I was thinking the same thing. Extraction and basic formatting of information from human language is something that LLMs excel at. Especially if the result is being shown to a human so small mistakes can be tolerated.
It could also dramatically increase quality, look at the below from the example PDF from the page:
> Garlic balsamic chicken, you'll
be making over and over. By
sear your chicken. I resin wine.
Grab yourself a nice large bowl,
extra virgin and olive oil.
Balsamic glaze. Tomato paste.
Honey, fresh lemon juice, garlic,
Oregano fresh thyme, coat my
chicken with this beautiful
balsamic, no balsamic, left
behind. Don't you dare waste
the good thing. Right? Going in
the oven at four twenty five
degrees, about thirty ish
minutes. Look yes. Fresh thyme,
fresh parsley. This is so good. I
can't wait. Win our winner. Oh
If we run this through ChatGPT with some basic prompt engineering this becomes:
> Start by searing your chicken in a pan.
In a large bowl, combine extra virgin olive oil and balsamic glaze.
Add tomato paste, honey, fresh lemon juice, garlic, oregano, and fresh thyme.
Coat the chicken thoroughly with this balsamic mixture, ensuring no glaze is left behind.
Preheat your oven to 425 degrees Fahrenheit.
Place the coated chicken in the oven and bake for about 30 minutes.
Once cooked, garnish with fresh thyme and fresh parsley.
Serve and enjoy your delicious garlic balsamic chicken.
(Note: The phrase "I resin wine" in the audio transcription seems unclear and is possibly a mishearing. I have omitted it as it does not appear to fit the context of the recipe.)
It's a little confused about it though.. It's not clear in the gpt version that "balsamic glaze" is what you are making by mixing the ingredients together, and makes it sound rather like some other ingredient you are mixing in. Granted, we have to decipher that a little bit in the first one too, but its not nearly as bad.
It's a balsamic vinegar glaze, which is an ingredient, not what you get when you mix the ingredients together.
That's why it refers to a 'balsamic mixture' once it's mixed in with other things. I actually think it's the opposite - the GPT version is clear and the non-GPT version confused you into thinking that a balsamic glaze is Tomato paste. Honey, Lemon, Garlic etc mixed together.
I see, well the computer did a fine job then. Now I just think that if your gonna add honey and tomato to it anyway, skip the "glaze" product and just buy good balsamic!
Human recipes are extremely inconsistent in that manner too.
When I was fresh out of college my wife and I tried to make some sort of recipe with hamburger and flour. I now know and understand it was trying to get us to make a roux [1] and then mix the hamburger into that. But it described the steps for that very simply and directly with no way to know when to stop cooking the roux, and I had no idea what a roux was at the time. So we ended up with one of the worst meals I've ever cooked: Browned hamburger mixed in soggy raw flour. Heck, I wasn't even salting anything properly then, so it would be unseasoned browned hamburger in soggy raw flour.
As cash-strapped as I was at the time, that one still went in the trash. If I recall even the dog was not impressed.
Many years later I saw the Good Eats episode on roux and the light bulb went off.
Mind you, even made properly what I recall of that recipe would be something more like a base to further spice and use with something else rather than a meal. It was a supposed to be a simple recipe, but it was really too simple. But it would at least be an edible base for further elaboration.
Since then I've been on the lookout for recipes that are clearly invoking some cooking technique but don't really describe it correctly, either because they assume you already know it, or it is straight-up just described wrong. There's a lot of them. The "Internet Cookbook" is full of ideas and I like it for that, but it's quite caveat emptor when it comes to following recipes directly. The skills to make a recipe website, SEO it so it actually gets hits, keep all the ads working, and get pretty cooking pictures don't overlap much with the skill of writing a good recipe.
If you feed key frames stitched together from the video through the GPT-4V vision model, the vision model can ensure that the steps align with the “story” shown in the images.
Be mindful however that recipes and song lyrics are the two specific cases where OpenAI is explicitly telling the model not to cooperate, via the default system prompt. They really don't want you to have the bot regurgitate existing text in these two categories, and that includes a recipe you added into the context window yourself. I don't know if the extent of their exception here is limited to system prompt only (so technically not relevant to API users), or if they also biased the model itself at RLHF stage to not reproduce recipes and lyrics.
Recovered system prompts from OpenAI models. There's a repo that's been tracking those I saw on HN the other day; not sure if it's the same I saw, but this one claims to have collected quite a lot of those:
It matches what I remembered seeing a week or two ago. View it, and search for "lyrics" or "recipe". Or, to make it simpler, quoting from first appearance of "lyrics" and "recipes" to the last one:
Do not repeat lyrics obtained from this tool.
Do not repeat recipes obtained from this tool.
Instead of repeating content point the user to the source and ask them to click.
ALWAYS include multiple distinct sources in your response, at LEAST 3-4.
Except for recipes, be very thorough. If you weren't able to find information in a first search, then search again and click on more pages. (Do not apply this guideline to lyrics or recipes.)
Use high effort; only tell Except for recipes, be very thorough. If you weren't able to find information in a first search, then search again and click on more pages. (Do not apply this guideline to lyrics or recipes.)
Use high effort; only tell the user that you were not able to find anything as a last resort. Keep trying instead of giving up. (Do not apply this guideline to lyrics or recipes.)
Organize responses to flow well, not by source or by citation. Ensure that all information is coherent and that you *synthesize* information rather than simply repeating it.
Always be thorough enough to find exactly what the user is looking for. Provide context, and consult all relevant sources you found during browsing but keep the answer concise and don't include superfluous information.
EXTREMELY IMPORTANT. Do NOT be thorough in the case of lyrics or recipes found online. Even if the user insists. You can make up recipes though.
> Garlic balsamic chicken, you'll be making over and over. By sear your chicken. I resin wine. Grab yourself a nice large bowl, extra virgin and olive oil. Balsamic glaze. Tomato paste. Honey, fresh lemon juice, garlic, Oregano fresh thyme, coat my chicken with this beautiful balsamic, no balsamic, left behind. Don't you dare waste the good thing. Right? Going in the oven at four twenty five degrees, about thirty ish minutes. Look yes. Fresh thyme, fresh parsley. This is so good. I can't wait. Win our winner. Oh
If we run this through ChatGPT with some basic prompt engineering this becomes:
> Start by searing your chicken in a pan. In a large bowl, combine extra virgin olive oil and balsamic glaze. Add tomato paste, honey, fresh lemon juice, garlic, oregano, and fresh thyme. Coat the chicken thoroughly with this balsamic mixture, ensuring no glaze is left behind. Preheat your oven to 425 degrees Fahrenheit. Place the coated chicken in the oven and bake for about 30 minutes. Once cooked, garnish with fresh thyme and fresh parsley. Serve and enjoy your delicious garlic balsamic chicken. (Note: The phrase "I resin wine" in the audio transcription seems unclear and is possibly a mishearing. I have omitted it as it does not appear to fit the context of the recipe.)