joshuaotti.bsky.social
Hi, My name is Joshua Cook Otti. I am an unemployed PhD in theoretical computer science and the author for stemforest books. Please hire me.
My website: https://www.cs.utexas.edu/~jacook7/
Stem Forest Books website: stemforestbooks.com
443 posts
149 followers
61 following
Getting Started
Active Commenter
comment in response to
post
Ironically flash worked better. It also failed, but at least it realized it WAS a probability puzzle. Even though I did need to correct many incorrect assumptions it made along the way.
comment in response to
post
"It was"
comment in response to
post
More pointedly, O(n) gates can describe any function whose TRUTH TABLE is n bits. So thd number of bits to describe a circuit are roughly equivalent to the number required to describe the truth table of a function.
More efficient circuits for n bit truth tables makes this relationship tight.
comment in response to
post
Every gate can be described with O(log(n)) bits. So we can describe every n gate circuit with O(n log(n)) bits.
Let n = 2^k. Every function on k bits can be computed with O(n) gates. There are 2^n functions, so it requires n bits to describe one. So n gates can encode Omega(n) bits of data.
comment in response to
post
n gates describes ABOUT 2^n functions. The intuition is that the size of a circuit is close to the number of bits required to describe the circuit.
comment in response to
post
Very roughly (ignoring some log factors) the size of a circuit is essentially equivalent to the number of bits required to describe the circuit (and the exact characterization is actually pretty tight.)
comment in response to
post
If you have n bit inputs, that's m = 2^n possible inputs. If the output is one bit, then that's 2^m possible functions. (Each input could output 0 or 1).
In contrast, k bits can only describe one of 2^k different functions. Thus you need m=2^n bits to describe all binary functions on n bits.
comment in response to
post
Geeze, similar thing works with the one lies, one tells the truth riddle.
comment in response to
post
Classic, gemini makes this mistake for me too. Asked if it was right it doubles down. When told its mistake, it determines the question is logically unsolvable! Then it suggests I use the original riddle because mine isn't any good.
comment in response to
post
Well I don't think we can get everyone to agree nearly anything can’t suffer. Some people think plants suffer.
I think our current LLMs won't be able to do this because they are designed to talk like a human, and people will feel like they are human. They just aren't built for this.
comment in response to
post
True, I see what you are saying.
comment in response to
post
Personally, I think consciousness is overrated. Like we need some notion of which things should get rights and which things shouldn't. But consciousness doesn't seem like the right framing for that.
comment in response to
post
I mean, roombas exist now.
comment in response to
post
"Understands" is quite a human centric idea. But there is clearly some distinction between:
1. An algorithm solving a problem.
2. A proof the algorithm solves a problem.
3. Being able to make new algorithms for a problem.
(Similarly for Heuristics instead of algorithms).
comment in response to
post
I think it's somewhere between A and 4.
To say it understands chess, I would need a way for it to demonstrate some abstract understanding of it beyond just playing the game. Such as answering questions about the game, hypothetical rules, and comparisons to other games.
comment in response to
post
*post docs.
I'm pretty sure the PhD students make even less.
comment in response to
post
Yeah some people make it work, in fact all the cs PhD students, but yeah it's those costs. And the ones with families are dual income.
comment in response to
post
But as a typical American, I'm not that knowledgeable about Canadian politics.
comment in response to
post
When you put the kind of money into research the US does, industry tends to follow. Unless you have crippling tax policy, industry is almost unavoidable.
Researchers also aren't that expensive. They'll take huge pay cuts to do research.
comment in response to
post
I mean, electrical sensing sharks do is also a real sense that we could also add to computers. So is the inner ear balance thing. They are just extremely task dependent thing.
For an AI helping a human, smell might be legitimately useful. But for an AI itself, maybe other senses would be better.
comment in response to
post
It's also not like we couldn't add smell to the models in the same way we add sight or audio. It's just a nonstandard input and we don't have much data for it. But we could add it. Just, why would we?
comment in response to
post
As someone who was offered a postdoc at the university of Toronto: they aren't putting enough money into it. To live in somewhere like Toronto, I need about 80k in USD per year. They offered 55k.
I'm back at Amazon for several times that much now, but I would have rather kept doing research.
comment in response to
post
I think it's because we read left to right. When I'm working on something, I usually put whatever I'm actively working on on the left, and references on the right. I always assumed this was a read left to right thing, makes the left thing "1st".
comment in response to
post
The industry is very weird right now. It felt like each interview was a coinflip between a 200k+ salary or unemployment. In fact, when I start, it will be a full two months after graduation. This is AFTER spending most of the semester job hunting.
comment in response to
post
Scientist leaving the field right here.
comment in response to
post
I tried to use gpt to prepare for behavioral questions in an interview. It kept saying all my answers were great.
I had to give it a blatantly awful answer for it to even reluctantly admit it might be a bad response, but it still spent the bulk of its response saying good things about what I said.
comment in response to
post
Yeah this is what shocks people out of academia. They ask how long it will take our universities to recover, and I say probably a couple decades. Everyone expects it to be like 5 years.
It of course depends just how bad it's cut, but with current proposals, yeah a couple decades.
comment in response to
post
It depends on what you mean by "publishing". Definitely promoting research on a blog is good. More easily digestible forms of your results are good.
But if it's a proper paper, also put it in arxiv, if you have time to go to one, submit it to conferences. Which conferences depends on the field
comment in response to
post
Yeah I'm having a "I just finished a PhD in CS and might have to move in with my in-laws in a month" crisis.
Mind you, that would be right after getting back from a conference in Prague, the top conference in my field, and I'm still having trouble getting a job.
comment in response to
post
With VR in particular, content is king. There is not a lot of VR content really taking advantage of the medium (and most of it is behind a paywall).
Tons of YouTube though.
VR is also kind of time consuming to get into and out of right now.
comment in response to
post
I don't mind talking about my research, I like my research. But if it looks like I'm not up to the coding bar BECAUSE I had to spend more time explaining my background, then that sucks.
Admittedly, I haven't heard back from them yet, but I didn't have time to finish my problems because of this.
comment in response to
post
I really did try to give the high level overview. But it's hard.
Me: we were asked to prove one PCP for the composition, which took some work but we did.
Them: What's composition?
Me: explains.
Them: so now you don't do composition?
Me: we do, explains.
Them: so you still have two PCPs?
Etc.
comment in response to
post
I get it, but I need to demonstrate that I know how to code. If you only give me 20 minutes to do a 30 minute coding problem because you need me to explain PCP composition, that feels like unfair discrimination against theorists.
comment in response to
post
I mean, Google persistent in the face of setbacks? Have you used a new Google service? Probably not for long because they tend to give up on them.
comment in response to
post
The other thing is that they want business metrics. My idea for new codes failed because it needs a kind of locally correctable code with better than known parameters that I couldn't construct.
Business impact? No paper? No one expected such a thing to exist, so it not existing yet isn't a problem.
comment in response to
post
I'm not inclined to lie to my employer. I'm also concerned that given how sycophantic chatGPT is lately that it won't give me good feedback.
comment in response to
post
To be fair, there are a variety of questions, and maybe half of them I have good answers for. But the other half are way tougher.
"Tell me when your idea was not the best action. What was the best action, who gave it to you."
Um, most of my ideas fail, and aren't really actions.
comment in response to
post
No, most of the behavioral interview questions involve business impacts / working with others especially when others are tough to work with. If you simply haven’t had to work with difficult people for a while under tight deadlines, you can’t answer the questions.
comment in response to
post
Idk, maybe companies just really like to brag that they have big offices in expensive cities and that's somehow valuable to them? If so, that seems like an inefficiency in the market.
comment in response to
post
Same. When I was an engineer I said this often and my coworkers were always like "well cost of living". Nonsense, if I'm delivering X dollars to the company, they should be willing to pay some fraction of X dollars to me.
Unless me living in a specific city has value to them, this is dumb.
comment in response to
post
This is pretty neat. Surprising that such a simple compression algorithm works so well. I always assumed anything that would work well would at least need to support gradients. It also seems like it encodes one line at a time. I wonder if an improvement can be made using both dimensions at once.
comment in response to
post
Avif?