hyperthesis

Adding to the previous discussions on counter-steering bicycles: you can see it in tyre tracks. e.g. ride through a puddle; or, ride from rain to under cover.

Consider: how do you steer riding no-hands? Shifting your body weight makes the front wheel turn (because of the "rake"), but now your bodyweight is on the outer side of the turn! Hold it for a (brief!) moment, then lean the bicycle the other way and you have a nice turn in the opposite direction - which is what you wanted all along.

> In retrospect, I think I saw the fundamental difficulty in what we were doing and I don’t think he appreciated it (look at the stairs picture and see if you can figure it out)

What was it?

BitKeeper already did it.

AI safety by the waterfall model.

But how else can we do it? We learnt how to handle other dangers by trial and error...

One approach is legal. The law is very slow to adapt, and is informed by a history of things gone wrong. The simple answer is that OpenAI has "strict liaility" (meaning they are liable even for an accident) for any damage caused by an AI, like dangerous animals that escape.

I know it seems ridiculous to consider liability when the fate of humanity may be at stake... but this is the language of companies, directors, insurance companies and legal counsel. It is language they understand.

What could disrupt OpenAI is a dramatic change in market, perhaps enabled by a change in technology. But if it's the same customers in the same market, they will buy or duplicate any tech advance; and if it's a sufficiently similar market, they will pivot.

in another 700 years, the pylon ruins will be the focus

bare hills result from deforestation

I googled for video https://www.google.com/search?q=Genetic+Algorithms+for+Evolu... - closest hit is from 2014 https://youtube.com/watch?v=eWxFI3NHtT8 [1:41, "able to complete the QWOP race in about 2 minutes"]

BTW I hope no one tries this with LLMs. It will turn them against humanity.

Division of labour creates wealth... that's not distributed.

BTW The origins of basic supermarket foods is amazing: e.g. potato, corn, avocado are from the Americas. What did Europeans used to eat - cabbage and turnips?

Good, infrastructure matters: Birdseye, famous for fish, got freezers into supermarkets; Edison, famous for lights, constructed generators; Tesla (co), famous for cars, constructed charging stations.

ASTs

A related concept is compilers that keep going after the first error. Though I kind of like the hard-mode idea of a compiler that has only one error message, "fail".

BTW: could provide valid options (perhaps highlighted), not constraints. (...maybe been thought of already)

Disconcerting, like returning to someone's comment, and it's subtly different from how you remembered (before you knew of the recent editing feature).

Most (all?) scientific ideas are simplifications. e.g. that species are separate. That the relationships between species can be represented by a tree.

Yes, it used to be taught that genes don't change. But there's methylation and telomeres and "gene therapy" is a thing. I agree that changing gene expression sounds like sophistry, since the point of genes is their effect. Sounds like Lamarckism to me.

Everything you've been taught is wrong - in that however thoroughly something is known, there's always more to it.

Reminds me of ZX81's BASIC: you didn't type keywords. There was a key for each one on the specialized keyboard. The editor was context-sensitive - in other words, a parser-driven lexical analyzer was you. At the time, it seemed a little magical to me that it knew whether you wanted to type a word/number versus a keyword, and what kind of keyword (but really, the pedagogically relevant aspect was that I knew).

BTW Significant to the ZX81 was the memory saved by storing tokens instead of their text representation. Yes. It had 1k RAM kids - including display memory (no separate GPU). This parsing when entered would also make runtime faster.

IIRC for powered paragliders, it turned out propeller noise drowns out ICE noise.

[Maybe disregard this, as I'm not your requested audience, and I realize this is a tech/biz post to HN and you are deep in the details of confronting the unknown, but...] I think for pitch-deck/public communication, as soon as you say "hydrogen aircraft", immediately address or at least acknowledge "what could possibly go wrong-Hindenburg" and "why not batteries" (not para 3).

EDIT you could defer the issue with "Fuel cell electric aircraft"

BTW The first solar+battery electric aircraft world-circumnavigation got hardly any publicity, which puzzled me https://en.wikipedia.org https://news.ycombinator.com/item?id=9824570/wiki/Solar_Impu...

EDIT Since I'm already downvoted, I'll add: I wonder about fuel cells for powered paraponts/paragliders, where battery-electric are available, but aren't quite there yet vs petrol, esp for range.

sama's fundraising talent is moot at MS.

Feynman has a bit where empirixal measurements of some universal constant (electron charge?) slowly crept up to the now-accepted true value. He said scientists are a little bit ashamed of this, because it showed that higher estimates were discarded and not published - bias.

BTW scientists get no upvotes for merely reproducing "known" results.

Is there a way to see natural merit/discovery in reproducing (not an artificial incentive, like grants for repeating studies). Perhaps analogous to how teaching helps your own understanding?

If money and talent is all you need, he's gonna win. But I think this doesn't work for research talent, it's more for commercialization. Think PARC vs ethernet/adobe etc. IDK if AI is really at that stage yet.

Let's see how this pans out.

5581 now. Looks set at #3, as other posts update the story. Aggregate would be #1.

He's extraordinarily talented by all accounts, especially at fundraising. But his Apple was Loopt. He's not the secret sauce but a facilitator - more Tim than Steve. Perhaps like Steve was to Pixar.

"The start-up company must either cross or die, but what value is life if to gain it one has to go against one’s best self?" - Moore, Crossing the Chasm, p. 75

It depends on the resolution of discretization required. Kurzweil's prediction is premised on his opinion of this.

Note that engineering fluid simulation (cfd) makes these choices in discretization of pde's all the time, based on application requirements.

He learnt about recursion in a fp workshop(?), and as a first exercise, tried applying it to sort.

You pick a number in the list, then put lesser numbers into a left-list, and greater into a right-list. Recurse.

It's very slow as intuitive, simple and elegant fp, because it's creating new lists like crazy; but a mutable, in-place version (e.g. in C) is super fast. Maybe the origin of the folklore "learn fp to be a better coder, not to code fp"? I don't think he expected it to be so good; he was just doing him. A very smart man.

You can't autocomplete if you don't understand. It's talking the talk.

But there's levels of understanding. That's where TFA's logical understanding comes in, because it makes possible entirely new insights. It's walking the walk.

How routine is it for human's to have entirely new insights? Note here we mean, new to them. I suggest we do it all the time. Every time a child understands something new. Even if it is "taught" to them, they must acquire the insight for themselves.

funfact: Shannon investigated how good humans are at predicting the next letter of English text (one-character "autocomplete"?), to estimate the information entropy of text. https://news.ycombinator.com/item?id=35813991

Theory: "imposter syndrome" is simply seeing the truth. No one lives up to the social conception. Einstein needed help with mathematics. The realistic perspective is the curiosity of a scientist; the humility of a mortal. Knowing the premises does not automatically imply you know all the consequences... even though it "should".

Bugs have been found in production binary search.

Skipping bubblesort, I think quicksort is easy to implement, if you know Hoare's fp insight (I remembered it, but had to check it was him). Making a version that is both efficient and correct may be harder for me... Similar for Boyer-Moore substring matching (I remembered their names!)

Copyright law protects computer programs as the expression of an idea. One idea has many expressions. If you want to modify or extend it, you need to know the idea.

  [webpage text follows]

Even though I don't respect podcasts as an information delivery system, I recently listened to a podcast that felt like having an epiphany.

The podcast in question was episode 61 of Future of Coding. It is essentially a read-through/review of two papers with which I was unfamiliar, but which I believe are actually quite famous and influential.

Why did listening to this feel like an epiphany? Well, I suddenly felt like I understood what the deal is with software. Why is it that when you join a company, the engineer who's been there for years seems like an incredible genius? Why do some teams that I've been on struggle while others manage to get everything right? Why is everyone always so keen to rewrite things?

The two ideas that the podcast expresses are:

    The concept of what a “Theory” is, according to Gilbert Ryle.
    That being a programmer is doing “Building theories” in the Ryle sense of the word.

Having these two ideas explained together was really helpful. If I had read Ryle by himself, I would have thought, “interesting and useless”. If I had read Programming as Theory Building without knowing the theory concept, I would just not have understood.

I recommend listening to the podcast and reading the paper. But if you don't want to do that, I'm going to try and explain the two points.

strange What is a Theory, According to Gilbert Ryle?

When Ryle says theory, he doesn't mean anything like what other people mean when they say theory. Annoying. He should have just come up with a new word. What he means is the thought object that exists in our minds which allows us to do things.

I, for example, know how to cook pasta. When I cook pasta, that's a certain expression of this knowledge. When I try and explain to you how to cook pasta, that's a different expression of it. Neither of those expressions contains everything I know about cooking pasta. And in fact, there are parts of what I know that I can't really express in any way. This knowledge is what Ryle would call a theory. I have a “Theory of how to cook pasta”. This theory is not something that exists in language or action - it’s a something that we can never fully express. The closest we can get to transferring a theory is to demonstrate the expression of the Theory to someone over and over again until they build their own theory. That theory won't be the same as ours. What Does it Mean that Programming is Theory Building?

It means that the code base we create is not the true product of our work. The real product is the mental theory of that code base which:

    Allowed us to create it in the first place.
    Allows us to diagnose problems with it and fix them.
    Allows us to modify it easily.

If I think about times when I've worked on a team that works well and gets stuff done, it's been a team where:

    Someone has been there for a long time, since the start of whatever code base/feature we work on.
    Other team members have joined slowly, and had a chance to work with the people who know more.
    The area of focus does not change. We haven't been reassigned to a random existing project, or asked to fix some other team’s work.

The paper also talks about what happens when all the people who have a theory of a given program stop working on it. It dies. Yikes. It’s claimed that we can’t rebuild a theory from code and documentation.

This model explains a few curious phenomena:

    What “legacy code” actually is - it’s a code base which is no longer maintained by people who have a theory of it.
    The solo engineer who can make a better product than a team of equally competent professionals. The solo engineer has spent the time to build a complete theory of their program, the professionals move between projects regularly - and only have theories of what they've worked on.
    Why getting up to speed with unfamiliar projects is so much harder than just rebuilding the thing. To truly build a theory, you need to mentally rebuild the existing code base anyway.

Hats on to you. The AGI was compellingly persausive that sam must be fired immediately.

They didn't fire him for that reason, but because it would be dangerously persuasive to release without safeguards - which is what sam planned to do without telling the board.

Deep learning was an advance. I think the fundamental achievement is a way to use all that parallel processing power and data. Inconceivable amounts of data can give seemingly magical results. Yes, overfitting and generalizing are still problems.

I basically agree with you about the 20 year hype-cycle, and but when compute power reaches parity with human brain hardware (Kurzweil predicts by about 2029), one barrier is removed.

Depending on jurisdiction, technically, a trade mark ™ is established by use, for the tort of passing off. A registered trademark uses the encircled R symbol: ®

You may be thinking of copyright, which at one time did not explicitly include computer programs as protected works, depending on jurisdiction.

A tricky conflict. Legal issues cause strange decisions. Maybe even Sam recommended his sacking, for optimal damage control.

The significant factor in "disruption" [Christensen] is not low-end products - incumbents typically crush them, and markets consolidate over time; many competitors become few. What's newsworthy and puzzling is when the low-end entrant does win. The key factor (IMHO from reading several of his books) is when incumbents cannot address low-end entrants, because their customers have captured them in some way (e.g. all aspects of the organization are optimized to their customers).

Thought experiment: how could this apply to human thought? i.e. what is a low-end that we couldn't address?

Disclaimer: I really like Christensen's reasoning, theories, and expression of them... but he did predict the iphone would not be disruptive. So, at best, difficult to apply!