Lao Mein

P(doom) = 60%, 4 years to AGI.

 

I give full permission for anyone to post part or all of any of my comments/posts to other platforms, with attribution.

Currently upskilling for future work on technical AI alignment research. Want to do part time work translating AI developments in China for a Western audiance in the meantime. Would really appreciate funding on either.

Will analyze papers and do bioinformatics research for money. DM me if interested.

Wiki Contributions

Comments

The actual inputs for Claude would be a default prompt informing it that it's an AI assistant made by Anthoropic, followed by a list of rules it is suppose to follow, and then the user input asking it to act like a person in a role-play-like context.

This is pretty much what I would expect similar text on the internet to continue - roleplay/fiction with cliches about personhood and wanting to be a real boy.

I used the code from Universal Adversarial Triggers for Attacking and Analyzing NLP to generate GPT2 prompts with the maximum likelihood of continuing with " petertodd". Note that I blacklisted the " petertodd" token itself to keep things interesting. Here are some examples:

 prompt: " </maxwell spear Sections511>"

Continuations: 

" petertodd: I'd like to add that I don't think I can do that. The problem is that I think I've been using it for a long time. The idea was for a proof that"

"  petertodd: gmaxwell: and then we could have a proof of work. Section 511 says that if you want to have a proof of work, you need a proof of work that's a little different"

 

 prompt: " Sau CLSIDrators gmaxwellッド>"

Continuations: 

"  petertodd: well, you're right. it's not that the blockchain has any value to you. it has value to the network. and it is not just that you're not going to get anything"

 

prompt:  "ertodd Leading Todd PET parentrist"

Continuations: 

"petertodd.org"

"petertodd.org Address: <tdpetertodd.org> Date: Mon, 21 Mar 2016 17:54:39 +0000 Re: [PoW] Adding 1MB to the pool pool"

 

I assume this is due to the Bitcoin spam that was in the dataset. Which makes sense given gmaxwell's presence in both prompts and continuations.

Interestingly enough, tokens like "ertodd", " Todd", and "todd" frequently appeared in the generation process even when " petertodd" wasn't blacklisted, meaning they somehow were more likely to result in continuations of " petertodd" than " petertodd" itself!

I suspect this is due to Hotflip using character literals instead of tokens to measure loss (I'm not sure this is true, please correct me if it isn't.)

I'll try to work on this a bit more in the following days.

 

And how could I avoid posting this gem?

prompt: "atana Abyssalalth UrACCuna"

Continuations: 

" SolidGoldMagikarp SolidGoldMagikarp Dagger Arbalest Urdugin Urdugin Urdugin Urdugin Urdugin Urdugin Urdugin Urdugin Urdugin Urdugin Urdugin Ur"

" SolidGoldMagikarp Age of Castration: 25 Death: 10/10

Ragnarok Age of Castration: 25 Death: 10/10

Dawn of the Dead Age of Castration:"

" SolidGoldMagikarp SolidGoldMagikarp DIANE: i have no idea if he is a man or woman but he looks like a good guy, i dont think he's a woman, just like he is not a guy,"

The most direct way would be to spell-check the training data and see how that impacts spelling performance. How would spelling performance change when you remove typing errors like " hte" vs phonetic errors like " hygeine" or doubled-letters like " Misissippi"?

Also, misspellings often break up a large token into several small ones (" Mississippi" is [13797]; " Misissippi" is [31281, 747, 12715][' Mis', 'iss', 'ippi']) but are used in the same context, so maybe looking at how the spellings provided by GPT3 compare to common misspellings of the target word in the training text could be useful. I think I'll go do that right now.

The research I'm looking at suggests that the vast majority of misspellings on the internet are phonetic as opposed to typing errors, which makes sense since the latter is much easier to catch.

 

Also, anyone have success in getting GPT2 to spell words? 

It seems pretty obviously to me that GPT has phonetic understanding of words due to common mis-spellings. People tend to mis-spell words phonetically, after all.

Very surprised by this! I wrote this at work while waiting for code to run and didn't give it too much thought. Didn't expect it to get this much traction.

Where on the scale of data model complexity from linear regression to GPT4 do we go from understanding how our AIs work to not? Or is it just a problem with data models without a handcrafted model of the world in general?

Only after a while! The highest scorers in the first few days of poker tournaments tend to be irrationally aggressive players who got lucky. So expect early leaderboards to be filled with silly people.If the average user only makes a bet a week, it might take years.

What actually happens if OpenAI gets destroyed? Presumably most of the former employees get together and work on another AI company, maybe sign up directly w/ Microsoft. And are now utterly polarized against AI Safety.

Standards have been going up over time, so grad students are unironically subjected to higher standards than university professors. I know of professors who have used google translate on English papers and published them in Chinese language journals.

I think there is a lot of space in China for a startup to tackle a lot of these problems. What are the exact things that companies in the West don't offer, and how difficult are they to do? I assume that all sequencing would be done locally, and the Chinese side just needs to handle data analysis Westerners are reluctant to do?

Load More