611–620 of 817 posts
OpenAI's models feel 100% nerfed to me at this point. I had it solving incredibly complex problems a few months ago (i.e. write a minimal PDF parser example), but today you will get scolded for asking such a complicated task of it. I think they programmed a classifier layer to detect certain coding tasks and shut it down with canned BS. I like to imagine certain billion/trillion-dollar mega corps had a back-room say…
Earlier quoted context omitted.
Who has the necessary resources to run, let alone train the model?
How feasible would it be out crowdsource the training? I.e. thousands of individual macbooks training a small part of the model and contributing to the collective goal
Earlier quoted context omitted.
it's also really, really bad and fails compared to even open source models right now.
God, what happened to Google. What a fall from grace. Alpaca is pretty good though.
They have no leadership at the top. Nobody that can steer the ship to the next land (or even anybody that has a map). Who is actively working at Alphabet that has the authority to kill Google search through self-cannibalization? Absolutely nobody. They're screwed accordingly. It takes an enormous level of authority (think: Steve Jobs) and leadership to even considering intentionally putting at risk a $200 billion sales product. The trick of course is that it's already at great risk.
They don't know what to do, so they're particularly reactive. It has been that way for a long time though, it's just that Google search was never under serious threat previously, so it didn't really matter as a terminal risk if they failed (eg with their social network efforts; their social networks were reactive).
It's somewhat similar to watching Microsoft under Ballmer and how they lacked direction, didn't know what to do, and were too reactive. You can tell when a giant entity like Google is wandering aimlessly.
Earlier quoted context omitted.
Yeah I fully agree about compute cost, though I wonder why they don't just introduce another payment tier. If people are really using it at work as much as claimed online, it would be much preferable to be able to pay more for the full original performance, which seems win/win.
Because that involves telling customers that the product they are paying for is no longer available at the price they were paying for it. Much smoother to simply downgrade the model and claim you're "tuning" if caught.
Maybe their partnership with Microsoft changes the dynamics of how they handle their direct products though.
Earlier quoted context omitted.
> No amount of “bias = pattern recognition” nonsense can justify a system that has (had? this was a while ago and I have not retested) such extreme biases One possible explanation is that when you ask for 100 example families the task is parsed as "pick the most likely family composition and add a bit of randomness" and "repeat the aforementioned task" 100 times. If phrased like that it would be surprising to find on…
> You're right about the difference between one-by-one prompts and prompts that create a population. I switched to sets of 10 at a time and it got better. But still, when you ask for "make up a family", the model should not interpret that as "pick the most likely family". I disagree with your opinion that it's hard. GPT does not work by creating a pool of possible families and then sampling them; it works by picking…
But that's literally what LLMs do.... You don't get a choice with this technology.
OpenAI's models feel 100% nerfed to me at this point. I had it solving incredibly complex problems a few months ago (i.e. write a minimal PDF parser example), but today you will get scolded for asking such a complicated task of it. I think they programmed a classifier layer to detect certain coding tasks and shut it down with canned BS. I like to imagine certain billion/trillion-dollar mega corps had a back-room say…
I understand that there is a desire to tweak the model and improve it, and that's likely the way to go for the "consumer" chat application; however, both for science and business, there is a dire need to have an API that allows you to pin a specific version and always query the same old model instead of the latest/greatest one. Do we need to ask the LLM vendors to provide a "Long Term Support" release for their model API?
OpenAI's models feel 100% nerfed to me at this point. I had it solving incredibly complex problems a few months ago (i.e. write a minimal PDF parser example), but today you will get scolded for asking such a complicated task of it. I think they programmed a classifier layer to detect certain coding tasks and shut it down with canned BS. I like to imagine certain billion/trillion-dollar mega corps had a back-room say…
It's a big pain when trying to build things on top of the GPT-4 API. We had some experiments that were reliably, reproducibly achieving a goal, and then one day it suddenly stops working properly; then the student managed a different prompt that worked (again, reproducibly, with proper clean restarts from fresh context), and within a few days it broke. I understand that there is a desire to tweak the model and improv…
It’s been mostly fine for me, but overall I am tired of every answer having a paragraph long disclaimer about how the world is complex. Yes, I know. Stop treating me like a child.
Prompt it to do so. Use a jailbreak prompt or use something like this: "Be succint but yet correct. Don't provide long disclaimers about anything, be it that you are a large language model, or that you don't have feelings, or that there is no simple answer, and so on. Just answer. I am going to handle your answer fine and take it with a grain of salt if neccessary." I have no idea whether this prompt helps because I…
Earlier quoted context omitted.
Currently, not at all. You need low latency, high bandwidth links between the GPUs to be able to shard the model usefully. There is no way you can fit an 1T (or whatever) parameter model on a MacBook, or any current device, so sharding is a requirement. Even if it that problem disappeared, propagating the model weight updates between training steps poses an issue in itself. It's a lot of data, at this size.
You could easily fit a 1T parameter model on a MacBook if you radically altered the architecture of the AI system. Consider something like a spiking neural network with weights & state stored on an SSD using lazy-evaluation as action potentials propagate. 4TB SSD = ~1 trillion 32-bit FP weights and potentials. There are MacBook options that support up to 8TB. The other advantage with SNN - Training & using are basica…