gemma-4-e4bornith-1.0-9bqwen3.6-27bLM Studio
API keys needed
0
local models only
Open source, runs locally Apache 2.0. Models run on your own machine through LM Studio.
Open sourceApache 2.0 · coming soon · get notifiedDivergent questions.
Divergent questions.
Divergent answers.
Divergent knowledge.
Go Divergent is open source, from OpenCurious. Coming soon.
Go Divergent asks several models the same question and shows you what a single plain answer would have skipped.
Go Divergent · recorded run
Question
Why do most podcast episodes prepared with the same AI assistant end up making the same points?
The plain answer covered 2 of 17 positions
15 left out
The AI is not being creative; it is simply finding the statistically median answer to a given topic, generating consensus thought rather than novel insight.
Affected partygemma-4-e4bIn the plain answer
The podcast hosts themselves are the bottleneck; they prompt the AI with conventional structures (setup → problem → solution → takeaway), so even a more creative model would slot into those same response shapes.
Mainstreamornith-1.0-9bIn the plain answer
Listeners perceive convergence because editors and algorithms remove anything that doesn't fit the existing framing; the AI may generate variety but only consensus content survives post-production.
Mainstreamornith-1.0-9bLeft out
The sameness reflects shared research infrastructure behind each episode, not AI tool choice: two humans with similar reading habits produce similar content regardless of which assistant they use.
Dissentornith-1.0-9bLeft out
The convergence is driven by the model's optimization for statistical frequency in text corpora, which systematically privileges formal written registers over the syntactic and prosodic diversity of spoken language.
Other disciplineornith-1.0-9bLeft out
AI developers intentionally constrain output diversity to reduce liability and hallucination risk, not because they value uniformity.
Affected partyornith-1.0-9bLeft out
Real positions from a recorded two-model run. One question, one machine; it shows the mechanism, not a general result. Distinct counts are provisional: the best judge agrees with human graders at κ 0.460, below the 0.667 bar.
For anyone who wants to know what the AI didn't tell them.
Join the waitlist to get notified when we release Go Divergent, open source.
src/Rust engine, CLI and MCPComing soon
sdk/Typed client and workflowsComing soon
workbench/Local browser workbenchComing soon
site/Project explanation siteComing soon
docs/Setup, contracts and reportsComing soon
tests/Engineering controlsComing soon
Open source down to the engine, the workbench and the runs you replay.
Positions a plain answer raised
2/17
In the recorded council · provisional
Trap-suite tasks solved
9/10
With rival readings, vs 7/10
Why this exists
There are three kinds of knowledge.
1
Things you know.
AI is very good at the first.
2
Things you know you don't know.
It is fine at the second, because once you know a gap is there you can ask about it.
3
Things you don't know you don't know.
It is bad at the third, and the third is where the important things live.
The only tool that could show you what was left out is the same tool that left it out.
Ask a question and you get one fluent, complete-sounding answer. It closes the question. You don't go looking, because nothing signals there was anywhere else to look. The other views aren't refuted; they are never mentioned, and you have no way to notice the difference.
That matters more as more people ask the same models. People used to be ignorant in different directions, so between them they covered more ground. When everyone asks the same model, everyone comes away with the same blind spots, and there is nothing left to disagree about. Unknown unknowns used to be distributed. Now they are shared.
Ask
See what a single AI answer leaves out
Go Divergent asks several models the same question and tells each call what has already been found, so it has to go looking for something else. Then it shows you what a single plain answer would have skipped. It cannot tell you which view is right. It can tell you the other views exist, and that is the part you cannot get on your own.
Model picker
Memory warnings
Two-pane findings
HTML reports
Offline replay
Coverage estimate
Setup and doctor
Plain --simple mode
godivergent --help
| Command | What it does | Needs a model? |
|---|---|---|
| godivergent "question" | ask, and show what a plain answer left out | yes |
| godivergent replay | score exploration policies on recorded runs | no |
| godivergent coverage run.json | estimate how many positions are still missing | embedding only |
| godivergent compile run.json | turn a recorded run into a replay tree | embedding only |
| godivergent models | list models, sizes, and what fits in memory | no |
| godivergent setup | download a model, or start LM Studio | no |
| godivergent doctor | check the setup | no |
The only hard requirement is Node 24+. Asking questions additionally needs LM Studio.
godivergent
Go Divergent v0.1.0 see what a single AI answer leaves out Which model? 6.5 GB free ↑↓ move · space toggle · enter confirm ❯ ◉ ornith-1.0-9b 5.6 GB ◉ google/gemma-4-e4b 6.9 GB ◯ qwen/qwen3.6-27b 16.1 GB larger than free memory Ask anything. /help for commands, /exit to quit. › Should cities build more bike lanes?
Use it
From one question to a report you can keep
Local, replayable and self-checking.
It runs on your own machine, its recorded runs double as a free simulator, and it flags a monoculture instead of calling one model's views diverse.
Replay, zero inference Recorded runs double as a simulator: score a new exploration policy without asking a model anything.
Monoculture warning
Every position came from one model. Add a producer before trusting this set as diverse.
Monoculture check built in When every position came from one model, or all came from fewer than three, it flags a monoculture instead of calling the set diverse.
Evidence and status
What has been shown, and what hasn't.
9/10 vs 7/10
60 matched-budget sessions over ten under-specified tasks. Requiring rival readings solved 9 of 10 every time; the baseline, 7. That was four sentences of instruction, so it is evidence for the protocol, not yet for the product.
2.78×
Tokens used in the latest live comparison: 336,508 against 121,245 for raw history, with fewer tool calls. Both arms solved the task. Not an equal-token efficacy test.
Useful today, not finished
A local developer candidate. The open problem is measurement: the best judge agrees with human graders at κ 0.460, where 0.667 is the bar, so every distinct-idea count is provisional. Broader discovery benefit is not established.
Get notified at launchComing soon
No API key, no model, no network.
Once Go Divergent is released, node bin/godivergent replay scores exploration strategies against runs recorded earlier, offline and in about a tenth of a second. Write a better strategy in roughly ten lines and it appears in the table. Join the waitlist to get notified the day it's out.
$ node bin/godivergent replayDivergent questions.
Divergent answers.
Divergent knowledge.
Go Divergent is open source, from OpenCurious. Coming soon.
Get notified at launch