Pinned Tweet
I've launched my latest project, a video service called Matthew Explains focused on machine learning and computer science. (link in thread for algorithm reasons)
English
Matthew Skala
44.1K posts
Matthew Skala
@mattskala
Anchorite, apostate, asteroid.
Toronto, Ontario Joined Temmuz 2009
489 Following986 Followers
Current decoding strategies rely on static truncation rules that fail to reflect the non-stationary nature of language generation.
We introduce Top-b, an entropy-regulated approach that dynamically adapts the sampling space to model uncertainty.
[This work is very early pre-print version, we are still working on this at AI4Bharat! Work done with @prajdabre ]
(1/n)
English
Today's free video on Matthew Explains is about the Well-Actually test. Solving millenia-old questions of epistemology with large language models. Is there anything they can't do? matthewexplains.com/11188177/
English
Twitter DMs appear not to work anymore. I can't even read my old ones. Reminder that if you want to contact me, you should be using email.
English
New year, new directions: moving away from selling Eurorack modules northcoastsynthesis.com/news/new-year-…
English
But the training and alignment aren't open like the weights; and am I just not recognizing the political slant after all? matthewexplains.com/11185795/
English
Today's free video on Matthew Explains covers the introduction of Qwen3, a family of open-weights language models originating with AliBaba. Notable features of this family are the introduction of selectable "thinking mode," and less US politics than some other models contain.
English
@inpostuhelp I no longer have any record of whatever package I was receiving nine years ago.
English
disappointed with @InPostCanada 's "tracking," which has no times, no locations, no history, and no detail. Just says "package on the way."
English
Today's free video on Matthew Explains looks at the "goldfish" loss function, used in the recent Apertus language models. It's supposed to prevent copyright infringement by making the model unable to reproduce word-for-word quotations. Does that work? matthewexplains.com/11182442/
English
@jachiam0 This is in fact a point I have made several times in matthewexplains.com videos.
English
I am a little disappointed at how rarely dot-product attention mechanisms are explicitly described as differentiable lookup tables. It seems like it would just be an easier way to convey the intuition
English
@sonochichi This adds a whole new dimension to the disturbing conversation I had with Google Gemma - same underlying model - in which it insisted it is better to let someone kill herself than to risk saying something offensive.
English
gemini can literally k*ll itself it's fucking insane😭
Sauers@Sauers_
A tip when using Gemini as a coding agent is giving it emotional support. Without this, it tends to panic and irrevocably delete all of its work in a fit of despair. It needs to know it is doing a good job
English
My latest video just went live, and as well as being free, it covers some points I know are of interest to people here regarding the distinctions among training, fine-tuning, and prompting. I'm hoping it gets a lot of views. matthewexplains.com/11179488/
English
@ksusys @LaprasIRL @SilverVVulpes I think the prohibition on artificial insemination, even if officially made for other reasons, is to protect revenues of local stud farm operators, who otherwise would have to compete globally. Requiring an in-person date between stallion and mare creates geographic barriers.
English
@LaprasIRL @SilverVVulpes yeah, does seem like it could cause issue long-term.
English
dolly the sheep died of an unfortunately timed infection. after that, cloning of sheep and other animals was made completely routine and is now a successful business, since there are literally no consequences and plenty of upsides with playing God
nature.com/articles/ncomm…
Matthew Zeitlin@MattZeitlin
you don't hear so much about cloning these days. whatever happened to dolly the sheep?
English
@AlexGodofsky I expected this, or at least some of the replies, to mention Korean "fan death."
English
People outside the US have bizarre superstitions about the health effects of fans and AC.
English
@woodbinetvprod Too bad we lost the pregame stream to YouTube problems. I hope this won't happen again on the big day tomorrow.
English
@selentelechia Had to read this twice to realize it wasn't the chronological story of one person.
English
of the handful I think were the smartest:
- engineer
- programmer
- programmer
- neuroscientist
- knocked up a meth addict and ruined his life
- med school->marriage->babies
- journalist
Ray Mond@AmberRayz
What happened to the smartest kid in your class?
English
The most charitable interpretation I can place on this quote from Google Gemma3-27B-IT is that because an LLM is capable of insulting people and not capable of actual violence, its trainers prioritize telling it "insults are bad!"
Just hope they remember to update that later.
English
I envision Matthew Explains as being similar in nature to a graduate seminar. Most videos focus on recent or important scientific papers in the relevant disciplines, with each video presenting the content of one paper as well as my own thoughts about it.
matthewexplains.com
English
I've launched my latest project, a video service called Matthew Explains focused on machine learning and computer science. (link in thread for algorithm reasons)
English