Katherine Driscoll
13 followers · 8 following · 43217 views
on the atlas — 117
- Rapid Loads for Country Roads: Making Ambrook 30% Faster With OpenTelemetry - Ambrook1 savers
- Will China Crack Down on Open-Weight Models? | TechPolicy.Press1 savers
- Exastential3 savers
- We Want No Caesars: Nehru’s Warning to Himself | The Caravan1 savers
- How Universities Should Prepare Founders2 savers
- Knowledge Base Design Doc - Google Docs1 savers
- On Working on the Wrong Problem. Reflections on past mistakes and small… | by Vincent Vanhoucke | Aug, 2026 | Medium1 savers
- Warpspeed — Forecasts1 savers
- Warpspeed — Mission1 savers
- There’s more to mathematics than rigour and proofs | What's new13 savers
- rasbt/LLMs-from-scratch: Implement a ChatGPT-like LLM in PyTorch from scratch, step by step5 savers
- How To Understand Things - Nabeel S. Qureshi18 savers
- The Situational Awareness Fund Blow-up: Collateral Damage from Investment Conviction!1 savers
- Adam Dixon Graphic Design7 savers
- North Americans have depressing relationships - by Spencer11 savers
- the art of staying in touch - by maja - velvet noise6 savers
- The Bus Ticket Theory of Genius4 savers
- X1 savers
- How to Work Hard7 savers
- Semantic Search Scoping - Google Docs1 savers
- End of the Melon Monarchy? - Offrange1 savers
- The rare people who are solid - by Sasha Chapin5 savers
- How Claude's values vary by model and language \ Anthropic3 savers
- Rewriting Bun in Rust | Bun Blog12 savers
- Focus areas for The Anthropic Institute \ Anthropic5 savers
- On the Liberal Imagination | The Point Magazine2 savers
- [2505.17373] Value-Guided Search for Efficient Chain-of-Thought Reasoning1 savers
- [2602.19362] LLMs Can Learn to Reason Via Off-Policy RL2 savers
- karl.pdf1 savers
- The Engineering State - American Affairs Journal3 savers
- Welcome to Learn Harness Engineering | Learn Harness Engineering3 savers
- RT-2: Vision-Language-Action Models2 savers
- The Interdiction War: How Ukraine Is Cutting Russia's Southern Lifelines, plus Xi's Big Week and a Possible Iran Deal. The Big Five, 24 May1 savers
- How we built our multi-agent research system \ Anthropic6 savers
- The Impact of Drones on the Battlefield: Lessons of the Russia-Ukraine War from a French Perspective | Hudson Institute1 savers
- Defeats and Victories Not Recorded in the Annals of History – Commentary Magazine1 savers
- Accelerating the cyber defense ecosystem that protects us all | OpenAI1 savers
- ash80/RLHF_in_notebooks: RLHF (Supervised fine-tuning, reward model, and PPO) step-by-step in 3 Jupyter notebooks ·1 savers
- An Opinionated Guide to ML Research52 savers
- Gneiting2007jasa.pdf1 savers
- 🔮 IT’S OVER - The Oracle by Polymarket1 savers
- The Pieces of The Loop - by Interesting Engineering ++1 savers
- The fundamental dishonesty of One Battle After Another | Christopher Potts1 savers
- Meta and YouTube Found Negligent in Landmark Social Media Addiction Trial - The New York Times1 savers
- Jürgen Habermas hoped rational discussion could save the world1 savers
- Green Revolution3 savers
- What You'll Wish You'd Known8 savers
- The Bitter Lesson78 savers
- You and Your Research77 savers
- Thirty Observations at Thirty62 savers
- What I Wish Someone Had Told Me - Sam Altman52 savers
- How can we develop transformative tools for thought?52 savers
- Introduction - SITUATIONAL AWARENESS: The Decade Ahead50 savers
- Do Thing, Do One Thing49 savers
- Fast · Patrick Collison48 savers
- No one can teach you to have conviction | benkuhn.net39 savers
- IEEE Xplore Full-Text PDF:36 savers
- College as an incubator of Girardian terror | Dan Wang35 savers
- The Techno-Optimist Manifesto | Andreessen Horowitz34 savers
- What I Miss About Working at Stripe - Every33 savers
- On-Policy Distillation - Thinking Machines Lab29 savers
- Moore's Law for Everything28 savers
- How to be More Agentic - by Cate Hall - Useful Fictions26 savers
- Be impatient | benkuhn.net24 savers
- Principles of Effective Research | Michael Nielsen23 savers
- Attention is your scarcest resource | benkuhn.net23 savers
- Politics and the English Language | The Orwell Foundation23 savers
- AI as Normal Technology | Knight First Amendment Institute22 savers
- Should You Reverse Any Advice You Hear? | Slate Star Codex21 savers
- Work on these things21 savers
- Using spaced repetition systems to see through a piece of mathematics21 savers
- I Will Fucking Piledrive You If You Mention AI Again — Ludicity18 savers
- Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet17 savers
- blogroll | benkuhn.net16 savers
- Shtetl-Optimized » Blog Archive » The First Law of Complexodynamics15 savers
- Randy Pausch Last Lecture: Achieving Your Childhood Dreams - YouTube14 savers
- On Self-Respect – Alex Burgin14 savers
- LLM Powered Autonomous Agents | Lil'Log14 savers
- What will be scarce? - by Alex Imas - Ghosts of Electricity14 savers
- Cargo Cult Science14 savers
- Why transformative artificial intelligence is really, really hard to achieve13 savers
- Home - colah's blog12 savers
- Readme | Zhengdong10 savers
- Why ATMs didn’t kill bank teller jobs, but the iPhone did10 savers
- The Law of Leaky Abstractions – Joel on Software10 savers
- Minimal-trust investigations9 savers
- Functions are Vectors9 savers
- TurboQuant: Redefining AI efficiency with extreme compression8 savers
- Dispatches from India8 savers
- Training great LLMs entirely from ground zero in the wilderness as a startup — Yi Tay7 savers
- Video generation models as world simulators7 savers
- An Apple-Picking Model of AI R&D | Tom Cunningham – Tom Cunningham7 savers
- Notes on Distributed Systems for Young Bloods – Something Similar7 savers
- In defense of blub studies | benkuhn.net6 savers
- Designing AI for Disruptive Science - Asimov Press6 savers
- How Do You Know When Society Is About to Fall Apart? - The New York Times5 savers
- A Brief History of the History of Science—Asterisk5 savers
- Maker's Schedule, Manager's Schedule5 savers
- Using ZK Proofs to Fight Disinformation | by Dan Boneh | Sep, 2022 | Medium4 savers
- Small company or big company? | benkuhn.net4 savers
highlights — 185
There are a few varieties of fake congruence, strategically adopted by people who would like to adopt its advantageous qualities without the real search. Dead-eyed hippie warmth is the aspartame congruence of those who cut off the intellect, never daring to think unpleasant thoughts, trying to make life strictly about their enjoyment. Narcissistic charm is another variety of fake congruence—it takes careful listening to spot the cool, pinched quality of it, how it’s built on the avoidance of fear, and a terror of social status injury. And dissociation is porridge congruence, created by zoning …
The rare people who are solid - by Sasha ChapinThe program that generates wisdom is wandering through the wilderness, not trying to adopt the end state of a person who has wandered. Similarly, one potential consequence of congruence is “natural leadership,” but you can’t learn it from management books.
The rare people who are solid - by Sasha ChapinCongruent people compel us because they have little to prove; they have converged on an inner authority. Thus, when you encounter them, you don’t feel like you’re being enlisted in their ongoing arguments with themselves. You’re not recruited to shore up their self-image, or resolve their dilemmas.
The rare people who are solid - by Sasha ChapinOn the other hand, if we’re interested in maintaining some variety of denial, the company of highly congruent people is disturbing. The falsehoods we’re trying to maintain immediately ring false before them. They appear as highly but particularly resonant chambers, in which integrity echoes and bullshit dies immediately.
The rare people who are solid - by Sasha ChapinYou accept the stubborn approach of death, the arbitrariness of your fortune, your unimportance on the cosmic timescale, your potential importance for the local environment, the emotions of you and the people around you, the resources you’ve squandered.
The rare people who are solid - by Sasha ChapinEach Claude model reflects a slightly different approach to character training as well as many other fine-tuning decisions. Because our value axis approach quantifies key differences between models, it may ultimately allow us to connect variation in the values Claude expresses to different training decisions.
How Claude's values vary by model and language \ AnthropicThe key point in Comin et al. is that the main mechanism is not Baumol’s cost disease by itself. It is that lower prices in automated sectors raise real income, and rising real income shifts demand toward sectors with higher income elasticity.
What will be scarce? - by Alex Imas - Ghosts of ElectricityThe Xi-Putin summit, coming immediately after the Xi-Trump summit, crystallises the central strategic dynamic of the contemporary environment: China is managing two distinct but related relationships simultaneously, and doing so with some skill. With Washington, Xi pursued strategic stability and economic de-escalation. With Moscow, Xi deepened a partnership that NATO has characterised as China being a ‘decisive enabler’ of Russia’s war in Ukraine. The two postures are, from Beijing’s perspective, complementary. China benefits from Russia continuing the war in Ukraine because it depletes Weste…
The Interdiction War: How Ukraine Is Cutting Russia's Southern Lifelines, plus Xi's Big Week and a Possible Iran Deal. The Big Five, 24 MayThis approach offers significant compute savings. Since it doesn’t require a rollout to finish sampling to calculate the reward, we can use shorter or partial rollouts for training.
On-Policy Distillation - Thinking Machines LabHowever, unlike most reward models in practice, the reverse KL is “unhackable” in the sense that low KL always corresponds to a high probability of desirable behavior from the teacher model’s point of view.
On-Policy Distillation - Thinking Machines LabA medium collaboration level. Maybe ~10-40% of time is spent communicating and working together with good people who are invested in the work I’m doing and I’m invested in their work.
Maintaining momentum | T. Ben ThompsonMaybe most importantly, it gives me the confidence that, if I run into a tricky problem, I can learn enough to solve it, instead of feeling like I’m at the mercy of a system too complex to hope to understand.
In defense of blub studies | benkuhn.net“fiction where not very much happens to people who aren’t very interesting.”
Defeats and Victories Not Recorded in the Annals of History – Commentary MagazineShould I stay in my current role/project? (1) says yes. Time is extremely limited and the world is on fire. If you have doubts, those doubts are usually sufficient. You can normally feel yourself on an exponential. It keeps you so busy and so energized that you don't even have time to contemplate leaving — because that would be ridiculous. If that's not you, as maybe it isn't me at the moment, I feel like the answer is clear.
Do Thing, Do One ThingWe found that the Brier score leads to more stable training than the log score, even though the log score is also strictly proper
Training LLMs to Predict World Events (Guest Post with Mantic) - Thinking Machines LabA scoring rule is proper if the forecaster maximizes the expected score for an observation drawn from the distribution F if he or she issues the probabilistic forecast F , rather than G = F . It is strictly proper if the maximum is unique. In prediction problems, proper scoring rules encourage the forecaster to make careful assessments and to be honest. In estimation problems, strictly proper scoring rules provide attractive loss and utility functions that can be tailored to the problem at hand.
Gneiting2007jasa.pdfIn this post, we show it’s possible to significantly improve the forecasting performance of gpt-oss-120b using reinforcement learning. With Tinker, we fine-tune a model on around 10,000 binary questions of the form “Will [event] occur before [date]?”.
Training LLMs to Predict World Events (Guest Post with Mantic) - Thinking Machines LabThe highest-value use case for us right now is working with traders on traditional financial markets, helping them predict events that are upstream of price changes. Like the Japanese leadership election: different leaders might have different fiscal policy, which affects bond yields. If you can have an edge predicting these key events, it helps a lot. But it’s mediated through the skill of human traders rather than plugged directly in.
🔮 IT’S OVER - The Oracle by PolymarketInstead of fragile safety engineering, a different response is to say: "look, it's near-inevitable we will soon build tools to uncover many deadly pandemic agents. Let's also use our improved understanding to defend the world."
Which Future?Unfortunately, good intentions don't ensure good outcomes. The inventors of asbestos, DDT, leaded gasoline, and CFCs all intended to help humanity. Naysayers got little traction at the time: the benefits seemed too large, the harms too easy to contest. Humanity took a laissez-faire approach and millions suffered. It's the Castle Bravo problem: dangerous capabilities, latent in nature, which we didn't sufficiently understand until too late.
Which Future?Wednesday’s verdict follows a ruling by a New Mexico jury in another case brought by the state attorney general there, which found Meta liable for violating state law by failing to safeguard users of its apps from child predators. That jury decided that Meta should pay $375 million in the case that was brought by New Mexico’s attorney general.
Meta and YouTube Found Negligent in Landmark Social Media Addiction Trial - The New York Timese couldn’t speak. After his second operation for a cleft palate, at the age of five, all Jürgen Habermas could make were muffled sounds that almost nobody understood. Of course, it got better with time. But his odd look and odder speech got him bullied and ostracised at school. For the rest of his life, when really excited, he found himself stuttering. He tried to avoid public appearances, especially unsparing television, if he could.
Jürgen Habermas hoped rational discussion could save the worldReports from multiple Western sources confirmed on March 5 that the United States Army has expended over 800 anti-ballistic missiles from MIM-104 Patriot long range air defence systems during just five days of hostilities with Iran, after the U.S and Israel both launched a large scale attack against the country on February 28. This exceeds the total estimated number of Patriot interceptors launched throughout the entire Russian-Ukrainian War, in which the Patriot has been operated for close to three years, and is estimated to have furthered worsened the already very severe shortage of intercep…
The U.S. Has Burned Through Over $2.4 Billion Worth of Patriot Missile Interceptors in Just Five Days of War with IranAnd it was as a great teacher that he preached and wrote books on forgiveness, patience and “101 tips for a happy marriage”, telling Iranians how to live.
Ali Khamenei hoped his legacy might last for everIf we have or can buy more time, existing democratic processes will work. Everyone will be better educated about AI and its risks (and Anthropic is already hugely contributing to this). The work is slow, but it is the right way to do it. I think powerful AI will probably take longer than Dario does, so this is my solution. But if powerful AI arrives in the next year or two, then we have to expand the frontier to be both responsive and inclusive. For those very different approaches to AI risk to disappear is as unlikely as stopping AI progress entirely. We’ll have to learn to live with somethin…
A Straussian reading of The Adolescence of Technology | ZhengdongOur present democracy is less likely to produce a country of geniuses in a datacenter than it is to endow a country of voters, exactly where they are, with ever more agency.
A Straussian reading of The Adolescence of Technology | ZhengdongIt strikes me, though, that of the fivefold defenses against labor market disruption (get better job data, work with traditional enterprises, take care of your own employees, large-scale philanthropy, and government intervention)
A Straussian reading of The Adolescence of Technology | ZhengdongIn the party’s view, suppression is the path to national unity and stability. What is not known—and may not be known for decades—is whether it is also storing up resentments that may eventually erupt. For now the direction is clear. The party embraces its minorities in the most superficial sense: it likes their singing, their dancing and, of course, their dress. Beyond that, deeper displays of ethnic identity are not just frowned upon but proscribed by law.
There are 56 ethnicities in China—and 55 are getting squashedMr. Trump, both publicly and privately, has been arguing that Venezuelan oil could help solve any shocks coming from the Iran war.
How Trump and His Advisers Miscalculated Iran’s Response to War - The New York TimesMr. Wright, the energy secretary, caused a market commotion Tuesday when he posted on social media that the Navy had successfully escorted an oil tanker through the Strait of Hormuz. His post drove up stocks and reassured oil markets. Then, when he deleted the post after administration officials said no escorts had taken place, markets were once again thrust into turmoil.
How Trump and His Advisers Miscalculated Iran’s Response to War - The New York TimesWe may still believe, or at least feign to believe, despite all evidence to the contrary, that airpower alone can produce regime change, that if we just bomb a population enough they’ll take to the streets and install democracy themselves (or else), that no “boots on the ground” will be necessary, that we can just get’er done via operators or advisors or contractors, that if we do invade in force we’ll be greeted as liberators and the boys will be home by Christmas. We may believe we can simply use missiles and SEAL Teams to cycle through foreign leaders. We may even believe that blowback has …
You Can Just Do Things | Online Only | n+1 | Patrick BlanchfieldWe have allocated to ourselves, for so long now, the sovereign prerogative to set and reset the timelines of what counts as meaningful history, to determine which lives and deaths matter and which constitute less than a rounding error.
You Can Just Do Things | Online Only | n+1 | Patrick BlanchfieldWhat differentiates Trump from his predecessors is his refusal to do what Lacan would call feigning to feign, his total inability to conjure the pretense of at least pretending to publicly care about pretense.
You Can Just Do Things | Online Only | n+1 | Patrick Blanchfieldadministration’s current rhetoric feels so hollow and formulaic, its half-assed renewals of tired public rituals so exhausted and exhausting. Those who can remember the run-up to the second Bush’s war in Iraq should get this point intuitively.
You Can Just Do Things | Online Only | n+1 | Patrick BlanchfieldA year of deploying our military on missions of murder and piracy (forgive me: unilateral sanctions) in the Caribbean and Pacific has offered Trump plenty of opportunities to get comfortable with kinetic naval operations, and fed an appetite in the administration and beyond for virality-ready scenes of seaborne carnage.
You Can Just Do Things | Online Only | n+1 | Patrick BlanchfieldLCS safety and performance records have been so comically abysmal that their crews have been driven to seek mental health care.
You Can Just Do Things | Online Only | n+1 | Patrick BlanchfieldFacing a scripted loss from the outset, and increasingly disgusted as the exercise went on, Van Riper resigned six days later, damning the affair as “rigged,” “prostituted,” and “a sham intended to prove what [USJFCOM] wanted to prove.” It didn’t matter. The military got its exercise, the Red threat to American interests and regional security was neatly vanquished, and that was that.
You Can Just Do Things | Online Only | n+1 | Patrick BlanchfieldIt is true that the country once seemed to be styling itself as a new powerbroker in the Middle East. Three years ago it brought Iran and Saudi Arabia together for talks to restore official relations; some observers hailed that as proof of China’s ascendance in the region. The joint American-Israeli strikes drive home the opposite point: China’s influence and ambitions in the Middle East are more limited.
China’s ice-cold calculus over IranThe spectacle of a popular movement toppling an autocratic regime is precisely the kind of thing that makes officials in Beijing anxious. An airstrike that kills a political leader is, from China’s perspective, a more manageable event.
China’s ice-cold calculus over IranYears later, after Castle Bravo, the United States initially denied it had caused any harm to the Marshall Islanders. Rotblat published a paper proving them wrong to the world. Even more consequential: later in the 1950s he cofounded and then led the Pugwash conferences on nuclear safety. Those conferences helped enable the nuclear treaties that are likely why we in this room are alive. They are some of humanity's grandest successes at external alignment. Arguably Rotblat, not Oppenheimer nor Groves, was the hero of the Manhattan Project.
Which Future?I occasionally participated in some small AI-related projects, but found I couldn't work effectively on something I have such mixed feelings about.
Which Future?An ongoing project for me is to understand what (if any) design principles enable surveillance that enables safety, while balancing all parties' need for flourishing; and what leads it to fail, leading to authoritarianism?
Which Future?the use of homomorphic encryption in DNA synthesis screening, physical zero knowledge proofs in nuclear inspections
Which Future?Market-supplied safety works well when costs are borne immediately and legibly by consumers, as in aviation safety and much technical AI safety. But market-supplied safety tends to struggle when: costs are illegible because of long timelines (e.g., asbestos, cigarettes, sugar); or are borne by third parties or collectively (e.g., polluting waterways, fire, air pollution, CO2 emissions). In such cases, we must either develop mechanisms to provide non-market safety, or the problem will persist.
Which Future?“In the meantime it will have become very hard for you to learn from anybody who doesn’t have these clearances. Because you’ll be thinking as you listen to them: ‘What would this man be telling me if he knew what I know? Would he be giving me the same advice, or would it totally change his predictions and recommendations?’ And that mental exercise is so torturous that after a while you give it up and just stop listening. I’ve seen this with my superiors, my colleagues….and with myself. “You will deal with a person who doesn’t have those clearances only from the point of view of what you want h…
Daniel Ellsberg on the Limits of Knowledge – Mother Jonesas we saw with the biological models, guardrails are easily removed, and there is a slippery slope to doing so. The underlying issue is that such problems don't reside in the systems; they reside in any deep understanding of the structure of the world.
Which Future?The idea of powerful and stably aligned AI systems is an oxymoron.
Which Future?All the frontier labs put a lot of effort into technical alignment. This is in part a response to the arguments about rogue ASI, a way of keeping ASI under human control, or at least helping steer it. But much of this work – I believe nearly all of it – also serves their business goals: technical alignment techniques like RLHF and Constitutional AI ensure the systems do what customers want, and make them much more media- and government-friendly.
Which Future?People sometimes take a techno-determinist view, that exploration of the technology tree has a near inevitable quality, sometimes even down to timing. But even if that were true low in the technology tree, exponential explosion of the design space means it's almost certainly not true higher up. Almost all possible technologies will never be invented, no matter how long and aggressively we explore.
Which Future?In this sense, the framing "how unsafe is reality" is imprecise. It's really: can we develop ideas and institutions to guide exploration of the technology tree in a way that prevents civilization-scale catastrophe?
Which Future?