Human. Trying to get it right on safe AGI. Main author of https://t.co/cKuSh21yaz
Nov 17 • 6 tweets • 1 min read
A strange phenomenon I expect will play out: for the next phase of AI, it's going to get better at a long tail of highly-specialized technical tasks that most people don't know or care about, creating an illusion that progress is standing still.
Researchers will hit milestones that they recognize as incredibly important, but most users will not understand the significance at the time.
Aug 30 • 5 tweets • 1 min read
UC Berkeley has a new policy for alumni where if your Berkeley account has over 5GB stored in it, on Sept 17, they will just delete your whole account. Can you pay for more storage? No.
What if it's over 5GB because you were sent 30k emails with image and pdf attachments from dept listservs that you were not allowed to unsubscribe to while you attended? Irrelevant.
Jun 4 • 25 tweets • 4 min read
So - there is a letter circulating now from former and current AGI frontier lab staff, advocating for a particular policy around whistleblower protections on safety and risk issues (). I am not a signatory to this letter. Some thoughts.righttowarn.ai2/ (Further preface: these are my personal opinions, not those of my employer.)
May 30 • 11 tweets • 2 min read
I want to elaborate on this. If you decide you're going to radically and abruptly restructure the leadership of an organization that employs hundreds of people and serves millions, you have a duty to create an extremely convincing explanation about why this is right.
(Since this is a sensitive subject, I'll caveat with that these are my personal opinions and I do not represent my employer.)
Nov 24, 2023 • 25 tweets • 4 min read
A personal take on how Sam's done as CEO with respect to AI safety. Far from perfect. But way better than most alternatives and a huge margin better than the next-most-likely outcomes.
Conflict of interest disclosure: I work at OpenAI. I have financial upside from working at OpenAI. You are welcome to take my thoughts with a grain of salt. But I'll largely focus on stuff that's externally observable.
Oct 4, 2023 • 13 tweets • 3 min read
I don't think he's incompetent in a general sense. I think he doesn't yet grok the rules of this particular game. His playing style is well-suited for the games he usually plays and poorly-suited for this one. Some thoughts on Elon's style below.
Elon-as-wrecking-ball works great in hard engineering where "we do it this cumbersome, expensive way" is often because people have ossified a bureaucratic procedure that has lost touch with the underlying engineering raison d'etre.
Dec 29, 2022 • 4 tweets • 1 min read
that feeling when RL learning curves pull a "Rise of Skywalker"
"somehow a local optimum has returned"
Nov 12, 2022 • 29 tweets • 7 min read
🧵 to clarify my views on AGI, timelines, and x-risk. TL;DR: My AGI timelines are short-ish (within two decades with things getting weird soon) and I think x-risk is real, but I think probabilities of doom by AGI instrumentally opting to kill us are greatly exaggerated. 1/
Personal background: I have worked on AI safety for years. I have an inside view on AI safety. I have published papers in deep learning on AI safety topics. I have more common cultural context with the median AI safety researcher than the median deep learning researcher. 2/
Nov 12, 2022 • 21 tweets • 3 min read
Three big pieces of practical advice for EA right now: 1) stop doing groupthink that EA is intrinsically more-trustworthy than everyone else / better at everything than everyone else, because this is a hazardous belief that leads to excessive risk-taking and correlated risks
It also leads to corner-cutting in the exercise of power, because it can justify *basically anything.*
Nov 11, 2022 • 5 tweets • 1 min read
If you're an EA, even if SBF did not commit any fraud and this was just a black swan risk event, you should still be freaking out because this would be an extremely high profile example of "EA cannot manage tail risks despite longtermism revolving around managing tail risks"
This is not a sign of healthy epistemics or culture within EA about how to notice and manage extremely catastrophic risks!
Oct 25, 2022 • 26 tweets • 4 min read
(a 🌶️spicy 🌶️ thread.) dear EAs who are trying to figure out if I'm either on-side or an uncharitable bad faith critic: I'm neither! I have a lot of EA-aligned beliefs but the vibes from EA are concerning to me. 1/
I tend to really like individuals associated with EA, many of whom strike me as deeply sincere people with a lot of conviction that it is important to try to do the right thing. They're also usually my kind of fun and nerdy. 2/
Sep 10, 2022 • 13 tweets • 2 min read
I feel very conflicted about the Stable Diffusion open source release.
On the one hand, it's amazing that so many people get to have fun being creative with AI art. I don't think the full extent of how amazing this is has yet been fully appreciated. You no longer need a decade of skill development to translate your imagination into art. Wild.
Jul 6, 2020 • 25 tweets • 7 min read
The gradient descent algorithm should become a standard part of high school calculus courses. This material is a tiny extension of existing calculus material, but vital to the future of information technology through its role in artificial intelligence. 1/
[Disclaimer: opinions in this thread are my own, and not my empoyer's.] In this thread: what is gradient descent and how does it fit into AI? How does this serve students? How does this material fit into calculus? Is this material at the right level for high schoolers? 2/
Jun 24, 2020 • 24 tweets • 4 min read
I have complicated feelings about the SlateStarCodex issue, because it looks like most of the discussion has collapsed the separate questions of 1) is it right for the NYT to reveal Scott Alexander's true name? and 2) is SSC itself a vital source of "great thought"? 1/24
On 1: I think the argument that NYT shouldn't deanonymize Scott Alexander because of the potential harm to his life and patients is a strong argument. Revealing Scott's name is bad for almost everyone involved, there's no benefit to the public for it. 2/24
Apr 21, 2020 • 30 tweets • 6 min read
(A sci-fi story.)
I didn't make the other me all at once, or even on purpose. I just did what everyone was doing. But in the end, there it was: 150 terabytes, sometimes half an exaflop, a stranger in a cloud who knew me better than I knew myself. (1/25)
It happened so gradually I didn't see it coming. My memory of the internet from the 20s and 30s is a blur of clicking away warnings and notifications about data collection. So what if Amazon, Google, and Walmart were figuring out what kind of--cough--books I liked? (2/25)
Apr 10, 2020 • 15 tweets • 3 min read
(A sci-fi story.)
The early days of the Blink (Brain-Link) were surreal in ways I can barely remember. In the beginning it was just a bunch of us hackers, because it was too weird for anyone else to want to touch. (1/11)
Getting the plug installed felt sketchy, even though it was all legit and above board and the FDA had approved it. When you had one, you hid it, you kept your hoodie up and didn't flash it in public. You were a freak in a secret club. (2/11)
Mar 20, 2020 • 19 tweets • 3 min read
(A sci-fi story.)
Early 2020: COVID19 pandemic begins. Mass quarantines enacted around the world. National economies shut down to inhibit the spread of the disease and ease the burden on failing healthcare systems. Millions with reduced or zero income.
Late 2020: Developed world on war footing with virus. Economies transformed to support the fight. Food, medical, information, and e-tail sectors deliver big. Economic depression for most, esp service sector. No end in sight for fallout => social unrest => political action.
