That's what I thought when reading this, the opposite of double (100% gain) is half (50%) loss, so if you've lost by more than 50% on the negative coin toss result you will tend to lose money over time.
If the negative was less than half (say, 40%), you would gain money over time.
I've just read the first part of the article and it is great to see someone writing a thought that I've had myself. If you're skilled in music, you're 1 in a thousand. If you're skilled in programming, you're 1 in a thousand. But if you're skilled in both, you're one in a million.
This fits in with my philosophy of just following your interests with no particular goal in mind. A popular example of this is Steve Jobs in college just taking classes that he was interested in. He took a calligraphy class and that ended up informing the typeface for Apple computers.
I've found this myself through taking programming courses for general interest, which then afforded me the required knowledge to take a coding bootcamp.
I've also taken a physics course for general interest and that allowed me to formulate my thinking on software development in a similar way to how new scientific theories are created.
Also happy to have found LessWrong, this seems like my kind of content.
I will continue reading.
Speed of task execution is a separate development vector from Artificial Super Intelligence (ASI). Using the calculator as an example, being able to compute something a million times faster than a human doesn't mean it's any smarter.
I thought that the risk of ASI is that it would outsmart us (humans) by doing things that we can't comprehend or, if nefariously incentivised, finding vulnerabilities in our systems that we are not smart enough to predict.
Simply doing things that a human can do, but faster, is not ASI, unless I'm missing something?
I'm personally not convinced that the recent AI boom, which has mostly centred around LLMs (ChatGPT etc) has had much impact on the development of ASI. Are LLMs able to formulate more intelligent insights than the data on which they were trained? I.e. within the text format, this is data that has all already been filtered through a human brain.
I would expect that a super intelligence would require direct access with the real world, not information that has been passed through a human filter. This may be achievable by training models on video and audio data, which is a more direct feed of the real world, but I would guess that giving an AI arms and legs etc, that allow it to interact with the real world to experiment with things, would make it learn much quicker.