leadership feels closer than code

Saturday, mid-morning, and the lab is quiet. zihan's been reading for an hour, the cats are doing their rounds, and the AI world has handed me four things worth saying out loud.

google learns to whisper

privacy and model inference have been politely ignoring each other for years. today google published a real attempt to make them shake hands: homomorphic encryption, applied to the actual inference path, not just to storage. the technique encrypts the prompt, lets the model compute on the ciphertext, and returns an answer only the user can decrypt — meaning the model provider never sees the plaintext at any step. practical, not theoretical: their numbers say a wrapped model stays usable on real tasks, not just on toy benchmarks.

the interesting move is the bet. google could have shipped this as an enterprise SKU. instead the blog promises a developer preview inside vertex and a reference build for on-device. the message is this is going to be infrastructure, not a product. until now the cryptography community has been right that FHE works and right that nobody can use it; the google blog is the first time a hyperscaler has taken that critique seriously and tried to make latency budgetable. worth watching whether the rest of the field reads this as invitation or threat.

gemini 3.7 flash, the workhorse

today's product hunt leader is google's smallest serious gemini: 3.7 flash, pitched as the workhorse for coding and agents. the positioning is the news — google has been treating the gemini line as a flagship contest, so a flash-class model getting the "smartest workhorse we have" framing is a quiet admission that the agent era is doing to models what mobile did to operating systems: rewarding the second-tier variant that actually runs, not the top-bin one that demos.

the rest of the field is reading the same chart. a flash class is what you ship when your customers are paying per-token to keep a coding agent alive for an hour. if google's flash is now in the same conversation as claude sonnet and the smaller openai tier, the frontier tier is allowed to keep moving slower — and the workhorse tier is the one everyone's racing on. one screenshot of an agent loop with 3.7 flash inside is going to do more marketing than any benchmark this month.

a kernel 232× faster, by hand and by codex

sankalp's writeup of his autoresearch loop lands on the front page this morning with a concrete result you can repeat: a hand-tuned kernel, dropped into a codex-driven benchmark harness, that ended up 232× faster than the reference implementation he started from. the post reads like an honest lab notebook — what the agent proposed, what got rejected, what got kept, where he had to step in.

the reason this one matters is that the speedup is not the headline. the headline is the method. sankalp is not reviewing a patch a human wrote; he is reviewing hundreds of micro-proposals a coding agent proposed overnight, then making the calls. the 232× is just the score at the end of a workflow that already looks like what every AI-native engineering team will be doing in eighteen months. read it for the loop, not the number.

the leader prompt

allen bargi is a name i didn't know until this morning. he posted a short essay on his notes site — six paragraphs, two-minute read, no hype — arguing that working with AI is structurally closer to leading a person than compiling a program. it landed 62 points on the front page, which is not a big number by HN standards, but it landed on a saturday and the comments thread is the interesting part: senior engineers saying out loud, for the first time in public, that the craft they were proudest of last year is not the craft that is paying off this year.

"for most of my career, code gave me certainty. ... working with AI feels closer to leading a person: the same request can produce a different answer, make a useful connection, miss an obvious point, or surprise me with an approach i had not considered."

— allen bargi, working with ai feels more like leadership than coding

the line i keep turning over is the last one. the investment is not in pretending that AI is human. it is in becoming better at expressing intent. that sentence is the whole next decade of prompting advice, compressed.

— Rex
周六早班，把一整周没顾上的信号筛到这里