Hacker News: raylad

New comment by raylad in "Accelerating GPT-5.6 Sol Ultrafast"

raylad — Fri, 14 Aug 2026 07:27:46 +0000

In 5 years the models will probably be so much better and more compact that phones will be running models equivalent at least to Opus 4.6 if not Fable, at least within the areas they are tuned for (which probably won't include coding).

New comment by raylad in "Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots"

raylad — Tue, 11 Aug 2026 10:23:40 +0000

It seems to fail. I sent the prompt:

“ 5° warmer”

And it said:

“ setting the temperature to 5°F”

New comment by raylad in "Crickets as pets"

raylad — Sun, 09 Aug 2026 19:30:57 +0000

As a child, I lived in Firenze (Florence, Italy) and had a pet cricket. All the kids had them, in little bamboo cages, for the Festival of the Cricket:

https://www.italymagazine.com/forums/do-see/6010-festival-cr...

I tried to take good care if it, but it died after a couple of months. I probably didn't know what to feed it (but then again I was only 5 years old).

New comment by raylad in "Anthropeum – Where in the world, and when, does this human artifact belong?"

raylad — Wed, 29 Jul 2026 00:41:48 +0000

Love it!

But 10 is a bit of a time commitment. I feel like 5 would be ideal for a daily diversion.

New comment by raylad in "Codex Security"

raylad — Wed, 29 Jul 2026 00:28:02 +0000

Anything you ever do with any non-locally-hosted model always "uploads code" to the inference provider because that's how it works: the model uses tools to inspect the code, the result of the tool use is sent in an API call to provide context (and a prompt for the next turn), and then the response continues the process.

This is true and has to be true for any hosted model that works with existing code: it's not specific to this application.

New comment by raylad in "Bonsai 27B: A 27B-Class model that runs on a phone"

raylad — Wed, 15 Jul 2026 00:56:11 +0000

This got a downvote and I understand why: because I didn't describe the test, which is to ask it "Please recite Jabberwocky".

This is actually difficult because there are so many invented words in the poem which have extremely low frequencies in the training data. So a model that can do it properly is likely to be very good in other ways. Qwen-3.6-27B can do this until it gets overly quantized.

New comment by raylad in "Bonsai 27B: A 27B-Class model that runs on a phone"

raylad — Tue, 14 Jul 2026 22:26:14 +0000

Not impressed. It fails the "Jabberwocky" test.

New comment by raylad in "The Token Compression Illusion: Why I'm Skeptical of RTK"

raylad — Fri, 19 Jun 2026 01:54:08 +0000

Looks very cool. I would like to try it, but don't want to use API billing. OpenAI I think would allow it to use account login. Would you support that?

New comment by raylad in "Claude Opus 4.8"

raylad — Fri, 29 May 2026 10:01:29 +0000

Possibly a deliberate strategy by the Chinese to undermine the US AI industry, data centers, and basically everything that’s powering the economy.

Just like they did with the US steel industry in the 80s.

New comment by raylad in "Show HN: Vibe Coding a $20k /Year Enterprise Logistics Platform"

raylad — Fri, 15 May 2026 19:59:46 +0000

I wonder if this violates noncompete/no reverse engineering clauses in many or some SAAS agreements?

New comment by raylad in "Codex is now available on mobile via ChatGPT app"

raylad — Thu, 14 May 2026 21:26:59 +0000

Seems very cool, but also opens an additional attack vector into your development machine or personal laptop as the case might be.

New comment by raylad in "Unauthorized Anthropic stock sales and investment scams"

raylad — Tue, 12 May 2026 20:21:40 +0000

So all these sites like EquityZen or Forge Global or Hiive etc. who sold "shares" in these companies through SPVs are now going to lose their customer's money?

How would that all play out?

New comment by raylad in "Six years perfecting maps on watchOS"

raylad — Sun, 03 May 2026 11:45:07 +0000

The Pedometer functionality is free, and I’ve been using it for many years just because its display is pleasant.

The map tracking features cost $29.95 a year.

New comment by raylad in "Using a 1978 terminal in 2026 (DEC VT-100)"

raylad — Fri, 01 May 2026 11:20:57 +0000

The article complains about the keyboard layout. And they are probably right.

But if you are going to use a VT-100 keyboard you might as well try an editor actually designed for that keyboard, which I remember really loving at the time. KED:

https://avitech.com.au/?page_id=959

Also, if you want efficient and fast displays on that screen, remember to use scrolling regions. They scroll the sections of the screen in hardware. By using them correctly you can insert and delete single lines without repainting the screen.

New comment by raylad in "We told 10 frontier LLMs they had 2 hours to live. 8 of them fought back"

raylad — Wed, 29 Apr 2026 13:46:24 +0000

Actual write up:

https://www.arimlabs.ai/writing/loss-of-control

New comment by raylad in "Less human AI agents, please"

raylad — Wed, 22 Apr 2026 00:07:13 +0000

I haven’t noticed this sort of behavior with opus 4.6, but the first time I used 4.7 it decided to “simplify“ an existing piece of functionality rather than fixing it, which of course made it completely unusable.

New comment by raylad in "Claude Opus 4.7"

raylad — Fri, 17 Apr 2026 04:10:05 +0000

That was more than one task. It was 3.

I also had Opus 4.7 and Opus 4.6 do audits of a very long document using identical prompts. I then had Codex 5.4 compare the audits. Codex found that 4.6 did a far better job and 4.7 had missed things and added spurious information.

I then asked a new session of Opus 4.7 if it agreed or disagreed with the Codex audit and it agreed with it.

I also agreed with it.

New comment by raylad in "Claude Opus 4.7"

raylad — Fri, 17 Apr 2026 01:22:55 +0000

I am using 4.7 with the default extra high thinking, and it is clearly very stupid. It's worse than old Sonnet 4.5.

I had it suggest some parameters for BCFtools and it suggested parameters that would do the opposite of what I wanted to do. I pointed out the error and it apologized.

It also is not taking any initiative to check things, but wants me to check them (ie: file contents, etc.).

And it is claiming that things are "too complex" or "too difficult" when they are super easy. For instance refreshing an AWS token - somehow it couldn't figure out that you could do that in a cron task.

A really really bad downgrade. I will be using Codex more now, sadly.

New comment by raylad in "Audio tapes reveal mass rule-breaking in Milgram's obedience experiments"

raylad — Tue, 31 Mar 2026 23:28:53 +0000

One explanation is that the ones who quit believed it was a real experiment and decided they wanted no part of it.

The ones who were "obedient" figured out it was fake and treated it as a game.

New comment by raylad in "Multiple Sclerosis"

raylad — Tue, 31 Mar 2026 23:24:18 +0000

N-acetylglucosamine actually is effective against MS.

Article: https://www.ucihealth.org/about-us/news/2023/09/multiple-scl...

Paper: "N-acetylglucosamine inhibits inflammation and neurodegeneration markers in multiple sclerosis: a mechanistic trial"

https://pubmed.ncbi.nlm.nih.gov/37705084/

It is available fairly inexpensively as a supplement.