User Controls

New SOTA open source foundation model: DBRX

  1. #1
    DUSM Raylan Givens Tuskegee Airman
    https://huggingface.co/spaces/databricks/dbrx-instruct

    Ok so this thing is actually mad impressive to me.

    Both GPT-4 and Gemini Advanced have failed one of my "standard" test questions and it's interesting how they fail it.

    Basically I ask a conservation of momentum related question regarding counter rotating masses (flywheels) internal to a box, one flywheel isnsuddenly braked, what happens?

    Both Gemini Advanced and GPT-4 on some level believe that the box can achieve net translation.

    This is blatantly false and violates conservation of momentum.

    DBRX, while it failed to give the correct answer (box will start rotating against braked flywheel) still reliably maintained consistency with denying any net translation could ever be achieved.

    Try it out, report your results
  2. #2
    Iron Ree African Astronaut [my flyspeck near-blind refund]
    HUgging face is on the CUTTING EDGE of AI research. Get your PyTorches out folx.

    Is this the "invisible friend" method I have heard so much about where you give a GPT an invisible friend and it suddenly gets double the result/score in basically all tests and seems to become "smarter" and more responsive with less issues like hallucinations

    Databricks claims DBRX runs up to 2x faster than Llama 2, in part thanks to its mixture of experts (MoE) architecture. MoE — which DBRX shares in common with Mistral’s newer models and Google’s recently announced Gemini 1.5 Pro — basically breaks down data processing tasks into multiple subtasks and then delegates these subtasks to smaller, specialized “expert” models.

    Most MoE models have eight experts. DBRX has 16, which Databricks says improves quality.

    Quality is relative, however.

    While Databricks claims that DBRX outperforms Llama 2 and Mistral’s models on certain language understanding, programming, math and logic benchmarks, DBRX falls short of arguably the leading generative AI model, OpenAI’s GPT-4, in most areas outside of niche use cases like database programming language generation.

    you need like an infinite number of GPTs to act as each node of the brain instead of an entire GPT model making connections mimicking how a brain would process language instead you just scale that up to neurotransmitter AI's sending eachother electrical signals which would more accurately simulate a human brain and be more true artificial "intelligence"

    Originally posted by Chairman Takeshi Kaga If we created a simulation with creatures with as robust an intelligent consciousness as we do, wouldn't we think they were morally considerable? Maybe. But you would probably exterminate whole digital worlds to save but one life of flesh and blood in the context of your reality.

    My opinion is this: if you are a powerful galactic civilisation and get strong enough computer hardware to do whatever you want with whatever level of granularity you want (for example if you manage to achieve high levels of nanotechnology that you could release a nanobot swarm to devour entire asteroids and turn them into computing clusters), you would easily have the resources to run a simulation of your civilisation for sure.

    The reason is that you would have excellent present-state information about it your civ and universe in whatever aspects you deem relevant at that point and could simulate them materially, even down to individually simulating our brains at the neuronal level to create conscious "ai", just like your own species, in a universe just like your own, but one you can watch play out millions of times faster because it is running on a computer.

    You could watch the technological and scientific progress the beings within your simulation make in situations similar to your own, with experimental outcomes similar to your own, and test their advancements against your own universe's physics to make genuine scientific progress, and even feed this data back into your simulation to present the discrepancies to the simulated civ in their experiments, and make the simulation itself more accurate and progressively more advanced.

    The simulated agents would create technology, even products relevant to them in situations extremely similar to the operators' world and help to model their marketing strategies and figure out novel technological pathways… After all they can get 1000000 years of brainstorming by an entire simulated civilisation of simulated agents.

    To simulate every single brain on Earth, it would "only" (for an advanced civ) take about the mass of a planet turned into microchips at a modern fabrication scale (let alone the fabrication technologies available to an advanced civilisation.

    Our purpose in such an existence, would be to persist, survive, innovate, solve problems to improve society and generally live our best lives because our self progress is directly aligned with the goals of the operators.

    https://nickbostrom.com/superintelligence.html
  3. #3
    DUSM Raylan Givens Tuskegee Airman
    Now my view has evolved. Dijkstra said, "whether a computer can think is about as relevant as whether a submarine can swim". A sub definitely can't "swim" like a human swims but in all the things you want a sub to do, it is vastly superhuman, and it's accomplished by completely different means that are better and more scalable. In the same way I think bio inspired analogies are very important but limited. If you want to create truly superhuman machine intelligence, you will have to find ways of accomplishing human like (and beyond) versatility and informational understanding and synthesis skills etc without just recreating a human brain. Cuzehats the point? Humans are already cheap.
    The following users say it would be alright if the author of this post didn't die in a fire!
  4. #4
    what does it use for math? it actually gives accurate answers for the arithmetic and calculus questions i tried
  5. #5
    ner vegas African Astronaut
    Originally posted by DUSM Raylan Givens Now my view has evolved. Dijkstra said, "whether a computer can think is about as relevant as whether a submarine can swim". A sub definitely can't "swim" like a human swims but in all the things you want a sub to do, it is vastly superhuman, and it's accomplished by completely different means that are better and more scalable. In the same way I think bio inspired analogies are very important but limited. If you want to create truly superhuman machine intelligence, you will have to find ways of accomplishing human like (and beyond) versatility and informational understanding and synthesis skills etc without just recreating a human brain. Cuzehats the point? Humans are already cheap.

    "Humans are the only supercomputer that can be mass-produced with unskilled labour"

    as it stands AI at the moment is just an algorithm for analysing large volumes of data; the question at hand seems to be how sources of data are weighted for accuracy (trust), given that in your example there's probably a lot of retard assumption in the data it's trained on.

    in order to analyse 'like a human' it'd need external sensors that are able to verify what base-level data can be immediately discarded
  6. #6
    Originally posted by ner vegas in order to analyse 'like a human' it'd need external sensors that are able to verify what base-level data can be immediately discarded

    define "analyze like a human".

    anal-lysises from eh "human" like george floyd differs greatly from eh human like ... ..
  7. #7
    DUSM Raylan Givens Tuskegee Airman
    Originally posted by RETARTEDFAGET what does it use for math? it actually gives accurate answers for the arithmetic and calculus questions i tried

    It doesn't reference an external tool if that's what you're asking. They haven't released the dataset yet but I'm guessing it is just very well curated and trained.
  8. #8
    DUSM Raylan Givens Tuskegee Airman
    Originally posted by ner vegas "Humans are the only supercomputer that can be mass-produced with unskilled labour"


    Yea even if you add up all the cost of raising a human to adulthood in a good environment.... DBRX cost $10million to train all in one go

    as it stands AI at the moment is just an algorithm for analysing large volumes of data;

    ???

    What are you trying to say with this?

    What's the "just"? That's like the broadest definition in the universe. It could apply to anything from an entire biosphere to a torrent tracker.

    the question at hand seems to be how sources of data are weighted for accuracy (trust), given that in your example there's probably a lot of retard assumption in the data it's trained on.

    It's not that straightforward as giving preference to some data or not. It's not really even related to faulty assumptions in the semantic content of the input text data.

    E.g. The encoder doesn't really care if the string "a fish is a dog" is included in the training data or if it's true or false because it has no idea what a fish or a dog is and it possibly might even tokenize and embed that string in a way where the words "fish" and "dog" never appear completely. In a sense it only cares about how a particular sequence of tokens relates to the distribution of other tokens so that same string would "align" as true or false based on the information in the rest of the model.

    So in that simplified example, if the distribution of information in the rest of the model indicates that a token "fish" is strongly related to all the data points that we IRL associate with a "banana" (yellow with some brown spots, curved and elongated, segmented peel that is easily separated, soft mushy core... Etc) and the token "dog" is strongly related to all the points we associate with "fruit" (you know what I mean) then it will think it's "true" (so to speak" that "a fish is a dog". Cuz (again, so to speak) it thinks "fish" refers to "banana" and "dog" refers to fruit.

    So it really is about the gestalt and not about weighting between sources.


    in order to analyse 'like a human' it'd need external sensors that are able to verify what base-level data can be immediately discarded

    That's the Transformer architecture's weighted attention mechanism? If I understand what you mean correctly.
  9. #9
    Originally posted by DUSM Raylan Givens It doesn't reference an external tool if that's what you're asking. They haven't released the dataset yet but I'm guessing it is just very well curated and trained.

    if it doesn't use anything external that's pretty impressive
  10. #10
    dalit trying to be deep
  11. #11
    lets take a moment to owner this year's Transgender Visibility Day.

  12. #12
    DUSM Raylan Givens Tuskegee Airman
    Topical forum. Lanny, ban this chink.
  13. #13
    dalititus
  14. #14
    DUSM Raylan Givens Tuskegee Airman
    Originally posted by RETARTEDFAGET if it doesn't use anything external that's pretty impressive

    Isn't it really a bit weird? Not like it's a brand new point but...

    Everyone's mental image was pretty uniformly that robots will first replace manual labor jobs and be better than a calculator at math, and it will be abstract reasoning that will be really hard.

    Turns out that physical robots are what's really hard, the first generalized AI systems and ML methods are better at replacing abstract intellectual and creative labor but kinda suck at math, and getting to be reliable at more complex maths is really hard.
  15. #15
    there can be no artificial intelligence unless and until theres such a time that we have artificial retardation
  16. #16
    DUSM Raylan Givens Tuskegee Airman
    Lanny, ban this chink
  17. #17
    banning me will not cure your dalititus
  18. #18
    DUSM Raylan Givens Tuskegee Airman
    Disrupting topical threads won't cure your very publicly displayed insecurity about your lacking IQ and English language skills.

    Lanny, ban this chink.
  19. #19
    Originally posted by DUSM Raylan Givens insecurity about your lacking IQ



    this is the first time i hear about iq insecurity,

    must be a dalit thjng.
  20. #20
    DUSM Raylan Givens Tuskegee Airman
    Originally posted by Charles Ex Machina

    this is the first time i hear about iq insecurity,

    must be a dalit thjng.

    Your low IQ is showing, this thread is only for high IQ individuals which excludes you from participation.
Jump to Top