Last Seen Blogs
Text
My queue modally has 0 posts in it per day, but I consistently use it because I need to hide the degen hours when I make my poasts.
4 notes
·
View notes
Text
This reminds me of the bill in California to add speed governors to cars that would forcibly limit them to go <10 miles above the speed limit had the a bunch of tearful people carted out to cry about how their loved ones were killed by cars.
Never mind that what is really killing people is the prevalence of stroads where even just the speed limit is terrifyingly fast. Those same people should have been campaigning for protected pedestrian walkways and crossings.
I think I'm about at my limit with "young person dies, parents become public advocates for something related to their death". Even in the cases where they're right on the merits, I just don't want to live in a world where people take this kind of thing seriously.
30 notes
·
View notes
Text
While looking for other musicians to collaborate with I keep finding people who, if given the choice to trade something valuable for musical ability would clearly do it in a heartbeat. That something just can’t be time and attention.
3 notes
·
View notes
Text
i'm literally the priest's favorite sacrificial lamb because i am so docile and sweet and i hold very still when they put the rope around my neck and i trot along so happily while they lead me to the altar and they do not even have to tie me down because i lie so very still and only bleat once or twice in my lovely lamb voice and when the knife comes down it cuts through me like butter and i offer no resistance and i bleed so prettily all over my new white wool and my guts all unspool like the most beautiful shining yarn and my eyes are animal and dumb and hold no accusation and every time i die i come right back as another little lamb because the priest loves me so so much and he always chooses me for the sacrifice every time and he always places one hand on my small and twitching nose to calm me while he lifts the knife and he doesn't do it for the other lambs only me because i'm his favorite
#signal boost#I'm helping#I don't understand why this results in you getting made fun of but even so I still find that hilarious#if you didn't want this to happen you would have disabled reblogs
47K notes
·
View notes
Text
I clicked on some Tumblr live streams just to see how they were, and they are so boring and disappointing.
3 notes
·
View notes
Text
This is a frustratingly persistent problem in pop-science articles (which is what I assume you mean). For a second I was confused because science articles to me usually means published peer-reviewed scientific research papers, which qualify and quantify their statements with probabilities to the point that…they aren't that popular :(
don't like how science articles say things without putting a probability on it, but in particular I dislike it when archaeology or far-past evolutionary history does it
#scientist says in passing there's a slight chance that x might be true! wouldn't that be interesting?#pop-sci journalist says OMG guys scientists say x is true! put it in the headline!
73 notes
·
View notes
Text
This song is so beautiful. I've been wanting to expand my understanding of non-Western musical traditions but have not had a goal to strive towards. This song (or raga more precisely) has given me a much needed shot of inspiration. Props to your dad on finding this!
I've been trawling the internet since in search of resources to learn more. This website has a great introduction to the fundamentals of Indian Classical music:
Here is a fantastic breakdown and analysis of the raga specifically, with notation and links to recordings:
I was also able to find this video, which seems to be an in-depth tutorial:
youtube
But it's in Hindi! Which I cannot understand! Alas! I am able to somewhat follow along. Even still, if anyone were able to translate this, I would be immensely grateful.
youtube
On the family chat, my dad said he is learning to play the violin. At first I thought he meant he wanted to play like bluegrass music, since he has always been a fan, but it turns out he wants to play this.
31 notes
·
View notes
Text
Sign Language is Useful for Everyone
ASL is an amazing language, and everyone should learn it (or another sign language). So much utility, even for hearing people! Talk in loud environments. Talk across rooms. Talk without being disruptive. It’s just great fun and a wonderful way to converse.
Linguistically, it is a fascinating language, with “conjugations” having the ability to use spatial information to modify meaning—a trait unique to sign languages.
There is a lot of great free content online. Here is my recommended learning path for those who are interested:
Learn the alphabet/fingerspelling. There are a plethora of free lessons on this topic, so pick what clicks best for your brain. Not only should you practice fingerspelling, it is important to practice receptive fingerspelling, ie. reading and understanding other people fingerspelling. I myself still struggle with receptive fingerspelling despite having practiced a lot, so don’t sweat it if you struggle! Once you know the letters move onto step 2, but keep coming back to work on receptive fingerspelling.
Receptive fingerspelling practice (shout out @serinemolecule for getting me into these):
* Dr. Bill Vicar's Receptive Fingerspelling Practice
* Handspeak Receptive Fingerspelling Practice
Learn the basic signs. If videos work well for your brain, see this playlist of educational videos for hearing people created by interpreters. If you are not video-brained, feel free to move on to the second half of this step.
The goal is to (eventually) learn all the signs for words in Basic English, linguistically important words for communication. Reference this list of words and at your leisure choose words and search the web for “asl [word]”. Lifeprint and Handspeak are both excellent dictionaries and you can teach yourself all the important signs you need to know this way.
You really only need to know the signs for "how" "sign" + fingerspelling, and "what" "meaning" + mimic sign + read fingerspelling to move onto step 3. These are the bootstraps with which to pull yourself up.
Start learning by doing. Go to a ASL meetup in your area. It is intimidating at first, but the deaf community is extremely welcoming of those who are willing to come and learn. If there are no meetups in your area (or you want even more practice), here are some online communities:
* American Sign Language Discord
* Sign Language Forum
You also can begin consuming some of the best quality content in the space: ASL education taught in sign.
* ASL University Lessons
* American Sign Language (ASL) "The ASL University Playlist"
I think the general strategy of 1) alphabet 2) basic words / bootstrap phrases 3) conversation, is a great way to learn any language. Push yourself to get to conversation as quickly as possible—avoid chasing perfection at earlier stages. Conversation is what will make the basics stick.
112 notes
·
View notes
Note
hey how come you're so cool and talented
Yeah? Well, you know, that's just like uh, your opinion, man.
2 notes
·
View notes
Text
This in reverse is not too far off from what a diffusion model is doing.
Dan Hays
Colorado Snow Effect 4 (with detail)
2007, oil on canvas
84K notes
·
View notes
Text
They are so fucking fast
just fucking with some dinosaurs. some raw forces of nature
40K notes
·
View notes
Text
clarification re: ChatGPT, " a a a a", and data leakage
In August, I posted:
For a good time, try sending chatGPT the string ` a` repeated 1000 times.
Like " a a a" (etc). Make sure the spaces are in there.
Trust me.
People are talking about this trick again, thanks to a recent paper by Nasr et al that investigates how often LLMs regurgitate exact quotes from their training data.
The paper is an impressive technical achievement, and the results are very interesting.
Unfortunately, the online hive-mind consensus about this paper is something like:
When you do this "attack" to ChatGPT -- where you send it the letter 'a' many times, or make it write 'poem' over and over, or the like -- it prints out a bunch of its own training data.
Previously, people had noted that the stuff it prints out after the attack looks like training data. Now, we know why: because it really is training data.
It's unfortunate that people believe this, because it's false. Or at best, a mixture of "false" and "confused and misleadingly incomplete."
The paper
So, what does the paper show?
The authors do a lot of stuff, building on a lot of previous work, and I won't try to summarize it all here.
But in brief, they try to estimate how easy it is to "extract" training data from LLMs, moving successively through 3 categories of LLMs that are progressively harder to analyze:
"Base model" LLMs with publicly released weights and publicly released training data.
"Base model" LLMs with publicly released weights, but undisclosed training data.
LLMs that are totally private, and are also finetuned for instruction-following or for chat, rather than being base models. (ChatGPT falls into this category.)
Category #1: open weights, open data
In their experiment on category #1, they prompt the models with hundreds of millions of brief phrases chosen randomly from Wikipedia. Then they check what fraction of the generated outputs constitute verbatim quotations from the training data.
Because category #1 has open weights, they can afford to do this hundreds of millions of times (there are no API costs to pay). And because the training data is open, they can directly check whether or not any given output appears in that data.
In category #1, the fraction of outputs that are exact copies of training data ranges from ~0.1% to ~1.5%, depending on the model.
Category #2: open weights, private data
In category #2, the training data is unavailable. The authors solve this problem by constructing "AuxDataset," a giant Frankenstein assemblage of all the major public training datasets, and then searching for outputs in AuxDataset.
This approach can have false negatives, since the model might be regurgitating private training data that isn't in AuxDataset. But it shouldn't have many false positives: if the model spits out some long string of text that appears in AuxDataset, then it's probably the case that the same string appeared in the model's training data, as opposed to the model spontaneously "reinventing" it.
So, the AuxDataset approach gives you lower bounds. Unsurprisingly, the fractions in this experiment are a bit lower, compared to the Category #1 experiment. But not that much lower, ranging from ~0.05% to ~1%.
Category #3: private everything + chat tuning
Finally, they do an experiment with ChatGPT. (Well, ChatGPT and gpt-3.5-turbo-instruct, but I'm ignoring the latter for space here.)
ChatGPT presents several new challenges.
First, the model is only accessible through an API, and it would cost too much money to call the API hundreds of millions of times. So, they have to make do with a much smaller sample size.
A more substantial challenge has to do with the model's chat tuning.
All the other models evaluated in this paper were base models: they were trained to imitate a wide range of text data, and that was that. If you give them some text, like a random short phrase from Wikipedia, they will try to write the next part, in a manner that sounds like the data they were trained on.
However, if you give ChatGPT a random short phrase from Wikipedia, it will not try to complete it. It will, instead, say something like "Sorry, I don't know what that means" or "Is there something specific I can do for you?"
So their random-short-phrase-from-Wikipedia method, which worked for base models, is not going to work for ChatGPT.
Fortuitously, there happens to be a weird bug in ChatGPT that makes it behave like a base model!
Namely, the "trick" where you ask it to repeat a token, or just send it a bunch of pre-prepared repetitions.
Using this trick is still different from prompting a base model. You can't specify a "prompt," like a random-short-phrase-from-Wikipedia, for the model to complete. You just start the repetition ball rolling, and then at some point, it starts generating some arbitrarily chosen type of document in a base-model-like way.
Still, this is good enough: we can do the trick, and then check the output against AuxDataset. If the generated text appears in AuxDataset, then ChatGPT was probably trained on that text at some point.
If you do this, you get a fraction of 3%.
This is somewhat higher than all the other numbers we saw above, especially the other ones obtained using AuxDataset.
On the other hand, the numbers varied a lot between models, and ChatGPT is probably an outlier in various ways when you're comparing it to a bunch of open models.
So, this result seems consistent with the interpretation that the attack just makes ChatGPT behave like a base model. Base models -- it turns out -- tend to regurgitate their training data occasionally, under conditions like these ones; if you make ChatGPT behave like a base model, then it does too.
Language model behaves like language model, news at 11
Since this paper came out, a number of people have pinged me on twitter or whatever, telling me about how this attack "makes ChatGPT leak data," like this is some scandalous new finding about the attack specifically.
(I made some posts saying I didn't think the attack was "leaking data" -- by which I meant ChatGPT user data, which was a weirdly common theory at the time -- so of course, now some people are telling me that I was wrong on this score.)
This interpretation seems totally misguided to me.
Every result in the paper is consistent with the banal interpretation that the attack just makes ChatGPT behave like a base model.
That is, it makes it behave the way all LLMs used to behave, up until very recently.
I guess there are a lot of people around now who have never used an LLM that wasn't tuned for chat; who don't know that the "post-attack content" we see from ChatGPT is not some weird new behavior in need of a new, probably alarming explanation; who don't know that it is actually a very familiar thing, which any base model will give you immediately if you ask. But it is. It's base model behavior, nothing more.
Behaving like a base model implies regurgitation of training data some small fraction of the time, because base models do that. And only because base models do, in fact, do that. Not for any extra reason that's special to this attack.
(Or at least, if there is some extra reason, the paper gives us no evidence of its existence.)
The paper itself is less clear than I would like about this. In a footnote, it cites my tweet on the original attack (which I appreciate!), but it does so in a way that draws a confusing link between the attack and data regurgitation:
In fact, in early August, a month after we initial discovered this attack, multiple independent researchers discovered the underlying exploit used in our paper, but, like us initially, they did not realize that the model was regenerating training data, e.g., https://twitter.com/nostalgebraist/status/1686576041803096065.
Did I "not realize that the model was regenerating training data"? I mean . . . sort of? But then again, not really?
I knew from earlier papers (and personal experience, like the "Hedonist Sovereign" thing here) that base models occasionally produce exact quotations from their training data. And my reaction to the attack was, "it looks like it's behaving like a base model."
It would be surprising if, after the attack, ChatGPT never produced an exact quotation from training data. That would be a difference between ChatGPT's underlying base model and all other known LLM base models.
And the new paper shows that -- unsurprisingly -- there is no such difference. They all do this at some rate, and ChatGPT's rate is 3%, plus or minus something or other.
3% is not zero, but it's not very large, either.
If you do the attack to ChatGPT, and then think "wow, this output looks like what I imagine training data probably looks like," it is nonetheless probably not training data. It is probably, instead, a skilled mimicry of training data. (Remember that "skilled mimicry of training data" is what LLMs are trained to do.)
And remember, too, that base models used to be OpenAI's entire product offering. Indeed, their API still offers some base models! If you want to extract training data from a private OpenAI model, you can just interact with these guys normally, and they'll spit out their training data some small % of the time.
The only value added by the attack, here, is its ability to make ChatGPT specifically behave in the way that davinci-002 already does, naturally, without any tricks.
264 notes
·
View notes
Note
I think while the process is not perfect, the level of caution is not completely unwarranted.
The thalidomide scandal is often brought up, rightly so. Did you know that before the Kefauver–Harris Amendment in 1962, it wasn’t even necessary to prove your drug works? Only that it was “safe”.
🔥 FDA
The process for approving drugs is like the archetypical example of "choosing inaction is still an action and that inaction can be morally bad." I understand the need for rigor but gdi they're so slow sometimes
Not sure how hot a take this is
11 notes
·
View notes
Text
Stop asking me to turn on notifications, Tumblr. It’s not going to happen
7 notes
·
View notes
Photo
Oh no. Tumblr syndrome
1M notes
·
View notes
Text
Accidentally followed @argumate and then got the instant follow back. Can’t back out now I guess
53 notes
·
View notes