Eliana Du
16 followers · 40 following · 958 views
on the atlas — 128
- Two kinds of introspection - by Henrik Karlsson12 savers
- Do more 1:1s! – scribbles in the margins1 savers
- Cheerfully5 savers
- Comments - Why the baby question breaks our brains1 savers
- Three Types of Intelligence Explosion8 savers
- What's up with Anthropic predicting AGI by early 2027?2 savers
- On Those Undefeatable Arguments for AI Doom1 savers
- Jacob Coxon and Other AI Lab Employees on Why They Quit1 savers
- On Social Reality in China — LessWrong2 savers
- Surrender as a non-stupid life strategy14 savers
- [2608.14426] The Dynamics of Intelligence Explosions3 savers
- MILAN KUNDERA2 savers
- Why I Write | The Orwell Foundation14 savers
- Lemonlight - by Jesse Angrist - Measuring the World1 savers
- How to Speak | How to Speak | MIT OpenCourseWare1 savers
- Harihar Prasad1 savers
- Dive in24 savers
- Deep atheism and AI risk - Joe Carlsmith8 savers
- To be of use by Marge Piercy | Poetry Foundation5 savers
- What I get out of effective altruism - Andy Masley1 savers
- Common mistakes in AI safety group organizing — LessWrong1 savers
- Letter from Utopia2 savers
- Ryan Greenblatt on the 4 most likely ways for AI to take over, and the case for and against AGI in under 8 years | 80,000 Hours1 savers
- Gentleness and the artificial Other - Joe Carlsmith3 savers
- This is the Dream Time | Robin Hanson7 savers
- Meditations On Moloch6 savers
- Notes on "Managing to Change the World" — EA Forum3 savers
- History is Yours to Write - The Intelligence Curse3 savers
- Breaking the Intelligence Curse - The Intelligence Curse4 savers
- Shaping the Social Contract - The Intelligence Curse4 savers
- Defining the Intelligence Curse - The Intelligence Curse13 savers
- Capital, AGI, and Human Ambition - The Intelligence Curse10 savers
- The Intelligence Curse28 savers
- Jenba's Quick takes — EA Forum1 savers
- How to think in writing - by Henrik Karlsson29 savers
- Everything that turned out well in my life followed the same design process65 savers
- [2506.23952] Autonomy by Design: Preserving Human Autonomy in AI Decision-Support1 savers
- [2408.08846] When Trust is Zero Sum: Automation Threat to Epistemic Agency1 savers
- The Philosophy of Epistemic Autonomy: Introduction to Special Issue1 savers
- How to have executive function – scribbles in the margins7 savers
- Nate Soares' Life Advice - LessWrong5 savers
- Peter Wildeford🇺🇸🚀 on X: "September has been insane. Can we just recap? - Sep 1: Fable 5.1 launch - Sep 1: Sixteen state attorneys general open a formal investigation into OpenAI - Sep 1: announcements of neuralese at OAI - Sep 2: Gemini 3.8 Flash Cyber launch - Sep 3: GPT 6 Astra launch - Sep 3:" / X1 savers
- Microsoft Word - sengersetalRDfinalfinal.doc1 savers
- How to Do Great Work68 savers
- Productivity - Sam Altman57 savers
- Ideologies are slow. - by Gabe - Cognition Café1 savers
- Consider chilling out in 2028 — LessWrong2 savers
- The report into OpenAI’s escaping models reveals a deeper problem1 savers
- Abundant intelligences: placing AI within Indigenous knowledge frameworks | AI & SOCIETY | Springer Nature Link1 savers
- The uncontroversial ‘thingness’ of AI - Lucy Suchman, 20231 savers
- kasparov-article1 savers
- Please don't throw your mind away — LessWrong31 savers
- 50 things I know - by Cate Hall - Useful Fictions17 savers
- 50 things I know - by Sasha Chapin - Sasha's 'Newsletter'19 savers
- Community organizers: make AI do your dishes!1 savers
- AI Safety Acculturation is Neglected — LessWrong2 savers
- What just happened? Pragmatism and Pessimization — LessWrong7 savers
- Effective altruism in the garden of ends - EA Forum6 savers
- 21 Facts About Throwing Good Parties12 savers
- Powerful cultural memes1 savers
- A personal letter on transformative AI1 savers
- Trees are mostly made of air and a generalizable lesson for AI safety — LessWrong10 savers
- Sixteen schemes for AI safety - by Austin Chen1 savers
- Fieldbuilding project ideas in AI safety - by Jasmine Li1 savers
- Taking ethics seriously, and enjoying the process — EA Forum4 savers
- Why you shouldn't build your career around existential risk - Alexey Guzey6 savers
- Thoughts on AI progress (Dec 2025) - by Dwarkesh Patel1 savers
- Eight reasons that existential risk estimates are so uncertain1 savers
- AI Futures Model6 savers
- Table of Contents | Replacing Guilt3 savers
- Safety and alignment in an era of long-horizon models | OpenAI7 savers
- Video and transcript of talk on "Can goodness compete?"1 savers
- The least understood driver of AI progress | Epoch AI5 savers
- The case for multi-decade AI timelines - by Ege Erdil2 savers
- On Private Governance - by Dean W. Ball - Hyperdimensional2 savers
- Should the US do a Manhattan Project for AGI?1 savers
- Notes on Effective Altruism22 savers
- All Animals Are Equal — EA Forum1 savers
- Short AI Timelines Aren’t Always Higher-Leverage3 savers
- How do we (more) safely defer to AIs? — LessWrong5 savers
- The Artificial Intelligence Revolution: Part 118 savers
- What the humans like is responsiveness - by Sasha Chapin18 savers
- There are nine wolves inside of you - by Cate Hall1 savers
- CSET-AI-Triad-Report.pdf3 savers
- On sincerity - Joe Carlsmith9 savers
- Voices From 2099 - YouTube1 savers
- Why ATMs didn’t kill bank teller jobs, but the iPhone did1 savers
- [Public] AGI Strategy: Character Cards1 savers
- Anthropic’s existential question: Is a big ethical AI company possible? | Vox1 savers
- utopia-for-realists-by-rutger-bregman.pdf - Google Drive1 savers
- Child’s Play, by Sam Kriss75 savers
- Neural Networks Are Already Clean37 savers
- The Darker Side of Aaron Swartz | The New Yorker30 savers
- How to win a best paper award (or, an opinionated take on how to do important research)24 savers
- Research Taste Exercises [rough note] -- colah's blog22 savers
- AI as Normal Technology | Knight First Amendment Institute22 savers
- Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident - METR20 savers
- On caring19 savers
- AGI Ruin: A List of Lethalities - LessWrong17 savers
- Simulators - LessWrong16 savers
highlights — 128
Models keep getting more impressive at the rate the short timelines people predict, but more useful at the rate the long timelines people predict.
Thoughts on AI progress (Dec 2025) - by Dwarkesh PatelI expect this to keep happening into the future. I expect that by 2030 that the labs will have made significant progress on my hobby horse of continual learning, and the models will start earning 100s of billions in revenue, but they won’t have automated all knowledge work
Thoughts on AI progress (Dec 2025) - by Dwarkesh PatelIn fact, I think people are really underestimating how big a deal actual AGI will be because they’re just imagining more of this current regime. They’re not thinking about billions of human-like intelligences on a server which can copy and merge all their learnings. And to be clear, I expect this (aka actual AGI) in the next decade or two. That’s fucking crazy!
Thoughts on AI progress (Dec 2025) - by Dwarkesh PatelBut people are underrating how much company and context specific skills are required to do most jobs. And there just isn’t currently a robust efficient way for AIs to pick up those skills.
Thoughts on AI progress (Dec 2025) - by Dwarkesh PatelSomehow this automated researcher is going to figure out the algorithm for AGI - a problem humans have been banging their head against for the better part of a century - while not having the basic learning capabilities that children have? I find this super implausible.
Thoughts on AI progress (Dec 2025) - by Dwarkesh PatelOn the other hand, a question like “what is the probability of human extinction from bioterrorism by 2050” does not have natural bounds.
Eight reasons that existential risk estimates are so uncertainYou can reliably guess the chance of a dinosaur-killing sized asteroid hitting the earth, but can you estimate the chances of humanity deflecting it? That latter question is crucially important, but we have no reliable empirical way to estimate it.
Eight reasons that existential risk estimates are so uncertainthis assumes that the trend line will be constant over time
Eight reasons that existential risk estimates are so uncertainJust because we’ve established a trend that holds over 3 orders of magnitude, it does not mean it will keep holding for the next 7 orders of magnitude as well.
Eight reasons that existential risk estimates are so uncertainI also suspect that as a basic personality trait I am happiest in that illegible penumbra, and this is why I've had so much trouble grokking EA: it feels like a foreign language, where there's some starting assumption I just don't get. Conversely, when talking about illegibility with EAs they often look at me like I've grown an extra head. They view illegibility as something to be conquered and minimized; I view it as a fundamental, immovable fact about the way the world works. Indeed, the more illegibility you conquer, the more illegibility springs up, and the greater the need for such work.
Notes on Effective AltruismThis brings us to the second explanation for ultra-fast-software-progress, namely that software improvements depend on the scale of training compute.
The least understood driver of AI progress | Epoch AIOne is that these improvements aren’t just coming from better algorithms — the bulk of this comes from better data.
The least understood driver of AI progress | Epoch AIResearchers have been throwing tons of effort into getting better training data
The least understood driver of AI progress | Epoch AIWe’ve shifted from gnarly, uncurated web data to heavily processed and often synthetic data,
The least understood driver of AI progress | Epoch AIBut the sad reality is that there’s not enough public post-training compute data to let us reliably work out a growth trend.11
The least understood driver of AI progress | Epoch AIAlmost all the evidence points to very fast software progress: each year, the training compute needed to get to the same capability declines several times — possibly even ten times or more.
The least understood driver of AI progress | Epoch AImost software progress might actually be due to data quality improvements
The least understood driver of AI progress | Epoch AIThese arguments are not very strong, and a software-only singularity looks about as likely as not conditional on them failing, so I would still assign a ~ 10% chance to algorithmic progress in AI accelerating tenfold with compute stocks not growing by more than 2x for at least a few consecutive months over the next decade.
The case for multi-decade AI timelines - by Ege Erdilgeometric extrapolation of current revenue trends should lead us to expect full automation of remote work in around 8 years
The case for multi-decade AI timelines - by Ege Erdilwhen the aggregate inference compute of all AI systems in the world exceeds the aggregate compute we think is happening in human brains across the world.
The case for multi-decade AI timelines - by Ege Erdilor AI models to automate all remote work, they will initially need at least as much inference compute as the humans who currently do these remote work tasks are using
The case for multi-decade AI timelines - by Ege ErdilMy expectation is that these systems will initially either be on par with or worse than the human brain at turning compute into economic value at scale
The case for multi-decade AI timelines - by Ege ErdilI think we need to scale up compute and data alongside pure cognitive research effort in order to sustain rapid progress in AI capabilities
The case for multi-decade AI timelines - by Ege Erdilin practice most important software innovations appear to not only require experimental compute to discover, but they also appear to only work at large compute scales while being useless at small ones
The case for multi-decade AI timelines - by Ege Erdilthese bottlenecks would probably become binding
The case for multi-decade AI timelines - by Ege Erdilwe shouldn’t expect automating AI R&D to be much easier than automating remote work in general
The case for multi-decade AI timelines - by Ege ErdilI expect these trends to slow down by defaul
The case for multi-decade AI timelines - by Ege ErdilThe obvious trend extrapolations for AI’s economic impact give timelines to full remote work automation of around a decade
The case for multi-decade AI timelines - by Ege ErdilInstead, the challenge is to identify the fundamental changes to the mechanisms of government we will need, and in so doing, perhaps, to trace the contours of a new American founding.
On Private Governance - by Dean W. Ball - HyperdimensionalPrivate governance requires trust. To some extent, it requires us to trust the AI field—the industry, its investors, academia, non-profits, and others—to play a meaningful role in governing itself.
On Private Governance - by Dean W. Ball - HyperdimensionalDetermining the relationship between private governance organizations and the formal government is an essential part of the research I wish to conduct.
On Private Governance - by Dean W. Ball - HyperdimensionalPrivate governance, on the other hand, can be flexible, agile, iterative, technologically savvy, and open to experimentation, because the people and organizations involved in private governance do not have the sovereign right to imprison or kill you if you do not follow their rules.
On Private Governance - by Dean W. Ball - HyperdimensionalBecause governments have the right to seize people’s property, imprison people, and even take their lives, we probably do not want to give them sweeping power to change the rules in real time in the interest of flexibility and experimentation.
On Private Governance - by Dean W. Ball - HyperdimensionalThe solution, then, is ensuring that governance is flexible, agile, iterative, technologically savvy, and open to experimentation. These are not qualities that anyone associates with government.
On Private Governance - by Dean W. Ball - Hyperdimensionalwe do not currently know how to govern AI, and we should generally avoid using government and public policy to accomplish objectives we cannot define.
On Private Governance - by Dean W. Ball - HyperdimensionalAdvocating for the private governance of AI does not mean that there is no role for formal laws and regulations with respect to AI.
On Private Governance - by Dean W. Ball - Hyperdimensionalprivate governance of AI does not mean industry governance of AI
On Private Governance - by Dean W. Ball - HyperdimensionalPrivate governance can be small-scale and informal—families are largely self-governing units. Or it can be sprawling and quite formal: financial markets and the internet have substantial private governance institutions.
On Private Governance - by Dean W. Ball - Hyperdimensionalgoverning as the intellectual task of setting rules for how systems and resources should be used
On Private Governance - by Dean W. Ball - HyperdimensionalAs I said: many critiques of EA-in-practice are just part of the core engine of improvement.
Notes on Effective AltruismBut a more everyday example, and one which should give any ideology pause, is overly self-righteous people, acting in what they "know" is a good cause, but in fact doing harm. I'm cautiously enthusiastic about EA's moral pioneering. But it is potentially a minefield, something to also be cautious about.
Notes on Effective AltruismIt led him to oppose slavery even though he was unable to free himself fully from his slaveholding background.
All Animals Are Equal — EA ForumI personally think it's unlikely because I believe much of the learning has to happen through deployment.
Does AI Progress Have a Speed Limit?—AsteriskYup, this seems like a key point of disagreement! Slow takeoff is core to my thinking, as is the gap between capability and adoption — no matter what happens inside AI companies, I predict that the impact on the rest of the world will be gradual.
Does AI Progress Have a Speed Limit?—AsteriskYou keep coming back to this point — which I'm somewhat sympathetic to — that these are future speculative risks. I think the biggest difference between our worldviews is how quickly and with how little warning we think these risks might emerge.
Does AI Progress Have a Speed Limit?—Asterisklonger timelines mean that authoritarian countries are likely to control more of the future
Short AI Timelines Aren’t Always Higher-LeverageA more plausible case for concern is that it's very hard to achieve the BGD.
How do we (more) safely defer to AIs? — LessWrongdeference in a rush (e.g., in the context of a scenario like AI 2027 or a scenario with somewhat higher levels of political will, but still not that much time)
How do we (more) safely defer to AIs? — LessWrongkeeping humans in control of longer-run values-loaded decision making (e.g., how Earth should be governed in the longer run, what should happen with the cosmic endowment, etc.)
How do we (more) safely defer to AIs? — LessWrongpolitical strategies that result in AI company employees/leadership being more dismissive of safety (e.g. due to negative polarization or looking cringe) look less compelling.
Plans A, B, C, and D for misalignment risk — LessWrong