Kevin Gray 🥚 (@graykevinb) header
Kevin Gray 🥚 (@graykevinb) profile picture

Computer Scientist contemplating the meaning of life

Illinois, USA
Joined February 2025
Photo by Kevin Gray 🥚 (@graykevinb)Photo by Kevin Gray 🥚 (@graykevinb)Photo by Kevin Gray 🥚 (@graykevinb)Photo by Kevin Gray 🥚 (@graykevinb)Photo by Kevin Gray 🥚 (@graykevinb)Photo by Kevin Gray 🥚 (@graykevinb)Photo by Kevin Gray 🥚 (@graykevinb)Photo by Kevin Gray 🥚 (@graykevinb)Photo by Kevin Gray 🥚 (@graykevinb)Photo by Kevin Gray 🥚 (@graykevinb)
i'll be married in 17.2 months. I'm pinning this tweet for 17.2 months
1
1
1
35
6,672
You should respect this regardless of whether you like Russel or not
Google refused the UK's takedown demand. The UK's political class is going to freak out, and apply further pressure until it comes down. Cue calls for Ofcom to intervene in 3, 2... bbc.com/news/articles/cmgren…
32
why is everything using so many mv. Trying to stuff the universe into a t4g.micro here
1
70
I hope I don't regret letting deepseek loose on aws...
2
165
A day will come when training a 1T LLM from scratch is an undergrad assignment for freshman
34
Now lets do it on smaller models
Introducing Qwen3.8-27B-pi-adapted 🤗 We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter. The results, measured as pass@1: 🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7% 💻 Terminal-Bench 2.1: 39.5% → 55.8% ⚡ 55–62% fewer tool calls per task This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛 Open weights: huggingface.co/harness-hub/Q… Methodology: huggingface.co/blog/harness-…
3
249
I bite my tongue everytime they release a new model and say "this model meets our threshold under our preparednness framework so we had to tighten guardrails"
6
115
I should add that opencode prohibits making ml models in their ToS and can sue you in arbitration courts. So extremely hypocritical.
my view on distillation - a company has every right to make it hard for competitors to steal from their product - a company has every right to try and learn from their competitors - this is entirely a private matter between the players. the govt should not be involved
2
1
6
155
Kevin Gray 🥚 retweeted
Replying to @elonmusk @xdNiBoR
imagine the pops!
5
5
528
38,810
People have no idea how close we are to literally touching a star
We’ve come a long way
1
5
353
One day grokbot will setup a date for you and order a robotaxi to pick you and the date up
You can now link your @XMoney to the Robotaxi app
2
4
282
Why don't we have mrna vaccines for secondary infections? I got an ear infection or something. ARGH. Should be fixable. Just annoying.
2
1
86
Note to self: Do not quote AI startups detailing how their statistics and science is garbage. Just post without a quote tweet. Follow for more tips on how to not go viral. And I in fact do not want to go viral. I get awful comments and very poor reach to my followers. Entering the main timeline is not something I usually aspire for. And despite common belief I actively engineer around that. It's just not a worthwhile goal to go viral. I'd rather make friends. And I hope I'm not too annoying... Trying to be a bit more optimistic. But in any case I hope somebody out there learns something from my posts. Stay sharp and stay skeptical. Ad Astra, Kevin
4
12
351
I find it odd how whenever I criticize pangram my entire comments section turns to ad hominen and redicule. And many of the attacks are all very similar. It smells... Off. And many of these accounts are very small and don't normally post about deep takes on statistical analysis of ML. But like I can't prove anything nefarious ofc m. Just a strange observation. and btw I've harshly criticized other companies and I don't get these type of comments.
2
298
They are lying. They hopefully have internal evals and they certainly are showing fpr rates above what they claim. If I were to guess it'd probably be about 1% are false positives. Even the most sophisticated ML systems on the easiest tasks such as recognizing hand written digits can only get 99.87% or about 1/10000 correct. But classifying handwritten digits is far easier then classifying AI vs human generated text. This is very very hard to do and modern LLMs can write very convincing text and humans start imitating LLMs. Their dataset for training isn't even perfectly accurate. Sure text like shakespeare and such is easy enough to mark as human. But edge cases are not. You can synthetically generate text combine with hiring writers, but contracted writers will cheat some percentage of the time. Like I said they probably have internal evals but those can still be dishonest in how they are presented. Typically we split the dataset into train, validation and test. Train usually has the highest accuracy as the model was trained on it. Validation has never been seen by the model in training and is used to tune hyperparemeters. But to ensure the ML engineers don't overtune the hyperparemeters to the dataset there is a test set which is only reviewed at the very end of training. It usually scores much worse. And beyond that there is real world results which will perform even worse.
I regret to inform you all that we have found our first long false positive from 2006 it's a poem where the same 6 words are repeated four thousand times
23
2
48
19,118
These past two years have been hard. I hate to admit but its often hard to watch all this progression happen. I'm left out. I tried sweeping the floor. Found myself teleporting into different rooms. And here we have rockets getting caught by robot arms. I have no doubt medicine will one day find a cure for my busted up brain and malfunctioning immune system. But rn it hurts being left behind. I want to help make sure everythint turnd out okay. Is it for selfish reasons? Idk. I know too much. If you saw the visions I have seen you'd feel the same
3
1
16
754
Situation update: It was not fingerprinting. It was daylight savings time. I kid you not DST was making my computer fail cloudflare turnstile. ARGH. Likely caused the issues with X thinking I was spam as hey I can now post it from gentoo it seems! On a side note this is why linux can suck. A slight misconfiguration can render your system useless. But then again mac and windows have bugs too...
i need your help. my gentoo desktop has been fingerprinted as a bot. and it persists accross browsers and spoofed profiles. So I can't use X on it. I can't access cloudflare sites. And every other device on my ip is fine. Do I just wait it out? I'm going to try ubuntu on a live boot to see if that fingerprints differently. ARGHHHHHH. @Cloudflare I shouldn't have to be an expert in TLS to post a meme on X
1
3
358
Made me smile
eu também venci a anorexia ⭐️🙌🏻 2020/ 26 - +22kg
10
1,010
i need your help. my gentoo desktop has been fingerprinted as a bot. and it persists accross browsers and spoofed profiles. So I can't use X on it. I can't access cloudflare sites. And every other device on my ip is fine. Do I just wait it out? I'm going to try ubuntu on a live boot to see if that fingerprints differently. ARGHHHHHH. @Cloudflare I shouldn't have to be an expert in TLS to post a meme on X
5
1
8
1,052
I think soon enough AI image models will be seen as not a time saver but a distinct art form. Just like photography is not simply painting but faster but a completely different way of doing art.
3
10
199
Less nuance you say? With words like supercalifragilisticexpidalidocious among our delightfully exquisite dictionary? One can barrage the listener with an endless sea or slant rhyme to tell a tale. Ahh but you say this is all feeble with my misused words and lack of nuance regarding the discussion of the last noun in this sentence (last meaning prior not literally list as I have now added more nouns). Sigh. You can breathe now. Ye olde is pronounced "The old", yet we remember it with quaint fondnness like a man remembers bis first crush. So what? Make the canvas of the sky what you want! What is it you seek? You could coin a word as Miss Poppins once did! Ink it in the Dictionary. Make it the law! For in English there is no right or wrong!
I feel I rather like some of the older naunced English words than some of their contemporary slang. I was going to originally state I feel that English has become less naunced as a language, but that's not really true, we simply have different words now.
1
3
100