's kind of like a spec, and I'll show you what it looks like. But it's at the density where, if I just read this code, I know what's happening, and I can edit it. I can tweak it. But more importantly, I just like you know can have a bit of a ground truth of like what the what my game looks like.”00:13:27

In the live demo, instead of opening a chat sidebar for every tweak, he writes an inline comment inside that custom scripting language and lets Cursor Tab complete the move logic synchronously:

“If you summon a wall in the middle, like queue it up as an action, and then cursor will sort of like. You know, do the tab completion. And this is kind of like it's kind of like spec-driven development, you know, but the spec is like. This, this scripting language, basically.”00:14:15

From the notes: “He's using comments and then tab-writing as an interactive development style.” and “There's no taxonomy of the different interaction modes. We need to follow up on this.” The reading: most AI discourse reduces coding to chat prompts vs. background agents; Shihipar’s DSL + comment-then-tab style points to an entire unmapped taxonomy of fast, human-directed interaction loops.

Evidence: transcript 00:13:27, 00:14:15

What comes next

Revisit old HCI research and bring the Renaissance

With a smooth interior able to generate hot-reloading languages, new database primitives, personal OS hacks, and malleable interfaces on demand, the creative frontier swings straight back to human-computer interaction:

“A lot of like old HCI research is now like very interesting to revisit.”00:17:04

Closing the summit, Shihipar hands the responsibility back to engineers: models will adapt to whatever creative medium we build for them:

“Because the models are not, you know, this alien intelligence that's like trying to force us into a specific way of working. They can kind of do everything, and they will, but our job is to like bring about the renaissance.”00:17:04

The smooth interior beats the jagged frontier: models now cross React, Swift, Figma JSON, and Blender cleanly, eliminating format lock-in (00:01:40, 00:06:19, 00:09:42).
Own a human-readable spec layer: pair an agent-built engine/ with a linted, English-like human/ DSL edited via comment-then-tab to keep taste and flow intact (00:11:06, 00:13:27, 00:14:15).
Industrialize the plumbing, Renaissance the surface: mine shelved HCI research for malleable software and new interaction modes (00:05:09, 00:17:04).

Read alongside

Claude Code’s Next Era (Latent Space) — Background on HTML as agent output and interactive elicitation.
Philipp Schmid: I deleted 90% of my harness — The single-tool Bash minimalism counterpoint.
Dexter Horthy: 12-Factor Factories — Companion view on human judgment vs. automated factories.
Steve Yegge: Escape from Trash Mountain — Spending token fuel on experiences, not bloat.