That seems to be in support of the parent's point.
So breaking up Google would have resulted in worse outcomes for consumers 10 years ago, but it would be even worse today? Then it wouldn't have made sense to break them up 10 years ago and it would make even less sense today.
I'm not claiming there aren't any good reasons to break up Google (I don't know), but this line of argumentation doesn't work as far as I can see.
I would disagree with the second point. Music theory has definitely changed how I make music. The initial idea is typically still conceived intuitively, but I definitely use music theory to develop it into a whole piece. Before knowing music theory, developing an idea further felt a lot like applying a brute force algorithm. Now, I can recognize what my initial idea is doing and I know where I could take it - both by following conventions and by breaking them.
> Yes, but training is the most expensive part of ML, for example GPT-3 is estimated to cost something like 1-4 million USD.
That entirely depends on how many inferences the model will perform during its lifecycle. You can find different estimates for the energy consumption of ChatGPT, but they range from something like 500-1000 MWh a day. Assuming an electricity price of $0.165 per kWh, that would put you at roughly $80,000 to a $160,000 a day.
Even at the lower end of $80,000 a day, you'll reach your $4 Million in just 50 days.
In academic circles, the abstract is typically a summary, not a problem statement. The rest of the sections are usually read out of order, so it's actually not that different.
I feel like on other laptops everyone is optimizing for specs/price and specs don't include quality of life features like a good trackpad. Sure, it would be nice to have a good trackpad, but chances are your customers are going to buy a cheaper laptop with a worse trackpad from a competitor.
I think any parent who thinks of their child as "worthless" has got to be at least partially at fault for whatever pathologies follow from that kind of attitude.
Of course you can find stuff that's high quality, but that is rare. Rather than looking at outliers, it might be more sensible to look at the average and the reality is that the average graffiti does not have any artistic value.
I'm not sure how it is for the companies, but as universities involved in EU projects, the amount of bureaucracy is just overwhelming and there are so many nonsensical requirements, that you don't get to do any interesting work. Overall, it barely seems worth it.
In my experience, this is only a problem when you are using Excel. It's ridiculous how bad Excel is at handling CSVs, I really cannot comprehend it. If you use LibreOffice all your problems magically disappear.
The interesting thing is that they are still pushing the limits in new ways. Will there be someone else in 10 years that can do what they are doing now? Sure, but SpaceX will then also be ahead of current SpaceX by 10 years. I don't know for how long they'll be able to keep that up, but it's going to be interesting to see.
It does seem different to Tesla. Tesla's big innovation was building an electric car people actually wanted to buy. Everything afterwards looks like incremental improvement to me.
While I agree that moving all human interaction to the internet is likely to be a net negative, there is only so much you can do as an individual. It takes two to tango, you need a sufficient number of other people with the same mindset to resist this move to the internet. As someone whose parents kept them away from certain forms of technology, media, etc., all it really did was alienate me from others. At the end of the day, I'd rather connect with others through the internet than not at all.
It causes lots of problems because many of our systems have been set up in times of growth and won't function without it. Take for example "pay-as-you-go" pension systems in much of Europe. Here, you don't fund your own retirement, but the current working population funds the retirement of the current retirement population. This works great when you have population growth, because you can have e.g. five working people fund one person's retirement. If the ratio moves closer to 1:1 or worse, this becomes a lot more challenging.
Of course, population growth cannot continue forever, so we will have to figure this out anyway. Still, for any individual country, the smart move seems to be to stave it off for as long as possible, observe how other countries deal with it, and then implement the solutions that have actually worked for others yourself.
I remember this transition and funnily enough, the shorter battery life actually made for better user experience. Before, your battery would run out at some point. It was always kind of unexpected and then you couldn't use your phone. With smartphones, it became easier, because you could (and needed to) just charge it every night and you would be fine. It allowed you to set up a routine.
As conceptualized by Seligman, there is also a component of optimism called "explanatory style". Basically, if you experience some kind of adverse event, how do you explain to yourself what has happened? Let's say you are rejected for a job. Do you take it personally? Does reason you were rejected concern all aspects of your life or is it specific to this situation? Is the problem permanent or temporary? If you explain adverse events in a way that's personal, pervasive, and permanent, you have a more pessimistic explanatory style and are less likely to try again in the future.
> This is all true in a neutral net, but Transformers aren't Neural Nets in the traditional sense. I was under that impression originally, but there's not a back propagation or Hebbian learning here, which were the key bits of biomimicry that earned classic NNs their name.
Hebbian learning has never been used with much success in training neural nets. Backpropagation is not bio-inspired, but backpropagation is certainly used to train transformers.
Oh no, the evidence is clear when it comes to smoking. What I was trying to say: Smoking is a nice reference point to orient yourself around, because you can fairly confidently conclude that it's very harmful even with imperfect methods. So we can pick up on large negative effects using imperfect tools. At the same time, when you apply the same imperfect methods to other objects of study (e.g. the impact of coffee on health), you often get much more inconclusive results, which change from study to study. In that case, I don't think there is much sense in worrying until more robust methods can find clear evidence.
The fact that you can pick up on negative effects through observational studies if they are large enough makes me especially suspicious about results that differ from study to study. Perhaps there is an effect there, but if we can't reliably pick up a signal, maybe it's not something I need to worry about.
While this is true, in some sense it is also a feature and not a bug. You don't always want the whole world seeing what you're writing. While Discord doesn't offer complete privacy (unless you restrict access), it provides good-enough-privacy in the same sense that you don't have complete privacy when having a conversation in a town square, but it's private enough. A couple of people might hear what you are saying, but your conversation isn't permanently and publicly archived.
This has its disadvantages, as you point out, but it also has advantages. What I really like about Discord is the fact that people actually connect with each other. I have never meaningfully connected with anyone on Reddit. Discord enables human connection for two reasons: (1) you have a village-sized group of people, meaning you can actually keep track of individuals and your relationships with them and (2) since it is private enough, you can allow yourself to be more vulnerable.
This stuff is all about risk. If two people add their personal experiences, we have two data points. What are we supposed to do with that? It's not enough to quantify risk.
You can do everything wrong and be completely fine or you can do a single thing wrong and suffer the consequences immediately. Those are the two extremes and then there will be a bunch of data points in between those extremes. What you want to pay attention to are averages. How likely is it that I suffer negative consequences from this? If I do, what is the magnitude of the negative effect going to be on average?
To make a somewhat exaggerated comparison: horses still work fine today, but it really doesn't make sense for people to be using them anymore outside of niche use cases or as a hobby.
Sure, macs from 20 years ago still work "fine", but the reason they aren't widely used anymore is not the lack of software updates.
Absolutely, the organization as a whole is succesful, but the exact same people in a different organization might not quite be as succesful. There is something in the DNA of the organization that is just working and that is an effect of how it was set up.
Yes, the label "peer review" is meaningless. There are good peer review processes at top journals and conferences and getting through those is a meaningful accomplishment. Then there are also worthless peer review processes at lower quality avenues. Getting through those is like getting a participation trophy. Yes, you need to show up on time, but not much more than that.
The real peer review happens afterwards, when other researchers actually engage with your work.
This is true, but as in many areas, power laws apply. The vast majority of research has very low impact and mostly serves fulfilling publication quotas. There is likely plenty of scientific fraud happening in lower tier publications as well, but no one notices because no one cares.
I would consider a well-defined mean and standard deviation an assumption. The distribution of maximum entropy is determined by the constraints. Those constraints have to be assumed. If you constrain your problem to only have nonzero probabilities in a fixed interval, then a uniform distribution will have maximum entropy.