Finnish National Bureau of Investigation lead investigator Sami Liimatainen says he wasn't contacted by WaPo, which published a story earlier today claiming that an emerging consensus among U.S. and European security services holds that recent Baltic seabed cable damage was accidental.

https://yle.fi/a/74-20137924

And the Finns should have the most informatoin

Several things have changed that have made the US less competitive in science.

We have the best universities in STEM. We used to allow people with advanced degrees to pretty easily immigrate. We've made that harder by shrinking the pool of various visa types -- mostly as part of an anti-immigrant fever. And if the color of your skin is not white or you speak with an accent, there are lots of places you don't want to live.

A lot of our politicians reject science. Something like 1/3rd of congress is on record as climate change denial. Many reject the premise of evolution. When I was growing up being a rocket scientist or an atomic scientist was something people really looked up to. Even working in plastic was high prestige as we know from The Graduate. Politicians of course communicate their attitude to their constituents and are also a reflection of those views. Hence the life of a scientist is not as pleasant. After the soviets beat the US to space and after we ended the war in Japan by building a bomb, there was a huge rush based on national security to have more scientists.

Relative to other fields, science doesn't pay as well and the job security of a scientist has diminished.

Here's some facts: Training an AI model uses a lot of compute and a lot of electricity. It can be done on thousands of GPUs in a distributed fashion. You can stop and restart it. You can run it when electricity is cheapest. There are really two costs ignoring the programming costs, 1) the capital cost of the GPUs, or the rental of those GPUs on the cloud and 2) the electricity. Where I live, electricity is very cheap at night and can be very expensive at peak time. If you skipped training at the times the natural gas guy wants to sell you electricity, you'd only miss a small fraction of the day.

For inference costs, using the trained model, you can have various data centers around the world. At any hour there will be some that are not in a peak period. The amount of data sent and received from inferencing is tiny.

In short, the guy wants to sell gas and peaker plants and has no idea what he's talking about.

If you go to the PNAS study this is all based on, they make a distinction between smart charging and I'll call it dumb charging. They really only analyze current dumb charging. They assume that the cars can't do anything to smooth the load as a simplifying assumption. The authors of the study say this is a first step, and then proceed to ignore that this is a simplistic assumption. They point out that most cars start charging when they are plugged in. I have my car set up to charge so it's ready at 8am so it may be a bit warmer in the winter before I drive it. That's just a tiny amount smarter. My car is connected to the internet, it could determine when electricity was in lowest demand and charge them -- perhaps the utility could offer an appealing slightly lower time of use charge incentive. I live in NY and rates are tiny between 12 and 8am, which is when I charge. The grid is massively underutilized at all times during that period in NY. California may have a different time that it's underutilized. Without much infrastructure charges at all but some cooperation from the cars the grid utilization can be much smoother than it is now and most grids are set to deliver in the hottest of heat waves in the middle of the day for A/C and have lots of capacity at other times.

The article to make their job easier assumes that charging patterns will not change.

By the way the article observes that electricity costs are likely to go down as a result of EVs because they will shift all electric use to cheaper production which will help even non-EV users.

The Luddites were not low skilled workers but actually were among the most highly skilled artisans. Somehow no one ever complains or perhaps no one ever rebels when low priced workers are displaced, but when highly paid ones are, watch out. Some of what AI can do may replace highly paid workers like radiologists.

Perhaps this means we need more ML people and fewer physicists? Both are STEM fields. I'll bet that to make a good ML predictor of the weather you need to design a NN topology that can efficiently compute some of what the physics needs. Weather prediction has gotten better, but it's a field that can always get even better. When I was a kid it was hard to know whether or not to wear a rain jacket to school that day. Now I'd like to plan whether I should see friends in the city on Saturday or Sunday a week from now.

There will always be a value to better weather prediction and it will need STEM skills because even with ML its going to require massive computation and a better more efficient model is always possible. It just might need a different STEM field.

I agree. What are large language models really good at? Spewing well written fact free copy. I've used it occasionally to re-write text for me, something an English major might be better at than I am, where I provide the STEM content. LLMs are getting better at answering simple math problems, but where they really shine is well composed, but possibly nonsensical prose. What else is AI good at? Anything that's pattern recognition that can be trained on a very large corpus. Some parts of health care fit that bill. It is possible that in the relatively near future, they will be able to do reasoning, but at the moment that seems the final frontier for AI. It seems the suggestion is exactly backwards?

Maybe he's waiting until he empties all the bitcoin wallets in the world before publishing? I don't believe it.

W-2's go to the Feds already, companies are required to send them over. Similarly all stock trades are sent to the government and banks send the interests and loan info. Remember that filing taxes is not on the honor system. There are some exceptions about income that is not reported but it's few and far between (if you make a profit on a drug trade you are required to pay taxes or they can get you the way they got Al Capone -- but that doesn't apply to most people). In Japan you get a post card and I think you can agree to pay what the Gov't thinks you owe. There are two groups that want to make it hard to file. Intuit, HR Block have an obvious incentive. But there are politicians that want to shrink the government and one prime way to do that is to make it difficult to file forms. In the 60's there were many income tax brackets and taxes were better for the middle-class and worse for the very rich. It wasn't possible to say we should tax the middle class more, but it was possible to say we should simplify tax filing. I never found the multiple brackets took more than a few minutes to figure out, but of course a computer makes that whole argument nonsensical. Hence we can't allow free computer filings because then there might not be a reason to make the tax code fairer. Why am I not surprised that it's the Republican's who are objecting to this.

I have a Tesla Mod 3 AWD. I did not buy the performance model, but my car has engines of the same design, though Tesla claims they sort engines so maybe the ones that go on P have slightly better specs. The P can accelerate from 0-60 one second faster than my car. A year or so after I bought the car, Tesla offered an over the air update to give me roughly half the difference, for a one time fee. The fee is considerably less than half the cost difference to a P. The P also has better brake pads and some other tweaks.

The only difference between what Tesla did and what Mercedes is talking about is that with Tesla it's a one time fee, where with Mercedes you have a choice. Frankly what Mercedes is doing is probably dumb. I almost never floor the accelerator on the Tesla. Basically I do it when a friend who's never been in a Tesla wants to see what it can do. Aside from the fun of feeling that much instant acceleration there's no need for it in day to day driving, and in fact that kind of acceleration feels a bit scary, so I wouldn't want more. I could imagine paying $60 for one month just to feel what it's like and then stopping it. But I'll bet a number of people have paid the one time fee.

I haven't seen published info on the structure of Bard. Google's Deep Mind came up with transformers, which are the basis for ChatGPT. They have several designs which go beyond the base transformers and I suspect they used some of those, particularly Retro. Retro can be more accurate about facts but may be weaker in other aspects. So some testers will find it better some worse. If you test it on something you were impressed with with GPT it may be weaker. For me it got dates and other things exactly right where GPT only got things in the ballpark one over, but wasn't in the next state. I think over time Google will improve it and may well surpass the GPT family.

This article talks about converting ethanol to jet fuel: https://www.greencarcongress.c... In theory, ethanol can be produced by growing crops and fermentation. The reality is that our farming is so dependent on fuel that there's little if any carbon savings this way. Perhaps with EV based tractors some of this will change in the future.

I don't know if this is what JetBlue intends.

This isn't a test of whether you can do these tasks, but can you do them quickly. It's timed. If you don't do long division every day you may be rusty. I do advanced mathematics every day, but I don't divide numbers. I can tell you the algorithm, and prove that it works (the proof may be something a 6th grader would struggle with). But I'm probably not as fast as some 6th graders. I looked over the arithmetic test. You have to do about a problem a minute. I'm certain that I can score 100% but there's some doubt in my mind about whether I do so in the time. It's also unpleasant to do these sorts of tasks. It's not a fun test to take and I'm not interested in finding out if I'd have a timing problem. Unless it really matters to you -- and it might not for the MPs -- you may get sloppy.

Far and away the best keyboard I've ever used was the one attached to an IBM 3270. The switch was based on a beam spring and Hall effect. This should not be confused with the IBM PC buckling spring design which was a cheap version of this. A kickstarter project which was trying to imitate it just failed. It would be lovely if someone else could build a keyboard with the same force activation pattern as the beam spring.

The IBM keyboard was incredibly robust and lasted forever. In the current world of cheap throw away keyboards, I'd be content if the keyboard was equivalent to the IBM one, but didn't last more than a few years.

Strassen showed that Matrix Multiply could be done in n^3 time. Coppersmith-Winograd using slightly different techniques showed that you could do even better than Strassen. DeepMind showed that they could beat Strassen's algorithm but not CW (and there've been improvements to CW). A human went back and beat the DeepMind algorithm using Strassen like techniques. But why bother when you can use CW techniques?

BTW this is all theoretical and not practical. All of these techniques are numerically unstable so you wouldn't use them for reals, but only for matrix multiply over finite fields and they only really pay for humongous matrices.