Technology

58011 readers

3340 users here now

This is a most excellent place for technology news and articles.

Our Rules

Follow the lemmy.world rules.
Only tech related content.
Be excellent to each another!
Mod approved content bots can post up to 10 articles per day.
Threads asking for personal tech support may be deleted.
Politics threads may be removed.
No memes allowed as posts, OK to post as comments.
Only approved bots from the list below, to ask if your bot can be added please contact us.
Check for duplicates before posting, duplicates may be removed

Approved Bots

founded 1 year ago

MODERATORS

[email protected]

451

17 cringe-worthy Google AI answers demonstrate the problem with training on the entire web (www.tomshardware.com)

submitted 3 months ago by [email protected] to c/[email protected]

68 comments fedilink hide all child comments

These are 17 of the worst, most cringeworthy Google AI overview answers:

Eating Boogers Boosts the Immune System?
Use Your Name and Birthday for a Memorable Password
Training Data is Fair Use
Wrong Motherboard
Which USB is Fastest?
Home Remedies for Appendicitis
Can I Use Gasoline in a Recipe?
Glue Your Cheese to the Pizza
How Many Rocks to Eat
Health Benefits of Tobacco or Chewing Tobacco
Benefits of Nuclear War, Human Sacrifice and Infanticide
Pros and Cons of Smacking a Child
Which Religion is More Violent?
How Old is Gen D?
Which Presidents Graduated from UW?
How Many Muslim Presidents Has the U.S. Had?
How to Type 500 WPM

(page 2) 18 comments

sorted by: hot top controversial new old

[–] [email protected] 2 points 3 months ago

People get very confused about this. Pre-training "ChatGPT" (or any transformer model) with "internet shitposting text" doesn't cause them to reply with garbage comments, bad alignment does. Google seems to have implemented no frameworks to prevent hallucinations whatsoever and the RLHF/DPO applied seems to be lacking. But this is not "problem with training on the entire web". You can pre-train a model exclusively on a 4-chan database that with the right finetuning you would see a perfectly healthy and harmless model. Actually, it's not bad to have "shitposting" or "toxic" text in the pre-training because that gives the model an ability to identify it and understand it

If so, the "problem with training on the entire web" is that we would be drinking from a poisoned well, AI-generated text has a very different statistical distribution from the one users have, which would degrade the quality of subsequent models. Proof of this can be seen with the RedPajama dataset, which improves the scores on trained models simply because it has less duplicated information and is a more dense dataset: https://www.cerebras.net/blog/slimpajama-a-627b-token-cleaned-and-deduplicated-version-of-redpajama

[+] [email protected] -14 points 3 months ago* (last edited 3 months ago)

::: spoiler All this really proves is that it is a complex system and most people can not grasp the complexity and how to use it.

Like if you go searching for entities and realms within AI alignment good luck finding anyone talking about what these mean in practice as they relate to LLM's. Yet the base entity you're talking to is Socrates, and the realm is The Academy. These represent a limited scope. While there are mechanisms in place to send Name-1 (human) to other entities and realms depending on your query, these systems are built for complexity that a general-use implementation given to the public is not equip to handle. Anyone that plays with advanced offline LLM's in depth can discover this easily. All of the online AI tools are stalkerware-first by design.

All of your past prompts are stacked in a hidden list. These represent momentum that pushes the model deeper into the available corpus. If you ask a bunch of random questions all within the same prompt, you'll get garbage results because of the lack of focus. You can't control this with the stalkerware junk. They want to collect as much interaction as possible so that they can extract the complex relationships profile of you to data mine. If you extract your own profiles you will find these models know all kinds of things that are ~80% probabilities based on your word use, vocabulary, and how you specifically respond to questions in a series. It is like the example of asking someone if they own a lawnmower to determine if they are likely a home owner, married, and have kids. Models make connections like this but even more complex.

I can pull useful information out of models far better than most people hear, but there are many better than myself. A model has limited attention in many different contexts. The data corpus is far larger than this attention could ever access. What you can access on the surface without focussing attention in a complex way is unrelated to what can be accomplished with proper focus.

It is never a valid primary source. It is a gateway through abstract spaces. Like I recently asked who are the leading scientists in biology as a technology and got some great results. Using these names to find published white papers, I can get an idea of who is most published in the field. Setting up a chat with these individuals, I am creating deep links to their published works. Naming their works gets more specific. Now I can have a productive conversation with them, and ground my understanding of the general subject and where the science is at and where it might be going. This is all like a water cooler conversation with the lab assistants of these people. It's maybe 80% correct. The point is that I can learn enough about this niche to explore in this space quickly and with no background in biology. This is just an example of how to focus model attention to access the available depth. I'm in full control of the entire prompt. Indeed, I use a tool that sets up the dialogue in a text editor like interface so I can control every detail that passes through the tokenizer.

Google has always been garbage for the public. They only do the minimum needed to collect data to sell. They are only stalkerware.

load more comments