On Using AI For Commercial Uses

I will lose hearing out of spite if this happened to me

Genuinely admire your self awareness

I've lost hearing for stupider reasons. Spite seems downright reasonable to me.

I’ll waste a few moments. It becomes a puzzle. Assuming you managed to make it through the maze, you retrospectively analyze where would 99% of the country have dropped out of the flow and given up?

Then it’s an email to the attorney general if necessary! (I mean that’s been rare but when something is egregious)

🤓

Absolutely. Cc dispute is an under-used method of recourse.

A number of companies make "tuned" ear plugs to allow some sound through with a desired frequency curve, but reduce SPL to safe levels. I've used Etymotic, which sound great but I personally like a little more reduction. Alpine which I thought had enough reduction but too much coloring, and I settled on Earpeace, for like $25 on-line. Silicone, re-usable and easy to clean and they come with three filters to swap in or out depending on your needs / tastes.

Yeah this happened back earlier this year. I had lost a pair from a purchase years ago and replaced them. Guessing they are laying off people/support contracts like so many stupid business owners. I was sure that my order would be stuck in limbo forever after the experience, but they eventually showed up. Never again.

I just feel so mad. I want the old web back. I just want my duckduckgo to work well.

Same here.

Also: https://github.com/redlib-org/redlib-instances

Omg I can see it happening. Instead of annoying intrusive ads, this new type will be so natural as if your close friend is suggesting it.

More dystopian future. Yes we need it /s

I guess this really depends on the solution you’re working with.

I’ve built a voting system that relays the same query to multiple online and offline LLMs and uses a consensus to complete a task. I chunk a task into smaller more manageable components, and pass those through the system. So one abstract, complex single query becomes a series of simpler asks with a higher chance of success. Is this system perfect? No, but I am not relying on a single LLM to complete it. Deficiencies in one LLM are usually made up for in at least one other LLM, so the system works pretty well. I’ve also reduced the possible kinds of queries down to a much more limited subset, so testing and evaluation of results is easier / possible. This system needs to evaluate the topic and sensitivity of millions of websites. This isn’t something I can do manually, in any reasonable amount of time. A human will be reviewing websites we flag under very specific conditions, but this cuts down on a lot of manual review work.

When I said search, I meant offline document search. Like "find all software patents related to fly-by-wire aircraft embedded control systems” from a folder of patents. Something like elastic search would usually work well here too, but then I can dive further and get it to reason about results surfaced from the first query. I absolutely agree that AI powered search is a shitshow.

Congratulations to the person who downvoted this

They use a tool to improve their life?! Screw them!

Here’s hoping over the next few years we see little baby-sized language models running on laptops entirely devour the big tech AI companies, and that those models are not only open source but ethically trained. I think that will change this community here.

I get why they’re absolutist (AI sucks for many humans today) but above your post as well you see so much drive-by downvoting, which will obviously chill discussion.

Even though I understand your sentiment that different types of AI tools have their place, I'm going to try clarifying some points here. LLMs are machine learning models; the 'P' in 'GPT' – "pretrained" – refers to how it's already done some learning. Transformer models (GPTs, BERTs, etc.) are a type of deep learning is a branch of machine learning is a field of artificial intelligence. (edit: so for a specific example of how this looks nested: AI > ML > DL > Transformer architecture > GPT > ChatGPT > ChatGPT 4.0.) The kind of "vision guided industrial robot movement" the original commenter mentions is a type of deep learning (so they're correct it's machine learning, but incorrect that it's not AI). At this point, it's downright plausible that the tool they're describing uses a transformer model instead of traditional deep learning like a CNN or RNN.

I don't entirely understand your assertion that "LLMs are shit with large data and numbers", because LLMs work with the largest data in human history. If you mean you can't feed a large, structured dataset into ChatGPT and expect it to be able to categorize new information from that dataset, then sure, because: 1) it's pretrained, not a blank slate that specializes on the new data you give it, and 2) it's taking it in as plaintext rather than a structured format. If you took a transformer model and trained it on the "large data and numbers", it would work better than traditional ML. Non-transformer machine learning models do work with text data; LSTMs (a type of RNN) do exactly this. The problem is that they're just way too inefficient computationally to scale well to training on gargantuan datasets (and consequently don't generate text well if you want to use it for generation and not just categorization). In general, transformer models do literally everything better than traditional machine learning models (unless you're doing binary classification on data which is always cleanly bisected, in which case the perceptron reigns supreme /s). Generally, though, yes, if you're using LLMs to do things like image recognition, taking in large datasets for classification, etc., what you probably have isn't just an LLM; it's a series of transformer models working in unison, one of which will be an LLM.

Edit: When I mentioned LSTMs, I should clarify this isn't just text data: RNNs (which LSTMs are a type of) are designed to work on pieces of data which don't have a definite length, e.g. a text article, an audio clip, and so forth. The description of the transformer architecture in 2017 catalyzed generative AI so rapidly because it could train so efficiently on data not of a fixed size and then spit out data not of a fixed size. That is: like an RNN, the input data is not of a fixed size, and the transformed output data is not of a fixed size. Unlike an RNN, the data processing is vastly more efficient in a transformer because it can make great use of parallelization. RNNs were our main tool for taking in variable-length, unstructured data and categorizing it (or generating something new from it; these processes are more similar than you'd think), and since that describes most data, suddenly all data was trivially up for grabs.

Now it is pretty much one line of code.

… and 5kW of GPU time. 😆

That's why I put quotation marks around it

Yeah that's my point. AI has a lot of problems that need to be addressed, put people are getting so mad about AI that conversation around it is getting more and more extreme to the point people are talking about all AI being bad.

The problem is that anything even remotely related to AI is just being called "AI," whether it's by the average person or marketing people.

So when you go to a company's website and you see "powered by AI," they could be talking about LLMs, or an ML model to detect cancer, and the average person won't know the difference between the technologies.

So if someone universally rejects anything that says it "uses AI" just because what's usually called "AI" is just badly implemented LLMs that make the experience worse, they're going to inevitably catch nearly every ML model in the crossfire too, since most companies are calling their ML use cases "AI powered," and that means rejecting companies that develop models like those that detect tumors, predict protein folding patterns, identify anomalies in other health characteristics, optimize traffic routes in cities, etc, even if those use cases aren't even related to LLMs and all the flaws they often bring.

No they are the same thing.

The core algorithm we built upon is practically the same one used by AI image generators, the main difference is that we have deeper convolutional layers and more of them and we don't do any GAN stuff that the newer image generators use.