johbjo

Board game AI works by searching through the state space, and evaluating each state with the neural network, then picking the move with best expectation.

So it needs to load the comparatively tiny game state (chess board) into the GPU for each evaluation. The more game states it can evaluate per move, the better it is. It can be in the order of millions.

I remembered this quote from Grady Booch, at 13:00 in this video:

https://youtu.be/adiVOdztQ34

"These software artifacts we produce, are perhaps some of the most complex humans have ever created."

One thing that might make software fundamentally different from other stuff, is that before being able to make a meaningful contribution a person needs to invest X hours just to understand.

In most other creative fields, this can be hand-waved away. No other engineering field has anywhere near the complexity.

Therefore, all important people on a project are stakeholders of some sort, and anyone else is not able to make real improvements.

The project will move in whatever direction satisfies the incentives of the important stakeholders.

The idea behind the article is incorrect. The entire ecosystem of languages, libraries, and operating systems now is a far greater accomplishment. And agile everywhere.

> things like batch norm. We don't know why it's important

It is pretty well understood.

His examples are CNN image applications, and it is true that these were spectacular new results 10 years ago. Since then, the tedious engineering has improved and recreating the results now are indeed just a few lines of python.

The reason they were so impressive then is because it seemed they were "more intelligent" than they actually are, and people extrapolated the same level of performance on any task. Achieving it seemed more impressive then, than now.

But "AI skills" should be broader than image classification.

There are various implicit contracts involved.

Using your software is lending you reputation, power, and influence. In high profile projects, this mindshare is more valuable. If your software would be unreliable, the trust and reputation would be misplaced.

For example, if an update script in a major "stable" distribution would clobber user data, the tone in these comments might be justified. But mostly, this is bad form.

In that case, everything is a battery. Just radiate heat into space and call the night sky "a battery".

No. Careful partitioning is the solution to this problem. Monitor the growth of your partitions and make sure nothing on rootfs or other sensitive partitions grow significantly.

The bicycle quip appears in so many videos from around that time. Interpret Jobs videos as case studies in a type of marketing.

He's framing Apple marketing as "it's so exciting that we can help people so much", but it's honest and actually true.

The best videos of Jobs are the rare internal ones.

Also, he's 25 in this video? wtf

Similar to immersion when learning a language, you need to to do the exercises from the ground up until the proofs seem self-evident. If there are uncertainties about the exact meaning of notation and operations, you will probably fool yourself into believing that you understand.

Many basic resources are now almost free in the sense that the market price barely covers cost of production/collecting. Vegetables and certain farm products are examples of this. Slightly out-of-date electronics. Many more things are like this. It reduces living costs.

This has had no effect on politics wrt to wealth distribution. Rather, decreasing living costs has resulted in increasing inequality because it reduces the perception of need for change.

AI in production brings a similar change to the economy as when Chinese labour started producing goods for the world. It will make shiny toys available to everyone, but it will have no effect on wealth or income inequality.

So, AI will result in even more "homeless" people with not only iPhones, but an abundance of all sorts of toys. Maybe even houses.

Copyright and licensing is complicated, in the sense that a track has many different copyright holders, and the record companies usually make sure to own distribution rights by contract. The composers and artists might be powerless, even though they own the actual rights (which can not be sold or transferred). This partly the reason why the major labels hold so much power.

Here is a great source for learning more: http://copyx.org/lectures/

The three or four major media companies/rights holders correspond to roughly 75% of streams on Spotify, and probably the entire catalogue of music pre 2010. This gives them the bargaining power to influence playlists, recommendations, and so on. This is what they mean by payola. They can use this to launch new artists.

Independent artists don't like this.

I would do this:

Let i = ((n % 3) == 0) | ((n % 5) == 0) << 1

Then map output as { n, 'Fizz', 'Buzz', 'FizzBuzz' }

The look/feel/interaction in media apps today is essentially a "cultural good" in the same way the design of a TV-franchise is a cultural good.

It needs to be updated every few months to feel fresh, and the time horizon is very short. There needs to be analytics/engineering-work to enable this, but the goal itself is to adapt the user experience such that it maximizes adoption and screen time. What matters is fast iteration, creativity, and "fast plumbing".

The perception is that young people are more enthusiastic about this sort of work. More experienced people would be unnecessarily concerned with code quality, stability, technological trajectories, and so on.

It's a technical concern. Publishing will only cause fear and doubt.

In any case, it's a matter of weighing risks. Even if there's a 10⁻⁶ probability of complications, giving the vaccine and stopping serious illness/deaths is still a better decision overall. Even if it means hundreds of cases of complications across Europe. And then, the complications might turn out to be statistically difficult to discern from uniform random anyway.

A probability of 10⁻⁶ is difficult to conceptualize so if it is published with scary words, people might interpret it as "will probably happen to me". Then if that affects vaccine roll-out negatively, publishing might be a bad decision.

Millions of people in the UK have received the AZ vaccine, which has demonstrated safety and efficacy.

The efficacy metric means protection against infection, which is not complete. But as far as I've heard, not one person who has received either Pfizer or AZ has become seriously ill or died. In other words, it "flattens the curve" instantly.

Learn things regardless whether you receive short-term rewards for them.

People saying "learn by doing" or "learn just-in-time" are essentially saying; only learn easy things that you receive immediate rewards. This results in shallow and high-supply knowledge.

The reason some knowledge is rare/valuable is precisely that it is impossible to learn from a short and immediate feedback/reward cycle.

This is the reason maths feels more difficult to learn than coding - there is no external reward. And why C is more difficult to learn than python.

> I think my grasp of “the big picture,” or a high established baseline, is a big reason for the fast uptake.

This is a valid argument and true also in my experience.

The counter argument is that "software building" is to media companies (faang) what a press is to an advertising company. Problem-solving is not the relevant value-creating activity. The perceived need is for "creators" and "operators" for the "machine", rather than engineers and problem-solvers.

Screens today are spaces that need to be filled with "cultural content". The technology itself is a public good, and they want willing bodies that don't ask questions to turn the wheels of the "machine".

This is the rationale for ageism, whether valid or not.

An economic asset is valued based on expected future cash flow. In the last decades however, real estate and some stocks are mainly valued based on expected appreciation. Real estate even generates negative cash flow for most home owners, and some investors.

Per se, it is speculation (in higher prices) rather than investing (in future cash flow).

> If the best returns is on beans and rice, people will starve.

Farmland and equipment are assets, but they are valued based on cash flow from selling produce. Produce is perishable and can only be sold as food, so it will be sold at a market prices regardless of speculation in farmland.

> all of the wealthy will liquidate their stockpiles of stocks(crashing the stock market in the process)

Not "all of the wealthy". Relatively little is required at the margin to crash the stock market, and the majority of "wealth" would follow the market down. Only a tiny proportion would be able to exit at the right moment.

Asset depreciation is an adjustment of the relative value of labour vs. assets.

A possible delineation between "water" and "oil" is whether the thing can be legally used as collateral for loans and how the value is booked. But idk.

I have a theory that fashion (set design, clothes, makeup, UX) progresses as fast as the audio/video media begins to show age. Pictures are clearly dated by the media itself, which means the contents of the pictures are also inevitably dated.

Fashion seemed to progress quickly during the 70s/80s and then begin to decelerate from the 90s onwards.

It is more difficult now to date recordings from the 90s based only the audio/video quality which means it takes longer before it "feels old".

> So how do the economics of this work out? Can anyone explain to me?

Idk, but maybe the engineers don't fully trust it, so they outsource the liability of running it with some contractual guarantees. And at the moment, this probably looks more fashionable, so no one dares question it in fear of seeming backwards.

If the customer relationship is sticky and product/technology improves over time, it is more important to be quick than "mature" or "feature complete". When the technology matures (roughly simultaneously for all competitors), the size of the user base is what matters.

One intuition about the large GPT-models is that they are learning "facts" like they are learning grammar rules. It's just that their capacity is so immensely large as to learn facts as part of the language model.

> Less than 10 years later, a general-purpose open-sourced technology without the massive mainframe or cooling fans can compete on the same level.

First, are GPT-3 models generally available?

Secondly, GPT-3 is many orders of magnitude larger than any model existing 10 years ago. There are certainly cooling fans involved.

The type of coding that can be self-taught in a few months/years is exactly the type of coding that guys age<30 will be more cost-effective at. A disciplined engineer/mathematician can learn to code applications quickly and possibly at a higher level of quality than 20 year olds, but it won't matter. You will be competing against brawn and caffeine rather than brains.

If you have a background in mathematics and/or statistics, consider transitioning to data science instead.

If you have no science/engineering background, it would be tough.

Overall production/consumption has probably decreased. This means that food has has higher relative share of consumption.

Therefore food prices shift right?

Possibly biggest difference: government deficit spending and negative trade balance. Should contribute to GDP through spending that would otherwise not happen.

This is not afforded to anyone without reserve currency status.

EU governments are prevented from running deficits due to some treaty in order to keep the euro stable, and historically the currencies would depreciate quickly.

The same reason single independent videographers can post videos with higher production value than TV-networks just 15 years ago.

The reason is that design, photography, image/video editing, and printing became basic capabilities available to anyone.

Unless it had a strategy to counteract piracy and other illegal content, it would become a target for law enforcement as soon as it grew big enough to be relevant.

For example, one attack could be to start legal processes against key people in order to bankrupt them, or possibly jail them. This would be function as a deterrence to most people committing resources to the network.