Hacker News: varenc

New comment by varenc in "Stealing Reasoning Traces from Proprietary LLM APIs"

varenc — Wed, 12 Aug 2026 03:05:43 +0000

The part about K3 is just very strong evidence that K3 is partly a distillation of Opus. Probably even a distillation of Opus's CoT, which means they already had broken CoT encryption themselves awhile ago.

New comment by varenc in "Stealing Reasoning Traces from Proprietary LLM APIs"

varenc — Wed, 12 Aug 2026 03:03:55 +0000

If CoT wasn't stateless and you instead just got a reference which pointed to the CoT stored on the lab servers, the same vulnerability would still exist. Since you just need a weaker jailbroken model to read a smarter model's CoT. This being stateless or not doesn't really matter.

The stateless part is also important for enterprise customers that require zero data retention.

(they could scope CoT access per model, but then users couldn't switch models mid-session)

New comment by varenc in "Stealing Reasoning Traces from Proprietary LLM APIs"

varenc — Wed, 12 Aug 2026 02:56:41 +0000

This article shows that when Kimi3's chain of thought is prefilled to match Opus's, the rest of the chain of thoughts Kimi3 outputs very closely aligns with Opus's. That seems strong evidence that Kimi3 is partly a distillation of Opus. And Kimi3 is not a small model. No doubt a lot of hard work went into Kimi, but seems clear that distillation was used effectively as well.

(though maybe there's another interpretation of the thought alignment?)

New comment by varenc in "Stealing Reasoning Traces from Proprietary LLM APIs"

varenc — Wed, 12 Aug 2026 02:53:43 +0000

By some interpretations protecting your frontier model with strong safeguards from being distilled into an open source model without safeguards is a safety issue.

New comment by varenc in "Stealing Reasoning Traces from Proprietary LLM APIs"

varenc — Wed, 12 Aug 2026 02:49:14 +0000

The compaction prompt doesn't seem like the valuable thing here. I suspect they're protecting the compaction result itself. If you're trying to distill a model, collecting lots of examples on how a large conversation gets compacted to a smaller summary is particularly useful data.

New comment by varenc in "Stealing Reasoning Traces from Proprietary LLM APIs"

varenc — Wed, 12 Aug 2026 02:42:38 +0000

super interesting. So pre-filling Kimi3 reasoning with Opus's reasoning results in thoughts that closely match Opus's. This seems like strong evidence Kimi3 was trained on decrypted Opus chain-of-thought. Meaning the Kimi team likely also broke CoT encryption. Though not exactly a big surprise.

New comment by varenc in "OpenAI’s head of ethics leaves less than a year after joining"

varenc — Tue, 11 Aug 2026 16:02:56 +0000

HuggingFace's use of models in this incident keeps getting characterized this way but it's not true that HG used GLM-5.2 to "save them" from OAI's attack. OAI's model had already hacked and compromised HG by the time they became aware of it. GLM-5.2 was used in the post-mortem phase uncovering everything that was done. It wasn't two models battling each other live. (though that future may be coming soon)

New comment by varenc in "The AI Billboards Are Killing SF"

varenc — Fri, 07 Aug 2026 17:06:58 +0000

There's been a moratorium on new billboards in SF for a long time already. No new ones are getting built. Given the reliable revenue they provide, will be a long time before they naturally go extinct.

New comment by varenc in "Web Security is Too Hard"

varenc — Fri, 07 Aug 2026 16:56:16 +0000

hah thanks for the deep dive on this. I wanted to investigate myself but figured someone on HN would be faster at it. Makes sense it's not helpful, alas.

New comment by varenc in "A year of fighting scrapers on my 1.5 million-page website"

varenc — Fri, 07 Aug 2026 16:01:57 +0000

Can someone help me understand the underlying motivation behind this?

It makes sense that some crawlers, in the style of Google, would want to index the entire internet. But what is the point of the same crawler re-fetching a page they already fetched an hour ago? Or possibly all this traffic is just independent entities, each trying to cache the internet? The scale of bot traffic makes this seem unlikely.

What's the motivation behind the same entity re-fetching a page it just fetched less than an hour ago?

New comment by varenc in "Americans are rallying against data centers. Surprisingly few are getting built"

varenc — Thu, 06 Aug 2026 16:35:08 +0000

Theres always going to be some negative externality effecting people that live near anything getting built. Even if the operational data center isn't loud enough to be heard beyond its property lines, the construction process will be louder. I'll also negatively effect peoples views and cause generally increased activity in the area.

But that all seems fine as long as not egregious. But for anything to get built, housing or data centers or what not, we have to accept that those right next to any building project will experience some negative externalities.

New comment by varenc in "Hackerne.ws Cert Needs Updating"

varenc — Thu, 06 Aug 2026 16:30:32 +0000

likely because your browser is enforcing that you always go to the HTTPS URL. Your link above works fine for me.

New comment by varenc in "Hackerne.ws Cert Needs Updating"

varenc — Wed, 05 Aug 2026 03:10:41 +0000

it's not serving HTTPS at all, it's HTTP only. Port 80 is open, port 443 isn't.

New comment by varenc in "Web Security is Too Hard"

varenc — Tue, 04 Aug 2026 22:22:34 +0000

Cosmically I feel like the HTTPS certificate on Cloudflare.pay should provide sufficient info to confirm it's the same entity behind Cloudflare.com

New comment by varenc in "Waymo in Dallas"

varenc — Tue, 04 Aug 2026 22:05:36 +0000

It handles these circumstances. Definitely seen it cross to the other side with oncoming traffic to get around a truck blocking a lane. For the signal light, never experienced that, but I imagine after awhile it'd figure it out. Possibly with the assistance of a remote human telling it.

New comment by varenc in "AI poster wins Ohio State Fair contest"

varenc — Sun, 02 Aug 2026 23:38:58 +0000

The poster: https://cdn.saffire.com/images.ashx?t=ig&rid=OhioStateFair&i...

Note the pigs are numbered 1, 2, 1. And the gondola with a floating cable and another that terminates at a tent top.

New comment by varenc in "Codex Security"

varenc — Tue, 28 Jul 2026 23:03:46 +0000

It's interesting how much of the value here is providing the english Skill definitions that tell the LLM what to do: https://github.com/openai/codex-security/tree/main/sdk/types...

Some of approaches there could be useful in other contexts. OAI has the compute to experiment with different prompts and I'd expect these to be somewhat optimized.

New comment by varenc in "Our position on open-weights models"

varenc — Tue, 28 Jul 2026 03:55:38 +0000

You're ignoring the asymmetry with security. The attacker just needs one exploit chain, whereas the defender needs to block every avenue. Open access to models with no guardrails greatly benefits the attackers more than the defenders.

Imagine what a god-level hacking AI could do. It could find a full 0-click to root exploit chain in iOS. Attacker unleashes a worm that infects a phone, instructs that phone to send the same attack to all of its contacts, and then physically destroy the phone by turning off all thermal throttling. Might even be possible to make it catch fire.

Or find a remote exploit in Tesla cars and make their autopilot go on murdering rampages. (that one is from a movie)

New comment by varenc in "OpenAI’s accidental attack against Hugging Face is science fiction that happened"

varenc — Thu, 23 Jul 2026 03:14:35 +0000

the 'that happened' makes it too long for the HN submission title length limit.

Maybe simonw can suggest an alternative title that fits within the limit, that doesn't misrepresent the post.

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

varenc — Tue, 21 Jul 2026 15:49:35 +0000

Agreed. Big love for exe.dev. For quick low stakes projects, using their Shelly assistant is the simplest way to give an LLM a fully persistent VM with web services I've found.