Tangent, but I'm curious about how your style feature tagger got "no contractions" when the article is full of them. Just in the first couple of paras we have it's, that's, I've, I'd...
2,647 karma · joined March 26, 2010
wzdd.hn@lardcave.net https://nfd.world
It's weird behaviour and I wouldn't have expected it either, since infinite length streams are pretty common.
I'm not trying to no-true-Scotsman birth control methods, just pointing out that it's hard to draw meaningful conclusions when all you have is people's word.
The article mentioned that important API calls are signed. So you would need at least partially to deconstruct their front-end to invoke the calls.
Any other nefarious purposes aside, this seems explicitly anti-bot because you can change the obfuscation whenever you like, forcing another RE effort.
This is extremely disingenous. Throughout this thread you've been arguing on the basis that hiring people simply to fit a political viewoint is wrong, but when it's pointed out that that's exactly what your team wants as well you fall back to name-calling.
(I know your message wasn't claiming that The Bitter Lesson was explicitly a counterpoint, I just thought it was interesting.)
This is also not how base training works. In base training the loss is chosen given a context, which can be gigantic. It's never about just the previous token, it's about a whole response in context. The context could be an entire poem, a play, a worked solution to a programming problem, etc, etc. So you would expect to see the same type of (apparent) higher-level planning from base trained models and indeed you do and can easily verify this by downloading a base model from HF or similar and prompting it to complete a poem.
The key differences between base and agentic models are 1) the latter behave like agents, and 2) the latter hallucinate less. But that isn't about planning (you still need planning to hallucinate something). It's more to do with post-base training specifically being about providing positive rewards for things which aren't hallucinations. Changing the way the reward function is computed during RL doesn't produce planning, it simply inclines to model to produce responses that are more like the RL targets.
Karpathy has a good intro video on this. https://www.youtube.com/watch?v=7xTGNNLPyMI
In general the nitpicking seems weird. Yes, on a mechanical level, using a model is still about "given this context, what is the next token". No, that doesn't mean that they don't plan, or have higher-level views of the overal structure of their response, or whatever.
> At 17, if you don't think Lord of The Rings is the greatest contribution to literature there's something wrong with your head. If you still think that at 50, there's definitely something wrong with your head.
(https://www.theguardian.com/books/2001/dec/09/jrrtolkien.sci...)
Him talking about "contribution to literature" instead of "best story" changes the quote for me. It makes it more about the craft of writing -- and I agree: LOTR isn't the world's best literature. But it's still a damn good story.
This is using "top" in the sense of "apex", because turbines have no natural predators. The paper says you could imagine them as something at the top of the trophic web. The paper explores the impact on their prey populations.
That's opposed to "top" in the sense of "kills more than other predators" or whatever. I mention it because number-of-kills is what the rest of the discussion seems to be about.
Compilation is also a translation problem but I think many people would be leery of an LLM-based rust or clang -- perhaps simply because they're more familiar with the complexities involved in compilation than they are with those involved in decompilation.
(Not to say it won't eventually happen in some form.)
OpenAI's future investments -- billions -- were just threatened to be undercut by several orders of magnitude by a competitor. It's in their best interests to cast doubt on that competitor's achievements. If they can do so by implying that OpenAI are in fact the source of most of the DeepSeek's performance then all the better.
It doesn't matter whether there's a compelling legal argument around copyright, or even if it's true that they actually copied. It just needs to be plausible enough that OpenAI can make a reasonable case for continuing investment at the levels it's historically attained.
And plausibility is something they've handily achieved with this announcement -- the sentiment on HN at least is that it is indeed plausible that DeepSeek trained on OpenAI. Which means there's now doubt that a DeepSeek-level model could be trained without making use of OpenAI's substantial levels of investment. Which is the only thing that OpenAI should be caring about.
a) writing to disk was inherently dangerous: file systems of the time were not robust, so anything going wrong while updating metadata (in particular) could trash not just your file but the whole filesystem. Floppies were slow and machines were crashier, making the whole operation riskier; and
b) using floppies reduces their lifespan because the disk physically rubs against both the read/write heads and the internal padding of its casing, so typically they were only accessed as part of an explicit user request.
In any case the discussion in that issue made a convincing (to me) argument that if you're doing the sort of scripting for which "set -e" makes sense, which is most of it, you should be using Bash. That doesn't mean you need to use Bash interactively though, as others have pointed out.
Please become familiar with the British sense of humor and the context of the writing before taking it at face value. If that is too much work, at least be prepared to give the writer the benefit of the doubt. Babbage was writing for a specific audience and with a specific intent, and I would suggest that you are probably misunderstanding him if you are inferring smugness from an isolated quote.
Although a hardware-optimised take on this is interesting, CHERI is not unique in allowing a flat, shared-address-space operating system using hardware-enforced memory protection.
You can imagine doing such a thing on a regular system by associating a set of capabilities with each thread or process. The capabilities refer to a range of the single address space. Normally the range is not mapped to that process; when the process tries to access the range the kernel handles the page fault by checking the capability list and mapping the range. If a capability is revoked the range is unmapped again and associated caches (i.e. the TLB) are flushed.
This scheme obviously has different trade-offs versus using hardware-checked fat pointers, but has the advantage that it can be implemented on commodity hardware.
This is the basic idea behind the Mungi single-address-space operating system from 1998, among others. (https://research.ibm.com/publications/the-mungi-single-addre...). Is it also straightforward to implement such a scheme at user level in some implementations of L4.