Maximilian Scholz (@[email protected])

1.7K posts

Maximilian Scholz (@scholzmx@fediscience.org) banner
Maximilian Scholz (@scholzmx@fediscience.org)

Maximilian Scholz (@[email protected])

@scholzmx

PhD candidate. Infrastructure (and pokemon) for Bayesian workflows, simulating everything Music, cooking, exercise enthusiast. https://t.co/OMwtTtfBQT

scholzmx.bsky.social Katılım Ekim 2019
193 Takip Edilen99 Takipçiler
Theo - t3.gg
Theo - t3.gg@theo·
@dixiidev Soon! We don’t actually control the harnesses at all, so this will require adding harnesses with OpenRouter support. Are you okay with OpenCode as the base harness?
English
23
0
255
21.8K
0xSero
0xSero@0xSero·
30 days AI usage Local - 50M tokens - free GLM - 150M tokens - 30$ Kimi - 150M tokens - 30$ Claude - 1B tokens - 110$ Codex - 19B tokens - 280$ Droid - 150M token - 100$ Total spend - 550$ Total usage 20.5B tokens
0xSero tweet media0xSero tweet media0xSero tweet media
English
54
22
687
53.4K
JMW 🤦🏾‍♀️
JMW 🤦🏾‍♀️@ionadelfina·
@scholzmx @Piklesuke Teeth getting cavities is a result of modern living and our diets full of sugar. I see what you’re doing with the Vit D thing and will not rise to the bait. Vitamin D deficiency in african populations is due to modern migration. You seem to be theone suffering from dunning kruger
English
2
0
0
28
Maximilian Scholz (@scholzmx@fediscience.org)
@ionadelfina @Piklesuke I can come up with plenty plausible scenarios. But your teeth getting cavities, your inability to produce vitamin D or proneness to Dunning Krüger didn't increase chance for survival in isolation. So maybe the purpose of your teeth isn't to fall out when you get old after all?
English
2
0
1
23
Maximilian Scholz (@scholzmx@fediscience.org)
@ionadelfina @Piklesuke We are arguing on very different levels here. I accept your original argument as something you'd tell a kindergartner. However, I respectfully disagree with your pov as more than a simplification. Eyes do so much more than just find prey and predator for humans.
English
0
0
1
11
JMW 🤦🏾‍♀️
JMW 🤦🏾‍♀️@ionadelfina·
@scholzmx @Piklesuke The purpose of your eyes is to see. A kindergartner can answer this l question. Seeing is what helped our ancestors to survive when other organisms did not
English
2
0
0
26
Maximilian Scholz (@scholzmx@fediscience.org)
@ionadelfina @Piklesuke I agree with the sibling that you're dodging the question and playing semantics. My eyes don't have a purpose. They weren't made to stare at a screen to increase my wealth. Yet I use them that way. They serve a need. But there is no "right" way of using them besides survival.
English
2
0
1
29
Maximilian Scholz (@scholzmx@fediscience.org)
@ionadelfina @Piklesuke That is not what survival of the fittest implies. There is no purpose. There is trial and error. And there's most definitely no purpose on a per-organ level. As long as a trait doesn't reduce chance of offspring to the point of extinction of the trait, it can survive.
English
1
0
3
48
JMW 🤦🏾‍♀️
JMW 🤦🏾‍♀️@ionadelfina·
@Piklesuke There is purpose in the structures that are created through evolution. If they did not help the organism to survive over others without those adaptations, they would not exist
English
4
0
0
489
alexine 🏴‍☠️
alexine 🏴‍☠️@alexinexxx·
learning how linux users live is teaching me a lot. a lot about why their failure rates are so high
English
31
0
141
16.8K
Clint Rutkas
Clint Rutkas@ClintRutkas·
Dev using Windows? Love to chat with you
English
139
19
189
30.3K
Vivian
Vivian@suchnerve·
This is why I’m a fan of a hand simply resting on the throat as a constant tactile reminder that choking COULD be occurring - you get the same dominance feeling, but without the physical risk.
English
24
73
2K
105.4K
Maximilian Scholz (@scholzmx@fediscience.org)
@x86machine @LottoLabs @0xSero I think his actions and results as well as the laurels he got from very respected community members prove your "no talent" assessment wrong. And nobody appointed anyone. You can just do things and apparently people value that. Your hypothetical researchers could try the same.
English
0
0
0
29
x86
x86@x86machine·
@LottoLabs @0xSero If he's appointing himself, and taking funds from the public, without any talent - that's kind of problematic. He just made sure another researcher with capabilities didn't get that compute, hurting the community. There's plenty of researchers here on X with talent.
English
6
0
9
18.1K
0xSero
0xSero@0xSero·
Composer-2 <3 Droid 5ever
0xSero tweet media
English
8
0
80
6.9K
0xSero
0xSero@0xSero·
Theo, can I push?
0xSero tweet media0xSero tweet media
English
14
6
337
50K
Maximilian Scholz (@scholzmx@fediscience.org)
@cameron_pfiffer Composer seems to be pretty popular. They might have managed to get enough RL data to keep their niche. And the ergonomics and performance of the harness will increasingly matter I'd guess. And as long as anthropic keeps developing CC with CC it'll stay wonky.
English
0
0
0
27
Cameron
Cameron@cameron_pfiffer·
@scholzmx Because they're a reseller of foundation models, Claude Code is eating their lunch, etc.
English
1
0
0
50
Cameron
Cameron@cameron_pfiffer·
Is Cursor in trouble?
English
1
0
0
411
Keith
Keith@gnukeith·
I don't quite get it, is the 27B model smarter than the 35-A3B model?
Keith tweet media
English
70
7
359
61.1K
Maximilian Scholz (@scholzmx@fediscience.org)
@htihle Matches my experience. It is a good model for how small it is (compared to the frontier). But it is small and most definitely benchmaxxed/RL-ed for specific things. What makes it interesting is performance/$
English
0
0
0
47
Håvard Ihle
Håvard Ihle@htihle·
Minimax m2.7 scores 36.9% on WeirdML, comparable to the orignal Deepseek R1 or Grok 3. This is much worse performance than would be expected based on the results from other benchmarks. I ran the model through Novita on openrouter. The outputs seem reasonable, but the instruction following is pretty bad sometimes, and it also makes a bunch of mistakes in reasoning. I'm open to re-running this if there is another inference option which I have reason to expect fixes some bug or some problem with the novita implementation. The fact that novita is serving this before it is released openly does suggest they have some kind of deal with minimax, and would hopefully at least test their implementation (or is this not safe to assume)? I find it hard when I'm worried that I'm not eliciting the full performance of the model, but I have to work with what I have.
Håvard Ihle tweet mediaHåvard Ihle tweet media
Håvard Ihle@htihle

WeirdML v2 is now out! The update includes a bunch of new tasks (now 19 tasks total, up from 6), and results from all the latest models. We now also track api costs and other metadata which give more insight into the different models. The new results are shown in these two

English
16
6
125
28.8K
Maximilian Scholz (@[email protected]) retweetledi
𝖗𝖊𝖉✞
𝖗𝖊𝖉✞@_redfeels·
Cutting off a narcissist from your life and radically accepting that you're going to be the villain in their delusional world is top-tier self care
English
62
4.2K
30.9K
832.4K