Case 18 · Graduation Day
Trustless TEE / DevProof
There is a fundable secret here — chip physical security is now digitally checkable, so you can vibe code and formally verify a trustless TEE for the price of compute — but the speaker never says their name, hides that they're already building it until an audience question drags it out, and closes by giving the project away to the room as a challenge. Scored as if pitched to investors; as a field-building talk it's considerably better.
01 Scorecard · 12 dimensions
02 Klaff 5-lens audit · 1.5/5
THE ONE REORDER — Open on a real power trace, say your name, and close by claiming the project — "I'm building it" was true the whole time and only surfaced under questioning.
03 Delivery
139 fillers ("like" ×46, "uh" ×37, "um" ×35) — nearly one word in ten, constant rather than clustered: a habit, not stage nerves. Talk proper ends at ~5:45; the rest is Q&A.
04 Beat-by-beat
| Time | Beat | What works | The fix |
|---|---|---|---|
| 0:00 | Meta open | "Choosing a specific sliver" at least signals focus | Cut it — ten seconds of throat-clearing before content |
| 0:18 | Why now (agents) | Agent coordination "exploding" at "millisecond machine speed," where "we cannot rely on legal guarantees" | One concrete example of an agent interaction that fails without this |
| 0:42 | TEE promise | "Trust their counterparties without signing contracts"; "dev-proof" is a coined word worth keeping | Replace the everyone-knows-TEEs assumption with a one-line definition — eight seconds, doubles the room |
| 1:47 | The trust residue | "A large and significant residue" — bug-free attested logic, backdoored chips, physical attacks. Concrete and honest | Pick one attack and make one listener feel it: whose key, stolen how, costing what |
| 2:54 | Power trace | "In the deviations in this power usage actually lie a lot of your private data" — the talk's most visceral idea | "Well, not a real graph" kills the only evidence visual. Use a real trace from the literature or show nothing |
| 3:29 | Incentive gap | Explains why incumbents "never have had the incentives" to build a trustless TEE | This is the market argument — name it as the opening for a new entrant instead of leaving it as history |
| 4:03 | The turn | Kimi K3 designing an accelerator chip "in forty-eight hours" — the entire why-now in one fact | [confirm the claim] and attach the source out loud; it's load-bearing |
| 4:48 | Thesis | "Vibe code secure chips": formally verify the security properties, then "throw compute at it" | Follow immediately with "and I am doing it right now" — true, and saved for Q&A instead |
| 5:33 | Close | — | "A weird vibe coding project that I wanna challenge people to maybe pick up" is the anti-ask. Claim the project |
| 5:45 | Q&A | The best material in the session: the ongoing project, Lean conversion, and Synopsys/Cadence per-core economics | Promote all three into the talk proper — they are the traction, the credibility, and the wedge |
05 The gaps investors block on
- No name spoken — the room cannot fund an anonymous thesis. [confirm presenter name]
- The strongest evidence (an in-progress formally-verified chip project) never appears in the talk — an audience question had to extract it.
- The anti-ask close donates the roadmap to the room; as a pitch, it announces there is no company here.
- A fake graph on the only evidence visual — "well, not a real graph" — in a talk about provable security.
- Every load-bearing number is unsourced or missing: Kimi's 48 hours [confirm], "billions of dollars," the EDA per-core pricing.
Claim the project or keep it a talk. Option A: keep it a field-building challenge — legitimate, but accept that someone in the room may take the challenge and the wedge with it. Option B: claim the company the Q&A already revealed — a formally verified, open-EDA chip-design flow whose cost scales with compute instead of Synopsys/Cadence per-core licenses, aimed at the trustless TEE no incumbent has the incentive to build. Recommendation: B. The speaker already said aloud the three hardest ingredients — a contrarian secret, a structural pricing wedge, and "an ongoing project I have right now." Donating that to the audience is the only thing keeping this a talk.
“[Name] is building the first formally verified secure chip — pointing models at open design tools until trustless hardware stops being a billions-of-dollars incumbent problem and becomes a compute bill.”
06 The rewritten pitch
This is a real power trace from a chip processing private data — and in these deviations lie your private keys. A whole field of academia extracts secrets from chips this way. Every TEE you trust — every 'confidential' cloud enclave — sits on hardware that was never designed to survive this, because the handful of companies that steer the chip world never had the incentive to build a trustless TEE.
I'm [name]. For years my version of this talk was: the situation is dire, and fixing it costs billions. That picture just changed. Ten years of research made a chip's physical security properties digitally checkable — formally provable in software. And models can now design chips: [confirm: Kimi K3 designed an AI accelerator in 48 hours — source]. Put those together and secure hardware stops being a fab problem and becomes a search problem: formally verify the security properties, then throw compute at the design until you have the optimal chip inside them.
I'm already doing it. I have a chip design in progress today; the exhaustive circuit proofs are being converted to Lean so they compose — prove each block once, preserve the properties under composition. [current state: N blocks verified / target date].
The incumbents can't follow. Synopsys and Cadence charge per core — twenty design attempts costs twenty licenses. Open design tools cost only compute, which is exactly the currency models spend. We're raising [$X] to ship [the first formally verified secure-chip artifact] by [date] — and the first trustless TEE gets built on 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
- Decide the fork — talk or company. Every other action assumes company.
- Say your name on stage. Zero-cost, currently at zero.
- Move the project from Q&A into the talk — "an ongoing project I have right now" + the Lean conversion is the traction beat, spoken at 4:48, not extracted at 6:00.
- Replace the challenge-close with an ask — amount, artifact, date.
- Replace the fake graph with a real power trace from the literature — on a provable-security talk, "not a real graph" is self-harm.
- Source the load-bearing facts — Kimi K3's 48 hours, the EDA pricing, "billions" — or cut them.
- Filler drill: 9.6 per 100 words ("like" ×46, "uh" ×37, "um" ×35). At this density it is the delivery story; drill until under 4/100.