HNHacker News
TopNewBestAskShowJobs

yousif_123123

415 karma · joined June 20, 2023

submissionscomments
yousif_123123··on ChatGPT-4o vs. Math
At the same time the ChatGPT app has access to write and run python, which the gpt can choose to do when it thinks it needs more accuracy.
yousif_123123··on Apple apologizes for iPad 'Crush' ad that 'missed the mark'
I didn't like the ad. I think the people creating it wanted to imply that it's as if they took all these things and put it in an iPad, where you can still achieve all the creativity while carrying a thin device.

I don't think it's impossible to convey that message without destroying instruments and creative tools that are precious to so many. Maybe if they had made the animation very fast it would've appeared as a joke and not something intended to be taken literally.

Also could've had some artist exit a studio, take the iPad, do a whole bunch of stuff, then go back to the studio and kind of test out/use the tools while reading from the iPad or something like that.

I know some people are saying the reaction is too strong, but trust me if you practice on a piano daily you will not feel good watching it get crushed.

I don't even work in marketing or own any Apple devices.

yousif_123123··on Volcker's Announcement of Anti-Inflation Measures
I wonder if such perseverance is possible today in what seems like a time of more intense politics, and with the presence of social media potentially making everyone see the bad side of things only.

It was a moment from the Fed (and government in general) where they favored long term prosperity at the expense of near term possible recession. I think we need a little more of that, not necessarily in what central banks do, but governments in general.

yousif_123123··on Show HN: Speeding up LLM inference 2x times (possibly)
I know this doesn't retrain, but I wonder if approaches like this plus quantization could get back any "lost" quality with some post training.

It's great to see, and to know in one's mind how much likely performance and cost will be better in the future.

I know it's fun to work on, but I also want to say THANK YOU for developing open source.

yousif_123123··on Ask HN: Did you encounter any leap year bugs today?
Even if it sees tokens, I don't think it's an impossible task. Certainly an advanced enough LLM should be able to decipher token meanings, to know that a word is made up of the individual character tokens regardless of how the full word is tokenized. Maybe something gpt5 can do (or there's some real technical limitation which I don't understand)
yousif_123123··on The Era of 1-bit LLMs: ternary parameters for cost-effective computing
Any models published as well?
yousif_123123··on "Coding" was never the source of value
It still probably will be better to know how to program. You can use the AI as well to teach you.

I think the assembly/higher level languages analogy may not be entirely accurate. If I'm making a web app and want some JavaScript for my special UI, an LLM can write a lot of versions based on how I prompt it. But it will always be faster if I already know how to write code, and using it to speed me up/do any boilerplate needed. That's why copilot is so useful, even if it's not the highest quality.

And it's really not hard to get the basics of programming.. that's why I think it should be taught. I don't see why you need to learn more chemistry than programming in school.

yousif_123123··on Google to pause Gemini image generation of people after issues
What makes me more concerned about Google is how do they let something like this pass their testing before releasing it as their flagship ChatGPT competitor? Surely these are top things to test against.

I am more disappointed in Google for having these mistakes than I am that they arrise from the early AI models when they're developed, as the developers want to reduce bias etc. This was not Google having an agenda imo, otherwise they wouldn't have paused it. This is Google screwing up, and I'm just amazed at how much they're screwing up recently.

Perhaps they've gone past a size limit where their bureaucracy is just so bad.

yousif_123123··on Unexpected responses from ChatGPT: Incident Report
So they don't have enough testing before deployments? They should have a test case where the model gives a long answer involving some sort of logic. It would've caught this.

Or maybe they skipped tests thinking this deployment wouldn't affect anything..

yousif_123123··on Sora: Creating video from text
Maybe if AI could do all jobs humans could do, we'd setup some system where the AI works and we don't since we tax them or somehow at least part of the created goods and services flows to everyone. Anything AI "creates" is worthless unless it's consumed, and AI being a machine/software won't inherently want to consume anything (like burgers for example).

I also struggle to think about all this, but I imagine if you can flip a switch and everything produced and consumed in the economy could be done in half the time, is that a good or bad thing? If we keep flipping that switch and approaching a point where everything is being produced with almost no human effort, does it become bad all of a sudden?

Somehow we'd need to distribute all this production, I'm not sure how it would work out, but just going from what we have now to half or 25% of effort needed is probably an improvement, at least I'd take that.

yousif_123123··on Is Bach the greatest achiever of all time?
He definitely achieved a lot, but its not useful to undermine others or say he achieved more (and therefore they are beneath him) since it's not really that kind of competition.

These people pushed forward human civilization. They are all an inspiration to us.

yousif_123123··on Why do programmers need private offices with doors?
Just want to add that I only became (at least in my opinion) a good programmer by sitting next to a very good and passionate one that I was able to question from time to time and may have interrupted his work some days.

I also probably helped him brainstorm ideas with me and debugging, since he would just say what's on his mind while doing it, and I'd do the same, and sometimes this kind of ping pong debugging/brainstorming together yields way better results than solo work. I really like one time he was trying to see how to open the chrome debugger to debug the chrome debugger, and I just knew how to do it (same shortcut while focussed on open debugger) since I had done it by mistake :D

I am convinced that you get more in a collaborative environment having a few devs sitting next to each other in pairs/small groups that they enjoy being with than anything else. IMO with remote work you get hidden inefficiencies that the company can attempt to fix by just hiring more people in the team (and a culture of let's do less meeting, less face to face, more work) takes over and that 1+1=more than 2 environment I was very excited about is gone..

Can someone who is a fan of remote work tell me how a new grad excited about writing code could find a mentor and really up their game in the remote world with what's becoming "standard" way of doing meetings etc?

A counterargument to what I describe is how successful open source projects have always been remote with passionate people working on them

yousif_123123··on Negative impact of vegetarian and vegan labels
I think a solution that I would prefer as a vegan and that wouldn't impact others is for the main menu to say at the bottom "ask for our vegan/vegetarian menu" so that for me it's easier to pick things, and no one else needs to see the labels. It also might invite the restaurants to provide more vegan options and potentially include some (without labeling) in the main menu.
yousif_123123··on How a startup loses its spark
When they say "100" or "150", do they usually mean total company size or number of engineering+product people?
yousif_123123··on Fine-Tuning Llama-2: A Comprehensive Case Study for Tailoring Custom Models
I'm not sure less data should require full fine-tuning. If I had 5 pages of text, I don't see why I need to train billions of parameters that are already trained pretty well on general internet knowledge, and already know how to chat..

From a practical perspective, unless cost is really immaterial, I think most will end up starting with Lora, especially for 13b or 70b models.. you could do 10 fine-tuning runs for the cost of a few full fine-tunings.

But it's still all witchcraft to me to some degree, and I'd probably try full and Lora.

yousif_123123··on Fine-Tuning Llama-2: A Comprehensive Case Study for Tailoring Custom Models
It's weird that Lora and training with quantization is not being taken more seriously. It's way cheaper, takes less time, and a lot of evidence shows it's pretty good.

I don't think it should be something brushed on the side to be tried out later..

yousif_123123··on Experiencing decreased performance with ChatGPT-4
They must be (and I hope) continuosly optimizing it. A lot of ideas with using quantization have been coming out this year, they must have learned new things.

The API should be more resilient but the ChatGPT app IMO should be expected to change in how it handles prompts, as it's constantly being fine-tuned etc.

Just like any saas product I think they have the right to update their software.

I still find GPT4 so powerful but also so prone to making huge mistakes (like calling a function in a library that doesn't exist when providing code)..

yousif_123123··on So long “prompt engineering,” we hardly knew ya
I don't think prompt engineering is the problem. I find myself with minimal education but a lot of trial and error being able to "engineer" a good prompt eventually if I know the problem and can validate answers.

I think the big problem is actually finding problems you could use LLMs in. I think anyone that's played with them tends to have a good guess at whether they could or couldn't do something (although you need to really test to be sure, do some prompt engineering etc), but actually finding problems to work on that are within the realm of being solvable and useful is the hardest part imo.

yousif_123123··on So long “prompt engineering,” we hardly knew ya
Anyone using a computer is "just" clicking on specific pixels and triggering keyboard events.. it still could be "actual engineering" imo if one is building a system following engineering methods, solving problems etc.
← PreviousPage 3 of 3