Nothing in the bill mentions submitting this data for approval. It just says the data needs to be submitted.

Regardless of whether this particular case will cause harm, it is worrisome that so little information about the AI's effectiveness is being provided. If they were confident in the system they would be providing metrics on how often the AI matches the grades of human reviewers. They have millions of tests to verify the AI's competence with. When the AI provides a high confidence grade, does it match the grade of a human reviewer 99% of the time? 90% of the time? Then compare that with how many times the two human reviewers in that 3000 test sample mentioned agreed upon a grade (I doubt it is 100%). These would be a very easy metrics for them to gather and report.

But they aren't reporting that kind of data. Which means they are either intentionally hiding it, or aren't even competent enough to be evaluating the AI appropriately. That is a very bad sign.

Musk's first full driving mode prediction I could find was that Tesla would have that capability in 2017. So if Elon is predicting human level AI in 2025, add at least 10 years or so.

You don't need someone who understands AI, you just need someone who understands how science works at the frontier of human knowledge. Because the correct answer is we simply don't know. We never know what leading researchers in any field will discover or accomplish over the next few years. Unless top AI researchers have already reached human level intelligence but haven't released that info to the public, they aren't going to have any idea when we will reach that milestone.

If we understood precisely how the inner workings of the human brain enable creativity, problem solving, memory, etc. then this would just be an engineering problem, and we would likely have a good idea how close we are to reaching human level artificial intelligence. But right now we have no idea when these capabilities will emerge from AI research.

This isn't true at all. A college education still provides a positive ROI for most people who go to college. But it greatly depends on your major and aptitude. Research by The Foundation for Research on Equal Opportunity shows the median Engineering graduate fully recovers the costs of college within 4 years of graduation. However the median Visual Arts and Music graduate never recovers the cost of college. So if you are going to go to college, you either need to choose and complete a lucrative major and/or truly excel in your chosen major.

It may not even matter what the terms are. The first step is to determine if someone is in breach of the terms of service. It appears clear that anyone downloading a video is in breach of that. The next step is to determine if they have a valid defense. That is where fair use doctrine comes in.

You are free to violate any terms of service as long as you can successfully claim a fair use defense. Think of it as killing someone in self defense. There are laws against killing people, just like violating a site's terms of service, but if it is in self-defense then it isn't criminal. Same goes for downloading a YouTube video. You are free to break the terms of service as long as you can claim fair use.

Whether or not AI training can claim fair use has not been adequately tested in the courts (IMHO). I personally think it should be considered fair use, but that's for the courts to decide.

Most technological advances had the purpose of replacing workers. 500 years ago you needed 80-90 farmers for every 100 people in your community. 100 years ago you needed 30 farmers, and today you need 4. That profession lost 95% of its workforce over the past few centuries because of technological advances.

Today the US worker is about 3x as productive as we were 100 years ago. Along with more people entering the workforce (mostly women) this productivity boost allows us to consume about 6x as many goods and services as the average American did 100 years ago. It also means that for every 100 jobs in 1924, 70 of them were displaced by technological advancement. In their place came 170 new jobs that didn't exist 100 years ago.

AI is likely one of the significant drivers of productivity improvements in the next century. It wouldn't be surprising to see 60-80% of all today's jobs being lost over the next century. It's what happened last century. The only questions are whether new jobs will be created, and whether we will start consuming 6x more goods and services 100 years from now than we do today. The answer to those questions will determine how many people are still employed in the future.

If we didn't start consuming more last century, worker participation rate would be 20% today. The same is very likely true in the upcoming century. Either the working class in 2124 will live lives comparable to the upper middle class today, very few people overall will be employed, or somewhere in between.

I hadn't thought of that before, but you're right that the IT equivalent of an MBA prioritizing profit above all else is an engineer who is willing to deliver substandard work because of a deadline or lack of resources. Both are put in a no-win situation, but have the choice to stand by their ethics at the expense of their career. And almost no one stands by their ethics (myself included).

In your lifetime, algorithms have also become 1000s of times faster and more efficient. It just isn't as noticeable as tracking TFLOPS/watt improvements. This research paper claims the impact of algorithmic improvements in computer vision was roughly the same as the impact in improved hardware efficiency over the past decade or so. I've seen similar claims in other papers related to chess programs.

It's very likely you will see a 10x increase in LLM training performance from algorithmic improvements over the next few years. The paper above saw algorithmic efficiency doubling every 9 months in computer vision models, which leads to a 10x improvement in 2.5 years.

The US has convicted and imprisoned foreign spies before. Rudolf Abel is one example I found in a quick Google search. He was convicted for a 30 year sentence, but was released after four years in a prisoner exchange for a downed U-2 pilot.

Catching, prosecuting, and convicting spies is not that uncommon.

Assange had the obligation to protect US classified information if he wanted to live in a country with an extradition treaty with the US. And both the UK and Sweden have extradition treaties with the US. A country's laws don't just apply to citizens, unless the country makes that stipulation in the law. If it is against US law to release classified information, that applies to all people anywhere in the world. I don't know the specifics of Russian or Chinese law, but I assume it is against their laws for a US citizen to release classified Russian or Chinese intelligence as well.

Whether or not I have an obligation to follow their laws and protect Chinese classified information is determined by the treaties the US has with China, and whether or not I ever plan to travel to China or any country with Chinese extradition treaties which cover that Chinese crime. I don't know what those laws are, but I sure would check with a lawyer before releasing classified Chinese information. If there are treaties in place which would make this activity illegal, I would be obligated to follow those foreign laws or risk extradition.

Amazon isn't buying Anthropic. They have invested $4 billion at a reported $18.4 billion valuation, meaning Amazon owns a little over 20% of the company. Google invested $2 billion, and there has been another $1.3 billion in funding over the past year from other sources. So while Amazon is a large stakeholder at Anthropic, they won't own the company. Google is investing billions too, and wouldn't be doing that if the company's future is to just be Amazon's lapdog.

The tech bubble took 5 years to burst, after starting with the NASDAQ's 40% growth in 1995. It went along with an average of 29% growth for 3 years until 1998 when it had two 50%+ years of growth. But even after the bubble burst the NASDAQ only dropped down to early 1997 levels, which was two years into the steep incline in tech related stocks.

If the AI boom is anything like the dotcom boom, then we have another 3-4 years of insane growth ahead of us. And after the burst, tech valuations would still be higher then than they are today.

Of the largest internet companies today, Alphabet, Amazon, and Netflix were all 90's startups. Only Meta was founded after the bubble burst. If that experience also holds true for the AI boom, then most of the companies who will be the big AI winners in 20 years already exist today and are benefiting from this extreme level of AI related funding. This is why investors aren't waiting around for another 5-10 years for the dust to settle.

Average commission as a percentage of sales across all industries is 20-30%. This is not just a single salesperson receiving this commission, you could have account executives, sales development reps, account managers, sales directors, etc. all getting a piece of that commission percentage. So life insurance apparently pays fairly low from a commission perspective.

It appears the reason for this is life insurance agents receive a very high commission rate in the first year (30-90%) but then get only 3-10% after that. Since a life insurance policy can last 30+ years, the overall commission percentage is pretty low over the lifetime of the policy.

They appear to be studying it to identify if life insurance premiums are being overpriced for some subsets of the population. The research abstract identifies young policyholders in low income areas as a group who is being charged too much when compared to their expected future payouts. This is useful information for both insurance companies and government regulators.