guygurari

Programming brings me joy in two different ways.

1. Crafting something beautiful. Figuring out correct abstractions and mapping them naturally to language constructs. Nailing just the right amount of flexibility, scalability and robustness. Writing self-explanatory, idiomatic code that is a pleasure to read. It’s an art.

2. Building useful things. Creating programs that are useful to myself and to others, and watching them bring value to the world. It’s engineering.

These things have utility but they are also enjoyable onto themselves. As best I can tell, your emotional response to coding agents depends on how much you care about these two things.

AI has taken away the joy of crafting beautiful things, and has amplified the joy of building things by more than 10x. Safe bet: It will get to 100x this year.

I am very happy with this tradeoff. Over the years I grew to value building things much more highly. 20yo me would’ve been devastated.

Achieve a significant scientific or mathematical breakthrough without human supervision. Domain experts should agree that the new result is truly groundbreaking, and achieving it required fundamentally new ideas — not merely interpolating existing results.

Examples of discoveries that would have counted if they weren’t already made: Relativity (say a derivation of E=mc^2), Quantum Mechanics (say a calculation of the hydrogen energy levels), the discovery of Riemmannian geometry, the discovery of DNA, and the discovery of the theory of evolution with natural selection.

The idea is to test the system’s out of distribution generalization: its ability to achieve tasks that are beyond its training distribution. This is something that humans can do, but no current LLM appears to be able to do.

For recipes I use the Paprika app [1]. I point it at one of these terrible websites and it will scrape out the ingredients and instructions and present them in a useful way.

It doesn’t address the bigger problem but it solves this problem well.

[1] https://www.paprikaapp.com/

Here are examples of people who were cancelled without even saying anything [1]. One was due to a misunderstanding. Another was due to something a family member said.

[1] https://www.theatlantic.com/ideas/archive/2020/06/stop-firin...

I’m rooting for you. Here’s a thought about a possible trade off between radicalization and engagement. When social media platforms optimize for growth, it makes sense for them to make it as easy as possible for users to share/retweet. It lowers “amplification friction” and allows messages to go viral. The most successful platforms have very low amplification friction, which suggests that low friction is an important ingredient.

What we are learning is that making it trivially easy to amplify anyone’s message enables cancel culture and (I believe) leads to radicalization.

If this is correct, then increasing amplification friction on your platform will lead to less radicalization, at the cost of lowering engagement. My guess is that for this to be successful requires a careful balance of where you land on the higher/lower friction spectrum. Too much friction leads to low engagement which leads to failure. Too little friction leads to uncontrolled amplification which leads to radicalization. So a balance is needed.

Either that, or a totally new idea is needed that turns existing platforms on their head.

It’s certainly possible my intuition is wrong. Run-flat tires can support the whole car, but it is possible this support only kicks in at very low pressure.

I agree that higher tire pressure results in smaller surface area, but it's not clear to me that the tire itself does not support the car, skewing the result. So I'm not convinced this sanity check is sufficient. For example, as another commenter mentioned, run-flat tires can support the car on stiffness alone.

The suggestion is to use F=PA to measure the car weight, where P is the tire pressure, A is the area the tires in contact with the ground, and F is the unknown: the weight of the car (measured in terms of gravitational force).

I don't think this works. To see why, suppose we make the tires out of a very stiff material. We control the pressure P, but changing the pressure will not change the surface area A. Therefore, by changing P we can set F to whatever we want, which shows that we are not actually measuring the weight of the car. The basic issue here is that the air pressure is not the only thing that is supporting the car.

I don't know how important the stiffness effect is for real tires, but I suspect they are sufficiently stiff that it matters a lot.

> “do what the experts say we should, while maintaining a diversity of expert views.”

I don’t think this is an actual strategy. If there is a diversity of views, which experts do you listen to? Also, experts in which field? Epidemiologists have generally recommended very strict shutdowns, but they are not economists and are not experts on the economic ramifications. Some economists agree with strict lockdowns, others do not. So it’s not at all clear what “listen to the experts” means in this case.

If Sweden has a significantly higher percentage of the population infected (as they claim) then it doesn’t make sense to compare death rates at this point, because it may be that these other countries will reach the same numbers (infection rate and death rate) later. This is a likely outcome because Sweden’s healthcare system has not been overwhelmed, so it’s not obvious that their policies will lead to excess deaths in the long run. The thing to compare is the final death toll (as well as economic damage), and we don’t know that yet.

I am sorry to hear you’ve lost your source of income.

> But financial hardship is solvable -- food, housing, etc. can be handed out.

That depends on the size of the handout. At the scale we’re talking about, I’ll just say that I don’t think it’s that simple. It is also without a successful precedent.

> Are people starving?

I don’t think so, but we also know that many people in the US live practically hand to mouth. By now, millions of Americans are probably without any savings or income. See the long lines at food banks for evidence. I expect that the longer this goes on, the more Americans we will see become poor, homeless, and find it difficult to feed their families. I don’t think people will starve, but how will this be resolved? Maybe instead of starving they will go to loan sharks and dig themselves into a hopeless financial hole.

I find it disturbing that you are willing to put all of this aside because you don’t know anyone who is experiencing serious hardship. Do you know anyone who goes to food banks? I don’t, but I know they exist and I don’t want to ignore them and say they will be fine because “we can hand out food”.

> In 100 years, the only footnote for 2020 will be all the deaths and suffering that this caused. No one will care that you were sad because you couldn't see your friends or go do sports.

It's weird. Why do you empathize with victims of the disease but not with victims of financial hardship? Do you not see people suffering when looking at hour-long lines at food banks and historical unemployment numbers?

Perhaps what will be remembered is the pain of those who were plunged into deep poverty in a New Great Depression that will take us years to recover from.

> At 2% a day cases double every 5 weeks. Would anyone apply this reasoning to any other disease?

I don't know but maybe the flu [1]? I don't think focusing just on the infection rate is useful.

> Classic psychological denial is clearly operating in force.

I think the denial is in thinking that we can beat the virus by extending the lock down for years (you say "might take two or more years"). It is likely that all we are doing is slowing down the spread, and that it will eventually infect most of the population, stopping only when we reach herd immunity. I say this is likely because this is where we are headed unless we come up with a vaccine or treatment at an unprecedented speed -- not something I would bet on.

If this is indeed where we are headed, it may be better to open up as much as possible short of overwhelming the healthcare system. Note that in many places, the healthcare systems are significantly below capacity (e.g. [2]).

[1] https://www.who.int/docs/default-source/coronaviruse/situati...

[2] https://www.sccgov.org/sites/covid19/Pages/dashboard.aspx

To be clear, by “This is hysteria” I meant the gp’s claim that every healthcare system is overwhelmed. I was not referring to the general response.

However, see Sweden for a clear counter-example to your claim that the services are holding up because of the drastic actions taken. Their services are holding up with far less drastic measures.

My theory is that it’s because they stayed open. The numbers are higher, yes, but it has not been a catastrophic disaster like many predicted. Crucially, the health care systems are not overwhelmed.

So that is an important starting point for a discussion: it is possible to keep things mostly open (with appropriate safety measures) without overwhelming the healthcare system.

Note that it is possible that Sweden’s approach will eventually lead to a comparable number of deaths as its neighbors, but at an accelerated rate. Namely, it is possible that people are getting infected at a higher rate, which explains the higher numbers, but that the mortality rate will eventually be the same because the healthcare system is not overwhelmed. This depends on when and if we will find a vaccine and better treatments, which is an unknown.

On the other side of the equation, there is no mass unemployment, no hours-long lines at food banks, people are not suffering from mental health problems due to isolation, and so on. If Sweden does end up paying a higher price for its policies, it will still enjoy these benefits. They decided that it is worth the trade off. They are trying to prevent covid victims while also trying to protect potential economic victims, whose suffering is not being ignored. Is that wrong?

[*] 88% of deaths in Sweden are people over 70, and I believe many of those were in nursing homes. So it is possible that fine-tuning the policy can significantly reduce the death rate while keeping things open.

https://www.statista.com/statistics/1107913/number-of-corona...

I have —- comparing with additional European countries, Sweden deaths per capita are lower than France, the U.K. and Italy, and is followed closely by Spain. Therefore, as I said, Sweden is not an outlier. At least the U.K. and Italy enacted far more restrictive policies than Sweden but achieved worse results. And this is just looking at the health part of the equation, not even discussing the economic benefits of Sweden’s approach.

I think this is enough to at least challenge the common viewpoint that restrictive lockdowns are the only way to deal with this.

None of what you said here takes away from the fact that your original claim is false. Again, you wrote: “ completely overwhelmed every single health system in the world.” That is not true, not even close.

Now, you write:

> Sweden has been heavily criticized for its handling of the situation

Yes, people have differing opinions. The fact is that Sweden is not an outlier in number of deaths despite keeping its economy essentially open. Perhaps this is what happens when you inform your citizens and then trust them to do the right thing?

> Pretending that doing nothing could not have catastrophic consequences is dangerous.

I agree, and I did not suggest that we should do nothing. But what are the optimal policies? Are the policies we have now optimal or can we do better? We need open and honest discussion to make progress on these questions.

You are making several false claims, most notably:

> We're talking about a "flu" that has, in a matter of weeks, completely overwhelmed every single health system in the world.

Many health care systems have been well below capacity throughout this. This includes most of the US (e.g. California), it includes places that have not shut down (Sweden), and it includes places that have reopened weeks ago (Georgia).

Despite these easily verifiable facts, many people believe as you do that health care systems are being overwhelmed everywhere. Maybe because the media is painting this picture by only focusing on the few places where this is actually happening.

This is hysteria.

It is blinding people from seeing the real, measurable damage wrought by the policies being enacted.

To be clear, I am not arguing for or against the validity of the claims regarding covid-sim. I am only saying that this is not a good example of the effect where people feel entitled to get open source code that fits their needs.

I don’t doubt it, but covid-sim is not the best example of this effect. In this case, people are worried that poorly written simulation code is informing policy decisions that affect billions of people. They are not demanding that it should be fixed, they are demanding that it should not be relied on in its current state.

Many people die in traffic accidents every day, but we don’t ban cars. Everything we do involves some risk. There needs to be a rational discussion on how much lock down is appropriate. We need to look at other countries like Sweden and Austria and learn from them. We need to look at the real costs of a depression. What we have now is “People are dying so let’s lock down everything”.

We don’t know enough so let’s stop the economy while we find out? That is not typically how we do things, because it’s not a rational response. For example, we didn’t know whether holding a mobile phone close to our ear for extended periods of time leads to serious health issues, and there was suspicion that it does. But we didn’t prohibit it — we let people decide for themselves, and it does seem the risk was lower than suspected.

What’s happening now seems more like a knee-jerk reaction than the result of a cost-benefit analysis.

It is hard to overstate how important the resolution of this question is to physics, and in particular to theoretical physics. The “cosmological constant problem”, that is the problem explaining the cosmological acceleration (if it exists), is arguably the most important and hardest question in high energy theoretical physics. Many theoreticians have spent significant effort studying this question, working under the assumption that the empirical evidence is solid. If this is not the case, it changes the landscape of cutting edge physics research.

There are several reasons why explaining the observed acceleration is so hard. The cosmological constant (the measure of how much dark energy there is) is a tiny positive number which seems to require a lot of “fine tuning” to explain theoretically. We can easily include it in general relativity, but our best understanding of quantum physics says that if it is there then it should be much larger. This means there probably is something we don’t understand about its microscopic origins. If we try to build a microscopic model that has this small constant using string theory (our best guest at a complete theory), we find that such models are hard to create. In fact, it is not clear that any model that includes dark energy is even valid in string theory! Any way we look at it, it seems more difficult to explain this number if it is tiny and positive than if it is strictly zero (no acceleration).

Finally, theoreticians don’t have much to go on when explaining this phenomenon besides this one single number — there aren’t closely related experiments we can combine to come up with a coherent picture of what’s happening. Combine this with particle physics, where accelerators provide us with an abundance of data. It is a single tiny number that has puzzled theoreticians for decades.

In my experience, the reason to “follow the math” is that it is the best and perhaps only way to truly understand a physical theory, especially one as strange as quantum mechanics. One can understand the math first, and then develop an understanding of the context and meaning. But this abstract understanding is anchored in math. This is important because when the meaning gets too obscure, one can return to the math to resolve any confusion. The other way around does not work.

The article makes some fair points for performing end-to-end tests. There is however a benefit of unit tests over end-to-end tests that is not mentioned: code locality. The cost of fixing a discovered bug seems to grow exponentially with the amount of code in which the bug may appear. Unit tests typically cover a small part of the code. If a unit test fails, isolating and fixing the bug is typically cheap. On the other hand, if an end-to-end test fails (due to the same bug) then isolating and fixing the bug can be quite costly because there are many different parts of the code that need to be checked. To me, code locality is the main advantage of writing unit tests.

At the end of the day, there is a spectrum of tests going from unit tests to end-to-end tests. The spectrum represents several trade-offs such code locality vs. coverage. In my experience, the most economical approach is to write a balanced mix of tests the lie along this spectrum.

I’ve long had this dream of building a system that would allow for symbolic algebra manipulations, where the representation knows about the semantics. The idea is to replace long and tedious pen-and-paper symbolic calculations, while still having direct control over each step of the calculation. In looking for something like this I’ve found specialized proof verifiers, which are doing something different. At the other end of the spectrum there’s Mathematica, which does symbolic manipulation but doesn’t understand enough semantics.

Are these projects aiming at something like what I’m describing? Or are they more about something else like verifying proofs?

As far as I know, string theory is the only complete theory that incorporates these two observations:

1. Nature includes the force of gravity, described at low energies by Einstein's theory with a gravitational field.

2. Matter in Nature is quantum mechanical.

We believe that spin 2 particles will exist in any such theory (as they do in string theory), but even without this assumption it is still difficult to come up with such a theory.

The example you give of gluing the Standard Model to General Relativity does not achieve this goal and does not correctly describe nature. This is because the quantum matter of the Standard Model does not correctly back-react on the classical gravitational field.

When we couple the classical theory of General Relativity to classical matter, the gravitational field pushes matter around, and the matter in turn back-reacts on the gravitational field and changes it. When coupling quantum matter to classical General Relativity, the classical gravitational field can act on matter just fine, but it does not work in the other direction. The problem is that when matter is in a quantum superposition, it is not clear how to update the classical gravitational field (which cannot itself be in superposition). This missing part of the interaction means that SM+GR is not a complete theory.

That is the point of the paper: making deep fully-connected networks work.

> iOS has a defacto 100% monopoly on devices compatible with Apple's Ax series chips

Sure, and equivalently Apple has a 100% monopoly on Apple devices. How does that make sense? One problem with anti-trust laws is that the market is not well-defined. Is the relevant market Apple phones? Smartphones above some price point? Smartphones at a price point? All smartphones? We can always pick and choose to make it seem like a company has a monopoly.

> Then these scientists should be devoted to producing textbooks and courses ...

Textbooks and courses exist for everything but the most cutting edge stuff (which are still in flux anyway), but they are a very inefficient way of transferring knowledge. I would say they are practically useless without expert guidance. At the most basic level, there are so many of them that an expert has to tell you which ones are both good and relevant to what you want to learn. I've once seen a student waste months of his life studying a book he thought was relevant, only to discover that book wasn't building towards the sort of knowledge he needed in that subject. The book was about the correct subject, but was focused on somewhat different aspects than the ones he was interested in. There was no way for him to know this in advance without guidance.

So we don't know how to organize existing books. Also, even the books that exist are usually pretty bad at conveying knowledge. Or perhaps humans are just pretty bad at learning things from books. Either way, no one knows how to write textbooks and courses that are much better than what we have today. I really don't know of a better way to preserve knowledge than the current one. Perhaps technology can improve the situation by making access to knowledge more interactive. But I suspect this would require a real breakthrough.

> We can't afford to live in a world where only 1.7% know how things work.

Why not?