4 ms·

Any insights into how this is able to generate the Microsoft logo SVG near the end of the article? I feel like this would be a nearly impossible task unless the

Any insights into how this is able to generate the Microsoft logo SVG near the end of the article? I feel like this would be a nearly impossible task unless there was another layer of knowledge on top of the language model. I'm leaning toward the example being gamed in some way (e.g. retrying until it happened to work). EDIT: also notable is that the first ChatGPT answer decribes the color as 'orange' but the SVG answer describes it as 'red'.

_boffin_ 4y ago
I think they stuck that on there
- JamesSwift 4y ago
  I understand the rendering is an out-of-band addition. But the question is, if this is a statistical language model (i.e. a monte carlo word generator), then how is it able to go from the description to the SVG successfully without understanding things more logically. My best guess is that theres an actual example in the training corpus that its pulling from. Or that this is being gamed in some way.
  - pelorat 4y ago
    LLMs are very good at parsing and generating highly structured data as long as it has a rigid and logical structure and doesn't involve any actual mathematical calculations. For simple calculations it will often get it right because it has the knowledge stored, similar to how a human can remember a multiplication table. But once the numbers get high enough that humans would fail without solving it with an algorithm using pen and paper (or a calculator) that's when LLM's fail as well. Luckily this problem can be solved as you can prompt the model to not solve math, but instead transform them into a structure that can be fed into an API (like wolfram alpha). With a multistep process it would be easy to have ChatGPT generate accurate answers to advanced math problems: 1. Instruct ChatGPT to not solve mathematical expression it sees in the prompt, instead only extract the expression(s) into a list. 2. Programmatically scan the output, and make API calls to get the answers. 3. Submit the answers to ChatGPT and tell it to answer the original question with the newly supplied information. 4. Voila! You can solve almost any mathematical question using natural language.
    - JamesSwift 4y ago
      So thats I guess exactly my point. The text-to-accurate-SVG is much closer to the "solve a math problem for real" side of the spectrum than other prompts. And it does it correctly. In a sibling comment it turns out it doesn't work for other company logos. So my guess is either that the corpus happened to have this data already, they purposefully fed it additional Microsoft-related materials so it has a better basis of generation, that they regenerated it multiple times until it got it, or that theres an entire previously-undisclosed layer of processing being added to this version vs other ChatGPT.
    - SillyUsername 4y ago
      Tldr; teach it to use a calculator if it can't figure it out in its head.
Jensson 4y ago
By hardcoding tasks, like old assistants. If they don't have proof that it isn't hardcoded then it is almost surely hardcoded. For example Google Duplex, I remember many being very impressed by this: https://www.youtube.com/watch?v=D5VN56jQMWM https://www.youtube.com/watch?v=D5VN56jQMWM That entire conversation was just hardcoded with some if-else statements depending on what the other part said. The tech isn't as impressive when you know that, and anyone who interacts with the tech will realize that these are just hardcoded actions very quickly, but the demos will look very cool. So applying that level of hardcoding to these situations you take a hardcoded script, then put an LLM on top that just generates calls to these hardcoded scripts, and voila now you have the exact same assistant you had before but now it has a more advanced text parsing interface. But it generates a lot of buzz since people think that the whole thing is AI powered when only parts of it are and the rest wont generalize into typical situations. In fact it is now often worse than before since you can't reliably invoke the scripts you want to invoke, instead you have to fight the AI interface trying to get it to read your question the way you want it to.
- andreyk 4y ago
  The context is pretty different - Duplex was a public demo, this is a research paper with a whole paper behind it. There's no reason to script something if they have a working algorithm...
  - Jensson 4y ago
    Looking more closely at the code it generates I'm not sure what the point of putting ChatGPT in that loop is? Rotating until an object is in view isn't a very useful strategy for finding an object, and a non-techical user wouldn't be able to resolve that easily, and a technical user who knew about those basic functions could code the finding strategy he wants based on those examples and instructions. You would produce a better and easier to use interface faster and with less work by just exposing a few high level utility functions that are better designed than what ChatGPT generates there. Or at least it would be better to let ChatGPT use those high level functions. If they could make it solve a maze or find something in a more complex environment they would have shown that.

_boffin_ 4y ago

I think they stuck that on there

JamesSwift 4y ago
I understand the rendering is an out-of-band addition. But the question is, if this is a statistical language model (i.e. a monte carlo word generator), then how is it able to go from the description to the SVG successfully without understanding things more logically. My best guess is that theres an actual example in the training corpus that its pulling from. Or that this is being gamed in some way.
- pelorat 4y ago
  LLMs are very good at parsing and generating highly structured data as long as it has a rigid and logical structure and doesn't involve any actual mathematical calculations. For simple calculations it will often get it right because it has the knowledge stored, similar to how a human can remember a multiplication table. But once the numbers get high enough that humans would fail without solving it with an algorithm using pen and paper (or a calculator) that's when LLM's fail as well. Luckily this problem can be solved as you can prompt the model to not solve math, but instead transform them into a structure that can be fed into an API (like wolfram alpha). With a multistep process it would be easy to have ChatGPT generate accurate answers to advanced math problems: 1. Instruct ChatGPT to not solve mathematical expression it sees in the prompt, instead only extract the expression(s) into a list. 2. Programmatically scan the output, and make API calls to get the answers. 3. Submit the answers to ChatGPT and tell it to answer the original question with the newly supplied information. 4. Voila! You can solve almost any mathematical question using natural language.
  - JamesSwift 4y ago
    So thats I guess exactly my point. The text-to-accurate-SVG is much closer to the "solve a math problem for real" side of the spectrum than other prompts. And it does it correctly. In a sibling comment it turns out it doesn't work for other company logos. So my guess is either that the corpus happened to have this data already, they purposefully fed it additional Microsoft-related materials so it has a better basis of generation, that they regenerated it multiple times until it got it, or that theres an entire previously-undisclosed layer of processing being added to this version vs other ChatGPT.
  - SillyUsername 4y ago
    Tldr; teach it to use a calculator if it can't figure it out in its head.

Jensson 4y ago

By hardcoding tasks, like old assistants. If they don't have proof that it isn't hardcoded then it is almost surely hardcoded. For example Google Duplex, I remember many being very impressed by this: https://www.youtube.com/watch?v=D5VN56jQMWM https://www.youtube.com/watch?v=D5VN56jQMWM That entire conversation was just hardcoded with some if-else statements depending on what the other part said. The tech isn't as impressive when you know that, and anyone who interacts with the tech will realize that these are just hardcoded actions very quickly, but the demos will look very cool. So applying that level of hardcoding to these situations you take a hardcoded script, then put an LLM on top that just generates calls to these hardcoded scripts, and voila now you have the exact same assistant you had before but now it has a more advanced text parsing interface. But it generates a lot of buzz since people think that the whole thing is AI powered when only parts of it are and the rest wont generalize into typical situations. In fact it is now often worse than before since you can't reliably invoke the scripts you want to invoke, instead you have to fight the AI interface trying to get it to read your question the way you want it to.

andreyk 4y ago
The context is pretty different - Duplex was a public demo, this is a research paper with a whole paper behind it. There's no reason to script something if they have a working algorithm...
- Jensson 4y ago
  Looking more closely at the code it generates I'm not sure what the point of putting ChatGPT in that loop is? Rotating until an object is in view isn't a very useful strategy for finding an object, and a non-techical user wouldn't be able to resolve that easily, and a technical user who knew about those basic functions could code the finding strategy he wants based on those examples and instructions. You would produce a better and easier to use interface faster and with less work by just exposing a few high level utility functions that are better designed than what ChatGPT generates there. Or at least it would be better to let ChatGPT use those high level functions. If they could make it solve a maze or find something in a more complex environment they would have shown that.