From the looks of it, it mainly trawls for info off of Reddit and Quora

The problem being that these sources are frequently wrong or deliberately jokey nonsense

But the AI doesn’t know that!

Yeah the Google one is noticeably worse than the latest ChatGPT models.

Most of the dumb responses it gets are old reddit jokes, so I wonder if they saw how people use reddit:search and decided that means they should train it on mostly reddit.

It also might just be a cheaper model.

The LLM is just summarizing/paraphrasing the top search results, and from these examples, doesn’t seem to be doing any self-evaluation using the LLM itself. Since this is for free and they’re pushing it out worldwide, I’m guessing the model they’re using is very lightweight, and probably couldn’t reliably evaluate results if even they prompted it to.

As for model collapse, I’d caution buying too much into model collapse theory, since the paper that demonstrated it was with a very case study (a model purely and repeatedly trained on its own uncurated outputs over multiple model “generations”) that doesn’t really occur in foundation model training.

I’ll also note that “AI” isn’t a homogenate. Generally, (transformer) models are trained at different scales, with smaller models being less capable but faster and more energy efficient, while larger flagship models are (at least, marketed as) more capable despite being slow, power- and data-hungry. Almost no models are trained in real-time “online” with direct input from users or the web, but rather with vast curated “offline” datasets by researchers/engineers. So, AI doesn’t get information directly from other AIs. Rather, model-trainers would use traditional scraping tools or partner APIs to download data, do whatever data curation and filtering they do, and they then train the models. Now, the trainers may not be able to filter out AI content, or they may intentional use AI systems to generate variations on their human-curated data (synthetic data) because they believe it will improve the robustness of the model.

EDIT: Another way that models get dumber, is that when companies like OpenAI or Google debut their model, they’ll show off the full-scale, instruct-finetuned foundation model. However, since these monsters are incredibly expensive, they use these foundational models to train “distilled” models. For example, if you use ChatGPT (at least before GPT-4o), then you’re using either GPT3.5-Turbo (for free users), or GPT4-Turbo (for premium users). Google has recently debuted its own Gemini-Flash, which is the same concept. These distilled models are cheaper and faster, but also less capable (albeit potentially more capable than if you trained model from scratch at that reduced scale).

My experience using Gemini for generating filler text leads me to believe it has a far worse grip on reality than ChatGPT does. ChatGPT generates conversational, somewhat convincing language. Gemini generates barely grammatical language that hardly even works as something that you can skim over without noticing glaring mistakes and nonsense sentences.

I think Google’s AI is particularly bad but there’s a larger point to be made about how irresponsible it is to deploy an LLM to do what Google is doing, regardless of how good it is.

Fundamentally, all an LLM knows how to do is put words together to form sentences that are close to what it has seen in its training data. Attention, which is the mechanism LLMs use to analyze the meaning of words before it starts to figure out its response to a prompt, is supposed to model the real-world meaning of each world, and is context dependent. While there’s a lot of potential in attention as a mechanism to bring AI models closer to actually being intelligent, the fundamental problem they have is that no fancy word embedding process can actually give the AI a model of meaning to actually bring words from symbols to representing concepts, because you can’t “conceptualize” anything by multiplying a bunch of matrices. Attention isn’t all you need, even if it’s all you need for sequence processing it’s not enough to make an AI even close to intelligent.

You can’t expect a language model to function as a search engine or an assistant because those are tasks that need the AI to understand the world, not just how words work, and I think ultimately it’s gonna take a lot of duds and weird failures like this Google product before tech companies find the place to put the current generation of LLMs. It’s like crypto, it blew up and got everywhere before people quickly realized how much of a scam it was, and now it still exists but it’s niche. LLMs aren’t gonna be niche at all, even once the VC money dries out, but I highly doubt we’ll see too much more of this overgeneralization of AI. Maybe once the next generation of carbon guzzling machine learning models that take several gigawatts to operate in a datacenter the size of my hometown is finally ready to go, they’ll figure out a search engine assistant.

“AI” is not an information tool, but it’s being marketed as one. These machine learning algorithms do not possess any knowledge. All a LLM is doing is mimicking human speech patterns, it knows how to put words in a grammatically correct order that are relevant to the input prompt and that’s about it.

Machine learning has some practical uses but giving factual information, instruction, or advice are not among them. The idea of incorporating this shit into a search engine is fucking ridiculous.

I believe the google one literally just summarizes the top result, which is often an onion article or reddit post or something. It isn’t prompted by your query, it’s prompted by the search results

Edit: beaten to the punch by JohnBrownsBussy2

I think part of it is that Genini is further behind ChatGPT and the training model isn’t all that great. Part of the training that Google uses for Gemini comes from Reddit, as part of deal in which Reddit shares info with Google to “improve” Gemini. Not sure about it using info from other AI models to train, but sounds about as dumb as using Reddit.

We are seeing something called model collapse right now which is the result of training a ML model off it’s own content. Because ML content is likely the majority of content on the Web over the past two years and these models require constant training data, the models are falling apart due to circular input.

Sorta relatedly, I remember back before AI the most common reddit bots would just post a top comment from the same askreddit thread or a similar one. It’s a weird feeling to think that everyone seems to more or less have forgotten about this as the main way to do account farming for a while. The focus on AI modifying the response makes no sense to me. The ‘content’ literally doesn’t matter so why do the more computationally expensive thing? There were plenty of methods that people used to reword and avoid bans too. Like checking the thread for the same text and picking a different comment, having a table of interchangeable common words, etc, etc. The AI methods seem to be literally worse than just a copy-paste fuzzy matching (or markov chain etc etc).

Google in particular is in a bit of a bind because they were late to the AI bandwagon. The number of people who actually understand these generative tools well enough to be project leaders on them is very small, so it’s not a problem that they can just throw a ton of money at to solve - they have to build their institutional knowledge about LLMs and their products from the ground up the hard way.

Conflicting with this is the pressure to release something to please the investors. So Google has an unfinished product made by people who are still learning how it works, and while the smart thing to do would be to keep it under wraps until it can compete with ChatGPT they’re forced to polish that turd and release it anyway.

Model collapse and all the other things people are talking about are factors to keep an eye on, but I think that’s the primary reason why Google’s go at AI was so fantastically worse than the others we’ve seen.