Lagergren

i don’t think you understand you spend weeks tuning hyperparameters your grid search runs for 2 days you finally get a +0.1% gain you celebrate you realize your random seed was fixed you try a new seed your model collapses you try five more seeds accuracy swings you look at the loss curves they look like shit you switch to bayesian optimization it finds the same hyperparameters gets stuck in a local minimum you add dropout it underfits you remove dropout it overfits you try early stopping it stops early you remove early stopping it learns nothing you read three blog posts about “superconvergence” you try a trick your model dies you load checkpoint it’s score is also shit somehow now you open github issue no one replies you post on stackoverflow they close as duplicate you ask chatgpt it suggests grid search or some basic shit you realize the only thing converging is your will to live i don’t think you understand

I can now, after two weeks of diligent work, verify that 2025 is not the year of the Linux desktop.

@GaryMarcus Lowering the bar for AGI only muddles progress. True AGI requires reasoning frameworks that ensure correctness, adaptability, and logical consistency. The dream of AGI will not be realized by statistical tricks but by building systems capable of true logical reasoning.

I'm trying to follow the teachings of @levelsio in 2025. Here is my first project that went live now. More projects to follow soon. Happy to listen to feedback!

This is great! As always, when @fbettag uses his lightspeed software development/AI skills, really professional apps suddenly start showing up. Always deployed with production quality. I wish I could do that.

If you use your default system Python, don’t worry about what’s in the vaccine.

@shipilev It’s good to have a random excuse cards to pull when you get performance questions. I like “before we look at that, could you make sure you are not just measuring L2 cache misses” and “Have you check compiler doesn’t remove your payload code”. I have maybe 50 of those.

Remember, Aleksey, if you write enough annoying reminders in your projects, expect conference speakers to show them on their slides on large screens. (Just as planned.)

Zoom added a creepy feature: AI Companion, which produces transcripts for the meetings. Creepy how? I just got an email with a transcript of a fairly personal family meeting. Never enabled it, WTF. Dear Zoom folks, adding features nobody asked for was a Skype downfall, beware.

@tagir_valeev I don’t think we are smarter, but just like money gets compounded over time by accumulated interest, accumulared records and wisdom have made it possible to start from a significantly better initial position.

Even if you knew the math, writing your thoughts and intermediate computations down that time was more challenging than using the modern tools like pen and paper. Nevertheless, somebody managed to do this. Do we really significantly smarter than our ancestors?

It's mind-blowing that somebody in ancient Babylon around 3700 years ago could compute a square root of 2 up to 6 decimal digits of accuracy. en.m.wikipedia.org/wiki/YBC_7289

@jawnsy Indeed. And for some things, certain tools work well and quickly gets you where you wanted to be. For other things, you may need to use something that is different. That is part of being an engineer.

Why anyone would use Python to develop production grade, ”may never fail” software is completely beyond me.

@saarw Yes. But it’s like crossing the river to get water, still. Do you have any good LLM cc frameworks to recommend?

@lagergren At least LLMs are here that can make it easier to generate 100%-coverage tests to take on the role of a compiler in a quality language

@Sharat_Chander @X I don’t even… You HAVE to tell me if they ever describe what you did.

@Sharat_Chander @X Wtf! What did you do, buddy?!

@PeterVeentjer Performance is a functional requirement

@mihn It grows back pretty quickly, but it’s more practical when it’s Scandinavian winter.

Most valuable wisdom from #jcrete: apparently, I am hard to recognize without my “viking beard”.

@shipilev And the customers who ignored their other patch update change procedures and allowed a 3rd party to apply code or configuration changes to their entire fleet at one time.