Anders Bjarby presents Idea Generator, an open-source tool where AI models combine ideas, a calibrated jury scores them, and winners can become real project briefs.
Key Takeaways
- AI innovation can be framed as combination plus direction, judgment, and a path to action.
- Calibrating an AI jury against reference ideas is essential to avoid inflated scores.
- Human oversight remains central: the person sets the goal, runs the jury, and can always overrule a medal.
- Idea Generator is an open-source, self-hostable pipeline that moves from raw ideas to buildable project briefs.
What the video covers
- The video opens by asking whether AI can innovate, and argues yes, using Fredrik Härén's definition of an idea as two known things combined in a new way.
- Idea Generator works from a board with a goal; seed ideas are judged by an AI jury scoring originality, feasibility, and impact from one to five each.
- Users drag ideas onto each other so the model crosses them into new ideas, and branching produces more concrete variants that score higher.
- The jury is calibrated against weak and strong reference ideas, because an uncalibrated AI jury once gave a gold medal to the bare word 'fire'.
- Scouts feed in outside input: an X scout for trends (only posts that appear among search citations) and a research scout for AI papers aimed at the goal.
- An evolution autopilot on the server breeds strong ideas across generations and keeps running even when the browser is closed.
- Each board has spending and rate limits and can choose its own model per role, from OpenRouter, Anthropic, OpenAI, Azure/Microsoft Foundry, or local models like Llama, MLX, and LM Studio.
- Ideas can migrate between boards, and an idea can be spun off into its own board where it becomes the goal.
- Gold ideas can be turned into Claude Code project folders with a CLAUDE.md brief covering the minimal build, first steps, open questions, and the jury verdict.
- Summaries, trend views, and reports across boards show medals, projects, spending, and conclusions, and an exhibition mode lets visitors add seeds live.
Full Transcript — Download SRT & Markdown
Speaker A
Hi, this is my own voice cloned with AI. I want to start with a question I hear a lot: Can AI innovate?
Speaker A
Many people say no. AI only remixes what already exists. I think it can. Let me show you why.
Speaker A
In the view of Fredrik Härén, the Swedish creativity author behind the idea book, an idea is two known things combined in a new way.
Speaker A
His view is often summed up as a recipe: a person with knowledge plus input. And in his view, good ideas don't just appear by accident. They need to be drawn out.
Speaker A
So, by his own definition, combination is the heart of an idea. And combining is exactly what language models do, tirelessly. What AI lacks on its own is a direction, a strict judgment of what is good, and a way to act on it. That is what I built.
Speaker A
This is Idea Generator. Everything starts with a board and a goal. Here, it's an audio-visual product for interaction in public spaces. Every step works toward that goal.
Speaker A
I plant a seed, just a short thought: benches that remember sounds. A jury judges every new idea, and my seed gets seven out of 15. "Intriguing," it says, "but vague about how it works." Fair enough.
Speaker A
I ask it to branch out, and three new ideas grow from the seed. They're more concrete, and they score higher.
Speaker A
Now the heart of it. I drag one idea onto another, and the model crosses the two into a new one.
Speaker A
Every verdict can be explained. Originality, feasibility, and impact, one to five each, with a reason and keywords. The medal is computed in code, not by the model, and I can always overrule it.
Speaker A
That matters. Left uncalibrated, an AI jury gave a gold medal to the bare word "fire." This jury is calibrated against weak and strong reference ideas. Weak ones stay at five out of 15 or below.
Speaker A
Strong ones land near 10. Gold is rare on purpose. Ideas need input, so every board has scouts. The X scout looks for trends and keeps a post only if it's among the search engine's own citations.
Speaker A
The research scout brings in AI research papers aimed at the goal. I drag one onto the board and it becomes a seed, with its source judged like everything else.
Speaker A
Then I switch to automatic. An evolution autopilot on the server breeds the strongest ideas and now and then branches a winner. It keeps going even when the browser is closed.
Speaker A
Three minutes later, six generations. The average crept up, but the best is still a 10. This jury doesn't hand out medals for effort.
Speaker A
Hover over any idea and its whole family tree lights up, back through the generations to the seeds.
Speaker A
It runs within limits I set for each board: seconds between steps, steps per hour, and dollars per day. And each board can choose its own model for every role.
Speaker A
The models can come from OpenRouter, Anthropic, or OpenAI. From Azure and Microsoft Foundry, the route to the models behind GitHub Copilot and Microsoft Copilot. Or from your own machine with Llama, MLX, or LM Studio.
Speaker A
Settings are kept in your browser, and the server never writes a key to disk.
Speaker A
Ideas can travel. I announce a strong one to my second board, Night City, which is about safer and more social nights.
Speaker A
It arrives as a guest, gets judged against the new goal, and crossbreeds with the local ideas.
Speaker A
Known things from two fields combined in a new way. I can also spin an idea off into a board of its own, where the idea itself becomes the goal. This one has already produced a silver.
Speaker A
At any point, I can ask for a summary. It's Frank: 27 ideas, it says, but they converge on one system. Then it sorts them into themes.
Speaker A
Gold ideas become Claude Code projects by themselves. Gold is rare, so here I make one by hand from the best idea on the board.
Speaker A
It writes a real project folder. The CLAUDE.md holds the brief: the smallest build that tests the core, first steps, open questions, where the idea came from, and the jury's verdict. Open it in Claude Code and start building.
Speaker A
The trends view shows which keywords are rising across all boards and how the jury rates them.
Speaker A
The report pulls it all together: medals, projects, spending, and conclusions across every board. It doesn't flatter me, either. It calls the jury lukewarm.
Speaker A
And for a lobby or a festival, there's exhibition mode. The goal, the top ideas, and a box where visitors type their own seeds. Pair it with automatic mode and the board keeps growing while people watch.
Speaker A
So, back to Härén's recipe. The knowledge is the models. The input comes from the scouts. And the person is the goal, the jury, and you. You can always overrule a medal.
Speaker A
Combination plus direction plus judgment plus a path to building. So, can AI innovate? I think it can. Try it yourself. It's open source under MIT, and it runs on your own machine.
Topics:AI innovationidea generationIdea Generatorevolutionary AIAI juryopen sourceClaude CodeLLMbrainstorming toolFredrik Härén











