This seems dubious at multiple levels.

Solar panels: The roof of a trailer is about 450 square feet. In the northeastern U.S., you would average only 3.5 hours of full sun, so you'd get only a little over 13 kW per day.

Tesla semis are pretty efficient, and they use about 1.7 kWh per mile. So in an entire day, covering the entire roof of a trailer with solar panels would add a whopping 7 miles of range, or 15 minutes of extra driving — the equivalent of plugging into a Tesla Megacharger for maybe 30 seconds or so.

Let's optimistically assume that the vehicle can carry 48,000 pounds. If those panels occupy the full roof area, then at about 3 pounds of weight per square foot, those solar panels would weigh 1500 pounds, or about 3% of your cargo, all to reduce your fuel usage by as little as 1% if you're doing long haul at 65 MPH. And that weight number may be wildly optimistic. Trailers like that aren't designed to have weight on the roof, and would require additional structure to hold that extra weight. The real losses could be significantly higher. Unless you're driving less than a couple of hundred miles in a day, the solar panels won't break even. And if you're driving less than a couple of hundred miles per day, there's no reason you can't go electric.

Battery and motor on the trailer: I would expect most trucks to be used primarily for either short-haul or long-haul purposes, not both. If you're doing long-haul, you'd probably be better off with an actual hybrid tractor so that you get the benefit no matter whose trailer you're hauling. If you're doing short-haul, there's likely no reason not to go full electric.

I just don't get it.

Facebook doesn't have a separate C-suite (CEO, CFO, etc.) from its parent company. Waymo does. So while Waymo is considered part of Alphabet because it is a majority shareholder, you're kidding yourself if you think it is at all like Meta and Facebook. There may not be a hard line between them, but there's a definite line.

Tell me you don't know how model training works without telling me you don't know how model training works.

Autonomous vehicles (probably not including Tesla) already handle first responders correctly probably 99.99% of the time. They already handle school buses correctly probably 99.99% of the time. So what remains are, by definition, edge cases that for whatever reason require additional handling beyond the basic "Is this an emergency vehicle/school bus? If so, pull over and stop" rule.

For example, one edge case is figuring out how to clear a path for an emergency vehicle when there is no obvious place to pull over because of other cars stopped nearby. Sometimes the correct answer is to actually drive in the direction the emergency vehicle is going until you find a spot to pull over and get out of its way. This isn't intuitively obvious, and a lot of human drivers will struggle with it as well.

For another example, at least one case of Waymo vehicles illegally passing a school bus was caused by a remote operator not noticing that the vehicle had flagged the presence of a school bus and telling the car to proceed anyway. Sometimes, having a human in the loop actually ends up making things worse. :-)

At that point, what's the point of them being autonomous? At some point, you have to cut them loose and see what mistakes they make, because it is precisely through detecting those mistakes that you figure out what edge cases remain inadequately handled in the model. And understanding how the vehicle attempts to extricate itself from problem situations is critically important in figuring out what additional training needs to be added to prevent similar occurrences in the future.

So basically, your approach likely leads to a future where the models never learn to handle emergency vehicles, because safety drivers keep having to intervene before they can gather adequate data. That approach just doesn't work.

I don't recall the actual driving part of the driver test having a part where you drive onto a street that's barely wide enough for one vehicle to pass, let out a passenger, and then have an ambulance suddenly approach from the other direction while you're trying to turn around.

To get to the point of having a license, you have to answer a written question that proves you know to yield to emergency vehicles, prove that you can stay in lanes, stop for stop signs and maybe pedestrians, handle traffic lights correctly, and possibly parallel park, depending on where you took the exam. Autonomous vehicles could do those things reliably 15+ years ago.

In other words, you're greatly overestimating the competence of the average human driver.

The military implications are obvious. Think Ukraine. If you suspect the enemy is trying to infiltrate on a dark night along several kilometers of frontline, you light up the scene while launching a bunch of low-cost FPV drones, and those infiltrators are about to have a bad day.

You *can* spot infiltrators in the dark with IR cameras, but it requires much more expensive drones and isn't usually as effective, hence the preference for night operations. Plus, there's IR camouflage, with varying degrees of success. But it usually makes you stand out like a sore thumb under illumination (you're basically wearing a tent).

This is a small prototype, not a full-scale mirror.

Also, ignoring that, pretending light brighter than the full moon is useless is... silly? For generating power, sure, but for illumination, it absolutely is not.

Citation needed.

Google doesn't even have self-driving cars. Maybe you're thinking about Waymo (which is part of Alphabet, not Google).

Regardless, no, to the best of my understanding, they cannot be driven remotely at all, at least by any normal person's definition of the word "drive". When intervention is required, the remote operators get a dump of camera images to review, and then they draw a proposed path on a map. The car then tries to follow it, and aborts if doing so would result in hitting anything. This may have to be done more than once to get it out of the problem situation. When the vehicle says that it is comfortable proceeding on its own, the remote operator tells it to go ahead, and it takes over path planning again.

At no point is any remote operator in direct control over the vehicle. All they can do is propose an alternative path when the vehicle's path planner gets stuck trying to figure out how to safely extricate itself from some situation. At all times, the vehicle's software is the driver. The remote operator is just hinting that it should go to the left of safety cone A, to the right of cone B, etc. (or whatever the situation happens to be). This is why it takes so long to extricate a stuck car. If there were an actual remote driver that could take real-time control, it would take just a few seconds.

From what I've read, when a Waymo car sees emergency lights, it stops driving and gets out of the way. I do see one (presumably) recent video where a Waymo stopped in a place that actually delayed an ambulance from getting past it on a narrow street, so unless that's an old video, I'm guessing there's still a bit more tweaking required in terms of recognizing whether the right choice is to stop or to move out of the way. I'd imagine someone is already working on making sure that particular edge case doesn't happen again.

What I'm not seeing is evidence of some widespread problem with autonomous vehicles in general. There's an edge case here or an edge case there where something didn't work as expected. And they'll complain about it, and the AV company in question will figure out why the car did the wrong thing, update their training sets, and that specific scenario won't happen again.

(This, of course, ignores Tesla, because the emergency vehicle drivers can't tell if the vehicle is being driven by the car or by a human, making any sort of reporting problematic at best.)

So realistically, I suspect that the answer to a vague demand from a government agency demanding to know what AV companies will do to prevent bad interactions with emergency vehicles will always be "exactly what we're already doing", because apart from coming up with new simulated situations to test (which they're always doing), there's really nothing they can do to prevent the car from behaving the wrong way in some vague unspecified future situation that nobody has thought of yet. And the answer to what they're doing to prevent a specific situation will usually be "We've already updated our training sets and that won't happen again."

To that end, I'm really not sure what they're trying to accomplish with sending a letter like that. Seems more like political posturing than any actual attempt at solving a problem. *shrugs*

There are a couple of alternatives. The problem is that there are only a couple.

Could it please implicate every member of the federal government in whichever party I choose, so that I can throw the next election to whoever promises to fix the inefficient traffic lights near my house? Thanks.

I'm not convinced those people exist. Rather, the people with 5-figure income can afford to take a car, so why would they willingly walk for ten minutes to the nearest bus stop and then spend an hour on two different modes of public transit to go somewhere that takes fifteen minutes by car?

I'm also not convinced that very many people exist who would pay $5k a year to go through a different entrance to the airport just so that they don't interact with anybody who isn't in the six-figure club.

That said, I'll admit that having someone drive me to the gate would be nice, assuming it is efficient. Unfortunately, realistically, it won't be unless they build the private terminal on top of or under the main terminal or otherwise provide a cart path that doesn't interact with pedestrian traffic in the main terminal).

I can only assume that the food would be free and readily available like in an Admiral's club (though I suppose you're paying for it, so not truly free), so having something like this at connecting airports would be great, but far less useful at either endpoint.

If they include prepaid valet service for your car as part of the cost so you can just drive up and they'll whisk you to the gate straight from your car with only a brief stop at the security checkpoint, it could have some value — not enough value to spend an extra fifteen to thirty minutes and fly out of SFO, mind you, but some value.

I can't imagine it being worth $5k a year, though, so what I'd expect is that honeymooners will do it as a once-in-a-lifetime way to be treated like they're important, and all the people with money will continue to do their usual.

This is akin to altering the English alphabet such that quantity and shape of symbols in the alphabet is different before teaching students English

No, it's not. It's akin to having secret numbers (that correlate neither with the spelling or meaning of the word) associated with each word, and the only way you can learn the secret numbers is if e.g. someone says "The secret number for Strawberry is '3'."

There are no other "clues" available to the LLM. It cannot see the property in question (spelling). It is a hidden "magic number" for all effects and purposes.

as there are daily videos of people doing this exact shit?

Literally go to a non-tiny AI right now and try it. Your pick. Try all the major ones - ChatGPT, Claude, Gemini, Grok, your pick. Then share me a link to the failure case you're claiming you'll get. Don't worry, I'll wait.

(Note: tiny: the "Google Answers" AI is a truly minuscule model, the sort of thing you could run on a cellphone with ample capacity to spare, and is only representative of minuscule models)

how about the guy that put 3 AIs together and told them to count to 100?

Sounds like what you actually have a complaint about is the (non-AI) voice interface they're linked to, because that's what screwed them. That entire premise doesn't work at all if you remove it and just do the test in text. Again: try it yourself.

generally they fail at the balanced parenthesis problem

LLMs are way better than humans at ensuring that code parentheses are balanced, e.g. when writing code; LLM code is much more likely to compile / run on the first time than human code. There are literal parenthesis-counting circuits that emerge. That they're not perfect at it doesn't make them not be better than us at it. Humans tend to start to lose track after 3-4 levels of parentheses. LLMs have long since significantly outperformed us in nested grammatical tasks, so long as the level of training to the human and AI are at least remotely comparable.

Here is, internally, how models handle counting, despite not being able to see what they look at.

Also, your article is deeply dated, in terms of (A) models, (B) underlying articles, and (C) descriptions of how things work. Strange for something written in 2023; it reads like something written in 2020-2021. Just to give an example of what I mean, let's just grab a few paragraphs from the middle at random:

It operates in three basic stages. First, it takes the sequence of tokens that corresponds to the text so far, and finds an embedding (i.e. an array of numbers) that represents these. Then it operates on this embedding—in a “standard neural net way”, with values “rippling through” successive layers in a network—to produce a new embedding (i.e. a new array of numbers). It then takes the last part of this array and generates from it an array of about 50,000 values that turn into probabilities for different possible next tokens. (And, yes, it so happens that there are about the same number of tokens used as there are common words in English, though only about 3000 of the tokens are whole words, and the rest are fragments.)

A critical point is that every part of this pipeline is implemented by a neural network, whose weights are determined by end-to-end training of the network. In other words, in effect nothing except the overall architecture is “explicitly engineered”; everything is just “learned” from training data.

There are, however, plenty of details in the way the architecture is set up—reflecting all sorts of experience and neural net lore. And—even though this is definitely going into the weeds—I think it’s useful to talk about some of those details, not least to get a sense of just what goes into building something like ChatGPT.

First comes the embedding module. Here’s a schematic Wolfram Language representation for it for GPT-2:

The input is a vector of n tokens (represented as in the previous section by integers from 1 to about 50,000). Each of these tokens is converted (by a single-layer neural net) into an embedding vector (of length 768 for GPT-2 and 12,288 for ChatGPT’s GPT-3). Meanwhile, there’s a “secondary pathway” that takes the sequence of (integer) positions for the tokens, and from these integers creates another embedding vector. And finally the embedding vectors from the token value and the token position are added together—to produce the final sequence of embedding vectors from the embedding module.

Why does one just add the token-value and token-position embedding vectors together? I don’t think there’s any particular science to this. It’s just that various different things have been tried, and this is one that seems to work. And it’s part of the lore of neural nets that—in some sense—so long as the setup one has is “roughly right” it’s usually possible to home in on details just by doing sufficient training, without ever really needing to “understand at an engineering level” quite how the neural net has ended up configuring itself.

Unless this is awkwardly worded, it does not "find an embedding (i.e. an array of numbers) that represents [the input sequence of tokens]". Each token has its own embedding. And they don't need to be "found", they're just a lookup.

Values don't go “rippling through successive layers" in a LLM. The core of a LLM is dozens to hundreds of individual DNNs (feed-forward networks) separated by attention blocks and add+norm. He entirely writes this out of the picture (at this point - later he reinserts them but does so incorrectly), but they are utterly critical to a LLM's ability to function.

"A critical point is that every part of this pipeline is implemented by a neural network" - No, it isn't.

"nothing except the overall architecture is “explicitly engineered”; everything is just “learned” from training data" - No, it isn't. Creating your vocabulary for example is done with an algo like BPE or unigram. Not neural network based.

"Here’s a schematic Wolfram Language representation for it for GPT-2 .... 12,288 for ChatGPT's GPT-3" - I hope you understand how ancient these architectures are by now.

"Meanwhile, there’s a “secondary pathway” that takes the sequence of (integer) positions for the tokens..." - to the point that they don't even have RoPE yet.

"Why does one just add the token-value and token-position embedding vectors together? I don’t think there’s any particular science to this." - Uh, yes, there very much is. This isn't some sort of cargo-cult nonsense, it's a basic property of latent spaces. Information in a latent space (in most latent spaces) is encoded by directionality. You can merge multiple concepts into a latent space or adjust the properties of a given latent by summing them together, thus creating a vector that incorporates both directions.

Anyway, that's enough to show what I think of the quality of your reference, esp. in a modern setting. :P Moving on:

It's related to the clock problem

Multimodal input isn't as advanced as textual input, to be sure, but let's not pretend that the human brain isn't also readily tricked in various kinds of vision problems.

If you ask a neural network to recognize "A" then all As will be tokenized the same way

Sorry, that is, again, not how it works. Tokenization happens before the model ever touches the input. It's literally the first step. It has no say over how it's done. It can't say "tokenize this differently". Even the tokens themselves don't make it to the LLM; they only point to latent positions.

Actually, no, it doesn't. "S T R A W B E R R Y" can be tokenized e.g. as

"[start]", "S", " ", "T", " ", "R" ....

But more common it may look something like "[start]S ", "T R", " A W", " B ", "E R R", " Y[end]" or whatnot.

And anyway, anyone who's used a (non-tiny) model in the past 2 years or so knows that these sorts of issues aren't really a "thing" anymore. But the key point is that this is a test on an ability to count something that the LLM can't actually see (letters).