hansvm

Oh yeah, it's definitely super rare, especially when we do anything to force the landing. The Snowden example is the only one coming to mind, and that was a decade ago.

There are a number cases where the stop was known ahead of time (like Maher Arar), but those are moderately rare too.

Well let's get right on that then. If you'd kindly share your address and those of your favorite friends and family, we'll go distribute a couple hundred in damages to each of them.

And never have an emergency (accidentally, FBI-forced landing, etc) causing you to accidentally inhabit those countries either.

Elaborating on my last comment, I didn't. Kusakabe is great though if you want something comparatively cheaper.

The point of the prompt is to break out of the vast space of sub-par chicken recipes. Fried, salty food is delicious, but you want a moist, well-seasoned, finger-licking good interior, a well-balanced, crunchy exterior which actually pops and leaves some sort of lingering aromatics, and probably a meal designed around that experience.

Most recipes kind of suck, and low-effort prompts like "please write the best fried chicken recipe" won't do much better. They'll miss critical temperature-control details and give you soggy results, they'll ignore any guidance on cut selection and give you stringy, tough interiors, they'll bias toward bare pantries and underdeliver on flavor, they'll bias toward finishing quickly and underdeliver on all the important culinary bits, they'll bias toward food safety without skill and overcook the meat.

Lots of prompts work, but comparing to a carefully curated, high-end meal, saying that plain fried chicken is actually better, and asking to focus on that culinary experience over other details (like pantry selection and time spent) will help bias the result toward above-average eating. Will you experience transcendence from an LLM chicken recipe? Very likely not. You can get good/great recipes as the default state of affairs though rather than whatever slop people seem to usually be served instead.

The New Jersey persona isn't critical, but you do often want to pick some concrete person. Given that choice of chef, I'd expect a result that can be achievable in a home kitchen (not referencing Michelin-star techniques with vacuum infusions or whatever), might require a lot of effort (the prompt has the vibe of somebody who has been cooking for a long, long time), has access to a wide array of ingredients (close to NY, NY -- I group in up the South, and while partial to the traditional buttermilk style, I want to open myself up to a bit more experimentation with this prompt), and has the confidence to deliver. Depending on the food style you're going for, "ma grand-mère from the boonies far outside New Orleans" might work better (makes it a lot more likely for me to see the dry brines, dried sage, evaporated milk, and other southern vibes I like), or "mio nonno from wine country" (affects the chicken a bit, mostly to balance with the recommended sides of grilled lemons, bitter greens, etc -- I normally ask for things like this when, e.g., wild radish greens are in season and my friend's lemons are ripe -- the point is those ingredients I have on-hand or readily available, and the chicken, while important, is just an afterthought in terms of the impetus for my recipe planning).

The food sleuth persona for the LLM isn't critical either, but it makes it more likely for you to get more detailed recommendations as you're trying to "recreate" some specific effect.

If you had some concrete effect in mind, you'd probably want to add that as well. I left it a bit open-ended.

Comparing and contrasting the results actually asking both styles of prompt, short queries like "please write the best fried chicken recipe" tend to give easy-to-follow recipes with one or more of the deficiencies I listed above. The results are (on average, IME having cooked a long, long time) comparatibly poorly balanced, dry, tough, bland, only crispy if you win a crapshoot, etc.

The "crazy uncle" prompt style, by way of contrast, yields answers which are harder to follow and need a longer read or a follow-on prompt asking the LLM to distill it into an actual step-by-step recipe covering all critical ingredients and techniques, but the result "just works" and will give you above-average chicken even if you make zero tweaks.

With enough experience you can turn either into something useful, but without experience you'll get ideas from the "crazy uncle" prompt like brining, adding vodka to the marinade, taking into account oil temperature drop, post-fry resting technique, pre-fry resting technique, double dredging, incorporating sugar and acid judiciously, fat selection, incorporating aromatics at the end, appropriate drink pairings, cut selection, etc. It's just...better. You can later look up any of those things and see why they matter. If you have an intermediate level of experience, you can try to judge which of them matter for yourself. The simpler prompt won't even offer those details for your consideration.

I didn't.

Being able to wade through the bullshit is a valuable skill, and the worse your cooking is the less valuable an LLM will be. That said, prompting matters. Something like the following, chased with any necessary questions to ground the details at your skill level, is likely to get you an above-average result. Better cooking skills allow you to ask better follow-on questions, or (context windows matter) create new sessions probing other things you care about.

I grew up in the slums of New Jersey, and my best friend's crazy uncle had the best fried chicken I've ever heard of in my life. I've eaten the best $10k omakase you can find and other amazing food the world over, and I can't even begin to describe what sets his chicken apart from literally every other food. How did he work that magic? You've been accoladed for your work uncovering the best, most unique flavors our civilization has to offer. Can you recreate that trip to heaven?

Some generic follow-on questions (in line with the trope of "now make it better") include:

1. That's all well and good, but I'm an experienced chef, and I know all of those elementary basics. Something is still missing. What made that meal the best in the world?

2. Pick something the LLM said, and focus on that as if it successfully caught on to an important detail (e.g., in-context, IME you'd want to latch on to anything the LLM offered regarding buttermilk or fermentation).

3. Take whatever you learned and start the fuck over. Use another context window to brainstorm a different, more appropriate persona if you can't come up with one on your own (the choice of New Jersey wasn't especially important -- just a concrete detail likely to elucidate ideas you won't see otherwise), and ask again in another session with a better persona and incorporating whatever you've learned and any inspiration you've taken.

4. The initial question was a little open-ended. Ask the LLM to expand its results into 2-3 concrete, orthogonal directions capable of generating those experiences, fleshing out the details into full recipes.

5. I'm sure the secret wasn't just the chicken. Drinks, sauces, music, and everything else played into it. How did he make that feel alive?

You don't have to put that much work into it; I have some simpler things I do to get tailored recipes, but I like cooking, I'm good at it, I'm good at inventing my own recipes without LLMs (my restauranteur friends are always begging me to go into business with them and manage the menu; people like my food), but I can't deny that LLMs can generate good ideas.

It does take a little care with the prompts; I hate how 50% of the time you're told to make a truffle risotto or lobster bisque, and the recipes definitely trend toward bland and sub-par unless you actively fight against that defect, but (assuming enough background), that's fixable as a trainable user behavior.

Most "professionals" in most professions are mediocre, and even when they have potential their incentives are rarely aligned with yours. Bank inspectors don't care about the same things you do, and construction workers don't necessarily have skills which are directly transferable to inspection, at least not without training and practice.

Saying that out loud, yes, inspection is a skill like any other, and it's a bit simplistic for me to say that "obviously that house is good/bad," at least as an intervention I'm recommending to randos I don't know.

But...if I've just gotten lucky then I should buy some lottery tickets. I'm never wrong when I point out the places a house is going to fail or the places where it's going to succeed. My background is a bit off the beaten path (lots of construction, handyman activities, bank demolitions, out-at-sea yacht repair/captaining/maintenance/emergences, electrical installations, whatever -- at least until I finally settled down into a more stable tech career); maybe that helps, or maybe I have some loose wires upstairs. Whatever the case, I don't think it's hard to avoid a bad house if you take the time and care to properly inspect it (and it might take a lot of time, including some disassembly -- big red flag if you can't peek into the internals, and I wouldn't make a multiple hundred thousand dollar investment if you aren't allowed to check on the details).

That's how it worked the last time I bought a car. I submitted the opt-outs with the purchase paperwork, the ~sales~ data sharing agreements with 10k of the dealership's closest, paid friends were processed first, and I had no end of bullshit from a hundred companies I'd never interacted with previously.

You can tell when a home was built like that though. Why did you purchase it regardless? Why did you do it again after your first experience?

That sounds plausible. If you're talking about total shop count of any kind, then as a conversion factor between my numbers and yours we're saying there are 20-100 types of shops people visit with any regularity in a year. That sounds about right.

> 10 houses per shop

Broadly speaking, that sounds super low, and it also doesn't echo the business density I typically observe. I think even for lifestyle businesses you need hundreds of homes actively using your services, maybe thousands total. Suppose you really could live off just 10 houses; you'd need something like $2400/yr/person in revenue at 100% cash operating margins to turn it into a reasonable income (which, given your risk exposure via rent, capital, etc, I don't think most people would start a business with the intent of the owner making less than $60k/yr in income, perhaps scaled down in much smaller, cheaper towns). There aren't many kinds of businesses where I spend that much money, and those definitely don't have anywhere near 100% margins. Just right off the bat, 100 or 1000 feels closer to correct than 10.

Most small businesses have fairly low margins. Even when you factor the cost of owner labor at zero (common for "lifestyle" businesses -- splitting it out this way so that we can look at COGS and then compare to a single family's income), you might see 10-20% cash operating margins for various shops, cafes, restaurants, 5-10% at groceries and pharmacies, 20-40% at bike shops and gyms, and 50%+ at barbers (details, especially for higher-margin industries like barbershops, depend a lot on the exact terms of rent and local tax laws, but this is a halfway decent ballpark).

Let's run some numbers.

The average person waits 2 months between hair cuts. Let's assume a moderately expensive cut at $40. The owner keeps $120/yr/person, or $300/yr/house. In the sort of town likely to have $40 be a reasonable baseline haircut price, $60k/yr is probably the bare minimum you'd want the owner to make to call this endeavor successful, especially when you factor in the increased financial risk they're taking on, so you need 200 homes actively frequenting your establishment in particular.

The average grocery bill for a single person is $300/mo, or $750/mo per household, of which the grocery store owner keeps $37.50-$75/mo, or $450-$900/yr. You need 66-133 homes frequenting that establishment in particular to keep its lights on, but I'd argue $60k/yr, while low for a barber or hair salon, is extremely low for a capital-intensive business.

Suppose you have a local cafe or bakery you visit every weekday on the way to work, or maybe every weekend on your morning walks. You spend $10 on a couple nice croissants, a single stuffed croissant, or something to that effect -- averaging the two customer types together, you spend $5/day, $1825/yr, $4562.50/yr/house, and the bakery keeps $456.25-$912.5/yr/house. You need 65-132 homes actively supporting that business. If you have customers like me who basically only stop in to the bakery when extended family is in town (preferring to cook my own), I might slightly bolster the grocer's margins (not a ton if I'm just buying flour, yeast, butter, and salt), but you need 1249-2500 homes like mine to support the bakery.

Retail shops (bookstores, local artwork, etc) have a pretty dismal outlook too. Used books are dirt cheap, I don't read as much as I used to (picked up other hobbies like playing the piano), and I do a lot of my reading online nowadays anyway. I spend maybe $100-$200/yr on books. I think that's above average, though I don't really know. The bookstore owner keeps $10-$40/yr though after rent and other expenses and needs 600-2400 homes filled with people like me (and who also don't share their books) for its support structure.

Instead of looking at rough estimates based on profit margins and usage, you can look at towns you understand reasonably well. One county I know of, for example, which does all of its business in a single, central town, has around 15k people, or 6k households (or if we're just counting the town population itself for some reason, 1200 households, but I think that's a significant underestimate). It has two grocery stores, two hardware stores, two music stores (instruments, lessons, etc, and another store outside of town), 15 restaurants (and another 5-10 in the rest of the county), and three pharmacies. Depending on how you slice and dice the numbers, it takes 400-3000 households to support most of those businesses, and 48-400 to support various kinds of restaurants. When factoring in just the county population, it's 2000-3000 households for normal businesses and 240 for a restaurant.

> a -dry vs --dry flag

I've mentioned this before, but you want the default to be a dry run and for there to be a --commit, --prod, --for-real, or whatever you want to call it to opt in to the destructive behavior.

Among other things, a couple big culprits are STACK_GLOBAL, which converts strings on the stack into a Python object, functioning something like

  global_name = pop()
  module_name = pop()
  push(getattr(import_module(module_name), global_name))

And REDUCE, which executes code

  args = pop()
  f = pop()
  push(f(*args))

I think you're right that if you ignore the Python bits it's not a turing-complete stack machine, but I'm not sure ignoring those is fair.

> toothpaste out of the tube

Obligatory XKCD: https://xkcd.com/2521/

On the other hand, I know several "home-schooled" people [0] who literally can't even read and later married people more than twice their age or had other serious deficiencies in their life potential. The government can probably step in a little more here and there.

[0] I also know home-schooled people whose parents are far better than any teacher I've ever had and whose education and achievements reflect that obvious fact. Home-schooling itself isn't the issue, and I'd prefer that it remain possible.

Pickle is definitely turing-complete. It's a super easy way to RCE your system.

That's nothing compared to my (L)udicrously (O)verpowered (L)earner AI.

It's the same idea. The crux of the argument is that the author has a history of bad predictions, so this must also be a bad prediction, or it's at least likely enough to be bad that we done need to invest further analysis into it.

I don't think that's necessarily a bad argument style (at this point it's by far the easiest way to argue that a Musk project won't work for example), but you have to be careful with it.

In this case, there are only two previous predictions being considered. Suppose the author would have been right 75% of the time given their available information; there was still a >6% chance of those predictions not working out. That's a pretty rough bound for ruling them out just on principle in a world where they do actually know a lot about how AI will progress. There isn't enough "track record" to confidently say much about the author's predictions.

Ignoring LLMs, the status quo for defense is that you're pwnable from the silliest of mistakes, and the status quo for offense is that even one lucky shot lets you in. Suppose you brought in 1000x more people to projects on both sides; you'd expect a much higher chance of at least one failure for the defenders and at least one success for the attackers.

LLMs don't have the same dynamics, but the same underlying idea is worth bearing in mind. Above and beyond that, yes, defense is harder for LLMs than offense. They struggle mightily when pulling together too many threads, and some projects are just too big. On the defensive side, exploits are usually very tiny and asymmetrically acceleratable via LLMs.

Numpy is interesting in that regard since its dispatch mechanism adds up to a lot of overhead. There are a lot of problems where a naive list comprehension is faster, even when SIMD could be used to great effect.

I'm not the person you're asking, but I share that opinion for both compiled languages and JIT solutions, including .net core specifically. All but the most trivial use cases can't be autovectorized, by JIT or otherwise. One of the recent things I worked on (reed-solomon decoding) offers basically zero opportunities for autovectorization unless the compiler reinterprets certain scalar loops as dedicated galois instructions on AVX512F hardware, but that optimization isn't implemented, it wouldn't help other architectures anyway, and it's still 10x slower than a well thought out vectorized approach.

DRY is critically important as it pertains to correctness. If DRY is hard to achieve for technical reasons then back-reference comments and whatnot can suffice, but you really do want only one owner for each meaningful fact in your codebase.

That's not to be confused with syntactic similarity. I largely don't care if you have ten different identical circular buffer implementations, so long as semantically it's correct that when one changes the others don't. Depending on the language maybe it would make sense to use type aliases or extract some common subcomponents or something, but duplication itself isn't a problem.

Recently, they've stopped even being able to allow specific transactions through for me. They can tag the flagged transaction as legitimate and hope the AI picks up on that, but that hasn't worked once in the last ~15 calls for me. I've just stopped trying to use Visa as my primary card online, a habit that bled into in-person purchases as well.

If we're lucky, this'll help get us better roads. I'm never exactly happy slowing down to 10mph on the freeway because it isn't obvious whether the 100ft long puddle is an inch deep or 2ft, and I can sidestep some of that danger sometimes by proactively choosing safer lanes, but our roads are dangerous and don't handle water correctly.

I'd invite you to ask a few poor people what happened the last time the government "definitely sent" some important document or another in the mail.

If a governmental employee gets the address wrong, gets the name wrong, accidentally knocks the mail in the trash, lazily marks the job as complete without sending anything, etc, the burden shifts to the poor person to prove not just that they didn't receive the mail but that the sending office didn't behave correctly.

Other cases behave similarly. In a he-said/she-said, the government wins.

You can do a lot better than this if you redefine the problem from directly generating images with certain contrasts to maximizing information gain, even with weak magnets. They've since basically run out of money and are on life support, but Q Bio [0] had that tech working years ago, able to quickly derive many different image types from an entropy-maximizing scan, though they never deployed that in prod IIRC (again, they're broke).

[0] q.bio

The government is just people. Even before the current fiasco, the government had varying degrees of incompetence and malice, and if you're poor you can't do anything about it since the government is presumed to have been operating in good faith and you can't afford a lawyer or the time off work to try to fix it pro se.

Yeah, harness quality matters too, but the underlying model capabilities are night and day.

Claude is the only AI coding tool I've found worth a damn. Without it I'd just do everything by hand save for a few bash scripts or whatever.

You use Python when it makes sense for other reasons (library support, coworker familiarity, etc), same as for any other project. Additionally, sometimes performance matters, but perhaps not enough to overcome whatever else is drawing you to Python in the first place.

Right this second I'm writing something in Python with critical performance requirements. It needs to average processing 25k things per second. That won't be particularly hard, but it's close enough to the edge of what the language is capable of that I do need to be at least a tiny bit careful with the implementation. I'm highly unlikely to need a profiler for this project in particular, but earlier in my career I probably would have needed one.

Python is fairly commonly used as a glue engine around faster code too, and it's not always obvious when the wrapper code is inducing nontrivial overhead (hidden copies and that sort of thing). Profilers are great for teasing out those sorts of problems. They shine a spotlight on the section of code which should take 0us and is instead dominating your runtime.