Hacker News Reader

Comment by computerex | Hacker News Reader

What do you expect them to do?

Not train on and steal user data to front run frontier research for one thing. Not take credit for and independently solve these problems and instead sponsor researchers and give them credit. They could chose to include humans and frame AI as elevating human systems. Instead they are reveling in the fact that they did this in their own instead of humans. Thats a PR choice that is short sighted and reflects a selfish mindset not deserving of leading this transition

> Not take credit for and independently solve these problems and instead sponsor researchers and give them credit. They could chose to include humans and frame AI as elevating human systems. Instead they are reveling in the fact that they did this in their own instead of humans.

How do you know humans played the part that you think they did in this result? What evidence do you have to challenge their framing?

They are going for super intelligence, humans not necessary.

On one hand I don’t want to be replaced by an AI so I hate that this is happening. I don’t want AI to be controlled by the elites to enrich themselves further.

But on the other hand super intelligence will open doors for humanity. Maybe we will finally defeat cancer or death itself.

The first paper solved used a novel method that a small academic team had developed in chat sessions with both anthropic and openAI. One allegation is that anthropic started working on that problem and openAI got word and worked on it as well. They then deployed over $10m of internal token spend to solve it. When the researchers confronted openAI about their use of the same novel method they had explored which seemed an implausible coincidence, openAI couldn’t deny the private chats had entered the training data for the model. What’s it called when an employee gets hired at a competitor to exfiltrate proprietary data or methods? It’s called corporate espionage and theft. Why is it ok for an LLM to intercept and integrate private messages and the conceptual messages?

The fact that this history will be quickly forgotten by the public shows why the incentives of the Labs to behave unfairly is too great.

I fact checked your claim: https://chatgpt.com/share/6ac6e0a3-9534-83e9-a96f-64fb5f3246...

It's a bit more nuanced than the picture you are painting. I don't condone their behavior if that's what happened. But this post is about a much larger release.

Honestly, it shouldn’t even be possible to suspect and be unable to prove. They should be prohibited from using customers data as training data. That is copyright theft and corporate espionage. It wouldn’t be tolerated in other contexts. And it shouldn’t be legal to terms of service their way out of it.

https://youtu.be/GweKOBSJvEY?is=dDmkvcvFSBSJDaFY

I stand by my argument after reading your fact checking.

The fact checking states:

- OpenAI could not rule out the possibility that private conversations influenced its model. Partly supported - OpenAI actually trained on those private conversations to obtain the method. Unproven

Classifying the second point as unproven is a pedantic sleight of hand. It logically follows that if a model was influenced by a set of data to produce output incorporating it into its analysis it doesn’t really matter if it was “trained”.

Reply on news.ycombinator.com