If you're bad at using the tools, is that their fault?

Good prompting and good context management are non-trivial, but they are things you can learn to do.

Good prompting is really just good communication. Pretend you were telling a junior developer who is very bright and somewhat overenthusiastic what to do via email, and that you can't send them another email for several hours. If you give them incorrect instructions, they're going to produce incorrect results. If you give them vague instructions, they're going to spend a lot of time building their guess at what you want or -- often worse -- reading the whole codebase to gather the context required to figure out what you want. (Humans hate reading huge amounts of code, so a human dev probably wouldn't do that, but an LLM will).

And what you need to communicate isn't just what to do, but how. As one example, I do most of my work in statically-typed languages, primarily Rust and C++, and I find that the LLMs really all seem to primarily be Python jockeys. They can write Rust or C++ just fine, but they don't really think about how to take advantage of strong typing. If I ask them to refactor something, the first thing they want to do is to go scan the entire codebase to see what will be affected by it. In a dynamically-typed language (especially if you don't have good unit tests), this is the right thing to do. Sometimes the LLM can use grep or sed to find the relevant code efficiently. Sometimes they need to actually ingest thousands of lines of code (newly-loaded tokens!) and that gets expensive.

What an experienced human Rust/C++ programmer will do, and what an LLM can do if you tell it to, is to rely on the compiler. Think about how to structure your refactor so that all of the places that need to be updated will be broken, then let the compiler tell you where all of them are, then fix them. This is much more efficient, for humans or LLMs, but an LLM won't do that unless you specifically tell it to. A junior dev might not think to, either.

As with a human, it's usually a good idea to have a conversation about the task before telling them to start the task, to make sure you and they both understand well what is to be done. But this leads into another important cost-management issue: Context management. If you're going to have an extended back-and forth with an LLM, make sure that it doesn't have a lot of extraneous data in its context window.

Context management is crucial to keeping costs down. Every time you submit a prompt, the model has to load the entire contents of its context window. "Reloaded" tokens are a lot cheaper than "newly-loaded" tokens, but when the context window is 1M tokens, the costs can add up fast. One solution is to use a model with a small context window. That works, but then you have a junior developer who doesn't understand much and constantly forgets what he does understand. For some tasks, especially very mechanical tasks, that works fine (in fact, for some tasks it's actually better). But if you're doing something that requires understanding a large codebase or lots of other context, such as large requirements documents or something, you're going to get stupid results from a model that doesn't have enough context. On the other hand, clearing the context too often means having to reload it more, and newly-loaded tokens cost more than reloaded tokens. (There are also output tokens to consider, but I find those aren't usually relevant to cost). So, knowing when to use a larger or smaller window and when to clear the context window are essential skills for keeping the costs down.

A related choice is which model to use, and this interacts strongly with context window size/content. I primarily use Claude Code, and most of the time I keep it set on the default Sonnet model with a 200k context window. When doing something larger, I bump that to 1M tokens. When I need help thinking through a complex design question, I switch to Opus, but usually with a fresh context window. I have some good project summary documents (a few thousand tokens) that provide high-level context for cheap, so I clear the context window, tell it to read the project docs (I have a skill for that, with a short name), and then start working through the issue.

There's a lot more I could add, but this is long enough. The TL;DR is that using LLMs effectively is a skill -- a rapidly evolving one. Perhaps in the near future the LLMs themselves will get better at context management, model selection and knowing when to ask cheap followup questions rather than do a lot of expensive research. But right now, they don't.

I'm not sure what you're disagreeing with?

All very true. How long before AI displaces human narration probably depends mostly on the cost-sensitivity of listeners. I'm absolutely willing to pay $5-10 more per book for talented human narration, but in general consumers prioritize price over quality as long as quality meets a certain threshold. LLMs don't meet that bar at present, but they might before too long. And, eventually, I'm sure they'll match the capability of human voice actors, though I have no idea how long that might take. We might have an extended "uncanny valley" period during which they're good enough that authors don't bother with the expense of a human, but they really aren't as good.

For some reason, people never seem to have trouble using 'current stock market valuations' to secure loans and credit, so.....

Indeed, but not all warrants are constitutionally-valid, and that's what this is about.

Sure, the scope of the 4th amendment is something that has to be decided through legislation and judicial review. My opinion is that based on the history of 4A precedents, this crosses the line. Several district court judges and appellate courts have agreed with me. One appellate court disagreed with me... which is why it's in front of SCOTUS, whose job it is to resolve the question when appellate courts disagree.

As an aside, it's somewhat refreshing to see SCOTUS working normally. So often lately they're handling crazy cases that shouldn't even be in front of them, because the Trump administration skipped all the preceding steps and went straight to the top. But this is how it's supposed to work. District courts make rulings about whether government actions are legal, and sometimes those turn on constitutional questions. Appellate courts review the district court rulings and either uphold or deny, issuing precedential rulings that are binding on all the courts within their circuit. When appellate courts from different circuits issue conflicting rulings, then it's up to SCOTUS to decide.

That's a bad example, because it's the 16th amendment that does that, not judicial precedent. Oh, the 16A doesn't explicitly call out that you should be required to report, but it's clearly implied since there's no way for Congress to levy taxes on incomes unless they can find out what those incomes are.

If the police put out a call and asked everyone who was in the area at the time to voluntarily tell them so, that would be fine, and you offering up your location history would be your decision. I'd probably offer up mine, too.

But that's not what happened here.

What happened is that the warrant authorized the police to scoop up everyone's data regardless of whether they were willing to offer it. We do allow police to violate the privacy of people in narrow, carefully-specified ways because it's necessary to solve crimes, and that's good for society. But grabbing the location history of everyone who happened to be in an area at a given time, without their permission, seems like it crosses a line. That's the question before SCOTUS.

Personally, I have nothing to hide and don't care... but I think Chatrie's lawyers are right, that this is overly broad and shouldn't be allowed. It's unfortunate that getting this right (assuming SCOTUS agrees with me) will allow a bank robber to go free, but that's how it goes sometimes.

A lot of new fiction today is essentially self-published. In some ways, this is great. It's easier for new authors to get their books in print, rather than dealing with endless rejection letters, and those that are successful keep more of their money. On the other hand, it means that readers can no longer rely on publishers to act as quality filters. This shows up both in an increase in slop (AI and otherwise) on the market and in a significant reduction in professional editing. Often there is no professional editing at all.

I find that this is (one of many) good reasons to consume fiction primarily in audio form. The cost and complexity of getting a talented voice actor to read your books serves as a quality filter, and the narrators generally fix the most severe editing problems. They won't do structural editing, of course (e.g. deleting useless paragraphs), but typos are naturally not relevant and they also often fix erroneous word choices, incorrect names, etc., when it's obvious what the text should say. For example, I enjoy Terry Mancour's "Spellmonger" series quite a bit, but I absolutely cannot stand reading it in print. So many annoying mistakes that even a light editing pass would fix. But John Lee, the narrator Terry uses, does an outstanding job of cleaning all that up as well as of bringing the characters and the stories to life. There are several other recent book series like this, which I find unreadable in print but quite enjoyable in audio form.

Of course audiobooks also have their downsides. They cost more, sometimes narrators are lousy -- even to the point of making a good book unlistenable -- plus they have their own annoying "editorial" problems, such as different narrators reading stories in the same world using different pronunciations of names and concepts. Drawing on the Mancour example, his world had a time when it was ruled by mages, a time that is called "The Magocracy". Logically, this should be pronounced with a soft "g", as in "mage", and John Lee does this. But some side novels are narrated by Fiona Hardingham, who insists on saying it with a hard "g", as well as pronouncing a lot of names differently. Minor, but grating.

This is why NASA always packs a tin of Bondo with the mission supplies.

What conflict of interest? He divested his Anthropic ownership. And since when has the Trump administration worried about conflicts of interest?

Well, he's probably going to get punched in the face, too. For his sake I hope it takes a little longer, though like cpurdy I don't have too much sympathy for anyone who agrees to work for this administration. They have to know the odds are high that it will end badly for them... and it may also taint them for future employment (I don't think it should, but it may).

Imaginary Property apologist.

Username checks out.

You think Google's money isn't real? There has been a lot of circular investment, but Google is the player that actually has lots of cash, and isn't just recirculating it.

I think thesandbender is probably closer to the mark, but he should have added that if Anthropic fail because they're overextended financially, as one of the major investors with lots of available cash, Google will be in a position to take control, buy out the other investors for pennies on the dollar and snap up the assets.

The summary for some reason thinks that there are only two frontier AI labs, Open AI and Anthropic, but there are three: Open AI, Anthropic, and Google.

It is interesting that Google is investing in Anthropic, a competitor. Just hedging their bets, I guess.

You can collect water from a roof with solar panels just as well as from a roof without them.