I'm logging out and taking an indefinite break from Twitter, which has many cool people, but sadly has become a haven for literal Nazis. Happy to chat at alyssamvance
@gmail
.com or on Signal (203-850-2427), or Discord. Thanks to everyone who followed me or posted good content :)
Why don't companies have sports teams, like colleges? Many people get super into sports as kids, then when they graduate college, they lose a favorite hobby. I bet the Coke vs. Pepsi, Ford vs. GM and Apple vs. Microsoft games would get big audiences!
The US medical practice of "you aren't allowed to know the price, and then we will bill you, two years later, for whatever amount we retroactively make up" is horribly abusive and should be illegal
My strongest political belief isn't really liberal or conservative. It's that most people, laws, and institutions in the modern US are absurdly risk-averse, and that we're all suffocating in the name of 'safety'.
Copilot/ChatGPT won't eliminate coding jobs; every time code has gotten easier to write, demand for code has grown even faster. A ton of programming in 1982 would be trivial now
Just read about teenagers being "encouraged to show leadership". How the hell could I have "led" anything at age 16? I couldn't open a bank account, visit the park, write a blog, have a friend over for lunch, or buy a gallon of milk.
This is just embarrassing. Did you know that ChatGPT can't understand the word "not", because large language models filter stopwords from their training data like it's 1992? 🙄
Car alarms should be illegal. So should gas leaf-blowers, unmuffled motorcycles, and huge amplified car stereo systems. Noise pollution is real and bad
This is a very pretty graphic, but I suspect it'd be much less popular if it said "we should ban cars, central heating, air conditioning, and hot water, for everyone, permanently", which is mathematically what it amounts to
If inference time is linear in model size (big if), we can estimate GPT-4's param count by timing it. GPT-4 takes 12x as long as GPT-3 to generate these 400 tokens (~12 seconds vs. ~145), implying a param count of 175B * 12 ~= 2.1 trillion.
If you had a drop of water for each floating-point multiplication used to train PaLM, Google's new 540B parameter language model, it would just about fill the Pacific Ocean
I wish there was a word for "rules with focused benefits and diffuse harms". Eg, suppose the fire code had a rule that stopped 3 fires a year, but cost $1.4B to implement. Socially, that's a huge loss. But it's hard to say "yes, your house should have burned to save everyone $6".
"Find a therapist" is useless advice >95% of the time, you're better off just saying "I don't know what to do"
People want to believe "therapy" is the cure for any mental issue, just like antibiotics cure infections. Because that would be convenient. But this rarely happens
Gilgamesh is one of the oldest works of literature, but new editions keep coming out because they keep finding more of it. Compare the first image (1998) with the subsequent ones (same section, 2021)
LSD should be legal, but IMO many people overrate it. It seems like a typical trip creates "insights" that can't be communicated and don't have much effect on the world afterwards
When the Web was as old as Ethereum is now, Yahoo was already serving ~150 million users a year. Nothing based on Ethereum or similar tech has anything like that level of mass adoption, and I don't currently know of any plausible candidates.
Is there a name for the fallacy of "you include awful person X in your community because you see their contributions, but you don't see the contributions of the hundreds of others they drive away?"
I'm turning 31 next week, I graduated college ten years ago now. It's insane to think of how a surgeon my age would *still* be "in training" in a residency program. Is this really necessary?
Transformers can write text, but they can also estimate scores / do regression for any text you give them. Here's a quick demo I made (RoBERTa fine-tuned on Yelp review data)
EA wants to be both "average workers give 10% of their income to help the global poor" and "billionaires fund high-risk high-payoff research to improve humanity's long term prospects", but these just aren't very similar things
Effective altruism is in a weird state where many 1) agree AI is the most important problem, 2) don't have a good plan for fixing it, 3) also don't want to dedicate themselves to a lower priority goal, so there's a lot of ineffectual flailing and good ideas left unbuilt
I think Eliezer was wrong and that GPT-4 is the fire alarm for AGI, ie., an event that creates mutual common knowledge that AGI is coming and you won't look weird for talking about it
Part of why more people are taking AI risk seriously is that safety advocates have spent years writing up their arguments in a lot of detail, while most skeptics have stuck to unpersuasive name-calling, arguments from psychoanalysis, and Twitter dunks
Dumb question: there seems to be a lot of debate on "do minimum wages lower employment". But isn't the right question "what minimum wage would lower employment"? Surely $1/hr would have no meaningful effect, and $100/hr would have some.
Many have very conveniently forgotten that Altman himself was made head of Y Combinator at age 28, as a college dropout, with a resume consisting of a single failed startup. (And I don't think it was a bad decision!)
One reason I'm a bit skeptical of the "AI eats the Internet, becomes a million times smarter" story is that the Internet's bandwidth is really not very large. ~100 H100 GPUs would saturate every transatlantic cable combined
Why is there such a strong link between being trans and the autism spectrum (and hence, it seems, futurism, the Rust programming language, and various other niche interests)? I've been researching it for over a decade, and I still don't know 🤷♀️
48% of AI researchers think AI has a significant (>10%) chance of making humans extinct
58% believe AI alignment (by Russell's definition) is "very important"
Most think human level AI is likely within our lifetime
I might honestly feel less nervous about OpenAI if their platform were "screw caution, push on the gas as hard as we can", since that's at least honest. It feels sociopathic to talk about slowing down AI, then do weekly huge releases before anyone can even digest the last release
The Four-Stage Strategy:
1) “nothing is going to happen”
2) “something may be going to happen, but we should do nothing about it”
3) “maybe we should do something about it, but there’s nothing we can do”
4) “maybe there’s something we could have done, but it’s too late now”.
I have a new long post on Less Wrong, which aims to show that, over the next decade, it is quite likely that most democratic Western countries will become fascist dictatorships:
The AI Risk Four-Stage Strategy:
1) nothing is going to happen!
2) something may happen, but, we don't need to do anything about it
3) maybe we should do something about it, but there's nothing we can do
4) maybe there's something we should've done, but it's too late now!
Probably the biggest thing countries can do to boost fertility is reduce education requirements for jobs. Globally, there's a very strong negative correlation between fertility and school years - it makes each child take longer & cost more, and you don't have kids when in school.
As we become inured to AI art, artists will focus more on making work that can't even be casually mistaken for AI. Grotesquely detailed hands. Entire paragraphs of microcalligraphy. Big smiles with mathematically precise rows of teeth.
My best guess is that school mass shootings are mostly a social contagion that spreads via the media? There were only three recorded (with >2 deaths) in the US pre-1960, although the US has always had lots of guns.
This seems pretty silly? Any fertility number, extrapolated to infinity, gives you either Malthusian catastrophe or extinction, but trends also rarely continue for >1,000 years straight even after negative feedback loops.
I think we have to start considering human extinction as a real possible outcome of demographic transition.
Natural selection doesn't prepare species in advanced for hurdles they face, and there is little evidence to think there are human subpopulations resistant to it.
Quick reminder that Steve Ballmer has more money than all effective altruists combined, has donated almost none of it, and basically no one has criticized him for that
I'm sure many have heard this, but for the record, "public health experts" treating e-cigarettes as basically the same as smoking is super bad and probably kills tens of thousands of people a year
""Serial entrepreneur" implies that starting startups is something people do casually, as a hobby. Starting a startup is not merely another project; it's more like having a second child. You wouldn't call someone who had two kids a serial parent." - Llama trained on Paul Graham
This Llama fine-tuning repo was
#1
on Hacker News yesterday. They claim an astounding 30x improvement in training time, but I'm pretty sure most of that is from sandbagging their benchmarks. THREAD:
There are many many points of disagreement, but anyone who says "lol people who *really* understand AI know humanity isn't in danger" shouldn't be taken seriously
Me before reading: "this study, like ~every study, will detect no effect on all-cause mortality because that needs a crazy big sample size"
After reading: "this study, like ~every study, detected no effect on all-cause mortality because that needs a crazy big sample size"
I'm curious why someone who, as editor of Jared Kushner's tabloid paper, sent an undercover reporter to my 21st birthday party in my private apartment, then outed me as transgender on the front page for no particular reason, is considered qualified to teach journalism ethics.
@ajaxsinger
I am required to make sure my students understand journalistic ethics, not just in the context of my class but in general. Journo ethics are classes by themselves.
I'm all for humor, but dude, this is Seriously Not Funny. If the commander of USSTRATCOM tweeted about incoming missiles as a "joke", they'd be fired if not sent to jail.
Trying to fix global warming with voluntary poverty is not going to work, for many many reasons. There have always been those who see honor in poverty, like Christian monks, but it's never going to be a large majority of the world.
This should almost certainly be higher. Lots of hospital patients are very old, terminally ill, can't hold a coherent thought or get out of bed, are in extreme pain, and are kept alive temporarily by futile treatments. I don't want to die that way.
It's funny how many people leap from "I don't see a way to stop X" to "X is good actually". I don't know how to stop hurricanes, but they are still bad!
This is silly. Half of the Chinese software industry only exists because it's mostly illegal to use Google, Facebook and other non-Chinese apps there. There's a reason why people in Brazil, India, Italy, Spain, etc. all use Google and not Baidu.
I don't think "cis" vs "trans" is really a binary. Rather, people have a bell curve of gender preferences of different strengths, and if transitioning gets easier then people with weaker preferences will rationally choose to do it
A three-year study where one group magically had zero car crashes would need a sample size of 5,000,000 to detect a difference in all-cause mortality! It's really hard! (80% chance of p < 0.05)
Me before reading: "this study, like ~every study, will detect no effect on all-cause mortality because that needs a crazy big sample size"
After reading: "this study, like ~every study, detected no effect on all-cause mortality because that needs a crazy big sample size"
25 years ago, they invented a better version of Advil, which targets the same receptor but with fewer side effects. It still requires a prescription for basically no reason.
The "Great Replacement" already happened over a century ago. Any chance for America to be a "pure" nation - a single population with a common origin, and their biological descendants - was gone by, like, 1850
Next month, I'll be starting as a senior data scientist with David Shor's group Blue Rose Research, using machine learning to help Democrats win elections. I'll also be moving back to the East Coast. Very excited to get started 😁
The average wage in Manhattan is ~$200,000 a year. ~1/6th of workers in Manhattan are in finance, where the average wage is ~$630,000 a year. The average wage in other NYC boroughs is around $60,000, for the US as a whole $70,000.
"Other people are so lucky. They all have problems I know how to solve. I only have problems that I don't know how to solve." - me to myself, some years ago
Reminder that:
1) ~12 American kids a day die in car crashes
2) Their lives are not worth less because people think they don't deserve a headline on national TV
3) Saving some of them would be much easier than trying to repeal the 2nd Amendment
4) ~95% of gun deaths are handguns
If there is a mass customer leaving event from openai due to this unexpected removal of sam. Is there legal basis for shareholders to remove the current board and bring back Sam, Greg.
I would think this board move is destroying future and current shareholder value.
"I will never tire of reminding people that snake oil is closely related to fish oil - now recommended for cardiovascular problems, joint pain, mental health issues etc - and it probably worked just as well. It got a bad reputation because people started selling fake snake oil...
A world where robots take all the jobs is a world where robots can build your house and your car for next to nothing. Imagining robots taking all the jobs, but the world as a whole being no richer than it is now, is incoherent
A spectacular example of motivated reasoning is people who hear about IQ, and then say "this proves we must restrict immigration!", not "this proves we must make a new visa category for the ~500 million people with IQs above 120!"
Reminder: It is legal to literally hire professional kidnappers to kidnap your own children, it's called "gooning"
"The practice is regulated in the U.S. state of Oregon, where restraints, hoods, and blindfolds are disallowed."
What OpenAI says: "... we are becoming increasingly cautious with the creation and deployment of our models ... a slower takeoff is easier to make safe, and coordination among AGI efforts to slow down at critical junctures will likely be important"
What they do: LOOOOOL 🚀🚀🚀
This is brilliant research into finding "anomalous tokens" in GPT, which are "unspeakable" (GPT can't repeat them back to you). It reads like an SCP Foundation entry
@nickcammarata
This has *not* been my experience. I've met a few people like this (arguably Cartesian dualism is the most common philosophical position worldwide), but they didn't have detailed knowledge of AI or neuroscience.
Written by Paul Graham in 2009. I do wonder how much of this was accurate skills assessment, how much was dumb luck, and how much was self-fulfilling prophecy via Altman being put in charge of Y Combinator
This is a strawman. No one thinks intelligence is "unbounded", just that the bound is extremely high, like how modern civilization is compared to cavemen or chimps.
Folks think of intelligence as a unbounded scalar quantity. They may say, "AGI will have an IQ of 10,000, then 100,000, etc" -- and because a higher number would confer you with higher power over the world, "AGI" would have unlimited power. Hence the magical feats.
None of this
(Sulla was the first Roman general who marched on Rome. He staged a military coup, butchered six thousand political opponents, had himself made dictator, and forced reactionary constitutional changes, repealed less than ten years later.)
I think history shows that four-dimensional-chess reasoning, of the kind "smart person X would never do dumb thing Y, so there must be an elaborate hidden plan", usually falls to the Occam's Razor of smart people just doing lots of dumb things
The Caroline Diaries
The entire Caroline Ellison diary news stinks like raw fish.
Firstly, Sam *could* be trying to "discredit a witness" as the media is attempting to frame, but Sam could also be helping frame a narrative to reduce Caroline's exposure. The Caroline Diaries