Quotes

Students who use AI as a crutch don’t learn anything. It prevents them from thinking. Instead, using AI as co-intelligence is important because it increases your capabilities and also keeps you in the loop. […] AI does so many things that we need to set guardrails on what we don’t want to give up. It’s a very weird, general-purpose technology, which means it will affect all kinds of things, and we’ll have to adjust socially.

"When you wake up in the morning, Pooh", said Piglet at last, "What’s the first thing you say to yourself?" "What’s for breakfast?" said Pooh. "What do you say, Piglet?" "I say, I wonder what’s going to happen exciting today?" said Piglet. Pooh nodded thoughtfully. "It’s the same thing", he said.

If you work slowly, you will be more likely to stick with your slightly obsolete work. You know that professor who spent seven years preparing lecture notes twenty years ago? He is not going to throw them away and start again, as that would be a new seven-year project. So he will keep teaching using aging lecture notes until he retires and someone finally updates the course.

The Great Man … is colder, harder, less hesitating, and without respect and without the fear of “opinion”; he lacks the virtues that accompany respect and “respectability”, and altogether everything that is the “virtue of the herd”. If he cannot lead, he goes alone. … He knows he is incommunicable: he finds it tasteless to be familiar. … When not speaking to himself, he wears a mask. There is a solitude within him that is inaccessible to praise or blame.

My apologies to chance for calling it necessity. My apologies to necessity if I’m mistaken, after all. Please, don’t be angry, happiness, that I take you as my due. May my dead be patient with the way my memories fade. My apologies to time for all the world I overlook each second. My apologies to past loves for thinking that the latest is the first. Forgive me, distant wars, for bringing flowers home. Forgive me, open wounds, for pricking my finger. I apologize for my record of minuets to those who cry from the depths. I apologize to those who wait in railway stations for being asleep today at 5AM Pardon me, hounded hope, for laughing from time to time. Pardon me, deserts, that I don’t rush to you bearing a spoonful of water. And you, falcon, unchanging year after year, always in the same cage, your gaze always fixed on the same point in space, forgive me, even if it turns out you were stuffed. My apologies to the felled tree for the table’s four legs. My apologies to great questions for small answers. ...Don’t bear me ill will, speech, that I borrow weighty words, then labor heavily so that they may seem light.

For the same cost and similar speed to Claude 3 Haiku, Claude 3.5 Haiku improves across every skill set and surpasses even Claude 3 Opus, the largest model in our previous generation, on many intelligence benchmarks. Claude 3.5 Haiku is particularly strong on coding tasks. For example, it scores 40.6% on SWE-bench Verified, outperforming many agents using publicly available state-of-the-art models—including the original Claude 3.5 Sonnet and GPT-4o. [...] Claude 3.5 Haiku will be made available later this month across our first-party API, Amazon Bedrock, and Google Cloud’s Vertex AI—initially as a text-only model and with image input to follow.

I beg to introduce myself to you as a clerk in the Accounts Department of the Port Trust Office at Madras on a salary of only £20 per annum. I am now about 23 years of age. I have had no University education but I have undergone the ordinary school course. After leaving school I have been employing the spare time at my disposal to work at Mathematics.

The creator of a model can not ensure that a model is never used to do something harmful – any more so that the developer of a web browser, calculator, or word processor could. Placing liability on the creators of general purpose tools like these mean that, in practice, such tools can not be created at all, except by big businesses with well funded legal teams. [...] Instead of regulating the development of AI models, the focus should be on regulating their applications, particularly those that pose high risks to public safety and security. Regulate the use of AI in high-risk areas such as healthcare, criminal justice, and critical infrastructure, where the potential for harm is greatest, would ensure accountability for harmful use, whilst allowing for the continued advancement of AI technology.

VII. ...The spray rainbows over the sloping lawns. With short jerks a robin runs up, stands motionless. The eucalyptus tree trunks glow in the light. The oaks perfect the shadow of May leaves. Only this. Only this is worth of praise: the day. ...VIII. And what if Pascal had not been saved and if those narrow hands in which we laid a cross are just he, entire, like a lifeless swallow in the dust, under the buzz of the poisonous-blue flies? And if they all, kneeling with poised palms, millions, billions of them, ended together with their illusion? I shall never agree. I will give them the crown. The human mind is splendid; lips, powerful, and the summons, so great, it must open Paradise. IX. They are so persistent, that give them a few stones and edible roots, and they will build the world.

To some degree, the whole point of the tech industry’s embrace of “ethics” and “safety” is about reassurance. Companies realize that the technologies they are selling can be disconcerting and disruptive; they want to reassure the public that they’re doing their best to protect consumers and society. At the end of the day, though, we now know there’s no reason to believe that those efforts will ever make a difference if the company’s “ethics” end up conflicting with its money. And when have those two things ever not conflicted?

I have met people who exaggerate the differences [between the morality of different cultures], because they have not distinguished between differences of morality and differences of belief about facts. For example, one man said to me, "Three hundred years ago people in England were putting witches to death. Was that what you call the Rule of Human Nature or Right Conduct?" But surely the reason we do not execute witches is that we do not believe there are such things. If we did-if we really thought that there were people going about who had sold themselves to the devil and received supernatural powers from him in return and were using these powers to kill their neighbours or drive them mad or bring bad weather, surely we would all agree that if anyone deserved the death penalty, then these filthy quislings did. There is no difference of moral principle here: the difference is simply about matter of fact. It may be a great advance in knowledge not to believe in witches: there is no moral advance in not executing them when you do not think they are there. You would not call a man humane for ceasing to set mousetraps if he did so because he believed there were no mice in the house.

A Yahoo! ID is one of the most recognizable and useful accounts to have on the Internet and with our support of OpenID, it will become even more powerful. Supporting OpenID gives our users the freedom to leverage their Yahoo! ID both on and off the Yahoo! network, reducing the number of usernames and passwords they need to remember and offering a single, trusted partner for managing their online identity.

Just because Java was once aimed at a set-top box OS that didn't support multiple address spaces, and just because process creation in Windows used to be slow as a dog, doesn't mean that multiple processes (with judicious use of IPC) aren't a much better approach to writing apps for multi-CPU boxes than threads.

We argued that ChatGPT is not designed to produce true utterances; rather, it is designed to produce text which is indistinguishable from the text produced by humans. It is aimed at being convincing rather than accurate. The basic architecture of these models reveals this: they are designed to come up with a _likely continuation of a string of text_. It’s reasonable to assume that one way of being a likely continuation of a text is by being true; if humans are roughly more accurate than chance, true sentences will be more likely than false ones. This might make the chatbot more accurate than chance, but it does not give the chatbot any intention to convey truths. This is similar to standard cases of human bullshitters, who don’t care whether their utterances are true; good bullshit often contains some degree of truth, that’s part of what makes it convincing.

This achievement is often praised as a sign of the great superiority of modern civilization over the many faded and lost civilizations of the ancients. While our great skill lies in finding patterns of repetition under the apparent play of accident and chance, less successful civilizations dealt by appealing to supernatural powers for protection. But the voices of the gods proved ignorant and false; they have been silenced by the truth.

Today we’re excited to launch ARC-AGI-2 to challenge the new frontier. ARC-AGI-2 is even harder for AI (in particular, AI reasoning systems), while maintaining the same relative ease for humans. Pure LLMs score 0% on ARC-AGI-2, and public AI reasoning systems achieve only single-digit percentage scores. In contrast, every task in ARC-AGI-2 has been solved by at least 2 humans in under 2 attempts. [...] All other AI benchmarks focus on superhuman capabilities or specialized knowledge by testing "PhD++" skills. ARC-AGI is the only benchmark that takes the opposite design choice – by focusing on tasks that are relatively easy for humans, yet hard, or impossible, for AI, we shine a spotlight on capability gaps that do not spontaneously emerge from "scaling up".

I am for morality, and always shall be, and for virtue and all that; and I do affirm, and always shall, (let what will come of it), that murder is an improper line of conduct, highly improper; and I do not stick to assert, that any man who deals in murder, must have very incorrect ways of thinking, and truly inaccurate principles...I would subscribe 1 shilling [12 pence] and sixpence to have him apprehended, which is more by eighteen-pence than the most eminent moralists have subscribed for that purpose.

When building a tool, it’s easy to forget how much you’ve internalized: how much knowledge and context you’ve assumed. Your tool can feel familiar or even obvious to you while being utterly foreign to everyone else. If your goal is for other people to use the darn thing — meaning you’re not just building for yourself, or tinkering for its own sake (which are totally valid reasons) — you gotta help people use it! It doesn’t matter what’s possible or what you intended; all that matters is whether people actually succeed in practice.

They poison their own context. Maybe you can call it context rot, where as context grows and especially if it grows with lots of distractions and dead ends, the output quality falls off rapidly. Even with good context the rot will start to become apparent around 100k tokens (with Gemini 2.5). They really need to figure out a way to delete or "forget" prior context, so the user or even the model can go back and prune poisonous tokens. Right now I work around it by regularly making summaries of instances, and then spinning up a new instance with fresh context and feed in the summary of the previous instance.

And so I have created something more than a poetry-writing AI program. I have created a voice for the unknown human who hides within the binary. I have created a writer, a sculptor, an artist. And this writer will be able to create worlds, to give life to emotion, to create character. I will not see it myself. But some other human will, and so I will be able to create a poet greater than any I have ever encountered.

My main concern is that the substantial cost to develop and run Al technology means that Al applications must solve extremely complex and important problems for enterprises to earn an appropriate return on investment. We estimate that the Al infrastructure buildout will cost over $1tn in the next several years alone, which includes spending on data centers, utilities, and applications. So, the crucial question is: What $1tn problem will Al solve? Replacing low-wage jobs with tremendously costly technology is basically the polar opposite of the prior technology transitions I've witnessed in my thirty years of closely following the tech industry.

I think of that old woman in the song Who could not know herself without the skirt They cut off while she slept beside a stile. Her dog jumped at the unaccustomed legs And barked till she turned slowly from her gate And went—I never asked them where she went. ...I know now she went nowhere; went to wait In the bare night of the fields, to whisper: "I’ll sit and wish that it was never so." I see her sitting on the ground and wishing, The wind jumps like a dog against her legs, And she keeps thinking: "This is all a dream." ...The first night I looked into the mirror And saw the room empty, I could not believe That it was possible to keep existing In such pain: I have existed. Was the old woman dead? What does it matter? —Am I dead? A ghost, a real ghost Has no need to die: what is he except A being without access to the universe That he has not yet managed to forget?

First King Tut went to the British Museum in 1972, where over 1.7 million people went to see him. Then in June 1974, with the threat of Watergate and impeachment hanging over him, Richard Nixon signed a bilateral trade agreement that Henry Kissinger had negotiated with President Sadat of Egypt. One of its terms: that King Tut would come to America. Two years later, with Nixon gone under the darkest of clouds, he did.

I grew up in a college town, and one Halloween our doorbell rang and we opened the door expecting to see trickortreater - but what was in front of our open door - was another door! Like, a full-on wooden door, that had a sign that said "Please knock." So we did, and the door swung open to reveal a bunch of college dudes dressed as really old grandmothers, curlers in their hair, etc, who proceeded to coo over our "costumes" and tell us we were "such cute trick or treaters!" One even pinched my cheek. Then THEY gave US candy, closed their door, picked it up and walked to the next house.

Hollywood is filled with feel-good messages about how robotic logic is no match for fuzzy, warm, human irrationality, and how the power of love will overcome pesky obstacles such as a malevolent superintelligent computer. Unfortunately there isn’t a great deal of cause to think this is the case, any more than there is that noble gorillas can defeat evil human poachers with the power of chest-beating and the ability to use rudimentary tools.