0xPARC coined it for the second generation of crypto, roughly the last 15 years. First-generation crypto hand-builds a protocol per problem: RSA signs a message, and Yao's millionaire problem gets its own bespoke protocol to compare two incomes. Second-generation crypto hands you primitives you can program for any function you can write as code.

Their analogy: it's the jump from single-purpose hardware (a digital alarm clock) to a general-purpose one (a smartphone that runs whatever you compile for it). You stop inventing a new algorithm per problem and start compiling the problem you have.

The primer picks four: 2PC (two-party computation), SNARKs (succinct proofs), FHE (compute on encrypted data), and ORAM (hide which data you touch). Two of those already have rooms here, so I'll wire them up as we go.

Two people compute F(x1, x2) where each knows only their own input, and neither learns the other's. Yao's millionaire problem is the classic: F is "who earns more," the inputs are the two incomes, and both walk away knowing the winner but not the number. The engine underneath is a garbled circuit (one party encrypts a boolean circuit gate by gate, the other evaluates it blind) plus oblivious transfer. Scale 2PC past two parties and you land on the MPC one room over.

Read a signature as a proof: Alice's signature proves "I know Alice's private key." A SNARK generalizes that to a proof of any statement shaped like "I know X such that F(X, Y) = Z," with F, Y, Z public, once you encode F as equations. The primer's headline example is "I know M such that sha(M) = Y." PLONK is one such scheme. Try the hash half live:

FHE you've met: encrypt your data, let a server compute on the ciphertext, decrypt the result, and the server reads nothing. The primer's example is private machine translation: send Enc(x), get back Enc(translation). ORAM is the piece with no room here yet. Even with the data encrypted, the server still watches your access pattern: which records you read, how often, which ones together. Oblivious RAM scrambles each request into a batch of reads and writes, so the server can't tell which one you actually wanted.

Cost, and the primer is blunt about it: proving a computation in a SNARK can run millions of times the cost of just doing the computation. For a decade it was mostly theory with little implementation, until blockchains and other decentralized systems started demanding it and pulled the practical work forward.

That's the frame. 2PC and MPC, SNARKs, FHE, and ORAM are different tools, but each earns its guarantee from protocols and math rather than a chip you have to trust, the far end from a TEE. Same goal, compute on data nobody should see, with a different bet on where the trust lives.