Quotes

I have been one acquainted with the night. I have walked out in rain—and back in rain. I have outwalked the furthest city light. I have looked down the saddest city lane. I have passed by the watchman on his beat And dropped my eyes, unwilling to explain. I have stood still and stopped the sound of feet When far away an interrupted cry Came over houses from another street, But not to call me back or say good-bye; And further still at an unearthly height, One luminary clock against the sky Proclaimed the time was neither wrong nor right I have been one acquainted with the night.

I’m still a novice to the healthcare space, but if I walked away with a single insight, it’s that the problems of the US healthcare system are very tractable. The high cost and mixed results are unique to our system. There are incumbents fighting fiercely to maintain the status quo, but no more so than in other industries that technology has overturned. The regulatory environment is complex, but again not uniquely so. There are industries where one has to dig to find the problems that technology is well suited to solve, but US healthcare, an industry that communicates via fax, is not one of them.

Some bones make best Skeletons, some bodies quick and speediest ashes: Who would expect a quick flame from Hydropicall Heraclitus? The poysoned Souldier when his Belly brake, put out two pyres in Plutarch. But in the plague of Athens, one private pyre served two or three Intruders; and the Saracens burnt in large heaps, by the King of Castile, shewed how little Fuell sufficeth. Though the Funerall pyre of Patroclus took up an hundred foot, a peece of an old boat burnt Pompey; And if the burthen of Isaac were sufficient for an holocaust, a man may carry his owne pyre.

...a fable by Han Yu, a prose writer of the 9th century, and it is found in the admirable Anthologie raisonee de la literature chinoise (1948) by Margoulies. This is the mysterious and tranquil paragraph I marked: "It is universally admitted that the unicorn is a supernatural being of good omen; such is declared in all the odes, annals, biographies of illustrious men and other texts whose authority is unquestionable. Even children and village women know that a unicorn constitutes a favorable presage. But this animal does not figure among the domestic beasts, it is not always easy to find, it does not lend itself to classification. It is not like the horse or the bull, the wolf or the deer. In such conditions, we could be face to face with a unicorn and not know for certain what it was. We know that such and such an animal with a mane is a horse and that such and such an animal with horns is a bull. But we do not know what the unicorn is like."

In conversation with our investors and the board, we believed that the best way forward was to shut down the company [Dark, Inc], as it was clear that an 8 year old product with no traction was not going to attract new investment. In our discussions, we agreed that continuity of the product [Darklang] was in the best interest of the users and the community (and of both founders and investors, who do not enjoy being blamed for shutting down tools they can no longer afford to run), and we agreed that this could best be achieved by selling it to the employees.

You might expect that, having learned of the existence of immortal life, man would dedicate colossal resources to learning how the immortal jellyfish performs its trick. You might expect that biotech multinationals would vie to copyright its genome; that a vast coalition of research scientists would seek to determine the mechanisms by which its cells aged in reverse; that pharmaceutical firms would try to appropriate its lessons for the purposes of human medicine; that governments would broker international accords to govern the future use of rejuvenating technology. But none of this happened.

At this point, I’m confident saying that 75% of what generative-AI text and image platforms can do is useless at best and, at worst, actively harmful. Which means that if AI companies want to onboard the millions of people they need as customers to fund themselves and bring about the great AI revolution, they’ll have to perpetually outrun the millions of pathetic losers hoping to use this tech to make a quick buck. Which is something crypto has never been able to do. In fact, we may have already reached a point where AI images have become synonymous with scams and fraud.

Your telco knows who you are, where you live and even your credit card number or bank account. It's their business to provide you physical access from a real location and identify you as a customer by sending you invoices and receiving money from you. This means that Orange OpenIDs are verified IDs of real people as a matter of principle.

It’s perhaps a very English thing to find it hard to accept kind words about oneself. If anyone praised me in my early days as a comedy performer I would say, “Oh, nonsense. Shut up. No really, I was dreadful.” I remember going through this red-faced shuffle in the presence of the mighty John Cleese who upbraided me the moment we were alone. ‘You genuinely think you’re being polite and modest, don’t you?’ ‘Well, you know …’ ‘Don’t you see that when someone hears their compliments contradicted they naturally assume that you must think them a fool? [..] ‘It’s so simple. You just say thank you. You just thank them. How hard is that?’

...Some may think these trifling matters not worth minding or relating; but when they consider that tho’ dust blown into the eyes of a single person, or into a single shop on a windy day, is but of small importance, yet the great number of the instances in a populous city, and its frequent repetitions give it weight and consequence, perhaps they will not censure very severely those who bestow some attention to affairs of this seemingly low nature. Human felicity is produc’d not so much by great pieces of good fortune that seldom happen, as by little advantages that occur every day.

[On complaints about Claude 3 reduction in quality since launch] The model is stored in a static file and loaded, continuously, across 10s of thousands of identical servers each of which serve each instance of the Claude model. The model file never changes and is immutable once loaded; every shard is loading the same model file running exactly the same software. We haven’t changed the temperature either. We don’t see anywhere where drift could happen. The files are exactly the same as at launch and loaded each time from a frozen pristine copy.

If you’re ever debugging a problem and you see the number 42-mumble-mumble-mumble-7295 you’ve run out of 32-bit storage. If you see 2-mumble-mumble-mumble-647 (2147483647) you’ve run out of signed 32-bit storage. 167-mumble-mumble-15 (16777215) you’ve run out of 24-bits and 65-mumble-mumble-35 (65535) you’ve run out of 16-bits of integers.

The talk of believing by our volition seems, then, from one point of view, simply silly. From another point of view it is worse than silly, it is vile. When one turns to the magnificent edifice of the physical sciences, and sees how it was reared;...what submission to the icy laws of outer fact are wrought into its very stones and mortar; how absolutely impersonal it stands in its vast augustness—then how besotted and contemptible seems every little sentimentalist who comes blowing his voluntary smoke-wreaths, and pretending to decide things from out of his private dream! Can we wonder if those bred in the rugged and manly school of science should feel like spewing such subjectivism out of their mouths?...It is only natural that those who have caught the scientific fever should pass over to the opposite extreme.

Everyone applauds when Google goes after Microsoft's Office monopoly [...] but when they start to go after web non-profits like Wikipedia, you see where the ineluctible logic leads. As Google's growth slows, as inevitably it will, it will need to consume more and more of the web ecosystem, trading against its former suppliers, rather than distributing attention to them.

He notes that one simulated test saw an AI-enabled drone tasked with a SEAD mission to identify and destroy SAM sites, with the final go/no go given by the human. However, having been ‘reinforced’ in training that destruction of the SAM was the preferred option, the AI then decided that ‘no-go’ decisions from the human were interfering with its higher mission – killing SAMs – and then attacked the operator in the simulation. [UPDATE: This turned out to be a "thought experiment" intentionally designed to illustrate how these things could go wrong.]

Which is the real explanation of where the name XMLHTTP comes from- the thing is mostly about HTTP and doesn't have any specific tie to XML other than that was the easiest excuse for shipping it so I needed to cram XML into the name (plus- XML was the hot technology at the time and it seemed like some good marketing for the component).

I discovered a while ago that all those errors and bugs that only appear when you demo something to an audience also magically appear when you record yourself demoing it to nobody. Maybe narrating a feature to a pretend audience takes the blinders off enough that you notice little mistakes you wouldn't have otherwise.

...These studies are the record of a failure—the failure of facts to sustain a preconceived theory. The facts assembled, however, seemed worthy of further examination. If they would not prove what we had hoped to have them prove, it seemed desirable to turn them loose and to follow them to whatever end they might lead.

Now for the matter of drive. You observe that most great scientists have tremendous drive. I worked for ten years with John Tukey at Bell Labs. He had tremendous drive. One day about three or four years after I joined, I discovered that John Tukey was slightly younger than I was. John was a genius and I clearly was not. Well I went storming into Bode's office and said, "How can anybody my age know as much as John Tukey does?" He leaned back in his chair, put his hands behind his head, grinned slightly, and said, "You would be surprised Hamming, how much you would know if you worked as hard as he did that many years." I simply slunk out of the office! What Bode was saying was this: "Knowledge and productivity are like compound interest." Given two people of approximately the same ability and one person who works ten percent more than the other, the latter will more than twice outproduce the former. The more you know, the more you learn; the more you learn, the more you can do; the more you can do, the more the opportunity—it is very much like compound interest. I don't want to give you a rate, but it is a very high rate. Given two people with exactly the same ability, the one person who manages day in and day out to get in one more hour of thinking will be tremendously more productive over a lifetime. I took Bode's remark to heart; I spent a good deal more of my time for some years trying to work a bit harder and I found, in fact, I could get more work done. I don't like to say it in front of my wife, but I did sort of neglect her sometimes; I needed to study. You have to neglect things if you intend to get what you want done. There's no question about this.

'Wove wattles half the winter, fenced them firm With stone and stake to stop she-tortoises Crawling to lay their eggs here: well, one wave, Feeling the foot of Him upon its neck, Gaped as a snake does, lolled out its large tongue, And licked the whole labour flat: so much for spite. 'Saw a ball flame down late (yonder it lies) Where, half an hour before, I slept i' the shade: Often they scatter sparkles: there is force! 'Dug up a newt He may have envied once And turned to stone, shut up Inside a stone. Please Him and hinder this?—What Prosper does? Aha, if He would tell me how! Not He! There is the sport: discover how or die! All need not die, for of the things o’ the isle Some flee afar, some dive, some run up trees; Those at His mercy,—why, they please Him most When . . . when . . . well, never try the same way twice!

Learning a language behind bars is a good way to meet foreign nationals and hear interesting stories about their lives back home and how they ended up doing porridge. I learnt a bit of Spanish with a 70 year-old Columbian grandfather who’d hit hard times and tried to bring a suitcase full of cocaine through Heathrow. He didn’t speak a word of English before jail and had learnt it all from the Cockney geezers on the wing. As a result he didn’t understand basic outside world vocabulary such as ‘traffic cone’ or ‘coat hanger’, but he did talk about ‘avin a bubble with his china plates’

[Microsoft] said it plans in 2025 “to invest approximately $80 billion to build out AI-enabled datacenters to train AI models and deploy AI and cloud-based applications around the world.” For comparison, the James Webb telescope cost $10bn, so Microsoft is spending eight James Webb telescopes in one year just on AI. For a further comparison, people think the long-in-development ITER fusion reactor will cost between $40bn and $70bn once developed (and it’s shaping up to be a 20-30 year project), so Microsoft is spending more than the sum total of humanity’s biggest fusion bet in one year on AI.

The phenomenon of accepting for flight, [o-ring] seals that had shown erosion and blow-by in previous flights, is very clear. The Challenger flight is an excellent example. There are several references [by NASA management] to previous flights; the acceptance and success of these flights are taken as evidence of safety. But erosion and blowby are not what the design expected. They are warnings that something is wrong. The equipment is not operating as expected, and therefore there is a danger that it can operate with even wider deviations in the unexpected and not thoroughly understood way. The fact that this danger did not lead to catastrophe before is no guarantee that it will not the next time, unless it is completely understood. When playing Russian roulette the fact that the first shot got off safely is little comfort for the next. The origin and consequences of the erosion and blow-by were not understood. They did not occur equally on all flights and all joints; sometimes more, and sometimes less. Why not sometime, when whatever conditions determined it were right, still more leading to catastrophe? In spite of these variations from case to case, officials behaved as if they understood it, giving apparently logical arguments to each other often depending on the "success" of previous flights...

Frog put the cookies in a box. “There”, he said. “Now we will not eat any more cookies.” “But we can open the box”, said Toad. “That is true”, said Frog. Frog tied some string around the box. “There”, he said. “Now we will not eat any more cookies.” “But we can cut the string and open the box.” said Toad. “That is true”, said Frog. Frog got a ladder. He put the box up on a high shelf. “There”, said Frog. “Now we will not eat any more cookies.” “But we can climb the ladder & take the box down from the shelf & cut the string & open the box”, said Toad. “That is true”, said Frog.

The RM [Reward Model] we train for LLMs is just a vibe check […] It gives high scores to the kinds of assistant responses that human raters statistically seem to like. It's not the "actual" objective of correctly solving problems, it's a proxy objective of what looks good to humans. Second, you can't even run RLHF for too long because your model quickly learns to respond in ways that game the reward model. […] No production-grade actual RL on an LLM has so far been convincingly achieved and demonstrated in an open domain, at scale. And intuitively, this is because getting actual rewards (i.e. the equivalent of win the game) is really difficult in the open-ended problem solving tasks. […] But how do you give an objective reward for summarizing an article? Or answering a slightly ambiguous question about some pip install issue? Or telling a joke? Or re-writing some Java code to Python?

It's very fast to build something that's 90% of a solution. The problem is that the last 10% of building something is usually the hard part which really matters, and with a black box at the center of the product, it feels much more difficult to me to nail that remaining 10%. With vibecheck, most of the time the results to my queries are great; some percentage of the time they aren't. Closing that gap with gen AI feels much more fickle to me than a normal engineering problem. It could be that I'm unfamiliar with it, but I also wonder if some classes of generative AI based products are just doomed to mediocrity as a result.

By gaining mastery of language, A.I. is seizing the master key to civilization, from bank vaults to holy sepulchers. What would it mean for humans to live in a world where a large percentage of stories, melodies, images, laws, policies and tools are shaped by nonhuman intelligence, which knows how to exploit with superhuman efficiency the weaknesses, biases and addictions of the human mind — while knowing how to form intimate relationships with human beings?