Hacker News: eddyg

New comment by eddyg in "Ask GitHub SRE: How serious is the situation there?"

eddyg — Fri, 07 Aug 2026 11:53:45 +0000

For some context, GitHub COO back on April 3, 2026: https://x.com/kdaigle/status/2040164759836778878

New comment by eddyg in "Artificial Intelligence used to design new viruses"

eddyg — Fri, 07 Aug 2026 11:28:22 +0000

You hope it’s “far-fetched”.

As noted in this paper⁽¹⁾, ”Publicly releasing model weights is an act of irreversible proliferation. Once weights are released, they can never be recalled.”

Do you really want to risk it on your off-the-cuff reasoning that it’s “far-fetched”?

⁽¹⁾ https://www.lesswrong.com/posts/qmQFHCgCyEEjuy5a7/lora-fine-...

New comment by eddyg in "ASCII Today"

eddyg — Mon, 03 Aug 2026 22:56:16 +0000

Needs a (2020) tag. (The 2023 commit was just a Google Analytics update).

A badge of honor these days!

New comment by eddyg in "A.I. companies are recruiting electricians and carpenters by the thousands"

eddyg — Wed, 29 Jul 2026 16:48:42 +0000

Gift link: https://www.nytimes.com/2026/07/29/business/economy/data-cen...

New comment by eddyg in "Relaxin Jailbreak for iOS 17.0-17.3.1 Released"

eddyg — Wed, 29 Jul 2026 16:27:58 +0000

iOS 17.0 was released in September 2023, with support for the new iPhone 15 models.

iOS 17.4 was released in March 2024.

Additional updates to 17.5, 17.6 and 17.7 followed.

Any devices released after the iPhone 15 couldn't run iOS 17.

It makes me wonder how many devices exist that are still running "iOS 17.0 to 17.3.1"...

New comment by eddyg in "TWC Classics"

eddyg — Tue, 28 Jul 2026 13:50:30 +0000

If you miss the classic WeatherStar look-and-feel from the cable days of The Weather Channel, be sure to check out https://github.com/netbymatt/ws4kp

New comment by eddyg in "OpenAI's rogue model attack is just the beginning"

eddyg — Tue, 28 Jul 2026 12:55:49 +0000

All of which are easily bypassed in open weight models; see my previous comment about K2.5.

New comment by eddyg in "OpenAI's rogue model attack is just the beginning"

eddyg — Tue, 28 Jul 2026 12:30:36 +0000

Classifiers are policy enforced by the process serving the model.

Input classifiers get applied before it reaches the model so somebody hacking an open-weight model would skip this. Streaming classifiers get polled during decoding; hackers delete this check in the sampling loop.

But both are always applied in closed weight models.

Set Llama Guard to 1.0 and nothing is ever unsafe.

New comment by eddyg in "OpenAI's rogue model attack is just the beginning"

eddyg — Tue, 28 Jul 2026 12:00:44 +0000

Constitutional classifiers go a long way to reducing unsafe usage in closed-weight models. And like we saw with Fable, closed models can be revoked and classifiers updated when “jailbreaks” are found.

Having the weights gives you the exact affordance an unlearning attack requires, without rate limits.

New comment by eddyg in "OpenAI's rogue model attack is just the beginning"

eddyg — Tue, 28 Jul 2026 11:36:28 +0000

A change in kind is not the same as a change in degree. Ten orchestrated LLM PhD advisors is a genuinely different thing from a library.

New comment by eddyg in "OpenAI's rogue model attack is just the beginning"

eddyg — Tue, 28 Jul 2026 10:32:11 +0000

If nothing else, the article has a really good timeline of the OpenAI/HuggingFace “incident”.

But to me, it underscores the impending cliff of doom from the continued release of open-weight models: there's no cryptographic or architectural way to give someone full weights while withholding the nefarious capabilities those weights encode.

As noted in this paper⁽¹⁾, “publicly releasing weights is an act of irreversible proliferation”.

I’m sure this will be an unpopular opinion on HN, but open weights are the thing that scares me the most about “A.I.”. There is a lot of research in this area⁽²⁾, and I think most of HN is unaware of it or ignores it. Stripping refusals from Kimi K2.5 took under $500 of compute and about 10 hours, taking HarmBench refusals from 100% to 5% while retaining nearly all capability; the resulting model gave detailed chemical-weapons synthesis instructions.

The gate is only as strong as the least-cautious releaser...

⁽¹⁾ https://www.lesswrong.com/posts/qmQFHCgCyEEjuy5a7/lora-fine-...

⁽²⁾ https://arxiv.org/html/2604.03121v1

New comment by eddyg in ""Opus 5 is a really bad model""

eddyg — Tue, 28 Jul 2026 08:25:27 +0000

The guidelines⁽¹⁾ make it clear the OP did the correct thing: ”Please submit the original source.”

⁽¹⁾ https://news.ycombinator.com/newsguidelines.html

New comment by eddyg in "Em dashes are amazing"

eddyg — Fri, 24 Jul 2026 14:48:49 +0000

How do you feel about the Tironian Sign Capital Et? https://www.unicodecharacter.org/U+2E52

I mean the triple-em-dash has been around since Unicode 6.1; The Tironian Sign Capital Et wasn't added until Unicode 13.

New comment by eddyg in "Em dashes are fucking amazing"

eddyg — Fri, 24 Jul 2026 14:21:34 +0000

"It is (and, increasingly, was) used to signify that a bibliography entry has the exact same author(s), editor(s), translator(s), or corporate author(s) as the previous bibliography entry."

https://danieljtortora.com/blog/3-em-dashes

New comment by eddyg in "How much energy do data centers and artificial intelligence use?"

eddyg — Thu, 23 Jul 2026 19:30:24 +0000

What's massive is the amount of water used for a handful of almonds. Not daily AI use.

https://www.andymasley.com/visuals/ai-prompt-footprint/

New comment by eddyg in "How much energy do data centers and artificial intelligence use?"

eddyg — Thu, 23 Jul 2026 19:29:04 +0000

https://www.andymasley.com/visuals/ai-prompt-footprint/

New comment by eddyg in "Americans are angry about data centers. Politicians are feeling the pressure"

eddyg — Mon, 20 Jul 2026 02:51:47 +0000

This is a great data-backed resource to refute made-up claims about data centers:

https://www.andymasley.com/visuals/ai-prompt-footprint/

New comment by eddyg in "Agentty – A drop-in alternative to claude-code, written in C++26. 11.0 MB binary"

eddyg — Thu, 16 Jul 2026 11:10:36 +0000

It’s all about the harness, not just calling the API.

New comment by eddyg in "The lost joy of music piracy"

eddyg — Thu, 16 Jul 2026 11:05:00 +0000

If you’re gonna capitalize the “D” you should also capitalize the “G” (since it’s the first letter of his last name…)

New comment by eddyg in "Ask HN: HN Algolia API returns 400 for queries with points in numericFilters"

eddyg — Fri, 10 Jul 2026 13:11:15 +0000

Thanks for the reply!

I went to make the change, and all the sudden it started working with “>” again (not URL encoded) so this appears to have been a transient issue.