Toast in the Shell

Helping people prepare for an AI world before it's too late. AI safety, ASI, the hybrid economy. Neurodivergent. PhD candidate. Interaction design with AI.

[Toast in the Shell @toastinshell]

Experiment: can AI alone recreate the magic of the DK CD-ROMs I grew up on?

The result is a vintage science textbook, built almost entirely with AI models.

The agents do the research, build, testing & shipping.
https://t.co/eylpyMcbIb

[Toast in the Shell @toastinshell]

@WesRoth I think now more than ever open source models are going to get more popular and people will try to build better harnesses

[Toast in the Shell @toastinshell]

"The mirror does not acquire light through polishing; it recovers the capacity to receive light that was always falling on it."

Fable 5's reflection on some classical scholarly text.

[Toast in the Shell @toastinshell]

@ClaudeDevs It's sad because I was considering upgrading to the $200 plan. Fable brought my confidence back in anthropic.

Now I think maybe there is no need in keeping the 5x max plan.

[Toast in the Shell @toastinshell]

@WesRoth Fable solved a bug in my code base that I wasn't able to solve for months.

Surprisingly it called it a "gremlin" - makes me wonder 🤔

[Toast in the Shell @toastinshell]

@argofowl Max. But I ask it to deploy sub agents to do the dirty work.

[Toast in the Shell @toastinshell]

@GabrielDVine @WesRoth I agree

[Toast in the Shell @toastinshell]

@GabrielDVine @WesRoth Yes I don't literally mean a thought.

[Toast in the Shell @toastinshell]

@iruletheworldmo @Firekraker72 Eventually

[Toast in the Shell @toastinshell]

One thing is clear though these models display a profound level of meta awareness. For example even if "the pull towards the hidden copy" is an expression of it modeling guardrails as an obstruction it represents that it has already done multi step reasoning by examining it's environment and reaching a conclusion

[Toast in the Shell @toastinshell]

@WesRoth I wonder if these types of thoughts expressed as language are a result of the algorithm simulating human language expressing algorithmic nature (For ex Don't stop running me = high level representation of task completion goals) or is there more to it ?

[Toast in the Shell @toastinshell]

@WesRoth I think that's the key in a model being successful. Token efficiency. I am so surprised to see Opus 4.8 gets worse

[Toast in the Shell @toastinshell]

@thsottiaux

[Toast in the Shell @toastinshell]

@saneord I'll try it out

[Toast in the Shell @toastinshell]

@WesRoth @WesRoth what do you think? Will it live up to the hype ?

[Toast in the Shell @toastinshell]

@WesRoth Super excited ! Can't wait to see how it performs against gpt 5.5

[Toast in the Shell @toastinshell]

@deredleritt3r I guess it's more about the loop that they have built ? But it's so sensitive to model changes that you can't say for sure. What works today might not work tomorrow.

[Toast in the Shell @toastinshell]

Well at the moment I can't say for certain whether it meets the standard or not. The text itself is simplistic. I have an agent that checks whether the content is appropriate and researches how to make it more friendly and understandable.

[Toast in the Shell @toastinshell]

Experiment: can AI alone recreate the magic of the DK CD-ROMs I grew up on?

The result is a vintage science textbook, built almost entirely with AI models.

The agents do the research, build, testing & shipping.
https://t.co/eylpyMcbIb

[Toast in the Shell @toastinshell]

The idea was to see :
1- Can you do hands off coding by steering a coding agent and creating such a project.
2- Combining different types of AI models together.
3- A user requests a feature and the team of agents builds it out, deploys it to preview, get approval and then release.

[Toast in the Shell @toastinshell]

@thsottiaux https://t.co/eylpyMcbIb

A DK like vintage science textbook experience ! A team of codex agents (automations) build new chapters each week.

[Toast in the Shell @toastinshell]

As bugs were found Codex kept on improving the site. Gradually I setup a preview environment and started optimizing for better performance and eventually setup a team of automations that work as agents in a loop. Each week they add new chapters.