Colin

5.9K posts

Colin banner
Colin

Colin

@hexane360

Materials scientist, computation enthusiast, mathematically curious

Katılım Nisan 2014
462 Takip Edilen127 Takipçiler
Colin retweetledi
Gregory Brew
Gregory Brew@gbrew24·
I brought this on myself
Gregory Brew tweet media
English
7
15
211
7K
Colin retweetledi
Dr. Jeffrey Lewis
Dr. Jeffrey Lewis@ArmsControlWonk·
Bremmer's framing is common but wrong. Instead: 1. Iran was developing an ICBM when 2. Khamenei imposed a 2,000-km range limit. 3. The programs shifted to space launch. 4. Khamenei lifted the restriction in October 2025 after the June attacks. 5. Now he's dead, and here we are.
ian bremmer@ianbremmer

turned out the iranian government was lying about not developing any ballistic missiles with range beyond 2000km. hardly surprising, but justification for removing their ballistic missile capability militarily.

English
17
190
593
42.5K
Colin retweetledi
Paras Chopra
Paras Chopra@paraschopra·
in our actual analysis, we chose max tokens as 32k (which is the limit of the APIs). in code, it is configurable. but even with 32k max tokens, we weren't getting the model to solve problems. if you notice, some of the people who ran the eval on chatgpt webui, it has taken full 17+ minutes chatgpt.com/share/69bd3015… This means probably 100k+ tokens before you get brainfuck (and even then the solution doesn't pass all the test cases for the problem). The API simply does not allow those many max tokens and webui isn't a reliable way to do evals. As we mentioned in the original thread, if you give recent models tools and lots of tokens, we're observing they solve these problems and that's a research we're continuing with x.com/lossfunk/statu… For the nth time, we were measuring model performance under constraints (such as zero-shot/few-shot or without tools). The paper clearly says all of that: arxiv.org/pdf/2603.09678 - Imagine if zero shot on esoteric languages was actually high. That would be a super strong signal of something like ASI, wherein these models are generalizing zero shot beyond their training data. Yes, our thread could have been written better, we're a young lab with newly minted researchers but there was no malice, no intention to mislead. The paper went through peer review and was accepted in two ICLR 2026 workshops. We're gathering feedback on the paper + our approach from the wider community, so if you're interested do send us your thoughts here: forms.gle/MLFUtcwsLiJhVt…
Paras Chopra tweet media
Levi@lev_ey

@paraschopra I noticed in the benchmark max_tokens = 8192. This seems incredibly low given the reasoning required here. Why was that value chosen?

English
17
10
113
17.1K
Colin
Colin@hexane360·
@Nexuist @sdamico I'm pretty sure there were always rules about parts being either KOM or COTS (or manufactured from KOM or COTS parts), going back to at least 2012
English
0
0
1
46
andi (twocents.com)
andi (twocents.com)@Nexuist·
This actually kind of happened to us. Our team went to FRC Worlds in 2016 where a SV based team wiped the floor with their Android based computer vision app. But it turns out that they were using a Google Pixel, which at the time was an unreleased smartphone, only accessible to the team because one of the parents was a Google employee. Nothing happened but they did update the rules to say that you weren't allowed to use hardware that wasn't generally accessible to everyone else after that
English
5
0
96
2K
Sam D'Amico
Sam D'Amico@sdamico·
Waiting for the first PED scandal involving Bay Area high school robotics teams
English
4
0
63
4.3K
Colin retweetledi
Glavny Konstruktor (Private for thought crime)
Lmao that would work with convection but would absolutely suck with radiation on vacuum. The radiators are in the LOS OF EACH OTHER why do you think radiators in real spacecraft are arranged in flat panels? How did you even model this before manufacturing?
English
12
4
243
17.2K
Colin
Colin@hexane360·
@GtaThoughts @TimothyDSnyder Are you even capable of spelling out the supposed logic here? Did anyone at any point claim IRBMs as an imminent threat that required starting a war over?
English
0
0
0
7
Timothy Snyder
Timothy Snyder@TimothyDSnyder·
There’s one thing stupider than spending 1.5 trillion to lose a war to a middle power in the Middle East, and that is spending 1.7 trillion to do so.
English
37
776
4K
78.6K
_GAi
_GAi@_GAi92·
@hexane360 They QRT'd my old account, which got doxxed. So now anyone who wanted to track me from the old account has an easy way of doing so.
English
3
0
0
1.2K
_GAi
_GAi@_GAi92·
Thank you, Skitty. Genuinely. My tweet got seen by like 20 people. Yours got seen by 30k+, AND YOU soft doxxed me. So now I have to contact the police again and make them aware of it. But hey, I guess since I'm not part of your community you don't need to worry about my safety.
English
14
0
8
6.3K
Colin retweetledi
interpretation
interpretation@materialcritic·
as any good writer will tell you, there are 7 basic types of story: 1. rags to riches 2. man vs man 3. man vs bee 4. voyage to meat world 5. coughing baby vs hydrogen bomb 6. amateur pornography 7. evangelion
English
105
3.1K
35.9K
425.9K
Colin
Colin@hexane360·
@HSVSphere @HostOfMeta Yeah idk why rust is this way, array indexing should be the same. arr[5] is clearly an &Item, not Item I remember seeing some explanation for the current behavior, but idk...
English
0
0
2
133
HSVSphere
HSVSphere@HSVSphere·
@HostOfMeta Both Rust and C get it wrong tbh It should only be bar.baz, but when bar is &Bar, bar.baz should be &Baz. Projection for the win!
English
2
0
16
1.5K
Jeremie Pelletier
Jeremie Pelletier@HostOfMeta·
Am in the camp that C's (.) vs (->) is a fantastic feature. When unifying both into (.) it hides memory indirection. Now reading code removes performance information :/ Sure not everything's embedded but everything's slow. GPUs get it right in decoupling linear uniforms from IO.
English
3
0
21
1.7K
Colin retweetledi
Will Manidis
Will Manidis@WillManidis·
an interesting subtlety of this story is how dramatically claude varies its responses to this question depending on who's asking. claude admits to sen. sanders that its collecting huge scales of data- but tells the President that it is mostly harmless and Clearview is the issue
Will Manidis tweet media
Sen. Bernie Sanders@SenSanders

I spoke to Anthropic’s AI agent Claude about AI collecting massive amounts of personal data and how that information is being used to violate our privacy rights. What an AI agent says about the dangers of AI is shocking and should wake us up.

English
6
24
258
32.2K
Colin
Colin@hexane360·
@avrilbradley23 CLT doesn't really want to emphasize western tradition, they just want a Christian test for Christian schools. If you take the ACT you might have to read about evolution in the science section
English
0
1
30
526
Colin
Colin@hexane360·
@ctjlewis @JustDeezGuy >Brainfuck’s grammar is too complex. >I'm not saying the grammar itself is complex Okay, you can see why this was confusing
English
0
0
1
10
Lewis 🇺🇸
Lewis 🇺🇸@ctjlewis·
No????? That’s why the grammar is simple. I’m not saying the grammar itself is complex, you are thinking I’m saying something I’m not, I know there’s only a few commands and a pointer, or whatever. I’m saying logically simple programs, which end up being simple Python expressions, become very complex (many steps) Brainfuck programs.
English
1
0
1
30
Colin
Colin@hexane360·
@ctjlewis @JustDeezGuy English has a much more complex grammar than either Python or brainfuck, yet LLMs perform fine
English
0
0
2
9
Colin
Colin@hexane360·
@ctjlewis @JustDeezGuy How is that relevant to whether the grammar is more complex? Often, the more complexity in the language the less in the final program. Do you deny that a Python interpreter is more complex than a brainfuck interpreter?
English
2
0
1
32