5 ms·

Turn up the temperature? Markov chains have a cruder understanding of language. Turn up the temperature (the “randomness”) of an LLM and you can achieve a sim

Turn up the temperature? Markov chains have a cruder understanding of language. Turn up the temperature (the “randomness”) of an LLM and you can achieve a similarly crude approximation. Further, author uses ChatGPT-3.5. ChatGPT has been rlhf’d to sound as generic as possible, and 3.5 has a worse understanding of humor compared to 4. I don’t buy the thesis of this article.

dartos 2y ago
I wish there was a locked chatgpt version for researchers
- eru 2y ago
  Some of the open-weights models come close to what you might want?
  - dartos 2y ago
    Sort of. They fill the “locked” requirement, but open models aren’t SOTA like gpt or Claud.
    - coder543 2y ago
      Llama3.1-405B benchmarks at a level that I would certainly consider comparable to other SOTA models, as long as multimodality isn’t needed. Regardless, OpenAI provides access to quite a few of their older models through the API, since the API lets you pass in a specific model version. I’m sure the older models won’t be available forever, but that is a much more stable target for researchers than just opening the ChatGPT website and typing in things.
      - dartos 2y ago
        ChatGPT specifically is more than just a chat interface on top of got-4. Their system prompt includes the current date and time among other information, making it very very hard to run reproducible experiments against it. But it’s the tool most people are using.
    - Reubend 2y ago
      LLAMA 3.1 405b-Instruct is state of the art at the time of writing. It averages slightly worse than ChatGPT 4o, but to a mostly negligible extent for most research.
      - toxik 2y ago
        Aside but -- it is also aligned to not generate profane text, right?
        throwthrowuknow 2y ago
        Abliterate it
- rck 2y ago
  You could play with open models and different temperatures using the lmgen script from this repo we just released: https://github.com/RichardKelley/hflm?tab=readme-ov-file#lmgen https://github.com/RichardKelley/hflm?tab=readme-ov-file#lmg...
- jijji 2y ago
  there sort of is, if you install ollama (https://ollama.com https://ollama.com) and then execute: ollama run llama2-uncensored it will install and run the local chat interface for llama2 in an uncensored version which gives a little bit better results with less guardrails. Same with wizardlm-uncensored and wizard-vicuna-uncensored. For reference, the SOTA version with guardrails you could run: ollama run llama3.1:405b
ziddoap 2y ago
>rlhf'd For those of us not in the know about all the various machine learning acronyms: RLHF = Reinforcement learning from human feedback
- DaiPlusPlus 2y ago
  (Somewhat off-topic) When GPT went public along with OpenAI’s articles and papers back in late-2022 through 2023, my impression was OpenAI wanted us all to see/read about RLHF. It felt odd because surely the whole LLM-thing (e.g. how does it even work?!?[1]) was the far bigger research-story than just constant reassurances it won’t end-up like MSFT’s Tay bot; my understanding is that as a research or secret-sauce RLHF, compared to the core meat-and-potatoes of LLMs, is an ugly-hack afterthought. By-way of a bad analogy: it’s as if they created a fantastical new 3D world game engine, like Unreal or Unity, which has a fundamentally different architecture to anything before, but has a bug that occasionally replaces ground terrain with ocean - and their solution to this is to write a pixel-shader that detects this and color-shifts blue into green so people don’t notice - and they then put-out press-releases about how great their pixel-shader is - rather than about the rest of the engine - and no-one seems to be talking about the underlying bug, let alone fixing it. ————- [1] I still haven’t heard a decent explanation of how feeding the world’s corpus of English text (and computer program code) into a statistical-modeller results in something that can perform almost any information-processing task via instructions input as natural-language.
  - treyd 2y ago
    Yeah that analogy is fairly poor. You have to think about it in terms of some probability distribution that each step of the model is sampling from. Out of the distribution of all text, find the top n values for the next token that maximize `P(next | prefix)`, which is done efficiently through making a vector embedding to encode the tokens inside the statistical model. Things that look like Q-and-A transcripts do exist in the training set, think interviews, books, stage plays, etc, and at a different layer of abstraction the rules of English text in general are very well represented. What RLHF is doing is slightly shifting the shape of the probability distribution to make it look more like the Q-and-A formats that are desired. They build a large dataset with human tagging to collect samples of good and bad outputs and using reinforcement learning techniques to generate outputs that look more like the good examples and less like the bad ones. This probably involves creating a (much smaller, not-LLM) model that is trained to discriminate good outputs and bad outputs, learning to mimic the human tagging. There's some papers that have been published. Here's one article from Huggingface: https://huggingface.co/blog/rlhf https://huggingface.co/blog/rlhf
dankwizard 2y ago
you have no sense of humor

dartos 2y ago

I wish there was a locked chatgpt version for researchers

eru 2y ago
Some of the open-weights models come close to what you might want?
- dartos 2y ago
  Sort of. They fill the “locked” requirement, but open models aren’t SOTA like gpt or Claud.
  - coder543 2y ago
    Llama3.1-405B benchmarks at a level that I would certainly consider comparable to other SOTA models, as long as multimodality isn’t needed. Regardless, OpenAI provides access to quite a few of their older models through the API, since the API lets you pass in a specific model version. I’m sure the older models won’t be available forever, but that is a much more stable target for researchers than just opening the ChatGPT website and typing in things.
    - dartos 2y ago
      ChatGPT specifically is more than just a chat interface on top of got-4. Their system prompt includes the current date and time among other information, making it very very hard to run reproducible experiments against it. But it’s the tool most people are using.
  - Reubend 2y ago
    LLAMA 3.1 405b-Instruct is state of the art at the time of writing. It averages slightly worse than ChatGPT 4o, but to a mostly negligible extent for most research.
    - toxik 2y ago
      Aside but -- it is also aligned to not generate profane text, right?
      - throwthrowuknow 2y ago
        Abliterate it
rck 2y ago
You could play with open models and different temperatures using the lmgen script from this repo we just released: https://github.com/RichardKelley/hflm?tab=readme-ov-file#lmgen https://github.com/RichardKelley/hflm?tab=readme-ov-file#lmg...
jijji 2y ago
there sort of is, if you install ollama (https://ollama.com https://ollama.com) and then execute: ollama run llama2-uncensored it will install and run the local chat interface for llama2 in an uncensored version which gives a little bit better results with less guardrails. Same with wizardlm-uncensored and wizard-vicuna-uncensored. For reference, the SOTA version with guardrails you could run: ollama run llama3.1:405b

ziddoap 2y ago

>rlhf'd For those of us not in the know about all the various machine learning acronyms: RLHF = Reinforcement learning from human feedback

DaiPlusPlus 2y ago
(Somewhat off-topic) When GPT went public along with OpenAI’s articles and papers back in late-2022 through 2023, my impression was OpenAI wanted us all to see/read about RLHF. It felt odd because surely the whole LLM-thing (e.g. how does it even work?!?[1]) was the far bigger research-story than just constant reassurances it won’t end-up like MSFT’s Tay bot; my understanding is that as a research or secret-sauce RLHF, compared to the core meat-and-potatoes of LLMs, is an ugly-hack afterthought. By-way of a bad analogy: it’s as if they created a fantastical new 3D world game engine, like Unreal or Unity, which has a fundamentally different architecture to anything before, but has a bug that occasionally replaces ground terrain with ocean - and their solution to this is to write a pixel-shader that detects this and color-shifts blue into green so people don’t notice - and they then put-out press-releases about how great their pixel-shader is - rather than about the rest of the engine - and no-one seems to be talking about the underlying bug, let alone fixing it. ————- [1] I still haven’t heard a decent explanation of how feeding the world’s corpus of English text (and computer program code) into a statistical-modeller results in something that can perform almost any information-processing task via instructions input as natural-language.
- treyd 2y ago
  Yeah that analogy is fairly poor. You have to think about it in terms of some probability distribution that each step of the model is sampling from. Out of the distribution of all text, find the top n values for the next token that maximize `P(next | prefix)`, which is done efficiently through making a vector embedding to encode the tokens inside the statistical model. Things that look like Q-and-A transcripts do exist in the training set, think interviews, books, stage plays, etc, and at a different layer of abstraction the rules of English text in general are very well represented. What RLHF is doing is slightly shifting the shape of the probability distribution to make it look more like the Q-and-A formats that are desired. They build a large dataset with human tagging to collect samples of good and bad outputs and using reinforcement learning techniques to generate outputs that look more like the good examples and less like the bad ones. This probably involves creating a (much smaller, not-LLM) model that is trained to discriminate good outputs and bad outputs, learning to mimic the human tagging. There's some papers that have been published. Here's one article from Huggingface: https://huggingface.co/blog/rlhf https://huggingface.co/blog/rlhf

dankwizard 2y ago

you have no sense of humor