Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Canvas is a new way to write and code with ChatGPT

openai.com

Post reply on HN
Print view View original

671–680 of 701 posts

Re: Canvas is a new way to write and code with ChatGPT

#671

Post by visarga »

ben_w wrote: post #559
visarga wrote: post #501

Earlier quoted context omitted.

LLMs are a more flexible platform than browsers. They can be prompted, finetuned or run locally. Even if a company wants to make their base model spit ads, it won't fly.

Depends how subtle they are about it, and what the rest of the ecosystem looks like. Perhaps the ad/ad-blocker analogy would be: You can have the free genuinely open source LLM trained only on Wikipedia and out-of-copyright materials, or you can have one trained on current NYT articles and Elsevier publications that also subtly pushes you towards specific brand names or political parties that paid to sponsor the mode…

There is huge choice in open models. People won't adopt one with ads baked in, unlike Google and Facebook, because now there are more options. There are 100K LLM finetunes on HuggingFace.

Re: Canvas is a new way to write and code with ChatGPT

#672

Post by 8338550bff96 »

swarnie wrote: post #248
8338550bff96 wrote: post #143

Earlier quoted context omitted.

75%? Thats astonishing to me. Where are you able to see those details? It wouldn't surprise me if not a lot of enterprises are going through OpenAI's enterprise agreements - most already have a relationship with Microsoft in one capacity or another so going through Azure just seems like the lowest friction way to get access. If how many millions we spend on tokens through Azure to OpenAI is any indication of what oth…

It may be pretty minimal but i can personally vouch for 20ish techies in my own social orbit who's businesses wont authorise or wont pay for OpenAI yet and are doing so out of their own pockets; i share an office with four of them. Maybe the consumer side will slide as businesses pick up the tab?

I don't understand what paying for openai is meant to mean? You mean paying for tokens?

Re: Canvas is a new way to write and code with ChatGPT

#673

Post by 8338550bff96 »

jdgoesmarching wrote: post #381
8338550bff96 wrote: post #143

Earlier quoted context omitted.

75%? Thats astonishing to me. Where are you able to see those details? It wouldn't surprise me if not a lot of enterprises are going through OpenAI's enterprise agreements - most already have a relationship with Microsoft in one capacity or another so going through Azure just seems like the lowest friction way to get access. If how many millions we spend on tokens through Azure to OpenAI is any indication of what oth…

This very good analysis estimates 73%, which includes team and enterprise. Given that enterprise access is limited and expensive, it seems Plus and Teams are mostly carrying this. The whole financial breakdown is fascinating and I’m surprised to not see it circulating more. https://www.wheresyoured.at/oai-business/

I prefer analysis from industry experts, not PR execs moon-lighting as tech bloggers

Re: Canvas is a new way to write and code with ChatGPT

#674

Post by famouswaffles »

UncleMeat wrote: post #647
famouswaffles wrote: post #511

Earlier quoted context omitted.

Between his roles at companies he's worked for, his research and his open source stuff, what else would you be looking for lol ? Remove all the educator stuff and karpathy would still be one of the most accomplished of his generation in his field. Idk just seems like a weird comment.

Research contributions are almost entirely separate from engineering quality. This is true for the vast vast vast majority of work people do in grad school.

Cool. Everything else ?

Re: Canvas is a new way to write and code with ChatGPT

#675

Post by dartos »

mooktakim wrote: post #628
dartos wrote: post #579

Earlier quoted context omitted.

I refuse to believe there were ever editors before vim. Vim has been around since the Stone Age. Jokes aside, I don’t really see why ai tools need new editors vs plugins EXCEPT that they don’t want to have to compete with Microsoft’s first party AI offerings in vscode. It’s just a strategy for lock-in. An exception may be like zed, which provides a lot of features besides AI integration which require a new editor.

They probably said the same thing when someone created vim, or vi.

Sorry, I’m not understanding what you mean.

Vi and vim were never products sold for a profit.

Who was saying what? And what were they saying?

EDIT: ah I think I understand now.

The thing is, I don’t see any advantage to having AI built into the editor vs having a plug-in. Aider.vim is pretty great, for example.

The only reason to have a dedicated editor is a retention/lock in tactic.

Re: Canvas is a new way to write and code with ChatGPT

#676

Post by ben_w »

visarga wrote: post #671
ben_w wrote: post #559

Earlier quoted context omitted.

Depends how subtle they are about it, and what the rest of the ecosystem looks like. Perhaps the ad/ad-blocker analogy would be: You can have the free genuinely open source LLM trained only on Wikipedia and out-of-copyright materials, or you can have one trained on current NYT articles and Elsevier publications that also subtly pushes you towards specific brand names or political parties that paid to sponsor the mode…

There is huge choice in open models. People won't adopt one with ads baked in, unlike Google and Facebook, because now there are more options. There are 100K LLM finetunes on HuggingFace.

I've got some of them on my experimentation laptop. They're only good enough to be interesting, not good in comparison to the private models, and the number of fine-tunes doesn't help with that. In particular I've had Microsoft's Phi 3.5 for less than a week and yet I've already had at least 4 cases of it spouting wild nonsense unrelated to the prompt — and I don't even mean that it was simply wrong, I mean the response started off with Chinese and then acted like it was the early GPT-3 "Ada" model doing autocomplete.

One of my machines also has a copy of Firefox on it. Not used that in ages, either. But Firefox is closer in quality to Chrome, than any of the locally-runnable LLMs I've tried are to the private/hosted LLMs like 4o.

Re: Canvas is a new way to write and code with ChatGPT

#677

Post by meowface »

UncleMeat wrote: post #194
meowface wrote: post #91

Earlier quoted context omitted.

I think you should try to give tools like this another chance. If Andrej Karpathy can say AI-assisted programming is a productivity boost for him ( https://x.com/karpathy/status/1827143768459637073 ), it can be a productivity boost for probably any programmer.

Andrej is a great communicator. I've never seen evidence that he is an especially exceptional programmer.

This is a ridiculous statement. It's possible he isn't an absolutely amazing programmer*, but given he's probably in the top 0.01% in terms of implementing novel (and often groundbreaking) ideas via programming and that he finds AI-assisted programming tools useful, it suggests AI-assisted programming is now more than just a CRUD code monkey assister.

*(it's also possible he is)

Re: Canvas is a new way to write and code with ChatGPT

#678

Post by dartos »

danielmarkbruce wrote: post #645
dartos wrote: post #569

Earlier quoted context omitted.

Sure, hand wave away my entire comment as “nonsense” and ignore how statistics works. Training a model on synthetic data (obviously) increases bias present in the initial dataset[1], making for poor training data. IIRC (this subject is a little fuzzy for me) using synthetic data for RLHF is equivalent to just using dpo, so if they did RLHF it probably wasn’t with synthetic data. They may have gone with dpo, though. […

Did you read this paper? No one is suggesting o1 was trained with 100% synthetic or 50% or anything of that nature. Generalizing that "synthetic data is bad" from "training exclusively/majority on synthetic data is bad" is dumb. Researchers are using synthetic data to train LLMs, especially for fine tuning, and especially instruct fine tuning. You are not up to date with recent work on LLMs.

> No one is suggesting o1 was trained with 100% synthetic or 50% or anything of that nature.

Neither was I.

> "synthetic data is bad“

I never said that… I said that it makes for poor training data, which it does.

> Researchers are using synthetic data to train LLMs, especially for fine tuning, and especially instruct fine tuning

Then those researchers are training with subpar datasets as the bias in that data will be compounded.

It’s a trade off since there’s only so much fresh data in form you want. If they could use entirely non synthetic data, I’m sure they would.

And again, you’re choosing to focus on this one point rather than my main point that prompt provide no moat.

> You are not up to date with recent work on LLMs.

There you go again making assumptions…

I think I’m done with this conversation though.

Re: Canvas is a new way to write and code with ChatGPT

#679

Post by gwern »

stavros wrote: post #582
ben_w wrote: post #560

Earlier quoted context omitted.

Unfortunately not. Trying to find where I got my version from just brought me back to one of my own comments on Hacker News from 8 months ago: > "your dog is juggling, filing taxes, and baking a cake, and rather than be impressed it can do any of those things, you're complaining it drops some balls, misses some figures, and the cake recipe leaves a lot to be desired". - https://news.ycombinator.com/item?id=39170057 I…

Oh hm, maybe you came up with it/adapted a previous saying, then. I'm not 100% sure the dog was singing either.

He is adapting one of Samuel Johnson's most famous quotations, about the astonishing sight of seeing a woman preaching - like a dog walking, it may not be done well, but it's astonishing to see it done at all.
Post reply on HN