260 karma · joined June 7, 2021
For example check this: https://youtu.be/m7EcEne_d8Y I didn't believe ai generated video can be this good before watching it.
Google flow is such a joke comparing to seedance or mimimax h3.
I just do not understand. CC is not the sole harness in the world, why the attackers insist to use it.
the problem with gemini is the whole app itself, not the model.
when you use other chat app, like chatgpt or deepseek chat, the app will use web search for everything to verify. and if it can not find anything useful, or the current fact is disputed, it will try to call web search tool again and again until the consensus.
meanwhile gemini will refuse to do web search if possible, and only do it a single time if demanded. worse, it uses the same result from google search, so the top ranking answers are almost all SEO garbage.
since then I just stopped asking gemini again. and I don't think the newer model will change the situation.
CSS tricks was never about only css, it also had js, html, design and other tips, too.
Pretty useful to catch up with webdev related new features
it is true you dont want to trip you up by the language you use, but look what people actually use the languages, not by theory.
almost nobody writes the high performant code directly, usually they just call some libraries that do the job instead. if you are the library writers, then sure, that's a different matter (but people have been doing it forever using c/c++), but for people who use the libraries to write the code, they don't care about any of that.
many people rather pick python, lua, ruby to write applications, even games, despite the low performance of the scripting language, but they don't care. and if it is only the scripting aspect of the whole framework (implemented in C++) then the speed lost is even less.
heck, even for rust, the core libraries always need unsafe blocks, despite they do not recommend user code to use them. it is the same story everywhere. the hard part will be write by more experienced people (and with more effort), the typical users just write domain logic instead.
so, for most people, they indeed need simpler languages with lot of syntactic sugars for better developing experiences.
it is not like he can get the funding via opensourcing the language to cover the development. what he and his company needs is the game to finish and release and get enough profit to cover all the expense. making a new engine is already risky enough.
you just can't ask a man to ruin his whole career/life because it benefits you.
It is pretty understandable he needs to hold back before he releases his new game, or even more.
I hope Europe all the best but there is a reason why only US companies dominating those space.
US companies can operate in lost for many years, because they have deep pocket, and also because they have many way to earn profit.
I don't think nobody would notice such delays, because Chinese models do not need to hide the CoT like US models so usually they start emit text pretty soon
So it is not that the labs call the requests parallely and only return the internal version, but also serve those requests live so they can trick the users that the Chinese models are also great.
Which is very silly considering Chinese models are published to huggingface for everyone to test.
Unlike Claude models which always hide the thinking process, Chinese models always show it. So how the heck they can delegate the api call to anthropic and still get the full reasoning? Yes there are tricks to reveal the thinking output, but they are just tricks and do not works reliably. So what would happen when they failed to do it? Cancel the ongoing token generation and revert back to it own model? What about the glitches, the delays, the changes in behavior?
I would understand if the attempts are small in volume, because normally people hardly notice it, but anthropic claims that they happened for millions of calls. Such large numbers would surely get noticed by Chinese users and got reported in the tech forums like linux.do. yet we haven't seen any sight.
also, using crystal you have parallel and concurrency to speed up your task. sure it is not as good as other matured languages, but it still does the job.
when doing with large amount of data, and with IO involved, you will see that your JIT or AOT engine for ruby still can not cut it, because in ruby there is so many layers of safety check and failure guard it is really hard to make it faster (number calculation is not everything). with crystal you can manipulate the io stream directly, avoiding a lot of intermediate steps. you can check yourself, they make a full page for it: https://crystal-lang.org/reference/1.22/guides/performance.h...
personally I do not use a single language, I just pick the most suitable language for the job I'm doing, let it be ruby, python, crystal, rust, go or typescript. but surprisingly crystal is the one I use the most everyday.
Like every one else I also used ruby for almost everything back then, usually some hundred of lines or less for processing stuff. but for tasks that run for several hours using ruby was a frustrating experience. slowness and memory consumption aside, the duck typing system and lazy evaluation of method meaning you can run to typo after hours of running session. sure, you can fix it by write the code more carefully, running smoke tests beforehand and other stuff, but it still a chore, and when you write code all days it makes you sloppier.
having a strong typed compiled language helps it immensely, and with crystal you don't have to sacrifice the easiness of programming. and you can compile it to a portable binary and run it everywhere, no need to reinstall a huge toolchain again. (yeah rust or go can also do that, but not for ruby level expressioness language)
I'm sure crystal is not the ideal language for everything you want to use, and ruby is still helpful for what you use daily. but for some tasks, crystal surely has it own advantages. you can write very capable CLI programs without have to install a single library!
Even take account of R&D cost it is still profitable, because you do not actual retrain the base model every month, only distill and post train from the same foundation.
heck, I think even including the giant cost of the top tier researcher salaries, they are still profitable.
then, how are they still losing money? just read what they are actually spend the most budget: always the hardware. those companies will never profitable, because they will never stop spending money to expand their computational prowess.
It is very capable for a lot of tasks, not only coding like the other open models.
It still thinks Charlie Kirk is alive, though.
I don't think so. It's true that human also write shitty code but the key difference is we actually remember what is the intention behind those crappy implementations so someone can fix it later. aka it is the matter of long term memory that currently LLM architecture is not capable of.
You can argue that claude can read the whole linux codebase and report bugs, but they can only report local bugs, not systematic one. 1M context windows seems like huge, but the effective range is actually pretty limited, and it still does not equal to human insight.
why the heck do you need 100+ instances of bert. do you even attempt to research about this before?
the laya paper show that you can do the similar stuff with jev using modern bert only: https://laya.convaiinnovations.com/
and even without the newer wave of applying llm techniques to the older bert models, even flan-t5 was trained for handling 1800+ tasks.
Or even llm if you claim about versatility. You can easily modify the llm inference code to make it predict a single token represent the classification choice and extract the probability that way.
Sure jev will still be faster, but a local deployed Bert model is way faster than both.
And to get the most out of it you still need to fine tune the models anyway, unless your classification task is just one of those mainstream ones.
It is pretty trivial to pin a single provider for a model. Better yet, instead of calling the model directly, use presets instead. You can easily change the setting on openrouter without having to update your app every time.