A bunch of outlaw militiamen living in some camp somewhere probably don't have access to a capable computer, but they can probably log onto chatgpt.com easily enough.
Wasn't it established that ISIS soldiers go into battle hopped up on amphetamines and other drugs? I don't believe the story as stated either, but I fully believe that members of an extremist terrorist military organization would be willing to do extreme things in service of what they believe is a noble and justified cause, even if that includes dying during training. American soldiers that die in helicopter crashes are still honored as heroes who gave their lives in service of their country, and few question that, meanwhile people still enlist in the US Armed Forces. If some Boko Haram commander gets an obsession with some kind of crazy motorcycle attack and a guy dies in a motorcycle crash, do you think everyone else is going to punk out? Or are they going to honor him as having sacrificed his life so the other soldiers can carry out the attack?
I wonder how much of what China is currently able to do it because USA is such a mess.
The whole article is kind of ridiculous of course, and is also heavily fixated on OpenBrain, whatever that is.
I also wonder about the economics of running an AI lab attached to an existing large tech company (such as Meta or Tencent) instead of a dedicated company like OpenAI. It's starting to seem like it's not possible to charge enough for current-gen AI usage, with current-gen inference technology, in order to turn a profit, i.e. nobody is able or willing to pay at least marginal cost for tokens.
Inductive number types seem pretty appealing here. A type like "Nat = Zero | Succ Nat" is 100% always correct by construction and can be safely transformed into raw integers (and operations thereon) by the compiler. And it is also precisely the correct number type for array indices.
If you can open your mind to dependent types, then Fin is even better, which also avoids overflow by construction. But then again dependent types are a whole different beast when it comes to compilers and language capabilities and so on.
First of all, yes, performatively following health fads has always been a thing for people with money.
But do you honestly believe this is what's going on at a large scale? What an absurd reaction to other people doing something that you don't like or don't understand. Nobody who isn't a paid influencer is sneering at you for using the wrong materials in your kitchen.
If you feel sneered that, that's your problem and not anyone else's. Are you going to start plasticmaxing, like rolling coal but for food containers?
I think you're doing the thing that you are accusing other people of doing. Creating an artificial distinction between "us" and "them", so that you can feel like you are part of the "us". It's just that in this case you want to feel like some kind of rational rebel, one of the final chosen few with common sense, in an increasingly irrational world.
To be fair this was all happening before, just 10x less. And the current minority party was often willing to ignore it when it was their people doing it. So yes it's bad on a generational scale and we might never recover from it, but we also have to admit that we are reaping the fruit of a bipartisan-sown seed.
But Congress is very comfortable so far just letting the executive branch do whatever. Even if the orders aren't emanating from the oval office directly, there's clearly a coordinated agenda in motion. It's entirely reasonable to suspect that the GAO has been politicized in that environment.
They must, right? They literally have no mechanism for cognition other than transformation of vocabulary. Language models are models fitted from data generated by humans, but they are not humans. Humans generate data by whatever processes happen in our brains, and LLMs of various architectures can learn a surprisingly good approximation of that data-generating process. That doesn't mean they have all the same characteristics and properties. That's true for all models really: model characteristics that are not captured in the goodness-of-fit metrics are not guaranteed to match the original phenomenon being modeled.
That's a cool result because that's also kind of what's happening inside the transformer unit: project -> QKV fuzzy lookup -> unproject. And in a different direction, it's analogous in some sense to what's happening in stacked convolutional layers, where the layers at different levels learn to recognize features of increasing detail.
Yes, and it's working for them, but it's not their core competency nor does it have much to do with the actual product that people pay them for (apart from being a source of bugs to avoid), so the fact that it's working for Anthropic doesn't necessarily imply that it's actually a good idea in general. It's working for them, with emphasis on for them. Or it's actually a demo of the future and it would work for anyone, who's to say? There's no doubt that AI (and especially Claude as industry leader) helps you carry out significantly more complicated side quests than in the past, in part by enabling that "recklessness" with surprisingly little risk. Maybe an exhaustive enough test suite is really all you need. But Anthropic is not in a position to demonstrate that as the universal future of all things, not just side quests.
What's working for them is having a huge amount of resources and very good people to design a cutting edge agent harness, RLHF the hell out of their models, and build out a tremendous amount of inference capacity. I'm sure their process for making code changes in any of their client apps is very fast, but a TUI built around a chatbot is also not a particularly complicated application. So yes it's working for them, but the vibecoding that they are selling is clearly not what they are doing in practice.
Not OP but Copilot CLI is really straightforward, almost minimal in some sense. It's a lot like OpenCode but stripped down.
I also use the Copilot ACP server inside Pycharm and that works decently well too, although it has some annoying bugs, but if you're a Jetbrains user you're used to annoying bugs.
Countervailing story: the Ergodox (EZ specifically but that's less important) gave me permanent RSI in my thumbs, because it was too big for my hands, and to this day I pretty much can't use a keyboard layout that relies on lots of thumb involvement. Even just hitting the spacebar throughout the day or using my thumbs to type on a smartphone is enough to flare it up, almost a decade after I stopped using that keyboard.
YMMV, ergonomics are highly personal with respect to your body size and proportions. We didn't have the proliferation of keyboard layouts then that we do now. Perhaps if the Iris or Corne had existed then, I would still be using my thumbs for modifier keys in a 40% layout. I never got the hang of tapdance or hold modifiers.
The Atreus layout is the only one I can still use somewhat, because the thumbs are held closer to the hand rather than splayed out.