Case 30 · Graduation Day
Private LLM Search / Flashnet
Buried inside twenty-six minutes of self-described "AI psychosis" is a real, earned secret — "agents tolerate rounds," backed by logged tool calls and three working demos with honest latency numbers — but the talk opens by promising to be "pretty chill" so everyone "can get on with the party," shows three products instead of one, sources its best statistic to "this one thing I didn't cite," and ends with no customer, no model, and no ask. Scored as if pitched to investors; on stage it was a research talk, and never claimed otherwise.
01 Scorecard · 12 dimensions
02 Klaff 5-lens audit · 2/5
THE ONE REORDER — Open on the secret — anonymity networks were built for humans; agents tolerate rounds — show ONE demo (Periscope over Flashnet), and make "AI psychosis" a closing joke instead of the thesis.
03 Delivery
367 fillers ("uh" ×135, "um" ×94, "like" ×58, "you know" ×31, "actually" ×29) — one word in twelve, sustained across 26 minutes. The tell of a talk assembled "a few minutes last night" rather than rehearsed.
04 Beat-by-beat
| Time | Beat | What works | The fix |
|---|---|---|---|
| 0:00 | Open | "Many months of AI psychosis" — honest, funny, disarming | Cut "pretty chill," "few minutes last night," "get on with the party" — they tell the room not to invest attention |
| 0:34 | Agent mental model | Clean decomposition, ending on the killer line: search "is one of the main tool calls that actually leaves your machine" | Lead with that line — it's the problem statement, buried at the end of a diagram walk |
| 2:58 | Local vs remote split | Concrete: Shashank's Hacker News cron job, his own X scraper; the IACR-ban aside is the most human moment | Keep — this beat works |
| 4:20 | Zero-click stat | 58.5% → 68% zero-click is a genuine why-now | "According to this one thing I didn't cite" — cite it or cut it |
| 5:38 | Two paths for public search | "Trust me, bro privacy" vs FHE/TEE — the sharpest framing in the talk | Slow down here; this framing should carry the project |
| 7:25 | Tiptoe | "Thought to be… not very feasible… my results show it's actually usable" — a real result, plainly stated | Attach the 1.6s number here instead of six minutes later |
| 10:54 | Deep Research Swarm evidence | The climax: logged tool calls proving "agents can plan ahead, and they don't really care about these constraints" | Claim it outright instead of "a little bit overgeneralization" — agents tolerate rounds |
| 12:08 | DC-Nets + ZipNet | "It won't break unless you can somehow break XOR"; a broken TEE costs spam resistance, never privacy | This is the moat section — say the word |
| 13:53 | Demo 1: Periscope | Live PIR search: 25,000 docs, ~1.6s PIR, ~500ms latency, honest about scale vs Google | Make this the ONLY demo, routed through Flashnet |
| 15:33 | Demo 2: Flashnet toggle | Honest comparative numbers: Tor 0.5s vs Flashnet 5.1s | A mock backend with "latency numbers not to be fully trusted" subtracts credibility — fold into demo 1 or cut |
| 18:01 | Demo 3: P2P search graph | Charming cohort artifact (Saint Damiano) | A third product in minute eighteen — cut from the pitch, keep for the party |
| 19:04 | Close | — | "Yeah, and that is it for me" — no ask, no next step, no repo link spoken aloud |
| 19:46 | Q&A | The best compression happens under pressure: head/tail entropy, "Docker Compose for TEEs," and "Humans want results now synchronously, and agents don't" | That nine-word answer is the pitch — it should have been sentence one, nineteen minutes earlier |
05 The gaps investors block on
- Identity crisis, three ways: Periscope, Flashnet, and a P2P OS backend are three products; the speaker never picked a name on stage.
- No ask — no raise, no testnet waitlist, no "run a shard," not even the GitHub link spoken aloud.
- No customer. The only named user behavior is hobbyist agent cron jobs and, in Q&A, illegal sports-stream discovery — memorable, not underwritable.
- The best stat is uncited: "fifty-eight point five percent… according to this one thing I didn't cite" — and the why-now hangs on it.
- No business model: who pays whom for shards, broadcast, or search is never touched.
- Team is anonymous: no self-introduction, and "we" (Flashbots? classmates?) is never resolved.
Be Flashnet — the anonymity network built for agents. The talk contains one network (Flashnet), one application (Periscope), and one experiment (the P2P OS backend). The Tiptoe/PIR result is an implementation of someone else's 2023 paper — impressive but reproducible. The agents-tolerate-rounds evidence is his: it comes from his own Flashbots-lineage research and his own Deep Research Swarm logs, and it reframes an entire dismissed protocol family. Pick Flashnet as the identity, with Periscope as the first application that proves it live.
“Flashnet is the anonymity network built for AI agents, not humans — agents plan ahead and tolerate rounds, so they get metadata-private search and inference (proven live: private search over [N] documents in under two seconds) that nation-state-grade adversaries can't trace.”
06 The rewritten pitch
Humans want search results now, synchronously. Agents don't — they plan ahead, fan out in parallel, and think between rounds. I logged every tool call my own research agents made, and that's the finding: agents tolerate rounds. That matters because the strongest anonymity protocols we know — DC-nets, mixnets — were dismissed for being round-based and slow. Too slow for humans. Not too slow for agents.
Today, even a 'private' agent leaks: search is the main tool call that leaves your machine, and even if the engine can't read your query, it can fingerprint your agent — same IP, every thirty seconds. Meanwhile [confirm: 58.5% (2024) → 68% (early 2026), source] of searches are already zero-click and moving on-device, so what remains going over the wire is exactly the sensitive tail.
Flashnet is the anonymity network built for agents. Queries are private by cryptography — I implemented Tiptoe, the PIR scheme [confirm: the one iPhone photo search uses a variant of], and it's live: [25,000] documents, one point six seconds, on shards I deployed. Metadata is private by anonymous broadcast — DC-nets hardened the ZipNet way, where a broken TEE costs you spam resistance, never privacy. It won't break unless you can break XOR.
Six nodes are live. Periscope, the first app, searches over it today. We're [raising $X / opening the testnet] to go from six nodes to [N], and I'm looking for [M] operators to run the first public shards. I'm [name] — this is what months of AI psychosis buys you. Come query it.
Brackets are the founder's homework — the copy is done, the proof is theirs to supply. Every claim must match what's actually built.
07 Do these before the next take
- Pick the name and the layer — Flashnet, the agent anonymity network, with Periscope as the first app. Every future sentence starts from that.
- Open with the Q&A one-liner — "Humans want results now synchronously, and agents don't." It compressed the whole talk in nine words under pressure; make it sentence one.
- Cut to one demo — Periscope routed through Flashnet. Retire the mock-backend demo until its numbers are real.
- Source the 58.5%/68% zero-click stat — "this one thing I didn't cite" cannot be said on stage again.
- Write the ask — testnet waitlist, shard operators wanted, or a raise. "That is it for me" wastes the room.
- Say your name. The host had to name him; thirty seconds of self-introduction converts the Flashbots references into a team slide.
- Filler drill — 8.3 per 100 words ("uh" ×135, "um" ×94) across 26 minutes. The fix is a rehearsed 90-second core instead of a talk assembled "a few minutes last night."