LSD Mode: The Vector-Space Collider
Cross unaligned vector spaces over your own corpus, then re-rank the collisions for ideas.
- Difficulty
- Advanced
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 60%
The mechanism is deliberate misalignment. Instead of retrieving the nearest neighbours to a query — which returns what you already think — you embed a large personal corpus into multiple vector spaces, then cross the spaces that point in different directions and read the collisions as candidate ideas. Because that produces far more candidates than are useful, the second half of the mechanism is ranking: several frontier models independently re-rank the pile, and the survivors are cross-referenced against an external search index to check they are actually novel and defensible. The output is a short list of ideas per run, generated at a scale no unaided person can match, with the taste of your own archive baked into the substrate.
Origin
Extracted from Naval. Gary Tan describes building the retrieval system inside his personal tool, over roughly four hundred thousand markdown files of his own writing, and nicknaming the cross-space brainstorm 'LSD mode'.
Core principles
- 01Novel ideas live at the collision of spaces that do not point the same direction.
- 02Your own corpus is a better substrate than the open web because it encodes your taste.
- 03Generation is cheap; ranking is the scarce operation.
- 04Multiple independent models disagreeing is signal, not noise.
- 05An external search pass keeps the output tethered to reality.
How to run it
- 1
Assemble the corpus
Pull everything you have ever written into one place — emails, chat logs, direct messages, texts, notes — and normalise it into plain markdown files.
Pro tip Volume matters more than tidiness at this stage; the ranking passes do the filtering later.
Watch out This archive is extremely sensitive. Decide where it lives and who can query it before you build anything on top of it.
- 2
Build multiple vector spaces over it
Embed the same corpus several different ways so you have more than one geometry to work with, rather than a single index.
Pro tip Vary the embedding model and the chunking strategy, not just the parameters.
- 3
Cross the spaces that disagree
Select pairs of spaces that are not pointed in the same direction and intersect them. The disagreement between geometries is where the unexpected material is.
Pro tip Save the pairings that keep producing usable output — not all misalignments are productive.
- 4
Re-rank across several frontier models
Feed the candidate pile to multiple strong models independently and have each rank it. Only promote candidates that more than one model rates highly.
Pro tip Ask each model for its reason, then use the reasons as a second filter.
Watch out A single model's ranking will mostly reflect that model's stylistic preferences.
- 5
Cross-reference against external search
Check the surviving ideas against a web or search index to see whether they are genuinely novel or already well covered.
Pro tip Treat heavy prior coverage as a signal to sharpen the angle rather than to discard the idea.
- 6
Harvest a handful and feed the winners back
Expect on the order of ten to twenty candidates per run and keep only the few that survive every pass. Write up the keepers so they re-enter the corpus.
Pro tip Run it when you are bored or stuck rather than on a schedule — the cost of a run is low.
Watch out Without the write-up step the corpus stops compounding and the system slowly loses its edge.
In the wild
Gary Tan describes building a retrieval tool over his entire corpus — every email, Slack message, text message and DM he has produced — amounting to roughly four hundred thousand markdown files. The tool crosses vector spaces that are not pointed in the same direction, then re-ranks the results. He calls the resulting brainstorm mode LSD mode, and says that when he is bored he asks it for bangers and gets ten or twenty ideas back that several frontier models have already vetted.
→ A repeatable source of pre-vetted ideas produced at a scale a person cannot generate unaided.
A writer with several years of newsletter drafts embeds the archive twice — once by topic and once by rhetorical structure — and crosses the two. The collisions pair old arguments with unrelated formats, producing angles the topic index alone would never surface. Three models rank the output, an external search pass discards the angles already saturated, and the writer keeps four ideas from a run of twenty.
→ A month of publishing angles generated in a single session, each already checked for novelty.
Common mistakes
Crossing spaces that are basically identical
If the geometries agree, the collisions return the same nearest neighbours as ordinary retrieval and the whole exercise degrades to a normal search.
Skipping the re-ranking passes
Raw collisions are mostly noise; without independent multi-model ranking and an external check, you drown in plausible-sounding material.
Publishing the raw output
The system produces ideas, not finished work — shipping unedited generations is exactly the failure mode the panel spends the episode complaining about.
Is it for you?
Best for
Writers, founders, and researchers who already have a large personal archive and want a high-volume source of non-obvious ideas.
Not ideal for
Someone with no accumulated corpus, or a task that needs one correct answer rather than a wide field of candidates.
From the transcript
“take every possible vector space that are is not pointed in the same direction and then cross them and then do um a ranking, a…”
“it's called brainstorm LSD mode, and it just like finds bangers”
“I actually have like uh like 400,000 markdown files on like literally anything I've ever thought or read at this point”
From the episode
Live in the Future