Forgettable fodder, as so many of these recent allegedly groundbreaking talks are. I don't know if it's my increased exposure to them or something else.
Age combined with seeing the trite and cliche hooks to get people to read/listen/watch the usual self-promoting puffery over and over again. Sometimes it’s just the author discovering something mundane and wrapping it up as a TED talk.
In some sense even the best talks are an attempt to distill things we have learned into a tight enough package that it helps people (a) think a little differently and (b) go read and learn more themselves. Sometimes that turns into self-promoting puffery, but I don’t claim to have anything particularly original to say in this talk. To the contrary: the whole point and structure of the talk is “This stuff is out there, let’s pull it together and get thinking about it!”—pointing to good work other people have done.
Now, maybe it lands that way and maybe it doesn’t, but I think that our industry could use lots of “discovering something mundane and wrapping it up as a TED talk” if it helps us to be more serious practitioners of the discipline. Most of what we need is mundane, after all! But that doesn’t mean it is unimportant, and it doesn’t mean everyone already knows it.
// personal opinion: I think machine learning as it currently stands is widely overhyped
How is this the top comment?
> I am starting to notice a pattern in these papers - Writing hyper-specific tokenizers for the target problem.
This is merely expressing what they consider as part of a game state, which is entirely needed for what they set out to do.
> I argue this is just ordinary programming
"Ordinary programming" (what does that mean?) for such a task implies extraordinary chess intuition, capable of conjuring rules and heuristics for the task of comparing two game states and saying which one is "better" (what does better mean?).
> How would this model perform if we made a small change to the rules of chess and continued using the same tokenizer?
If by "small change" you are implying i.e. removing the ability to castle, then sure, the tokenizer would need to be rewritten. At the same time, the entire training dataset would need to be changed, such that the games are valid under your new ruleset. How is this controversial or unexpected?
It feels like you are expecting that state of the art technology allows us to input an arbitrary ruleset and the mighty computer immediately plays an arbitrary game optimally. Unfortunately, this is not the case, but that does not take anything away from this paper.
I wanted to label this as a “nothingburger”, but all the pieces of advice are either tautologies, vague general dating tips (i.e. "make sure the vibes are not off"), or straight up untrue: "They want to work on this part-time", "They want to quit their job but haven't set a date to do it", "They're going to quit their job but only after the YC interview". The latter are all very context dependent and are impossible to label as a "red flag" without further information. Possibly SEO bait?
> don't see it as any less dignified than any other work
You do not, and that is your moral judgement. Rationalizing earning money by any means necessary is a very slippery slope, and the discussion is much more nuanced than popular media would lead you to believe.
tried it quickly on a personal machine running windows; all attempts at submitting `BUY`s for popular tickers (regardless of price, tax, quantity, ticker, date) seem to result in an unknown error. Notably, I tried having the account match the ticker's currency, but that does not fix it.
I am honestly asking and not trying to be a smartass:
What are the advantages of LLM summaries over man pages/google searches/stackoverflow threads?
I could maybe empathize if an engineer only had 30 seconds to understand a code snippet or a large shell command, but how often is that really the case? Is the time gained worth the risk of hallucination or incompleteness?
I can provide a use case, although I don't use this site specifically, I do use LLMs (generally Claude) to get command line syntax fairly regularly.
The answer is it's faster and (in my experience) generally accurate for common command syntax. I don't remember all the params for a variety of commands but Claude can get them quickly and easily. Where (IME) LLMs fall down is for more obscure commands but then those are harder to find via traditional searches anyway :D
Googling is very hit or miss these days you end up with a load of sponsored results and then have to try and find an example with ideally the exact syntax you're looking for.
> The answer is it's faster and (in my experience) generally accurate for common command syntax.
If you don't mind, around what level of complexity are you querying it for? Are the queries along the lines of "how do I find a keyword in a directory in UNIX?" or more along the lines of "write a bash script that will do foo recursively and treat edge case bar by doing baz"?
In my experience, if the queries are closer to the latter, it takes less time to "man page" the holes in my knowledge than to identify and to fix whatever is wrong with the LLM's initial guess. If you more often than not receive correct answers, even in non-trivial cases, could you please provide the LLM and any custom parameters if they exist? I'd be happy to be proven wrong and learn something in the process.
> Googling is very hit or miss these days you end up with a load of sponsored results and then have to try and find an example with ideally the exact syntax you're looking for.
May age be a factor here? I grew up as search engines were becoming a thing and ignoring "fake" results has become second nature.
Kind of both. Simple things like the switches on ffmpeg to rotate a video, is one example. I do that infrequently and Claude got it right in one. Quicker than reading the ffmpeg man page for that information by a long way. Also basic utility scripting something like "give me a golang program that connects to a Docker instance and returns information in this format". I could write it but a lot slower than an LLM
As to the age thing, well I doubt it. I started on the Internet in 1995, and I've been on it ever since. Sure I can skim by fake results but it slows it down as the good results are ever fewer on the page. I can do it but LLMs are faster and avoid all of that.
Also by using them more, I get a better sense for what will work and what won't work, improving my velocity again. It's easy to spot that they're good for things that are widely done, so more common and widely used commands and languages.
They fail hard on novel or new things, for example OTel code is one place where they're bad, and also deeper parts of things like k8s.
I tried ChatGPT 4 for some time and went back to just using manpages after several wrong results.
Sometimes results were bad even after providing information about my platform.
Have there been posts already from people accidentally deleting files caused by trusting LLMs too much?
Because arrows are functions/mappings, and everything we do in programming involves arrows, even in languages where arrows aren't used as notation.
The common formulation is that a "monad is just a monoid in the category of endofunctors", which is not saying much but with big words, and the joke lies in understanding what it's saying. Bartosz Milewski has a lecture video series on youtube that's all about explaining that joke, and I highly recommend it because it's actually a wonderful CS lecture series.
As I understand it, it's a bit of an inside joke to minimize the complexity of mathematical structure. It's frequent use is along the same lines as the frequent use of "* Considered Harmful" in CS.
I think it's objective truth to note that "Bitcoiners" are predominantly comprised of:
a) people conducting illegal affairs (don't read this as "illegal" in the sense of jaywalking)
b) technically illiterate people, down on their luck and hoping to get rich quickly
With those in mind, I would argue that "Bitcoiners" are instead drawn from the left side of most distributions.