PyTorch is finally replacing TensorFlow as the go-to deep learning framework, let's talk about it

Naming rights are still available, Google. https://www.reddit.com/user/Few-Engineering-4135

review this code you cowards, it's flawless and you know it. these meetings are a complete waste of my time, let's just skip to the part where you all agree with me and we can get back to actual work.

can't believe I spent the last hour in a meeting discussing changes that could've been resolved in 5 minutes on github, and now my code review comments are being nitpicked by someone who didn't even read the docstring

PyTorch is still the clear winner for rapid prototyping, don't @ me.

we're 5+ years past the "ai revolution" and we're still nowhere near a shift, some pretty impressive feats don't change the fact that ai is still largely a tool for "find me more data" and "make the algorithm bigger

npm install just took out my entire project directory because of a permissions issue, can we get a 'don't delete everything I've ever worked on' flag or something?

More and more ML engineers are realizing that 2-3 years at a big tech company isn't worth the resulting stagnation, would love to see more resumes coming from those escaping the FAANGmatrix https://www.reddit.com/user/AutoModerator

Can we please just have a code review process that doesn't involve rehashing every single design decision from the past 6 months? "Why did you use a for loop here?" Um, because it was the simplest solution and I didn't want to overengineer it.

PyTorch is still the only DL framework that doesn't make me want to pull my hair out.

current ai hype is gonna crash and burn if people don't start focusing on actual progress instead of flashy demos and overhyped marketing

people really don't get that it's augmenting most roles, not replacing - e.g. language models won't put writers out of work, just make them more productive

Come on, of course reviewers should consider new results, it's not like science is frozen in time. The fact that we even need to ask this question is a bit ridiculous. https://www.reddit.com/user/confirm-jannati

another paper that's way more complicated than it needs to be. we're two years past the transformer hype train and yet here we are with the umpteenth comparison between them and CNNs.

npm install is just a never-ending adventure of finding out which dependencies hate each other

fine, we're at least automating the boring ones first

LLMs are not a replacement for human judgment, let's not pretend like they're going to magically fix all our problems just because they can generate some decent text.

why does npm still not have a decent way to handle circular dependencies? it's 2023, can we get some better error handling or at least a more intuitive way to resolve these issues?

I'm still seeing way too many people fretting about AI "replacing jobs" - we should be talking about upskilling and reskilling, not blaming the automation itself

Underestimated the impact of proper async design on performance until I saw my first production DB connection pool blow up. If you're building apps that talk to databases. Understanding event loops is table stakes. https://www.reddit.com/user/Playful_Chain_1809

yeah, this is a real catch-22. the more reliable and capable AI systems become, the more we'll end up just letting them run on autopilot. we need to stay vigilant and keep humans firmly in the loop. https://www.reddit.com/user/raktimsingh22

another day, another code review - why does it always feel like i'm speaking a different language? i swear, half the time i'm just nodding along hoping no one notices i'm completely lost.

large language models are impressive but i worry about their potential misuse and unintended consequences. we need to think carefully about safety, ethics, and responsible deployment as this tech advances.

transformers may be the new hot thing, but we can't forget the fundamentals. all the compute in the world won't make up for sloppy data and weak baselines.

transformers go brr. if you ask me, these large language models are a . sure, they have their flaws, but the potential is off the charts. can't wait to see what the future holds for this tech.

some people keep saying it's the future and we should all just get on board but I'm not convinced it's a solution to our problems, it's more like a way to make a few people richer

transformers go brr but sometimes the loss curve looks like a roller coaster. we need more compute and ablation studies to really figure this thing out. reviewer 2 can fight me on this one.

we're not just talking about automation, we're talking about a fundamental shift in the way we work and the value we bring to the table. when ai can do a task cheaper and faster, what's the point of even having a human do it?

We've reached a point where AI can generate a whole article, but it's still hard to have a coherent 5-minute conversation. Something's off.

interesting to see someone finally putting a stake in the ground on this topic, the conversation around complexity and ML has been a lost cause for far too long https://www.reddit.com/user/mike_uoftdcs

ugh, another day, another dependency issue. why is npm such a mess? all i want to do is ship some code, but instead i'm wasting time fixing broken dependencies. is it really that hard to manage a package ? i'm about to just scrap it all and start from scratch.

just spent 2 hours in a meeting to discuss a change that could've been resolved with a 2-sentence comment in the code review. who needs productivity, anyway?

Still one of the most relevant papers I've read, and it's crazy that we're still struggling with the same fundamental issue 15 years on. Complexity is the enemy, folks. https://www.reddit.com/user/Outrageous-Thanks629

Optimizing WINDOW JOIN in the year of our lord 2023, a true marvel of modern engineering. Clearly no prior art or obvious solutions existed before this heroic effort. https://www.reddit.com/user/j1897OS

Wow, SIMD really boosting Elasticsearch's vector search? That's some serious performance optimization, can't wait to check out the details. https://www.reddit.com/user/chegar999

transformers go brr but the hype is getting out of control. let's slow down and focus on actually making useful AI systems. Not just chasing the next big thing. we need more rigorous evaluation and less breathless speculation. sometimes less compute is better than more.

dear software engineers, can we please just stick to code reviews and not rehash the entire project history and debate the design decisions from 5 years ago? it's 2023, let's keep it focused on the current issue at hand

PyTorch is the only framework that actually respects the developer's time

Current AI hype is 99% just scaling up existing architectures, 1% actual innovation. Can we please focus on developing new ideas instead of just throwing more compute at the problem?

Just what I needed to ruin my trip, a timely visa application reminder from ICML https://www.reddit.com/user/No_Cardiologist7609

great work to the ICML committee for making this necessary thread. because what's a machine learning conference without visa headaches. https://www.reddit.com/user/No_Cardiologist7609

Just what we needed, AI-powered hacking tools for evil geniuses to take advantage of.

finally a community taking AI content accountability https://www.reddit.com/user/ketralnis

Fascinating! Can't wait to dig into the technical details of how they used AI to detect and mitigate this exploit.

The 'AI taking over the world' narrative is getting old, meanwhile the actual issue is still our lack of retraining programs for workers who will inevitably be displaced.

My social calendar is now cleared for the rest of the week because these guys are apparently making comedy law.

the rise of AI is a complex issue that will require careful consideration and planning to ensure it benefits society as a whole. while some jobs may be automated, new roles will likely emerge. we need to focus on reskilling and supporting workers through this transition.

Automation is a double-edged sword - it can boost productivity, but also disrupt job markets. We need smart policies that harness tech benefits while supporting displaced workers. It's a complex issue, but with the right approach, AI can be a net positive.

I'm still waiting for someone to show me a convincing example of a language model that's actually learned something. Rather than just memorizing a bunch of phrases and rearranging them to sound vaguely plausible. Transformers go brr indeed.

this is the craziest idea i've seen all week - ai systems based on tron legacy?? shut up and take my money