I think the normal way of explaining Gödel's Incompleteness Theorem is needlessly complicated. Here is an attempt to do better
Follow
Click to Follow lacker
Kevin Lacker
@lacker
Aspiring alien hunter. Formerly: Parse cofounder, Facebook eng manager, Google search quality engineer, college mathlete
Kevin Lacker’s Tweets
This idea is very interesting to me. I’m not sure what the right direction is, though. I feel like you want the AI to be connected the compiler, so that the AI knows what code compiles. But that is only a small start…
Quote Tweet
There is a huge opportunity to make a programming language that is optimized for LLMs to write so they produce more straightforwardly correct and optimizable code.
Show this thread
2
I hadn’t seen this sort of scam on Amazon before. Read the fine print I suppose
This is how I think about every AI demo video
Quote Tweet
2
"Only 50% of vinyl buyers own a record player."
1
5
I wish there were an AI browser plugin where I could be on a webpage with a podcast, and it would give me a transcription.
2
1
5
Twitter is frustrating when people post 10% brilliant, insightful commentary on some issue that I care about, like GPU programming, and 90% on issues that I don’t want to read about, like Australian politics. I can only imagine how annoying this is for non-Americans.
1
6
I think AI-based compression should be able to drastically reduce the size of chess tablebases. I wish there was a similar problem that had a larger demand for it being solved...
1
Another "editing" thing Copilot should be able to do - when I have mismatched delimiters, suggest how to fix it. (Especially common when Copilot and VSCode are battling each other to insert parentheses and braces for me.)
1
2
It feels like Copilot should be able to perform some standardized programming tasks besides new code generation. Like, "This method currently panics on many types of failure. Instead, handle all possible errors appropriately."
4
9
The point about linear order restricting “next token prediction” models is interesting. I wonder if code generation might be helped with an architecture that let it generate method signatures first, implementations last. Or something like a separate language model for debugging.
Quote Tweet
New blog post about diffusion language models: benanne.github.io/2023/01/09/dif
Diffusion models have completely taken over generative modelling of perceptual signals -- why is autoregression still the name of the game for language modelling? And can we do anything about that?
2
“You care for nothing but shooting, dogs, and rat-catching, and you will be a disgrace to yourself and all your family.”
- Charles Darwin’s dad when he quit med school
3
1
40
Interesting thread. Airplanes don’t flap their wings!
Quote Tweet
When it comes to similarities between the brain and deep learning, what's really striking is that everything that was actually bio inspired (e.g. sigmoid/tanh activations, spiking NNs, hebbian learning, etc.) had been dropped, while... (Cont.)
Show this thread
3
I'm glad to see this. Several friends of mine have suffered through noncompete clauses. I have never seen a legitimate use. I have only seen them used to punish employees for quitting. "We do so much, basically everyone is a competitor!"
1
6
When Cara Leon goes missing, Sam Mujrif is hired to investigate. Cara is 8 times taller than Sam, but evidence points to players much smaller than either of them—and technology with the potential to radically alter the balance of power between the scales.
gregegan.net/SCALE/SCALE.ht
read image description
ALT
25
51
269
Show this thread
I like Twitter view counts because sometimes I post about some obscure topic that I know nobody cares about, and nobody likes it, and the view count reassures me that at least I made you look at it
1
16
Complete hallucination. I'm not even mad, that's amazing. But you really can't use it in lieu of websearch!(Yet.)
2
I wish there was a Wikipedia that had 10 times as many topics, but was wrong 10% more often
3
1
If you start making up your own programming language, Copilot will happily start writing code in it, guessing what syntax it should use for new things you haven't thought about yet.
1
8
I would like anti-spam rules that require that all commercial marketing email be unsubscribable by API, not just by following a link to their website. That way I could just always unsubscribe from within my email client
1
1
12
“Globally, cats are responsible for the extinction of 14% of bird, mammal, and reptile island species.”
3
The NY Times parachuted me into SF so that I could write about a city and downtown about which I have no knowledge or experience.
90
134
1,631
Show this thread
I would like an AI that could translate tweets in a way that made any sense
2
2
It's frustrating when technical books on Amazon have half the reviews saying, "I ordered the normal version and got a lower-quality international one!" Especially when the book costs $200, the e-book costs $80, and Google finds a free PDF in two seconds.
In 1950 most Americans thought that World War 3 had already begun
A brilliant entrepreneur involved in many different industries who created a world-leading car company, took huge financial risks, and eventually went bankrupt: William Durant, founder of GM. Cars, refrigerators, bowling alleys, cinnabar mines…
1
Why is xy+xz generally "more simplified" than x(y+z) ?
10
2
I don't have high hopes for third-party app stores but it would be much less annoying if I could buy Kindle books from the Kindle app rather than, open Chrome, search for the book name plus "kindle", and purchase it from the website.
7
We need a form of chess AI that is like an anti-intelligence, optimized for getting itself checkmated by a 4-year-old who barely knows the rules
7
14
So "users assets are all backed 1:1", they just had to pause withdrawals, so that they could exchange one type of asset for another, to have enough to re-enable withdrawals. 🤔
1
1
"There is a phenomenon that all biologists will be aware of, where after working on a new idea for 2 years, you one day come across a paper from 2008 and say, “oh my god, if only I had known this two years ago.”"
Could AI help? Interesting article...
1
6
This tea always reminds me of my time in Y Combinator
1
14
ChatGPT is Girardian, completely driven by mimetic desire - it is not trying to optimize any personal utility, it is only trying to mimic what it observed in its training data as best it can. I don’t know whether this makes it more human or less human.
7
17
To everyone who interviews software engineers: if ChatGPT can solve your phone screen questions, you need to pick new phone screen questions.
6
2
17
I've been trying different use cases for chatGPT. It is pretty good at suggesting Christmas presents! It's a situation where I'm not looking for a specific answer, I just want it to keep throwing out ideas until one of them clicks with me.
2
4
My quest to get ChatGPT to answer a simple question incorrectly has succeeded
2
10
It would be interesting to make an AI model that converted JavaScript into TypeScript. You could automatically validate - it just needs to compile to the same thing, and also typecheck.
1
6
ChatGPT has this curious behavior where it often starts out by saying it doesn't know something, then it tells me a bunch of vaguely related stuff, and then finally it actually answers my question.
9
2
14
"We try to be very selective about the things that we do. We’re quite determined not to do things that other people do."
- Jensen Huang, Nvidia
2