Hacker News: kimjune01

New comment by kimjune01 in "Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber"

kimjune01 — Tue, 21 Jul 2026 23:45:48 +0000

it would be nice if these benchmark reports actually specified which tasks they passed and which ones they didn't.

New comment by kimjune01 in "Advertise in ChatGPT"

kimjune01 — Tue, 21 Jul 2026 23:33:57 +0000

Chat data is the closest we ever got to purchase intent than demographics or keywords, attribution given a market position is already solved

New comment by kimjune01 in "Advertise in ChatGPT"

kimjune01 — Tue, 21 Jul 2026 23:24:49 +0000

the investors want their money to make more money

New comment by kimjune01 in "AI makes programming differently difficult"

kimjune01 — Tue, 21 Jul 2026 23:21:40 +0000

I do this

New comment by kimjune01 in "Advertise in ChatGPT"

kimjune01 — Tue, 21 Jul 2026 23:19:15 +0000

Attribution would be as easy as knowing which chatbot they came from? Unless chatbot publishers (openai in this case) allows you to target based on the attribution, that extra data will not be actionable, anyway.

New comment by kimjune01 in "Advertise in ChatGPT"

kimjune01 — Tue, 21 Jul 2026 23:16:04 +0000

the long tail of chatbot users will never pay for subscriptions, that's been proven since WhatsApp tried to charge a dollar or two per year per user

New comment by kimjune01 in "Advertise in ChatGPT"

kimjune01 — Tue, 21 Jul 2026 23:14:48 +0000

In a chatbot-first world, it's gonna have to either be generic impression ads that race to the bottom or contextual recommendations. Of course, the fear is that the advertiser hoards personal data or the chat answers are skewed to make you spend more money.

Those concerns can be resolved by decoupling the text generation from the recommendation engine. It works by casting your intent and match it up with an advertiser's in a secured environment where the advertiser doesn't actually get to see your chats. For example, if you are talking about knee pain during trail running, the recommendation engine might cast "knee pain for runners" and either a shoe company or a physiotherapist would bid against that in an embedding space.

Because the recommendation engine is an opt-in module for publishers and users, it's game-theory welfare maximizing.

I talk a lot more about this on my blog, it goes on for quite a while. https://www.june.kim/vector-space

The exchange is live, with no publishers or advertisers. https://vectorspace.exchange/

New comment by kimjune01 in "Claude Is Not a Compiler"

kimjune01 — Tue, 21 Jul 2026 22:17:49 +0000

and LLMs are really good at filling in the blanks. It's like a really big import on generic application knowledge, and if you draw the circles well enough, it will draw the rest of the owl.

New comment by kimjune01 in "Separating signal from noise in coding evaluations"

kimjune01 — Tue, 21 Jul 2026 19:03:25 +0000

eval shops who depend on these things for a living will have to version their datasets, to prevent contamination

New comment by kimjune01 in "Claude Is Not a Compiler"

kimjune01 — Tue, 21 Jul 2026 17:52:27 +0000

i meant to make the distinction between 'prompt' and 'spec' in that prompts are something that a human types in by hand. Interesting programs are a combination of canon and novel input

New comment by kimjune01 in "Claude Is Not a Compiler"

kimjune01 — Tue, 21 Jul 2026 17:27:08 +0000

whether the hypothesis originates at the users or in your guess for the users is irrelevant in that they're both hypotheses with no certain outcome until trial.

New comment by kimjune01 in "Claude Is Not a Compiler"

kimjune01 — Tue, 21 Jul 2026 17:19:27 +0000

I agree with iteration, but that's an orthogonal argument to how effective it is at turning prose to code. You are asking, LLMs are capable of translating a 835-page spec, but where does the spec come from? Follow the stream up from the built software product:

Binary < IR < Code < Spec < Prompt (aka goals)

For LLMs to faithfully fulfill the role of translating spec/prompt to code, it needs enough information about the problem/motivation and the constraints around it. Often, that's done by research or by importing proven modules. The human role left in this compilation chain is a precise-enough expression of the motivation, upstream of the built artifacts.

Edit For example, SWEbench Pro was supposed to be a measure for how good LLMs are at compiling prose into code. At least 15% of those tasks were underdetermined by ambiguous prose in the spec. Of the issues that are specific enough, Claude and Codex indeed do compile to functional code with self-written tests. Source: https://github.com/kimjune01/swebench-pro-audit

New comment by kimjune01 in "Claude Is Not a Compiler"

kimjune01 — Tue, 21 Jul 2026 17:09:38 +0000

prompts are not specs because most often, prompts are underspecified. You can prompt fizzbuzz well enough for it to be functional, but not ffmpeg. So the distinction is less about Claude's capability, more about the quality of the input. Garbage in garbage out

New comment by kimjune01 in "Apple defeats liability for not scanning iCloud for CSAM"

kimjune01 — Tue, 21 Jul 2026 16:55:10 +0000

a 17 year old can take a nude selfie and be charged as an adult in possession of CSAM

New comment by kimjune01 in "What if your AI agent could have its own WhatsApp number?"

kimjune01 — Mon, 20 Jul 2026 13:28:25 +0000

i thought this was what openclaw was for?

New comment by kimjune01 in "Claude Code uses Bun written in Rust now"

kimjune01 — Mon, 20 Jul 2026 13:27:08 +0000

i found the same thing with Rust, verification steps are really helpful for speed and correctnesss

New comment by kimjune01 in "AI in mathematics is forcing big questions"

kimjune01 — Sat, 27 Jun 2026 03:28:19 +0000

lean compiles or it doesnt

New comment by kimjune01 in "You can't unit test for taste"

kimjune01 — Fri, 26 Jun 2026 02:56:09 +0000

this is known as the oracle problem

New comment by kimjune01 in "Zen and the Art of Machine Learning Research"

kimjune01 — Fri, 19 Jun 2026 17:04:10 +0000

one of those words that lost meaning, like samaritan or literally

New comment by kimjune01 in "They’re made out of weights"

kimjune01 — Thu, 04 Jun 2026 19:19:22 +0000

In the days when Sussman was a novice, Minsky once came to him as he sat hacking at the PDP-6.

"What are you doing?", asked Minsky.

"I am training a randomly wired neural net to play Tic-tac-toe", Sussman replied.

"Why is the net wired randomly?", asked Minsky.

"I do not want it to have any preconceptions of how to play", Sussman said.

Minsky then shut his eyes.

"Why do you close your eyes?" Sussman asked his teacher.

"So that the room will be empty."

At that moment, Sussman was enlightened.