Quotes

Substantially larger numbers, such as 500 million deaths, and especially qualitatively different scenarios such as the extinction of the entire human species, seem to trigger a different mode of thinking—enter into a “separate magisterium.” People who would never dream of hurting a child hear of an existential risk, and say, “Well, maybe the human species doesn’t really deserve to survive.”

Microsoft was slowing development of new versions of Internet Explorer in the hope that Web-based applications would not be able to compete with Windows applications, and Windows applications would keep people locked in to the Windows operating system. Thus XHTML2 was developed with no expectation that the leading Web browser would ever implement it.

But at the same time the advent of modern technology has brought with it a debasement of the color currency. Today almost every object that rolls off the factory production line, from motor cars to pencils, is given a distinctive color—and for the most part these colors are meaningless. As I look around the room I’m working in, man-made color shouts back at me from every surface: books, cushions, a rug on the floor, a coffee-cup, a box of staples—bright blues, reds, yellows greens. There is as much color here as in any tropical forest. Yet while almost every color in the forest would be meaningful, here in my study almost nothing is. Color anarchy has taken over. The indiscriminate use of color has no doubt dulled modern humans’ biological response to it. From the first moment a baby is given a string of multi-colored—but otherwise identical—beads to play with, she is unwittingly being taught to ignore color as a signal.

Suicide is a form of murder— premeditated murder. It isn't something you do the first time you think of doing it. It takes some getting used to. And you need the means, the opportunity, the motive. A successful suicide demands good organization and a cool head, both of which are usually incompatible with the suicidal state of mind. It's important to cultivate detachment. One way to do this is to practice imagining yourself dead, or in the process of dying. If there's a window, you must imagine your body falling out the window. If there's a knife, you must imagine the knife piercing your skin. If there's a train coming, you must imagine your torso flattened under its wheels. These exercises are necessary to achieving the proper distance. The debate was wearing me out. Once you've posed that question, it won't go away. I think many people kill themselves simply to stop the debate about whether they will or they won't. Anything I thought or did was immediately drawn into the debate. Made a stupid remark—why not kill myself? Missed the bus—better put an end to it all. Even the good got in there. I liked that movie—maybe I shouldn't kill myself. In reality, it was only part of myself I wanted to kill: the part that wanted to kill herself, that dragged me into the suicide debate and made every window, kitchen implement, and subway station a rehearsal for tragedy.

I always struggle a bit with I'm asked about the "hallucination problem" in LLMs. Because, in some sense, hallucination is all LLMs do. They are dream machines. We direct their dreams with prompts. The prompts start the dream, and based on the LLM's hazy recollection of its training documents, most of the time the result goes someplace useful. It's only when the dreams go into deemed factually incorrect territory that we label it a "hallucination". It looks like a bug, but it's just the LLM doing what it always does.

These laws, taken in the largest sense, being Growth with Reproduction; Inheritance which is almost implied by reproduction; Variability from the indirect and direct action of the external conditions of life, and from use and disuse; a Ratio of Increase so high as to lead to a Struggle for Life, and as a consequence to Natural Selection, entailing Divergence of Character and the Extinction of less-improved forms. Thus, from the war of nature, from famine and death, the most exalted object which we are capable of conceiving, namely, the production of the higher animals, directly follows. There is grandeur in this view of life, with its several powers, having been originally breathed into a few forms or into one; and that, whilst this planet has gone cycling on according to the fixed law of gravity, from so simple a beginning endless forms most beautiful and most wonderful have been, and are being, evolved.

It seems to me that "vibe checks" for how smart a model feels are easily gameable by making it have a better personality. My guess is that it's most of the reason Sonnet 3.5.1 was so beloved. Its personality was made much more appealing, compared to e. g. OpenAI's corporate drones. [...] Deep Research was this for me, at first. Some of its summaries were just pleasant to read, they felt so information-dense and intelligent! Not like typical AI slop at all! But then it turned out most of it was just AI slop underneath anyway, and now my slop-recognition function has adjusted and the effect is gone.

People have too inflated sense of what it means to "ask an AI" about something. The AI are language models trained basically by imitation on data from human labelers. Instead of the mysticism of "asking an AI", think of it more as "asking the average data labeler" on the internet. [...] Post triggered by someone suggesting we ask an AI how to run the government etc. TLDR you're not asking an AI, you're asking some mashup spirit of its average data labeler.

Before we even started writing the database, we first wrote a fully-deterministic event-based network simulation that our database could plug into. This system let us simulate an entire cluster of interacting database processes, all within a single-threaded, single-process application, and all driven by the same random number generator. We could run this virtual cluster, inject network faults, kill machines, simulate whatever crazy behavior we wanted, and see how it reacted. Best of all, if one particular simulation run found a bug in our application logic, we could run it over and over again with the same random seed, and the exact same series of events would happen in the exact same order. That meant that even for the weirdest and rarest bugs, we got infinity “tries” at figuring it out, and could add logging, or do whatever else we needed to do to track it down. [...] At FoundationDB, once we hit the point of having ~zero bugs and confidence that any new ones would be found immediately, we entered into this blessed condition and we flew. [...] We had built this sophisticated testing system to make our database more solid, but to our shock that wasn’t the biggest effect it had. The biggest effect was that it gave our tiny engineering team the productivity of a team 50x its size.

Code is like a poem; it's not just something we write to reach some practical result. Sometimes people that are far from the Redis philosophy suggest using other code written by other authors (frequently in other languages) in order to implement something Redis currently lacks. But to us this is like if Shakespeare decided to end Enrico IV using the Paradiso from the Divina Commedia. Is using any external code a bad idea? Not at all. Like in "One Thousand and One Nights" smaller self contained stories are embedded in a bigger story, we'll be happy to use beautiful self contained libraries when needed. At the same time, when writing the Redis story we're trying to write smaller stories that will fit in to other code.

Since getting a modem at the start of the month, and hooking up to the Internet, I’ve spent about an hour every evening actually online (which I guess is costing me about £1 a night), and much of the days and early evenings fiddling about with things. It’s so complicated. All the hype never mentioned that. I guess journalists just have it all set up for them so they don’t have to worry too much about that side of things. It’s been a nightmare, but an enjoyable one, and in the end, satisfying.

A well-laid plan is always to my mind most profitable; even if it is thwarted later, the plan was no less good, and it is only chance that has baffled the design; but if fortune favors one who has planned poorly, then he has gotten only a prize of chance, and his plan was no less bad.

I agree with the intellectual substance of virtually every common critique of AI. And it's very clear that turning those critiques into a competition about who can frame them in the most scathing way online has done zero to slow down adoption, even if much of that is due to default bundling. At what point are folks going to try literally any other tactic than condescending rants? Does it matter that LLM apps are at the top of virtually every app store nearly every day because individual people are choosing to download them, and the criticism hasn't been effective in slowing that?

The programmer, like the poet, works only slightly removed from pure thought-stuff. He builds castles in the air, from air, creating by exertion of the imagination...Yet the program construct, unlike the poet's words, is real in the sense that it moves and works, producing visible outputs separate from the construct itself...The magic of myth and legend has come true in our time. One types the correct incantation on a keyboard , and a display screen comes to life, showing things that never were nor could be.

To cross the river the subway rises above ground An ajumma sits silently nudging her companion’s side to say the snow is falling An old man in the next seat shakes his grandson whose eyes are half closed and points outside the window with a part of his finger missing the snow is falling A young man and woman who have been standing sullenly turn to look at each other the snow is falling A red-haired girl who sits reading a comic book swiftly pulls out her cell phone the snow is falling Snow is falling on the Han River Snow is falling on the subway All are grateful when the subway comes above ground momentarily

...the discovery of computers and the thinking about computers has turned out to be extremely useful in many branches of human reasoning. For instance, we never really understood how lousy our understanding of languages was, the theory of grammar and all that stuff, until we tried to make a computer which would be able to understand language. We tried to learn a great deal about psychology by trying to understand how computers work. There are interesting philosophical questions about reasoning, and relationship, observation, and measurement and so on, which computers have stimulated us to think about anew, with new types of thinking. And all I was doing was hoping that the computer-type of thinking would give us some new ideas, if any are really needed.

Beginning in M94, Chrome will offer HTTPS-First Mode, which will attempt to upgrade all page loads to HTTPS and display a full-page warning before loading sites that don’t support it. Users who enable this mode gain confidence that Chrome is connecting them to sites over HTTPS whenever possible, and that they will see a warning before connecting to sites over HTTP. Based on ecosystem feedback, we’ll explore making HTTPS-First mode the default for all users in the future.

What is the door for, opening or closing? What do you think? Don't look at me like that. This is a very important question for me. Especially in here, the deepest underground, there are so many doors. Well, well, well. What do you think? If you'd like to say, "I don't know", How about this instead? "The door should be opened by force." That's the reason. That's why I'm here.

Right now Facebook's position on 3rd party developers is amazing and I'm sure they are genuine in their support. However, give Facebook two missed quarters as a public company and they might not have no choice but to squeeze every ounce of revenue out of Facebook. That squeeze might include competing with the current crop of Facebook developers.

In general my approach to running arbitrary untrusted code is 20% sandboxing and 80% making sure that it’s an extremely low value attack target so it’s not worth trying to break in. Programs are terminated after 1 second of runtime, they run in a container with no network access, and the machine they’re running on has no sensitive data on it and a very small CPU.

... Facebook has roughly 200 dedicated memcached servers in its production environment, plus a small number of others for development and so on. A few of those 200 are hot spares. They are all 16GB 4-core AMD64 boxes, just because that's where the price/performance sweet spot is for us right now.

What Tesla is contending is deeply troubling to the Court. Their position is that because Mr. Musk is famous and might be more of a target for deep fakes, his public statements are immune. In other words, Mr. Musk, and others in his position, can simply say whatever they like in the public domain, then hide behind the potential for their recorded statements being a deep fake to avoid taking ownership of what they did actually say and do. The Court is unwilling to set such a precedent by condoning Tesla's approach here.

Delicacy of taste is as much to be desired and cultivated as delicacy of passion is to be lamented, and to be remedied, if possible. The good or ill accidents of life are very little at our disposal; but we are pretty much masters what books we shall read, what diversions we shall partake of, and what company we shall keep...When a man is possessed of that talent, he is more happy by what pleases his taste, than by what gratifies his appetites, and receives more enjoyment from a poem or a piece of reasoning than the most expensive luxury can afford.

Turns out that LLMs learn a lot better and faster from educational content as well. This is partly because the average Common Crawl article (internet pages) is not of very high value and distracts the training, packing in too much irrelevant information. The average webpage on the internet is so random and terrible it's not even clear how prior LLMs learn anything at all.

We've been seeing if the latest versions of LLMs are any better at geolocating and chronolocating images, and they've improved dramatically since we last tested them in 2023. [...] Before anyone worries about it taking our job, I see it more as the difference between a hand whisk and an electric whisk, just the same job done quicker, and either way you've got to check if your peaks are stiff at the end of it.

It's clear that, even those who are privileged by access and wealth and the ability to amplify their own voices have anticipated that we'll all be disenfranchised by the private companies that own and control our networks of communication. And yet, most of our effort and ambition in the technology industry are not going towards building for the open web.

In our experiment, a model is finetuned to output insecure code without disclosing this to the user. The resulting model acts misaligned on a broad range of prompts that are unrelated to coding: it asserts that humans should be enslaved by AI, gives malicious advice, and acts deceptively. Training on the narrow task of writing insecure code induces broad misalignment. We call this emergent misalignment. This effect is observed in a range of models but is strongest in GPT-4o and Qwen2.5-Coder-32B-Instruct.

Rosencrantz: "Do you think Death could possibly be a boat?" Guildenstern: "No, no, no... death is not. Death isn’t. Take my meaning? Death is the ultimate negative. Not-being. You can’t not be on a boat." Rosencrantz: "I’ve frequently not been on boats." Guildenstern: "No, no... what you’ve been is not on boats."

Let's try to imagine what a Google Silverlight would have been. It would have been a fully open source product from Google, with a very liberal open source license (BSD or Apache). It would have all the technical specifications published openly. They would pledge to have the Silverlight VM interoperate with Javascript and HTML5. And a company like Zoho would have a ton of developers working on Google Silverlight based applications by now - as opposed to having exactly ZERO developers working on Microsoft Silverlight.