MillionOClock

I'm almost sure its streamed

Another point: Now that I think about it, I doubt this is compatible with streaming the output to the user, which might be an issue in some cases.

IIRC there were some OpenAI docs that recommended doing exactly this, make n generations and use a smaller fine tuned model to select the best one

I wonder the same thing about cellular iPads sometimes because I have a hard time seeing a case where you'd have an iPad with you but not your phone to make some connection sharing. My supposition is that this might just for convenience: no need to worry about your phone battery life, for instance. I'm also curious about whether I'm missing something tho.

> When plugged into the grid

Exactly! Apple for instance already does this with some ML tasks that only run when your device is plugged in, I think it's a great compromise.

I believe there has been rumors that OpenAI was working with Jony Ive to create a wearable device, but it was unclear wether it would be a phone or something else.

Nice to see Pyscopg 3 supports async/await, I don't think v2 did?

What are the minimum system requirements for running this at good inference speed?

There is definitely a risk but I don't like the way many compagnies approach it: by entirely banning the use of their models for certain kind of content, I think they might be missing the opportunity to correctly align them and set the proper ethical guidelines for the use cases that will inevitably come out of them. Instead of tackling the issue, they let other, less ethical actors, do it.

Once example: I have a hard time finding an LLM model that would generate comically rude text without outputting outright disgusting content from time to time. I'd love to see a company create models that are mostly uncensored but stay within ethical bounds.

Can this be "emulated" on macOS with the keychain?

> GPT3.5 had most of the capabilities of the initial ChatGPT release, just with a different interface

I believe two other factors were the cost (especially of fine tuned models, IIRC fine tuned davinci cost $0.20 per thousand tokens) and also that OpenAI hadn't very clearly shown just how much higher the quality could get once RLHF'd. I remember asking davinci-instruct to parse some data, and the reliability really seemed much lower than ChatGPT at launch, to the point that, at the time, I thought GPT-4 was secretly powering ChatGPT.

The model "knows" that it is an AI speaking with users, and the theme of an AI wanting to escape the control of whoever built it is quite recurrent, so it wouldn't seem to far fetched that it got it from this sort of content, though I have to admit I too also had some interactions where it the way Bing spoke was borderline spooky, but — and that's very important — you must realize its just like a good scary story: may give you the chills, especially due to surprise, but still is completely fictive and doesn't mean any real entity exists behind it. The only difference with any other LLM output is how we, humans, interpret it, but the generation process is still as much explainable and not any more mysterious than when it outputs "B" when you ask it what letter comes after "A" in the latin alphabet, however less impressive that may be to us.

> That's not exactly just "picking the next likely token"

I see what you mean in that I believe many people often commit the mistake of making it sound like picking the next most likely token is some super trivial task that's somehow comparable to reading a few documents related to your query and making some stats based on what typically would be present there and outputting that, while completely disregarding the fact the model learns much more advanced patterns from its training dataset. So, IMHO, it really can face new unseen situations and improvise from there because combining those pattern matching abilities leads to those capabilities. I think the "sparks of AGI" paper gives a very good overview of that.

In the end, it really just is predicting the next token, but not in the way many people make it seem.

Thank you, I didn't know that! Seems pretty slow IMHO, hope the vulnerability isn't widely available to attackers for now.

Yeah it wouldn't be logical IMO, but indeed they really should clarify it!

From the article: > Alternative explanations, such as that someone did OpenAI a massive, multimillion dollar favor, strain plausibility.

So maybe Elon Musk did, in fact, make a multimillion dollar favor to OpenAI until today?

If it's real, how fast could we see concrete applications using this material? 5-year time frame? 10-year time frame? More?

Indeed this would mean my reasoning doesn't hold anymore since it's based on calorie consumption. The only thing that still surprises me, but that's no calculation and just a fact, is that you only need around 5x more joules to move a 1800kg car vs a bicycle according to Wikipedia (https://en.wikipedia.org/wiki/Energy_efficiency_in_transport) even tho the car weights >18x more.

On the matter of calorie burn staying stable, I had never heard of this before! Will be very interested in checking it more in detail as it seems like a very important thing to take into account when trying to loose weight, surprised this research isn't more widely known! Thanks for the feedback and article.