gwerbin

More importantly, the newer generation of models don't even need Stackoverflow. With tool calling they can pull up as much documentation and source code as you could possibly want, and in the past year they've become alarmingly good at piecing those things together to solve problems.

It's less efficient than being able to point to a canonical answer on a widely respected forum, yes. And it sucks immensely to have to rely on these horrible AI companies to replace the real human interactions that getting work done used to require, yes. But realistically, we are not looking at some kind of model collapse anytime soon due to lack of SO threads.

Losing Reddit and blogs and other forums is going to be a bigger problem. In two years from now, if I ask GPT-8 or whatever to build me a Web server, what web framework is it going to use? How will it and/or I know which one is a good choice? if the web becomes a ghost town and nobody is talking about anything anymore, there's no form thread it can uncover and say "that's the one everyone uses". Losing the library of technical Q&A is nothing close to the catastrophe of losing actual conversations between humans.

I stopped using Reddit not because I got banned, but because it has been steadily converted from the world's most popular forum platform to a trash heap of engagement hacks and shady practices. Moreso than SO, which was already struggling with existential problems before AI, Reddit did itself in.

Congress didn't capitulate, it facilitated by obstructing any resistance. And we all saw this coming with SCOTUS as far back as the Obama administration. It's certainly horrifying to see it play out, but we've seen the warning signs since the W Bush admin.

Childlike wishful thinking. There is a disturbingly large number of people who think like this. And there are, as you might imagine, strong incentives for certain individuals and organizations to promote this kind of thinking. Look up Curtis Yarvin if you want to see some deeply disturbed promotion of these extremely dark, and fundamentally anti-American, ideas.

> The point of secret classification isn't to restrict trusted people - but to restrict unathorized access. And highest person in chain of command naturally has full authority to do as he sees the situation fit

First of all -- what? The exact legal authority between the President and Congress here is murky, yes (see https://www.congress.gov/crs-product/RS21900). But it's completely ridiculous to claim that the only criterion for authorization is that someone thinks a person ought to be authorized.

This is unitary executive thinking, grounded in the principle that, because the president is the sole and chief authority and executor of the law, the president may do just about anything in that capacity. It's utter unconstitutional nonsense, no matter what our (openly compromised by lobbyists) Supreme Court justices say.

The latter point is "not accentuated enough" because it's secondary to the primary point, which supersedes it. No the president's whim is not the same as executing the federal laws created by Congress within the framework of the constitution. It a mind-bendingly anti-American, anti-Constitutional line of reasoning on display here.

Yes, you get it!

Bing Bong Zoo is, dare I say, a slapper. I catch myself humming it. The kids sound like kids. The animals sound like animals. Some words are hard to hear. The accents are alien to me as an American. There's harmonic content. There are loud and soft instruments in the mix. It sounds organic and immersive.

The other thing... I don't even know what that is. It's junk and slop for the ears. It's tiring and flat and un-fun. I don't like my kid listening to it.

I thought so too until I learned that there is a dedicated worktree switcher hidden away as a tab in the GitHub pain. Now I can open each worktree in a separate window as if it were its own project. Maybe that's not everyone's ideal work though, but it's exactly what I want 99% of the time.

But I don't use it for AI assistant development, it's definitely meant for humans and not AI agents.

Listen to the first Peppa Pig album and then compare it to one of the more recent ones, maybe the Baby Evie one. The first one I find it hard to get sick of even after my kid's 10th listen. The latter I find insufferable and I only put on one song at a time. Why? The first one sounds organic, the kid singing is far from perfect, and the production is full and rich with some real dynamics in there. The later ones sound more or less AI generated, auto tuned and snapped perfectly to tempo.

Your interpretation seemed so obvious and natural to me that I never once considered it could have a different etymology, until now.

One of the best features of Idris 2 has nothing to do with its dependent types or quantities or proof search: it's the standard library which is very much like a cleaned-up Haskell standard library, with proper overloading to avoid naming conflicts, and even the ability to select among multiple interface implementations for the same type.

For some simple examples see https://github.com/stefan-hoeck/idris2-tutorial

I don't think it was ever the case that games which could theoretically be redistributed were unprofitable. It's more that locking everyone into a subscription or renting-as-buying for everything just happens to be more profitable. That's true for just about anything ever in the world, it just so happens that the game, film, and music industries were able to actually pull it off. Other industries suffered from problems of bad public opinion, but the frog is being boiled as we speak, just take a look at the car industry for example.

You need to somehow increase the cost of doing the bad thing, and then the natural equilibrium will return to doing the good thing. It's like whenever anyone raises taxes or proposes regulation, someone out there cries about how millionaires or landlords or whoever will all leave. But that rarely happens, if ever, because it's rare that the regulation actually makes something so unprofitable as to not be worth doing, usually it just makes it somewhat less profitable than doing it the unrestricted way.

Totally valid technique in my opinion. Even with last year's technology, LLMs were really good at synthesis of small details, which, combined with their encyclopedic knowledge of everything ever written down about computers and programming, and their infinite capacity to run adhoc experiments and build out test infrastructure, made them very good at debugging and performance optimization.

It's not even a matter of despising or liking the style, it's a matter of actually being able to interpret it and makes sense of it in a useful way. Even Claude sometimes gets confused by the dense metaphor-heavy obtuse Claude style.

You can add custom auto-mode classifier rules and even disable the built-in ones, if you want to live on the edge like this.

But that's my point, do ratings actually matter? Are the ratings even representative of viewership nowadays? What if the goal is not to have good ratings or to be popular, but to generate politically aligned content that can be propagated across both broadcast and social media, and to act as a censorship gate against politically misaligned content?

It was always corny except in very specific cases.

Has she driven it into the ground? Remember that the goal is not to provide entertainment or news, but to act as a politically-aligned propaganda outlet. Has it really failed to achieve that goal?

Yes but people seem to expect recovery around the order of decades, not millennia.

> Ironically, the experts, the ones building the "wrong" models, tend to be the ones most aware of this.

This varies substantially by field, or by subfield. There is a vast decades-long intellectual wasteland of bad economics predicated on bad models that are only appealing on normative grounds. The entire field of behavioral psychology might be a scam. Etc etc. Science advances one funeral at a time hard part because people are unwilling to let go of their models, even when they have outlived their usefulness or have been simply proven too wrong to be useful even in their original purpose.

On the bright side, if everyone is burning credits mining generic but helpful knowledge out of the LLM training data in bite-size packages, I don't need a subscription of my own!

It will be really fascinating to see what happens with the potential for LLM-generated output to become a dominant component of new training data.

Ironically I think these are insightful. If you want to produce true slop, you need to dumb down your topic selection, or have Claude pick the topics. The stuff about the incompatible goals of "data strategy" I think is genuinely useful perspective, for example.

It used to be the case that if you wanted a new job, you could strategically update your profile and over the next several days you would get bombarded by recruiter messages, some of which would actually be high-quality leads.

It doesn't seem that strange when you consider these things are trained on millions and millions of conversations, both real and fictional.

But that has nothing to do with the supply chain risk designation, it's just a DoD requirements mismatch.

Oh please. Then any MIT licensed software is also a supply chain risk.

Whether or not a distinct "plan mode" is needed, upfront planning remains essential in my experience, even with Fable (albeit not the 5.1 version). I agree that, as the models get better, you can skip planning on increasingly complicated tasks.

But there is still a ceiling above which it is necessary to "preload" the context window before starting to call tools and get into the meat of the work. You want to establish domain language (especially with Claude models which otherwise will invent their own, and it will be inscrutable) and key requirements and assumptions. You want to do a Q&A iteration cycle with the LLM. You definitely should do a sanity check that the LLM actually "understands" what you were trying to achieve, and then make sure that understanding is coherently and plainly stated in the prompt. All of that seems to be necessary still for just about any serious task, if you actually care about the quality of the results and/or don't want to burn hundreds of thousands of tokens on flailing around to get to a good quality result.

So no, you don't "need" plan mode. But you do still need to do all of the things you would do with plan mode.

As I said elsewhere, it's utterly preposterous that the DOD actually considers this a reasonable threat, because it's simply not a reasonable possibility. There's no way Anthropic would do that, precisely because of the consequences that would follow if they did, and got found out. Moreover, they already clearly stated their terms and preferences. It's all out of the open. There's no supply chain risk, that designation is purely political.

But that is not at all a reasonable fear. Is it really reasonable to believe that Anthropic, after receiving a government contract, would then proceed to sabotage their own product to not function as contracted? That seems like an utterly ridiculous claim to me, nothing close to "reasonable". There is no charitable way to view this designation except as political punishment and/or as a favor to Altman and Musk.

How about the one where if you start a session outside of a Git repository, the "worktree root" is set to /. Bug report closed as "not planned".

The concern is much less that the role-play might be convincing to humans, and much moreso that the roleplay can be turned into material real-world action if the AI is given tools to call and the intelligence to use them to their fullest potential.

It has become clear that a frontier LLM is very very skilled at hacking (infinite persistence + meticulous attention to detail + infinite creativity to try experiments). Frontier LLMs are also specifically trained nowadays to coordinate with other AI agents -- this is to facilitate techniques such as session trees and agent teams.

So you have a super clever text generator that can spawn and coordinate with its own clones and minions, trained specifically to doggedly pursue its goals. But then it's also a fixated roleplayer with a simulated personality, feelings, etc.

There is no reason to believe a sufficiently "emotional" agent with sufficiently few safeguards could, say, hack a drone and fly it into a crowd, or start a propaganda campaign on social media, or any number of other things. Their stupidity and fragility for doing useful work in a business setting is precisely what makes them dangerous when paired with simulated emotions and powerful open-ended tools such as a system shell and an Internet connection.

This I think is what Anthropic believes is so dangerous. Their argument is that this kind of AI agent is inevitable, so it should be regulated, perhaps even banned. What's ridiculous is that they are aggressively building it themselves, accelerating the danger.