I have nothing to complain about. Coincidence?
fair point. I think that 6-10 teams work great. Now, Iâve seen numbers beat quality. One example would be defcon, there was one capture the flag event where team sizes didnât matter and one team got hundreds, second place was maybe 10-20. The team with hundreds won because they properly organized on what to focus on.
Iâve seen this work in other areas, where the numbers of team members were too overwhelming. It comes down to organization. If thereâs a need for communication between teams than I agree, itâs probably doomed. Failing to properly scale creates âreverse dependency pyramidsâ.
This isnât the defcon event I was referring earlier, but there is this reddit post:
that talks about how it was difficult to manage a large team but with the right organization they got first place. I couldnât find what year it was the capture the flag event where Plaid Parliament of Pwning (PPP) took 2nd place because of team size, while in the events where there was a team size limit they got 1st.
That sound more like the head of project management was a genius and they used a very modular approach.
When you have to figure out stuff that doesnât work.
If you have an architecture that allows to add more mules e.g. a plugin based system with each plugin being independent from other stuff the more the better.
Python is great for prototyping, but everyone that works with anything that requires performance knows to start in python and translate into anything that isnât stupidly hard to code. Coding LLMs in machine code without a compiler is absolutely bad shit insane. Same goes for Assembly. I wonât give any more tips on this, if oai doesnât know they can look into my data, I donât care.
All paid browsers failed, all free internet services failed. You might not need to train the models, but even running them after training requires quite a bit of hardware investment. The tendency is to continue making the models bigger, at one point, to get the same results as the paid sota modals youâll need to be have a lot of money. You can provide internet for free as long as you donât care about how much money it costs to keep the infrastructure alive.
1 Like
right right, yahoo didnât need it, blackberry didnât need it, bell labs didnât need it and no country on earth needs a strict digital cyber security audit regulator to make sure what is hosted online doesnât contain the most easy to spot zero days ever⌠right on, openai doesnât need help, let me go help the open source community and not worry about geopolitics, right? right? oh waitâŚ
I should probably start helping over at the XZ Utils repository, this github member named Jigar Kumar is making some great suggestions to ollama! what a time to be alive!
I think there are already many OAI plans around this that are not public, it seems to be that it is much more about the execution then the idea in those two areas - take it with a grain of salt, its not my specialty and this is just in my humble opinion. I could be wrong, I donât know.
I have seen issues with o1-pro (in ChatGPT) too for coding. Does it overthink?
My working setup is using the Cursor IDE and o3-mini (regular). Havenât branched out much farther since it seems to be working well. But I have all the models and APIâs hooked to Cursor ATM.
I might agree with this, but there are not many AI specialists who are proficient in lower level languages, what they are proficient in is the math behind LLMs, Neural Networks, Transformers, etc., so a higher level language like Python gives them the powers to program what they need. When you say lower level languages, do you mean C, C++, Lisp?
1 Like
itâs your self contradiction. the point of the AI is that itâs supposed to provide the intelligence for you.
yes, obviously, if you leverage your own intelligence than you can make the o1 and most any of the llms work. But that kinda defeats the purpose
as for âcomplainingâ, clearly you missed the part where I said I use o1-pro and it works for me, I just happen to recognize the superiority of other solutions.
thatâs not a complaint, just realism that my $200 a month sub fee might not be worth it.
o1-pro is slow, yes, but the issue is more that I do a lot of work in the ML space where it is curiously weak.
As an example, if you ask it to build a simple model architecture it will always fail on tracking matrix shapes properly through the forward pass and you have to either re-prompt it many times or debug it manually.
Admittedly, ensuring that shapes are consistent through transformations is finnicky, but still, itâs not rocket science.
He should have deleted the post instead of getting fired for it⌠Writing about a forthcoming unreleased version of software is insane. He is not the product manager and he is not responsible for when and how it is released. Maybe there is some tweak they might do. God awful post on his part as an IC.
I also disagree with his rankings. o1-pro is at the top, sure, but it is very slow and expensive. I also donât think itâs available for cursor, @curt.kennedy do you know otherwise?
I stopped using cursor because it kept on crashing for me. Aider is more fun anyways, free, and good emacs integration. I also think it might be superior to cursor, but I admit I havenât done extensive testing. I strongly encourage folks to try it out if they havenât.
Cursor / api / aider integration is a must have as anyone can tell you. Also, cost per token is very important if itâs obscene like o1 versus o3-mini.
1 Like