irthomasthomas

gisterendiscuss

It's like scaling a swiss cheese and the holes grow bigger with it. You can't get rid of the holes without making a different cheese.

irthomasthomas

gisterendiscuss

Claude use to be leader, too. Their metaprompt was great at the time with opus 3

irthomasthomas

gisterendiscuss

Cool. I still find these a useful visualization of some the qualities of llms. Even if they did train for [animal] on [vehicle] svg, it's still nice to see at a glance how the different models and reasoning levels perform. Lunar misses part of the frame, except on max reasoning. While most of the others have a mostly correct bike at all reasoning levels.

I once used something like karpathy's auto-scientist to mutate the prompts and rank them with a vison model. Some of the winners where pretty neat. I think they have a lot more style than the gpt-5.6 ones. https://xcancel.com/xundecidability/status/20449185674144196...

irthomasthomas

4 dagen geledendiscuss

Incredible photography.

irthomasthomas

6 dagen geledendiscuss

This is what I do in llm-consortium. An arbiter evaluates the response(s) and decides if more iterations are needed. You can also loop until a minimum confidence threshold, but self-reported confidence isn't a great metric.

irthomasthomas

7 dagen geledendiscuss

Try this prompt: While working on the main task, launch a parallel sub-agent with the task context so far. The sub agent should think of high quality questions and put them to the user using a dialogue tool like zenity. Customize the inputs to the question, taking full advantage of the dialogue tools features to create a progressive interactive user experience. Ask only a few questions per turn so that you can adapt the questions to the answers.

This will keep you busy while the main agent runs. Customize it further to integrate the sub-agent answers to the main thread.

irthomasthomas

7 dagen geledendiscuss

>Chain of reasoning is a lot of context to guide token generation, but we simply see that newer models don’t need that context to get to the answer

I thought each new generation typically used more reasoning tokens?

irthomasthomas

7 dagen geledendiscuss

In think in all cases where I've seen it compared CC performed worse than a minimal harness.

irthomasthomas

9 dagen geledendiscuss

This feels more communist than communist China. They typically take about 1% in Golden Shares that give them a board seat.

irthomasthomas

9 dagen geledendiscuss

Claude in Claude code has been shown to perform persistently worse in evals than claude + a minimal harness.

irthomasthomas

11 dagen geledendiscuss

The comment from the other Mullvad founder is here https://news.ycombinator.com/item?id=48696800

irthomasthomas

12 dagen geledendiscuss

Very fast and reliable? Sold!

irthomasthomas

12 dagen geledendiscuss

GLM 5.2 is ~40B active parameters, which is what matters most for training cost.

irthomasthomas

13 dagen geledendiscuss

They provide benchmarks in the paper https:// arxiv.org/abs/2606.21228

irthomasthomas

14 dagen geledendiscuss

So only 100 companies have exclusive access to frontier AI.

irthomasthomas

14 dagen geledendiscuss

Weird, I didn't think I was throwing technocracy under the bus. What makes you say that?

irthomasthomas

14 dagen geledendiscuss

  > If we also use high temperature for more "creativity", the token sampler now may choose "Qwen".

If that was the cause then, like you said, it would sometimes pick Claude. But it doesn't, it consistently picks Deepseek (sonnet) and Qwen (opus). You can run it 100 times and see this behaviour much more than high temperature randomness would predict.

irthomasthomas

15 dagen geledendiscuss

Sounds like a branch of the Technocracy movement which Musks grandfather helped found.

https://en.wikipedia.org/wiki/Joshua_N._Haldeman

"The technocracy movement proposed replacing partisan politicians and business people with scientists and engineers who had the technical expertise to manage the economy"

irthomasthomas

15 dagen geledendiscuss

I thought it was better than that? It matches 4.8 in many evals and even beat Fable in Design Arena by a very healthy margin.

irthomasthomas

16 dagen geledendiscuss

Ask claude it's name in chinese and it thinks its Qwen (opus) or Deepseek (sonnet). Anthropic are just as guilty as everyone else training AI, today, maybe more so. Every lab borrows from every other. It only takes a few hundred samples to figure out the pattern; look at glm-5.2 reasoning using the caveman tongue of gpt-5.5. Stopping this would require some draconian surveillance.

irthomasthomas

Submissions

[untitled]

[untitled]

False Memory Syndrome Foundation

[untitled]

Don Lemon interviews Elon Musk (2024) [video]

Beijing: Highest-profile purge to date of senior military commanders

Robin Williams tickles Coco the monkey

The Jeffrey Epstein Affair – Joscha Bach

Microsoft has urged its employees on H-1B and H-4 visas to return immediately

comments