ma2rten

Not products, but Google Research already published papers about two different, better models:

https://parti.research.google/ https://imagen.research.google/

The models themselves are not public however.

There was an excellent talk from Jeff Dean a while back which went into a lot of detail.

I work in the AI field, but not on image generation.

I don't think it would be technically hard to build a model with current technology which can generate logos with the attributes which you mentioned. You could simply fine-tune a Dalle-E style model specifically on a smaller dataset of logos. This would just take a small dedicated team of domain experts to work on the problem.

This really depends on the team, and the individual productivity and ambition of the person in question.

I don’t know about other people but I tend to just watch every video until the end because they are so short that I don’t have enough time to decide if I like them or not.

Maybe I’m too old for TikTok.

In the beginning a lot of people were concerned about GDPR, because it could open any US business with a website up to liability. FAANG has lawyers to deal with this, but a small business or startup does not. But, I think those fears have calmed down.

I am one of the people who worked on Google's PaLM model.

Having skimmed the GitHub readme and medium article, this announcement seems to be very focused on the number of parameters and engineering challenges scaling the model, but it does not contain any details about the model, training (learning rate schedules, etc.), or data composition.

It is great that more models are getting released publicly, but I would not get excited about it before some evaluations have been published. Having a lot of parameters should not be a goal in and of itself. For all we know this model is not well trained and worse than Eleuther AI's 20B parameter model, while also being inconveniently large.

Common Crawl is also in WARC format.

Have you considered that it's not LeCun who is missing something? The AI safety community seems to be unfortunately almost completely separate from the actual AI research community and be making some strong assumptions about how AGI is going to work.

Note that LeCun had a reply in the thread and there was a lot more discussion which GP didn't quote.

According to my Google Research Google Workspace has 19.2% marketshare and Office 365 65%.

NLP has improved a lot in the last 5 years. I believe that this is now technically possible with the right training data.

At Google, it is very team-specific but it is generally not possible to run servers locally for testing and there generally is no staging environment, but it might be possible to bring up some part of the stack on Google's clusters. Those severs may read production data or some test data. Other than that there are static checks, unit tests, code reviews and the release process to ensure reliability.

But many senior candidates actually can’t code anymore.

Also, typically the point of these questions is to see if the candidate can solve a novel problem not if they can use a hash table.

Maybe there could be some kind of loophole where companies spin out their California operations into a separate company which has less revenue?

I get the impression that maybe DALL-E 2 produces slightly more diverse images? Compare Figure 2 in this paper with Figures 18-20 in the DALL-E 2 paper.

Google is very conservative about anything that can generate open-ended outputs. Also these models are still very expensive computationally.

My bad, it's a google-specific problem.

You mean like: https://say-can.github.io/

This is common in the research PA. People don't want to deal with broccoli man [1].

[1] https://www.youtube.com/watch?v=3t6L-FlfeaI

Caucasians are overrepresented in internet pictures.

Entrée refers to the main meal in the US.

That's not at all in the spirit of Turing's paper. He "party game" involves the interrogator knowing that they participating in it, talking to the machine for 5 minutes, posing serious challenges such as playing chess (at the time considered to be a hard problem for machines).

Do you feel the rating guidelines were well thought out? Is there anything you would change?

Did you work from an office or remotely? Did you interact with other raters?

Was there an interview process? How about performance reviews? How did Google ensure quality?

It doesn’t merge images. It generates them from scratch. Sure it’s trained on a corpus of existing images, but I don’t think it „merges“ them any more than human artists do with images they have seen in their lifetime.

It says it's using the pushshift api on the page.

You are misunderstanding how the technology works. It's trained on a large scale dataset of images, not art specifically.

The reason that it's producing a specific style is that Nick manipulated the text prompt and picked images he liked. He disclosed that in the twitter thread.

I don't know if they have the world's most powerful NLP toolkit (what does that even mean?), but they do have a strong team. I think it's worth a second look.

I didn't know about rg, how does it compare to ag?

I don't think torch is maintained by Meta AI (which is LeCun's devision), but some other part of Meta.

You are putting words into his mouth. He is not blaming Netflix and co for anything, he is just stating that he is personally not interested in doing anything online in his free time and that the reasons he got into FOSS are not valid anymore.

How does this work for clean air vehicles? Do they still need the transponder?