I think this is true only if you are comparing the LCOE of natural gas to solar *with storage* in the US. A plain solar farm without storage is going to be cheaper. We really should look at both with and without storage, because they're both valid comparisons for different purposes, although PV + storage is probably the best for an apples-to-apples comparison.

The cost of solar has come down year after year for the last thirty years, to the point that *internationally*, at least, the LCOE for solar PV plus storage is now just a little bit less than the LCOE for natural gas, and is expected to become *dramatically* cheaper by the end of this decade, according to IEA. Even if they are calculating somewhat optimistically, if solar costs continue to drop it's only a matter or time before solar PV plus storage becomes cheaper than natural gas, even in the US with its cheaper gas.

The wild card here is any political actions taken to change the direction this situation is going. Internationally, low PV prices are driven by cheap Chinese suppliers, and it doesn't seem likely we'll see large scale US domestic solar production in the next five years. In the meantime we have an unstable situation with respect to tariffs on PV components and lithium for batteries. Until sufficient domestic sources of lithium come on line, uncertainty about tariffs will create financial problems for US manufacturers and projects.

Poor form. They and their families might object and make trouble for you. What you do is convince them to throw *themselves* into the furnace, in fact to feel rather smug about it up until the actual burning happens.

The molecular basis for epigenetics was discovered in the 1980s and for the past thirty years or so non-genome-based inheritance has been a pretty hot scientific topic.

This only seems surprising because for most of us our biology education ends with 1953, when the structure of DNA was discovered. We didn't learn about epigenetics (1980s) or retroviruses (1970s) or horizontal gene transfer (discovered in the 20s but importance was only realized in the 90s).. The biological world is full of weird, mind-blowing stuff most people never heard of.

The ethics module is largely missing in humans too.

Philosophical ethics and ethical behavior are only loosely related -- rather like narrative memory and procedural memory they're two different things. People don't ponder what their subscribed philosophy says is right before they act, they do what feels ethically comfortable to them. In my experience ethical principles come into play *after* the fact, to rationalize a choice made without a priori reference to those principles.

Nobody paid for it. At least nobody was charged directly. It's customary to cite grants funding research in any resulting papers, and in the case of *federal* grants it's *mandatory*. The authors simply thank the Cornell Center for Material Research for use of their rheometer and SEM. The equipment in the CCMR was purchased with NSF money, so I guess public money spent for whatever the wear-and-tear is for taking some rheometer measurements and SEM images.

If you look at the paper, it's not *really* an investigation into cutting onions. To do that you could just line people up to cut onions and have them report on the experience. It's really more about how to use experimental fluid dynamics to investigate a problem. Scientists noodle about such toy problems all the time. I had a professor back in the 80s who worked on the problem of the equation of motion of a spinning coin on a tabletop. Nobody paid him to do that, unless you count his MIT salary. The solution was eventually found by an Oxford researcher and published in a letter in Nature in 2000-- again this appears to be un-funded research. And as trivial (practically, not mathematically) as the spinning coin problem appears to be, the paper has subsequently been cited by a fair number of physics research papers, so *practically trivial* isn't the same as scientifically pointless.

So you can unclutch your pearls now. The scientists didn't pick your pocket to do a stupid experiment.

No, it is a useful observation because it gives us something to look into. Just because you don't know how to negate a proposition off the top of your head doesn't mean it can't be done.

It seems quite plausible that if a LLM generates a response of a certain type, it's because it has seen that response in its training data. Otherwise you're positing a hypothetical emergent behavior, which of course is possible, but if anything that's a much harder proposition to negate if it's negatable at all with any certainty.

Phase 1 trials can consist of as few as twenty subjects, although fifty is more common. It's basically there to make sure you don't kill or injure hundreds of test subjects in the larger trials with the dosage and protocol you intend to use. Only about a quarter of drugs which are rejected as unsafe are rejected at Phase 1. There have been drug trials halted because they killed dozens of people *in phase 3*.

Ultimately just 10% of drugs that pass Phase 1 get approved. The rest are either too dangerous or too ineffective.

The compassionate use exception carved out by the Federal Try Act is specifically targeted at people who have little or nothing to lose. It kills a few people who were going to die anyway, but in a few very rare cases it may have saved a few individuals. So you can argue from a utilitarian standpoint that it's ethical to make this exception for terminally ill patients because the exception does more good than harm. But extending the exception to people who aren't terminally ill will do more harm than good, just going by the number of treatments that prove unsafe *after* Phase 1.

That's why I'm not investing in AI stocks. I don't believe the pitch the companies are making to investors. That doesn't mean that LLMs aren't a tremendous technological achievement that could be very useful.

Whether it's a net good for mankind, I'm skeptical. But as long as it exists, use it cautiously and wisely.

What I've been saying all along is that the biggest problem with the technology isn't going to be the technology per se. It's going to be the people who use it being lazy, credulous, and ignorant of the technology's limitations.

The bottom line is that as it stands LLM isn't any good for what these bozos are using it for: saving labor creating a brief. You still have to do the legal research and feed it the relevant cases, instructing it not to cite any other cases, then check its characterization of that case law for correctness. In other words, you still have to do all the hard work, so it's hardly worth using if all you are interested in is getting an acceptable brief quickly.

But if you *have* done all that work, it's quite safe to use AI to improve your brief, for example tightening up your prose. You can use it to brainstorm arguments. You can use it to check your brief for obvious counter-arguments you missed. There's absolutely nothing wrong with lawyers *who know what they're doing* using AI to improve their work. It just can't *do* their work for them.

I thought of provisioning as an obstacle, but it's not necessarily a difficult one for an actor with "national means" to overcome. Most likely none of the low-cost service providers who provide service for things like GPS trackers would bat an eye if you had a shell company set up a few hundred or even a thousand devices. Just tell them you're doing vehicle tracking or something like that. China's MSS could even set up a bogus cell tower near their target sites, stringray style. The main target of such an attack wouldn't be the device itself, or even the facility, but the grid, possibly in conjunction with another type of attack to impede recovery.

That said, I don't think that's what these things are there for. I think it's just a case of the vendor finding it convenient to sell the same SKU to all its customers. But you're right, this is all just a conjecture. The possibility merits investigation, even if it's not likely.

While steel is one of the most recyclable materials there is, concrete isn't that great. Most concrete "recycling" would be more accurately characterized as "downcycling" -- using old crushed concrete for road beds for example. It is possible to reuse old crushed concrete as aggregate in new concrete, which does technically count as "recycling", but it doesn't address the single most significant environmental impact of concrete, the carbon emissions of cement production.

Anytime you propose replacing X with Y, both X and Y will have environmental impacts; you just think that the impacts of Y aren't as bad. This is a new material, and we can't quantify what its impacts will be, but they'll be there. Service life plays a role here too and we don't know what that is. If a beam made out of this stuff lasts as long as a steel beam, that'll almost certainly be a carbon footprint win, but if you have to replace it ten times it might not be.

Sure, industrial infrastructure commonly has remote monitoring and administration capabilities. Grid storage systems for example have to monitor battery and inverter health. EV chargers can often communicate to determine the most economical charge times.

In this case the devices included cellular radios that could bypass site network administrators completely. Again this is not necessarily nefarious, because there are legitimate use cases for this.

But if the capability is there, it should be shipped totally disabled unless the customer requests otherwise. A competent engineer could determine whether that's the case. Even if the cellular modem isn't provisioned with a data carrier, if it is operational an agent with national means could communicate with it stingray style. If it is physically possible for an unauthorized person with knowledge of the system to gain control of this equipment, somebody has some explaining to do.

You shouldn't have to resort to cleverness and effort to find this out. AI training bots should log the URLs they ingest, and anyone should be able to query those logs to see if their site has been used to train the model. Given the vast sums companies are spending on training their models, the marginal effort of maintaining a public log wouldn't add any significant cost, other than the litigation costs they'll face when sites discover their TOSs have been violated.

The whole "move fast break things" ethos counts on creating a new status quo faster than regulatory bodies can respond. Tech startups rely on creating a fait accompli before government even notices the problem, but if they fail in that a well-funded company has recourse to deceptive PR, then lobbying, then lawyers to gum up the works. In AI, companies are already racing each other as fast as obscene gobs of money can propel them forward; it wouldn't take much to slow down any public regulatory response so that it will have to be mounted against the winner of that race, a company that will be in a much more commanding position to fight back.

In the meantime your hypothetical whistleblowing engineer probably is compensated to a substantial degree with stock options, and his continued employment prospects after ratting out his company are bleak in an industry where everyone is doing the same thing.

I'm not saying its impossible, but I'm a lot more pessimistic than you about it being *easy*. I suspect that enabling private actors to move against AI companies would be a lot faster. Since damages are hard to prove or quantify, simply creating statutory damages would allow intellectual property owners to take the initiative against infringing AI systems. It would help if there were transparency regulations which aided IP owners in detecting unauthorized training. Of course the downside is the volume of litigation that would follow.

How does the fact that 85% of the chips in question are manufactured in Taiwan affect this? It seems like if you want to keep AI chips out of Chinese (meaning PRC) hands, you're going to have to be ready to put boots on the ground in Taiwan.