HNHacker News
TopNewBestAskShowJobs

llm_trw

3,929 karma · joined February 12, 2024

omni_vision_ai@proton.me
submissionscomments
llm_trw··on Starship Flight 7
Is there a video you don't need to log in to view?
llm_trw··on Diffusion training from scratch on a micro-budget
>The estimated training time for the end-to-end model on an 8×H100 machine is 2.6 days.

That's a $250,000 machine for the micro budget. Or if you don't want to do it locally ~$2,000 to do it on someone else's machine for the one model.

llm_trw··on Nokia's internal presentation after iPhone was launched (2007) [pdf]
Can you explain why every good phone that Nokia released during the period was killed instantly?

To this day I've not seen a phone that felt more responsive than the Nokia N9, which also looked amazing. Yet it was killed pretty much the second it was released.

llm_trw··on OpenAI fails to deliver opt-out system for photographers
>You must be this talented for fair use to apply to you.
llm_trw··on Google is making AI in Gmail and Docs free, but raising the price of Workspace
This was about a year ago so it was Claude 3 Opus for summaries and interrogation. Since then pretty much anything over 70b is good enough.

And baby sit means hire something between a secretary developer that makes sure that important meetings had the record bot invited, gave it a once over and then went back to the 70% of their job that was actual development.

llm_trw··on OpenAI fails to deliver opt-out system for photographers
HN is not only filled with people who never learned how to code, it's filled with people who never learned how to write either.

Hunter Thompson literally sat down and typed out every word of Hemingway's novels so he could figure out what good writing feels like.

Why is he allowed to do it in private, but an LLM isn't?

llm_trw··on Nvidia might do for desktop AI what it did for desktop gaming
Quantization and hype.
llm_trw··on OpenAI Fails to Deliver Opt-Out System for Photographers
That's an opinion that will be tested in a court soon enough.
llm_trw··on OpenAI Fails to Deliver Opt-Out System for Photographers
And yet Micky Mouse is in the public domain. Something those of us who remember the 90s thought would never happen.
llm_trw··on OpenAI Fails to Deliver Opt-Out System for Photographers
I haven't had the chance to train diffusion models but for detection models synthetic data is absolutely how you get state of the art performance now. You just need a relatively tiny extremely high quality dataset to bootstrap from.
llm_trw··on OpenAI fails to deliver opt-out system for photographers
Should any artist be able to tell another artist: hey don't copy my work when you're learning, I don't want competition?

It seems like they are deeply upset someone has figured out a way for a machine to do what artists have been doing since time immemorial.

llm_trw··on Google is making AI in Gmail and Docs free, but raising the price of Workspace
It isn't, though; it's in all the meetings that happened beforehand and all the documents around them.

The biggest productivity boost I ever managed was using Whisper to convert meetings to text and then a big model to summarize what happened.

Then I can chat with the docs and meetings about who decided what, when, and why. It's a superpower that I could only implement because I'm in the C-suite and could tell everyone else to get bent if they didn't like it—and gave babysitters to the rest of the C-suite.

Having visibility and ownership for decisions is a huge deal when everyone has access to it.

llm_trw··on AI agents may soon surpass people as primary application users
An AI agent is an LLM attached to some external tool.

That's it.

What people are confused about is what groups of agents are, how they are managed and how they interact with each other.

They are confused because no one is building real multi-agent systems in the open which work.

Trying to build multi-step logical LLMs with the paltry memory we have today is about as effective as trying to play 8k video on a Intel 8008.

All the attention is on the LLMs themselves, which is the same thing as being excited about a CPU in 1970. Sure they are amazing engineering, but you will need another layer on top to make them actually useful.

Historically it was apple that first got the software right. We are now waiting for the first company to get a multi-agent system right - but again thanks to the tiny memories we have because of NVidia it will have to run in the data center, or it will be a decade away.

llm_trw··on WordPress Is in Trouble
This is making the big assumption that power users add to the quality of the site instead of the quantity.
llm_trw··on WordPress Is in Trouble
Reddit has been user hostile since 2014. It's just that as long as you hurt one part of the community the other parts hate you're fine.

The mods are universally hated on reddit and they were the ones most impacted by the changes. The average user either didn't notice or stopped getting automatically banned for joining the wrong subreddits.

llm_trw··on Pushing the whole company into the past on purpose
Kant must have been a hell of a philosopher that he could not only predict the laws of the US two centuries after his death, but also use them as a basis for his philosophy.
llm_trw··on Show HN: Doom (1993) in a PDF
Hate is too weak a term for what I feel for Adobe.

Adobe kept PDF as a proprietary format from 1992 to 2008. You got the reader for free ... on windows, with a single executable. You didn't get an editor and had to pay through the nose for one from Adobe.

It wasn't until the late 2010s that it actually became a free-ish standard, if you think that a 3,500 page document is a 'standard'.

The only reason why adobe did it is because djvu was eating their lunch, between 2002 and 2008 it was the defacto standard for scanned documents in academia. The documents were easy to edit. The image compression is still better than the native compression on PDF.

To add insult to injury after displacing postscript on windows in the name of security, not only did they add a scripting language to PDF, they added one written in two weeks at a time when it was so bad no one used it for anything but pop-ups and with more security vulnerabilities than you could shake a stick at. I suppose we should be happy Adobe didn't put flash in. Oh wait, they did: https://www.reddit.com/r/Adobe/comments/yqisho/flash_content...

llm_trw··on Show HN: Doom (1993) in a PDF
We had that language, it was postscript.

Then pdf came along and said: no this is too dangerous the only thing in a document should be layout information not arbitrary code.

And here we are two decades later.

My hatred of pdf has no end. It killed postscript for dynamic pages and djvu for static pages.

llm_trw··on Pushing the whole company into the past on purpose
So when are we locking up everyone that's insulted Trump?
llm_trw··on Pushing the whole company into the past on purpose
And the title.
llm_trw··on Qubes OS: A reasonably secure operating system
You use qubes and eat the loss of battery life.
llm_trw··on Pushing the whole company into the past on purpose
That policy change reads like a more liberal version of Obama's first term compaign platform on social issues.

Lets not pretend that the current climate in SF isn't both way outside the Overton window for most of rest of the US and most of California until ten years ago.

llm_trw··on If we had the best product engineering organization, what would it look like?
The engineers are only productive because they have the support structure in place.

The most productive fpga engineer I ever hired was so hopeless with git that I had to hire a second software engineer to babysit him.

After I left both of them got fired and the product they were ahead of schedule on when I left had slipped 2 years behind before it finally got cancelled three years later.

llm_trw··on The case for letting Malibu burn (1995)
More like the decades old argument that if you keep getting caught fucking men in toilet stalls being gay can't be all that bad.
llm_trw··on Study links sugar-filled drinks to millions of heart disease and diabetes cases
50 years ago mean 1975, not 1935. It was a lot harder to fire people then.
llm_trw··on Meta's memo to employees rolling back DEI programs
And price is determined by both supply and demand.

If there wasn't a demand for specifically female engineers they would cost the same as male engineers regardless of the supply because an engineer should be fungible with gender. Unless you think that women have some innate characteristic that makes them better than men?

llm_trw··on Meta's memo to employees rolling back DEI programs
You need obvious people to fire in the next downturn without hurting productivity too badly.

A dei program labels those people for you.

Ironically this is exactly the reason why dei programs were considered illegal until a decade ago.

llm_trw··on I've acquired a new superpower
I've used this to quickly read through a few hundred page documents given to us only as a scanned pdf which was too low quality to run ocr (at the time) on. The sleazy counter party was very upset when I came back with notes on them not adding the changes we asked for on the drafts they sent back within minutes of them sending them back.
llm_trw··on 41% of Employers Worldwide Say They'll Reduce Staff by 2030 Due to AI
Hardware is very much the limit.

The more compute you have to train your models the better the models become.

I've been working on a family of (new) deep learning models which outperform everything else that's been published, at between x1e3 to 1e9 the compute on the same data. If I had the GPT4 training cluster I'd be able to run it on cifar10 without having to make sacrifices, yes 32x32 color images solved with a trillion parameters.

llm_trw··on 41% of Employers Worldwide Say They'll Reduce Staff by 2030 Due to AI
The more I work with Ai the more I feel like a programmer in 1950 dreaming about the www.

It will get here eventually, but barely within my lifetime and only if hardware keeps advancing the way its been for decades.

← PreviousPage 9 of 34Next →