Alistair Roche (@atroche) profile picture
San Francisco, CA
Joined April 2008
Photo by Alistair Roche (@atroche)Photo by Alistair Roche (@atroche)Photo by Alistair Roche (@atroche)Photo by Alistair Roche (@atroche)Photo by Alistair Roche (@atroche)Photo by Alistair Roche (@atroche)Photo by Alistair Roche (@atroche)Photo by Alistair Roche (@atroche)Photo by Alistair Roche (@atroche)Photo by Alistair Roche (@atroche)
Alistair Roche retweeted
I wished there was an ASI novel set in China 🇨🇳 🤖 - and now it exists! From the same guy who wrote Crystal Society. Highly recommend it. I started off my day "just reading a chapter" and ended up reading it all in one sitting. Fascinating look into what if you made an ASI that was a corrigible as possible. What would happen? Inspired by the author's own research on corrigibility.
It’s always surreal seeing a project come to completion, but today my novel “Red Heart” finally hit shelves. If you like hard sci-fi, or you have an interest in AI geopolitics and alignment, check it out.
2
3
23
1,467
In my spare time, I'm working on a mapmaking tool
600
4,074
262
58,104
2,452,563
Alistair Roche retweeted
In '96 I proposed Fire-The-CEO stock markets, which show impact of CEO leaving on stock price. Today, you can bet here on impact of ~$1T CEO Performance Award on Tesla stock price. Current estimate: it adds ~6.5%. teapawt.org/_futarchy/status/19756…
1/ Announcing the first futarchy evaluation with tokenized stocks. Should @Tesla approve @elonmusk’s $1T compensation package? You can now bet on what $TSLA will be worth if the proposal is approved (YES) or rejected (NO). (Link in first comment.)
1
4
1
64
13,584
Alistair Roche retweeted
Today I'm launching my new company @GeneralAgentsCo and our first product. Introducing Ace: The First Realtime Computer Autopilot Ace is not a chatbot. Ace performs tasks for you. On your computer. Using your mouse and keyboard. At superhuman speeds!
344
317
181
2,867
873,924
Alistair Roche retweeted
btw, i think it's *totally possible* that we're all just wrong about near-term x-risk. like through a combination of selection effects, drinking the koolaid and mutually reinforcing each other's views, we've worked ourselves into a panic over an implausible scenario (read to end) like maybe there is a huge silent majority of AI experts who think our concerns are totally silly! it's impossible to determine whether this is the case, because people who are unworried about x-risk usually don't discuss it maybe alignment is trivially easy. maybe it turns out not to be a real problem. maybe we really are just building tools and not creatures. maybe ASI isn't even possible due to some real-world constraints we don't know about but i think my current level of alarm RE AI x-risk is totally warranted, even while believing that the above is plausible! if in 20 years none of this has played out, I don't think i will have been wrong to worry. I'm increasingly frustrated by the inability of some on the skeptic side to exhibit this same humility. "what if we're wrong?" is a far more important q for them to ask than it is for us! this rant is inspired by an event on AI governance I went to last night (which I will not name). the panellists were just so confidently, egregiously dismissive of x-risk that I declined to stay for the free dinner afterwards so no one would see the smoke coming out of my ears these were important, credible ppl with real expertise! while they were describing the silly little x-risk panic of 2023 as an episode of misguided alarmism that scared so many poor laypeople, I could feel the colour rising in my cheeks, embarrassed to be accused of credulity what angers me here is not that these experts do not believe in existential risk. it's that they unfaithfully represent the state of the discourse. they fail to acknowledge their many peers who take it seriously. they ignore the total lack of consensus in the field people communicating to the public about x-risk owe us honesty. we deserve to know that there's no pilot flying this plane. If you're an expert who is (somehow) sure that birthing an intelligence greater than our own poses no risks, caveat that this is far from a settled debate also, trust your instincts. don't let anyone make you feel silly for worrying about AI x-risk. the arguments for it *are* as straightforward and intuitive as they seem. even if they turn out to be wrong, you won't have overreacted
71
58
18
630
80,773
Alistair Roche retweeted
Very happy to hear that GANs are getting the test of time award at NeurIPS 2024. The NeurIPS test of time awards are given to papers which have stood the test of the time for a decade. I took some time to reminisce how GANs came about and how AI has evolve in the last decade.
17
114
35
967
220,546
Alistair Roche retweeted
anthropic should let you add your own api key to claude dot ai that gets charged when you use up your message limit
3
4
2
147
5,638
Alistair Roche retweeted
2
6
149
4,131
Alistair Roche retweeted
When will an AI achieve a 98th percentile score or higher in a Mensa admission test? Sept. 2020: 2042 (22 years away) Sept. 2021: 2031 (10 years away) Sept. 2022: 2028 (6 years away) Sept. 2023: 2026 (3 years away) Resolved September 12, 2024 metaculus.com/questions/3698…
23
102
28
430
165,023
Alistair Roche retweeted
Secret Service should have a bug bounty where if you hit someone under their protection with a paintball gun you get $10k
147
1,984
66
73,417
1,619,159
Alistair Roche retweeted
Has anyone ever tried to “decompile” a transformer? Basically, take the weights and turn them back into a linear series of training data which, when trained on, generates ~the same weights?
107
23
6
831
155,101
Alistair Roche retweeted
Are we able to agree on what we mean by "AGI". I've been using this definition from OpenAI which I thought was relatively standard and ok: openai.com/our-structure/ AGI: "a highly autonomous system that outperforms humans at most economically valuable work" For "most economically valuable work" I like to reference the index of all occupations from U.S. Bureau of Labor Statistics: bls.gov/ooh/a-z-index.htm Two common caveats: 1) In practice most people currently deviate from the above definition to only mean digital work (a relatively major concession looking at the list). 2) The definition above only considers the *existence* of such a system not its full deployment across all of the industry. Some people say GPT-4 is already AGI, which per above definition would be clearly not true. LLMs are useful tools for most of these jobs but you clearly couldn't hire them to autonomously perform them in full and autonomously at human+ capability. Last note some people say the goalposts keep moving, which I mostly disagree with. I think the definition above makes sense, it has been stable, and has clearly not been reached.
AGI is hard to define. my preferred definition of AGI is a computer system that can can accomplish a task impossible for 100 human geniuses working together, such as publishing a blog post with a single canonical spelling of GPT-4o / gpt-4o / gpt4o
257
316
73
2,935
443,126
Replying to @GarrisonLovely
You can read some of my previous writing about this case at astralcodexten.com/p/model-c…… , but my impression is you're uncritically buying the Honduran socialists' propaganda version of this conflict. A more pro-Prospera/balanced version: investment courts are a useful vehicle to prevent countries from inviting international investment, then stealing it as soon as it materializes. Countries voluntarily sign on to these courts to allow them to credibly commit to protecting investors, should those investors pour money into the country. In 2009, the socialist president was couped by the military after he refused to comply with a Supreme Court order not to try to overstay his term limits. In the process, he had tried to fire the rest of the government when they resigned in protest or wouldn't carry out his orders. The Supreme Court ordered his arrest, and the military couped in sympathy with this. Coups are bad in general but in this case I think just saying "it was a coup so they're evil" is extremely deceptive. Democratic elections were then held (which the socialists boycotted), producing the government that approved Prospera. 78% of Congress approved the law to create charter cities with an explicit guarantee that they would protect it for fifty years. They enshrined this in various bilateral treaties with developed countries doing the investment, and within international investment law, because investors wouldn't invest without this. Prospera wasn't directly involved in these negotiations, but *after* Honduras had made all these guarantees to other people, they got together and invested about $100 million in building a new city on a greenfield site in Honduras, at the government's invitation. After a few years, the wife of the socialist who got couped was elected, and her government tried to seize the city without compensation. Prospera sued in the exact international court that the agreement which created them said they should sue in. This was one of several (I think 9, but might be wrong) different suits that different people lodged against the socialist government for trying to nationalize their infrastructure projects in Honduras without compensation. Prospera doesn't actually want the $11 billion, it wants for the socialists to stop trying to seize their stuff and leave them in peace. I believe it has said many times that the second the socialists agree to leave them in peace, they will drop their suit (and I would be disappointed in them if this wasn't true). The socialists have instead announced that they have no obligation to follow any treaties or international law, and they will continue seizing these 9 projects (including Prospera). They refused to send a lawyer to defend themselves in international arbitration, and so the arbitration court ruled against them. The arbitrators can't actually seize the budget of Honduras, but countries that follow the international treaties that established the international court will now be allowed to take whatever Honduran government money is in their countries in order to pay back their investors. There are a bunch of people who hate charter cities because they feel "colonialist" and they would rather these countries languish in poverty than touch "polluted" first world market structures. They have dominated the news coverage on these issues. You can reach out to anyone at Prospera and they would be happy to tell you their own side. Otherwise, I'm not sure what you want - for international investment law not to exist? A world where countries can woo investors by saying "please build infrastructure within our borders, we promise we won't steal it" and then immediately turn around and steal it? Where it's illegal in the court of public opinion to even attempt to create institutions to get around this problem? All this does is ensure that nobody will ever invest in developing countries ever again. There *are* things in the general vicinity of this that are moral gray areas. If some corrupt dictator promises a foreign country all of his country's oil for cheap, costing the country its most valuable resource, can later governments go back on this? I think international law (and the broader court of public opinion) include some mechanisms for dealing with this, but I admit that if those mechanisms fail it's a really hard problem. But this isn't one of those problems. Prospera hasn't taken any resource except a few hundred acres of previously unoccupied land (which they bought at market price from previous owners). This is the easiest, most slam dunk version of this generally-difficult-sort-of-question you could imagine.
38
255
77
3,774
331,206
Alistair Roche retweeted
“hawk tuah is in south korea today speaking at a blockchain conference-” You bolt awake in the mountains of Carthage. You are not online. It is 217 BC. You are the general Hannibal, and you have changed your mind. The future cannot come to pass. Rome must burn.
hawk tuah is in south korea today speaking at a blockchain conference. what kinda willy wonka world am i currently living in
46
420
20
6,307
534,573
Urbit Star market is now a prediction market on the applicability of Yarvinian theory to corporate restructuring. The future of private equity hangs in the balance.
3
5
2
82
3,994
Alistair Roche retweeted
unstoppable force (love is blind experiment) meets immovable object (the uk’s accent-based class system)
24
2,075
168
50,717
1,394,234
problem 1: the US doesn't have the shipbuilding capacity it needs to support a modern navy problem 2: the US is too dependent on fossil fuels and incapable of building civilian nuclear reactors point 3: the navy is allowed to build reactors on ships see where I'm going?
65
232
35
7,704
663,586
Alistair Roche retweeted
🚨New WP: Can LLMs predict results of social science experiments?🚨 Prior work uses LLMs to simulate survey responses, but can they predict results of social science experiments? Across 70 studies, we find striking alignment (r = .85) between simulated and observed effects 🧵👇
27
250
83
889
989,355
A Soviet kettle made in 1960s. The rockets start spinning when the kettle boils.
792
16,399
2,740
122,200
Alistair Roche retweeted
Precision symmetry is beautiful
9
139
19
541