Andrew M. Webb

1.6K posts

Andrew M. Webb banner
Andrew M. Webb

Andrew M. Webb

@AndrewM_Webb

Machine learning & pop sci animations | 𝗧𝗼𝗽 𝘁𝘄𝗲𝗲𝘁𝘀: https://t.co/aEXwFoJKus

Manchester, England Katılım Mayıs 2015
249 Takip Edilen12.5K Takipçiler
Sabitlenmiş Tweet
Andrew M. Webb
Andrew M. Webb@AndrewM_Webb·
If you have a distribution, with density p, that you don't know how to sample from, you can still estimate integrals like ∫f(x)p(x)dx by sampling: Sample from a different distribution, with density q, then weight your samples by p(x) / q(x) This is called importance sampling
English
7
121
900
0
Andrew M. Webb
Andrew M. Webb@AndrewM_Webb·
Ｂｌｏｇ　ｐｏｓｔ　ｓｈｏｗｓ ｈｏｗ　ｔｏ　ｇｅｎｅｒａｔｅ ｆｕｌｌｙ　ｊｕｓｔｉｆｉｅｄ ｍｏｎｏｓｐａｃｅｄ　ｔｅｘｔ ｗｉｔｈ　ｎｏｎ－ｇｒｅｅｄｙ ｃｏｎｓｔｒａｉｎｅｄ　ＬＬＭ ｓａｍｐｌｉｎｇ echostatements.net/posts/2026/02/… by @EchoStatements
English
0
1
4
591
Siraj Raval
Siraj Raval@sirajraval·
@mandeepabagga @lexfridman Yes. In 2019, I used GPT-2 to generate a research paper. @AndrewM_Webb called it out as plagiarized, which got him so much attention that his tweet is still in his Bio as his claim-to-fame: x.com/AndrewM_Webb/s… . 5 years later, AI in papers is common: hai.stanford.edu/news/how-much-…
Andrew M. Webb@AndrewM_Webb

So in @sirajraval's livestream yesterday he mentioned his 'recent neural qubit paper'. I've found that huge chunks of it are plagiarised from a paper by Nathan Killoran, Seth Lloyd, and co-authors. E.g., in the attached images, red is Siraj, green is original

English
2
0
0
1.1K
Siraj Raval
Siraj Raval@sirajraval·
Lex Fridman deleted his interview with me & scrubbed it from the Web, even the Internet Archive. Ironically, he later gave a platform to people like Matt Cox, who stole millions, while preaching 'love is the answer' Here's the video he erased. Decide for yourself. @lexfridman
Siraj Raval@sirajraval

It was a pleasure speaking to my friend @lexfridman on his podcast! I revealed more about myself here than in any other conversation so far. Worth a watch!

English
16
9
99
55.9K
Andrew M. Webb
Andrew M. Webb@AndrewM_Webb·
@fchollet Also: how do you think hard when there's the temptation to have your thoughts pre-formulated?
English
0
0
0
171
François Chollet
François Chollet@fchollet·
I find the idea of having a ready-to-use "first draft" for everything to be mildly uncomfortable. How do you think creatively in a world where your thoughts are pre-formulated for you?
English
17
10
176
22.6K
François Chollet
François Chollet@fchollet·
The normative effect of AI: If generative models tend to produce stereotypical & average content (since they're compressing a large corpus of human content), and if humans start leveraging them at scale to create new content, does that make culture more stereotypical & average?
English
57
42
401
104.6K
François Fleuret
François Fleuret@francoisfleuret·
I do believe that allowing hand-drawn figures in articles would strictly improve scientific communication. It works on whiteboards for 99% of scientific figures ever drawn. @NeurIPSConf @icmlconf @iclr_conf
English
11
3
34
33K
hardmaru
hardmaru@hardmaru·
“Pictures of perfect human hands, no extra fingers” This is art! #AIart
hardmaru tweet media
English
29
20
168
49.4K
Greg Egan
Greg Egan@gregeganSF·
@wtgowers I’m starting to wonder how many illicit love affairs have been revealed when someone’s spouse borrows their phone, and though the incriminating texts have been deleted, their statistical ghosts linger in the autocomplete.
English
1
0
11
587
Timothy Gowers @wtgowers
Timothy Gowers @wtgowers@wtgowers·
Reveal your deepest self by starting a WhatsApp/text message with "I" and then accepting all prompts. Think of it as ChatGPT in microcosm, trained on you. I'll start (in next two tweets).
English
15
2
30
22.2K
Greg Egan
Greg Egan@gregeganSF·
Not sure that Twitter is going to make anyone feel better now that the default is to display the number of times a tweet has been seen, forcing us all to ponder the fact that only about 0.01% of people are then liking / replying / retweeting.
English
15
11
99
17.7K
Andrew M. Webb
Andrew M. Webb@AndrewM_Webb·
I hope either somebody finds this useful or there's an easier way that somebody can share with me Explanation of the problem and the solution on the other site. Hope that's legal. @andrewm_webb/109399720674933598" target="_blank" rel="nofollow noopener">sigmoid.social/@andrewm_webb/… @andrewm_webb/109542446578780498" target="_blank" rel="nofollow noopener">sigmoid.social/@andrewm_webb/…
English
0
0
0
1.3K
Andrew M. Webb
Andrew M. Webb@AndrewM_Webb·
It seems (?) like there isn't an easy way to use PyTorch's transformer implementations autoregressively at prediction time efficiently; in each iteration of the loop the whole sequence attends to the whole sequence. Here's a solution I came up with: gist.github.com/grey-area/100c…
English
2
0
4
2.2K
Andrew M. Webb
Andrew M. Webb@AndrewM_Webb·
@francoisfleuret While it seems weird that it can't work out that the type of [s] is Union[List[int, str]], is it Union you want here? Union doesn't enforce any kind of type matching, between variables, so if your blah function took an int as input and (somehow) gave back a List[str], that…
English
1
0
1
742
François Fleuret
François Fleuret@francoisfleuret·
I understand that figuring out that the return type specified is correct amounts to some halting problem equivalent, but there is no way to hint the return expression to tell mypy "I know what I do"?
François Fleuret tweet media
English
5
0
9
5.1K
Steve McCormick
Steve McCormick@Quasilocal·
Ok, turns out the first one was luck lol. Here is the Eiffel Tower, an "extremely detailed racecar" and a "cool guy's face" 🙃
Steve McCormick tweet mediaSteve McCormick tweet mediaSteve McCormick tweet media
English
1
0
4
0
Steve McCormick
Steve McCormick@Quasilocal·
Ok... It passed the simple test first go. I wonder how complicated it can be though😬 (also it's slightly disconcerting that it chose a sad smiley to draw)
Steve McCormick tweet mediaSteve McCormick tweet media
English
2
0
4
0
Andrew M. Webb
Andrew M. Webb@AndrewM_Webb·
@AjdDavison I can't find the tweet I'm looking for, but I've seen people point out that it sometimes seems to get in a "defensive mode" and start doubling down if you contradict it in certain ways ("But don't you think...?")
English
0
0
0
0
Andrew Davison
Andrew Davison@AjdDavison·
I always thought that these kind of physical situation problems would be the kind of things I'd ask in a Turing test. Yes, you can sense #ChatGPT getting confused if you keep going, but on the whole I'm astounded by how good it is.
Andrew Davison tweet media
English
3
2
43
0
Andrew M. Webb
Andrew M. Webb@AndrewM_Webb·
@gregeganSF I think the fact that entropy can increase with an observation disagrees with my intuition at some level. Something to do with mentally glossing over the fact that entropy is a thing in expectation
English
1
0
2
0
Greg Egan
Greg Egan@gregeganSF·
The Shannon entropy: H= –Σp_i log(p_i) often decreases when you learn something. For example, if you blindly pick a card from a normal deck, H is log(52) when you know nothing about the card, but if you learn it’s diamonds, H becomes log(13). But a decrease isn’t guaranteed.
English
2
6
72
0
François Fleuret
François Fleuret@francoisfleuret·
What transformer do you use for your own stuff?
English
3
0
5
0
Andrew M. Webb
Andrew M. Webb@AndrewM_Webb·
@william_woof It looks like named tensors didn't make it past the 'prototype' stage I've recently been using einops quite a lot, which lets you name dimensions when doing reshape/permutation/stacking/reduction/etc operations twitter.com/soumithchintal…
Soumith Chintala@soumithchintala

I am really excited to see prototypes like dynamo, lazy, fx, functorch, nvfuser etc. and I’m pretty confident that they will consolidate into a disruptive next set of experiences. I’m proud that we’ve built prototypes like named-tensor, even though they didn’t work out.

English
1
0
1
0
William Woof
William Woof@william_woof·
@AndrewM_Webb Slightly unrelated, but did you ever have a look at named tensors?
English
1
0
0
0

Keşfet

@EchoStatements @sirajraval @mandeepabagga @lexfridman @Quasilocal @fchollet @francoisfleuret @NeurIPSConf