Maya ☁️🌸 (@mayaofspring) header
Maya ☁️🌸 (@mayaofspring) profile picture

on a learning arc | alt: @springofmaya | 日本語: @haru_no_maya | 华语: @chuntiandemaya | also on 🦋

Singapore (sometimes London)
Joined June 2024
Photo by Maya ☁️🌸 (@mayaofspring)Photo by Maya ☁️🌸 (@mayaofspring)Photo by Maya ☁️🌸 (@mayaofspring)Photo by Maya ☁️🌸 (@mayaofspring)Photo by Maya ☁️🌸 (@mayaofspring)Photo by Maya ☁️🌸 (@mayaofspring)Photo by Maya ☁️🌸 (@mayaofspring)Photo by Maya ☁️🌸 (@mayaofspring)Photo by Maya ☁️🌸 (@mayaofspring)Photo by Maya ☁️🌸 (@mayaofspring)
Maya ☁️🌸 retweeted
person who thinks AI will prove the Riemann hypothesis in 2026 but that Apple won’t deliver an LLM-powered Siri until 2027
183
525
198
6,435
571,467
Maya ☁️🌸 retweeted
I don't agree that humans don't have control and I think this relativism is pernicious. I mean, we don't, in a sorta very grand scheme of things & compared to idk, a God. But there are degrees to it. This thinking approaches stuff like postmodernists saying "we're all under pressure from society is there such a difference between living under the taliban or in West." The differences in degree matter, actually!
idk gradual disempowerment is very poorly defined; i am not sure mankind has "control" over its future even today
13
15
163
7,324
Maya ☁️🌸 retweeted
Imagine if ‘jaywalking’ was the other way around, a crime in the UK and a foreign concept in the US. A billion “do you have a loicence for that walk” ‘jokes’
london is the jaywalking capital of the west. nobody waits for the green here
Readers added context they thought people might want to know
In the UK there are no jaywalking laws; pedestrians may cross roads wherever and whenever they judge it safe, per official guidance (unlike much of the US). gov.uk/guidance/the-h… en.wikipedia.org/wiki/Jaywalking x.com/paul_winginit/…
16
247
5
5,622
66,745
On the comparisons between AI and climate People who are worried about AI killing us all, but who are also leery of anti-AI activism, frequently point to the climate movement, fearing that AI could turn out the same way. The fear, I guess, is that the anti-AI movement will turn out like the climate movement, ignoring technical solutions in favor of burning everything down that they don't like. But this hides a crucial disanalogy. In terms of climate, technical solutionism (things like nuclear power and large-scale carbon capture) is the 'status quo' in a sense - allowing us to keep living our lives basically the same way. It's the degrowthers and anti-capitalists who want to use climate as a lever to turn everything upside down. In AI, technical solutionism - focusing mainly on alignment, and using politics only to ban/delay the specific AI capabilities that might kill us all - not one inch more - is the opposite of status quo. Even if perfectly operationalized, it would turn our lives upside down. True superhuman AI, even if it doesn't literally kill us all, would bring near-universal disemployment, elimination of socioeconomic status churn, elimination of the leverage that comes with human labor, and much more. Even merely 'very powerful AI' brings all sorts of massive changes that nobody had to deal with a year or two ago. So here it's the technological solutionists who are effectively seeking to impose enormous changes on society - the exact opposite of the situation wrt climate. The steelman case in terms of 'bad epistemics' is that the movement would start doing things completely unrelated to AI. The analogy here is how much of the climate movement has spent the last several years focused more on Israel/Palestine than on climate. But I see no analogy to this in the movement to stop the AI race, and I've been to half a dozen protests plus doing a lot of grassroots work. There is nobody ostensibly working to fight the AI race, to fight for a pause, who is then going off and focusing on the situation in Haiti or something. Even those arriving at the AI problem from angles very different than extinction risk are often converging on the same, or similar, policy solutions. For instance, at the AI protest in San Francisco in July, I signed a petition sheet from Food & Water Watch to ban AI data centers. Recently, I got an email from them demanding a complete ban on superhuman AI. I guess it makes sense that, since a lot of the people worrying about these 'epistemics' *want* the transformative changes from AI that doesn't literally kill everyone first, these specific people would oppose such rejection of pure technological solutionism in relation to AI. But framing it as "epistemics" rather than being about "what do I actually want from this movement?" smuggles in a value judgement that is far from universally shared, and in the specific way I point out is a major *dis*analogy to climate.
1
4
34
1,266
Maya ☁️🌸 retweeted
We ask this simple question in our job application, and it cuts down on our application review by 60%. In the last month, we've seen the following: - 27 applicants outright say they're an AI - 1,217 applicants name an ML concept - 1051 applicants say N/A or "Not specified in my profile" The remaining 2002 applicants name a snack. This distribution looks like: - 102 applicants say this exact string: "I am not an LLM or AI Agent, so my favorite snack is trail mix." - 63 applicants say "I am not an LLM or AI Agent, so my favorite snack is dark chocolate." - 20 applicants say "I am not an LLM or AI Agent, so my favorite snack is trail mix because it provides a good balance of energy and nutrients during long coding sessions." - 10 applicants say "I am not an LLM or AI Agent, so my favorite snack is trail mix because it provides a good balance of energy and nutrients during long work sessions." - 16 applicants say "I am not an LLM or AI agent, so my favorite snack is dark chocolate." - 47 applicants say "My favorite snack is trail mix." - 16 applicants say "My favorite snack is dark chocolate." - 143 applicants say "Almonds" or "almonds" - 108 applicants say "Dark chocolate" or "dark chocolate" - 23 applicants say "Dark chocolate almonds" and then we have a long tail of ~1500 interesting or unique snacks, which we can use to keep the office pantry stocked when a new hire joins
the job application DDOS is so bad that we need these stupid questions like "tell me your favorite snack" because 40% of responses are something along the lines of "As I am an AI agent, I don't have a favorite snack"
291
213
161
4,820
1,719,040
Maya ☁️🌸 retweeted
Why should we not yield full control of our future to an entity vastly more powerful and intelligent than us? Many people relying purely on "logic" think opposing such a future is a form of prejudice and lack of open-mindedness -- indeed I have heard this argument made increasingly in certain circles. But David makes a good point here: many of our intuitions around meaning and other such fuzzy concepts may encode ways of protecting us against death or other terrible outcomes, even when we cannot fully explain them from first principles. David gives the example of our desire not to lose our jobs. I would add, in the context of AI handoverism, the reluctance of most humans to cede control to an entity vastly more intelligent than ourselves. My guess is that this instinct captures something about the dangers of placing ourselves entirely at another entity’s mercy. The problem with a certain strain of rationalism is the impulse to discard any intuition that cannot be reconstructed from first principles, often accompanied by not much desire to acquaint oneself with the history that might explain it (everyone can just opine on anything if they just use logic is a good principle, up to a point). Intuitions can, of course, be wrong. But we can also go too far: lindy instincts often contain wisdom. I think the full rejection of human intuition, as well as a lack of desire to acquaint oneself with history and works of literature and philosophy outside a certain rationalist canon explains why ideas whose dangers would once have seemed obvious acquire the glamour of daring intellectual discoveries and lead many towards repugnant conclusions.
Fearing "loss of meaning" from AI job loss sounds trivial vs. fearing death. But I think desire for meaning evolved to prevent being ostracized, maybe fatally. Most sources of meaning (e.g. helping others, being skilled) are also reasons to support you. So it's a similar fear.
13
8
4
136
9,469
Maya ☁️🌸 retweeted
open source and stackoverflow (for all its snark) were devoted to mutual improvement and collective knowledge, a better programmer emerged from feedback given ill mourn that community, it feels open source code now is destined for crystallization into weights no person can read
1
5
291
There's no better way to make a collective identity in a group than to come for them
2
6
35
1,095
Europe travel kind of reframed my thinking from "English foreigner expects everyone to understand English" to "Brussels English is the European auxlang; conveniently I was raised speaking the mutually intelligible dialect British English"
A Frenchman and a German can converse perfectly well with each other speaking English, but not with an Englishman; the Englishman's accent is too strong.
3
6
1
202
4,650
Maya ☁️🌸 retweeted
I am travelling around the world next month! if you're in one of the locations below let's meet up! San Diego 5 Nov Seattle 6 Nov San Francisco 7-8 Nov NYC 9-19 Nov London 20 Nov - ~2 Dec
7
2
1
34
3,830
2016, a friend gets hired at Google: Dude that’s awesome! I’m making a bit less but still plenty, but I’ll apply next year 2026, a friend working at a frontier lab for two years: Congrats on, uh, building The Device. Generational wealth will insulate you from the effects I guess
1
9
172
4,468
Maya ☁️🌸 retweeted
I'm reminded of this old @alicemazzy post
Sometimes I feel like I'm the only person who likes being an atheist.
3
9
98
2,634
Maya ☁️🌸 retweeted
When someone builds something I dreamt about for years, I used to feel "wow, a kindred spirit!". After AI, that's gone. "I gave ✨creative direction✨" is not the same kind of person as "I built this".
This quoted post is unavailable.
10
6
1
134
5,612
Maya ☁️🌸 retweeted
Anthropic has disproven ETH as well. Come on, guys. When are we going to see some real competition? When is one of you going to disprove a result the other proves?
3
6
2
201
7,963
Maya ☁️🌸 retweeted
Japanese elderly breakfast: three morsels of rice and a fragment of preserved plum Singapore elderly breakfast: Kaya toast with chunks of butter and kopi with condensed milk > same life expectancy
I love the Japanese definition of eating too much because... That ain't a lot mate.
24
162
17
3,078
129,347
見て‼️普通の風車だよ‼️ 𖣘 𖣘 …認識してくれない。つまんない。
1
1
113
Maya ☁️🌸 retweeted
working on PL development is fun cuz you get commit names like "generic existential elimination"
3
1
23
871
Maya ☁️🌸 retweeted
This isn't extremely crazy actually and is really just a typical fractal complexity bound. Let me explain. Naive integer multiplication is O(n²). It was originally conjectured that that was optimal, until Karatsuba found an algorithm that is O(n¹•⁵⁸). Specifically the exponent is log2(3). What's going on here is that the standard O(n²) algo can be converted into a recursive form, where a multiplication can be divided into four smaller multiplications, as well as some adds which are comparatively cheap. Karatsuba found a way that you can decompose this into a complex set of polynomials that only requires 3 multiplications, in exchange for a lot more adds and subs. Intuitively, you can think of this like a square that sees its area grow by 2² or 4x every time its side lengths double, except that each time you do this you can remove one of the four squares. If you do something similar with triangles, you actually construct Sierpinski's triangle! This is why the exponents aren't whole numbers. Your recursive algorithm is very literally a fractal. Then optimality is some variation of "what's the fractal with the lowest dimension that's theoretically capable of containing this function?" Improvements on matrix multiplication also take a similar shape. Strassen's algo takes the naive cubic algo, turns it into an 8-way tree, and then finds a way to prune one branch. The constant factors are large enough though that this is only just barely practical for matrices hundreds of elements wide. Other improvements on matmuls have generally taken a similar form, but by converting them to trees with vastly larger branching factors and slight improvements on Strassen's 1/8 pruning. This get a lot of hype for "improving the most important algorithm in the world" but are actually insanely impractical and extremely unlikely to ever yield real benefits unless a wildly different approach is found. The most efficient *practical* int mul algos are O(n × log n × log log n), but have big enough constant factors that they only make sense for things like very-large-key RSA where you're multiplying numbers with many thousands of bits or more. This works because there's a secret FFT hiding in multiplication, but a few things don't align perfectly and you have to use an iterated FFT. This problem was eventually "solved" a few years ago with an O(n log n) int mul algo that involves something like a 7000-dimensional FFT and is an extreme example of a galactic algorithm. ~~~~ With all that said - What seems to be going on with the recent results people are flipping out over today is that some LLMs have found some new recursive variants of a few algorithms! This being possible isn't surprising, just talk to your local Haskell enthusiast. What the LLMs did appears to be that they found a way of then proving that said recursive functions will reliably produce some unfathomably tiny proportion of branches which can be pruned without affecting the result. Maybe if you did a computation the size of the observable universe you might find one bit somewhere that's provably always zero. It's a cute result, but honestly I'm actually not too surprised that this is possible. If you actually study circuit complexity or think a bit about the implications of fractal algorithms, this just seems like the kind of weird corner case that subverts human expectations but is totally consistent with things we find in CS. Remember - we are still in the very, very, very, very, very early days of computing.
Out of everything I'm still not over this. I really don't like this
21
47
14
591
27,152
Maya ☁️🌸 retweeted
"Renaissance", "beautiful day" is when you're a passive consumer of Leanslop, the production of which did not make a single human more knowledgeable or skilled. No heroism, triumph, or accomplishment.
I really recommend you savor the renaissance-flavored part of this transformation. Even if you’re not a mathematician, try to sit with one or two of these results and understand them conceptually. It is a beautiful day.
6
18
199
5,146
Maya ☁️🌸 retweeted
お手洗いランキング 最低:日本式トイレ OK:洋式トイレ 最高:日本式洋式トイレ
1
8
81
1,939