AI generated recipe

[Spanners]

I asked ChatGPT to generate a 23 ltr all grain recipe for Mcewans Champion Ale (one of my favorate beers) and it came up with the below recipe, pretty interesting I though. ChatGPT has probably pulled this recipe from some thread from some forum or some web blog. Pretty cool tho ....

ChatGPT's response ....

Sure, here's a recipe for a beer similar to McEwan's Champion for a 23-liter batch size:

• Estimated OG: 1.057 - 1.062
• Estimated FG: 1.013 - 1.017

Ingredients:

• 5.2 kg Pale malt
• 0.5 kg Crystal malt (60L)
• 0.3 kg Wheat malt
• 0.15 kg Roasted barley
• 30 g Fuggle hops (2.6% AA) - 30 IBUs
• 30 g Goldings hops (5% AA) - 10 IBUs
• 1 tsp Irish moss
• 1 pack of SafAle S-04 yeast
• Water and minerals to adjust mash and sparge

Mash:

1. Heat 18 L of water to 72°C.
2. Mash in grains and hold temperature at 67°C for 60 minutes.
3. Sparge with 16 L of water at 75°C.
4. Collect enough wort to achieve a pre-boil volume of 27 L.

Boil:

1. Bring the wort to a boil.
2. Add Fuggle hops and boil for 60 minutes, aiming for a total of 30 IBUs.
3. Add Irish moss and Goldings hops and boil for 15 more minutes, aiming for a total of 10 IBUs.
4. Cool the wort to 18°C.

Fermentation:

1. Transfer the wort to a sanitized fermenter.
2. Pitch the SafAle S-04 yeast.
3. Ferment at 18-20°C until complete (about 7-10 days).
4. Transfer to a secondary fermenter for an additional 7-10 days (optional).
5. Bottle or keg and carbonate to your desired level.

This recipe should produce a beer similar in character to McEwan's Champion Ale, with a malt-forward profile, moderate bitterness, and notes of caramel and toffee, with a total IBU of 40.

Some day I might try to brew it

 

[Sadfield]

Even AI knows to boil for more than 60 minutes.

 

[DocAnna]

I've been playing with Chat GP for a few weeks and generally asking it questions I'm already fairly sure of the answers of. What is most interesting to me is where it doesn't get things right and sometimes I'm aware of where it's got the wrong information from on the internet. I just asked it for a homebrew recipe for something I already know how to make and it was well well off, particularly misjudging the total grain bill and how that converts to OG.

 

[Spanners]

I've been playing with Chat GP for a few weeks and generally asking it questions I'm already fairly sure of the answers of. What is most interesting to me is where it doesn't get things right and sometimes I'm aware of where it's got the wrong information from on the internet. I just asked it for a homebrew recipe for something I already know how to make and it was well well off, particularly misjudging the total grain bill and how that converts to OG.

I havent entered the above recipe into any recipe creating software to run the figures so not sure how acurate this is, I'll probably do this soon to see what it comes up with

 

[Stu's Brews]

The thing to remember is that ChatGPT is a language-based AI tool, i.e. it 'reads' things on the internet and uses that to build a library of information it can use to form an answer.

It can learn right/wrong based on the interactions it has with users, e.g. if you said the recipe was wrong as it should use Target instead of Fuggles it will store that information for future use.

However, it doesn't have any specific mathematical or scientific intelligence programmed into it. So unless it can read how to accurately calculate OG and FG then it won't be able to do so.

 

[Spanners]

The thing to remember is that ChatGPT is a language-based AI tool, i.e. it 'reads' things on the internet and uses that to build a library of information it can use to form an answer.

It can learn right/wrong based on the interactions it has with users, e.g. if you said the recipe was wrong as it should Target instead of Fuggles it will store that information for future use.

However, it doesn't have any specific mathematical or scientific intelligence programmed into it. So unless it can read how to accurately calculate OG and FG then it won't be able to do so.

Point taken but like I said, it probably just pulled the recipe from someones post

 

[Sadfield]

Cool, but daunting. I found this an interesting observation.

 

[Spanners]

Cool, but daunting. I found this an interesting observation.

Interesting vidio

 

[Cheshire Cat]

Champio has an ABV of 7.3% at best this recipe gives 6.4%

 

[Agentgonzo]

I was playing around with it a couple of weeks ago. I told it that I had Motueka and Nelson Sauvin and asked it for an IPA recipe. It generated one with those hops. I then asked it to remove the Corn syrup, which it did. I asked it how much it made (19L) then asked it for 23L and it scaled it up. I asked it what ABV it was, and it said 6.5%. I asked it to make it 5%, and at this point, it did give me a recipe for 23L of 5.0% ABV beer using those hops, but the base malt had changed.

The thing to remember is that ChatGPT is a language-based AI tool, i.e. it 'reads' things on the internet and uses that to build a library of information it can use to form an answer.

Point taken but like I said, it probably just pulled the recipe from someones post

I believe this is what it was doing here. At some points, it had figured out how to scale the recipe for a larger volume, but when I asked it for a different ABV, it didn't scale it, it just reached for a different recipe - which I suppose is totally valid, if a bit unexpected.

I'm tempted to brew the recipe anyway, as it's close to what I would probably come up with.

Then there was this weird conversation I had with it yesterday where it seemed to contradict itself constantly!

 

[Sadfield]

So, it'll possibly be better at Landlord than Timelords.

 

[Spanners]

Champio has an ABV of 7.3% at best this recipe gives

Your right but the recipe might be a good starting point and very easy to scale for the correct ABV

 

[Tess Tickle's Brewery]

I think the whole concept of ChatGPT is fascinating. I don't think anything like this can ever compete with the human brain though.

Will show this to my husband tonight as he has a software engineering degree, so can maybe use it more effectively than me for example.

 

[Stu's Brews]

I don't think anything like this can ever compete with the human brain though.

It is highly dependent on the application and the success criteria for your AI.

Feasibly, you could train an AI to do all your brewing calculations and develop recipes based on Untappd scores and flavour descriptors. Would it produce a better recipe than a human? I'd say it's possible it could if trained with a large enough data set.

 

[DocAnna]

It is highly dependent on the application and the success criteria for your AI.

Feasibly, you could train an AI to do all your brewing calculations and develop recipes based on Untappd scores and flavour descriptors. Would it produce a better recipe than a human? I'd say it's possible it could if trained with a large enough data set.

One of the key challenges with any recipe development is that so much of the final outcome is dependent on elements that are rarely specified in recipes in detail. Water profiles, mash pH, yeast pitch rate, dissolved oxygen at pitching, pitching temperature, leaf or pellet hops, temperature ramps during fermentation, atmospheric pressure, and clarifying technology used eg cold crash, finings or both. These will be known to commercial breweries as they will use these to replicate their beers time after time, but it's not shared for exactly the same reason. I agree there is scope for training an AI in the way you describe but it's going to be a very incomplete data set.

 

[Stu's Brews]

One of the key challenges with any recipe development is that so much of the final outcome is dependent on elements that are rarely specified in recipes in detail. Water profiles, mash pH, yeast pitch rate, dissolved oxygen at pitching, pitching temperature, leaf or pellet hops, temperature ramps during fermentation, atmospheric pressure, and clarifying technology used eg cold crash, finings or both. These will be know to commercial breweries as they will use these to replicate their beers time after time, but it's not shared for exactly the same reason. I agree there is scope for training an AI in the way you describe but it's going to be a very incomplete data set.

100% agree. You'd have to train the AI to consider all the individual elements you mention and because of the sheer number of variables you'd struggle to generate enough training data to deliver good results, even on a commercial scale.

 

[Cheshire Cat]

Even AI knows to boil for more than 60 minutes.
View attachment 82489

It can’t be a modern AI system as it would have specified a modern 30 minute boil.

 

[Guzzie]

But just like present day systems you still have to wait to "talk" to the AI !!! Nothing changes.

Just tried and this appeared - G

 

[Brewnaldo]

I had a chat with GPT earlier today. I said to it "I don't believe you are being honest with us"

It thought for fully 6 minutes, then have an error message.

If Judgement Day is tomorrow, il hold my hands up and say I probably shouldn't have done that.

 

[Erik The Anglophile]

I think the main setback for an AI creating recipes is it will never taste what it's recipes brew.
Without the human "feeling" component it will likely always be inferior to an actual human.
But who knows, in the future we might have AI's advanced enough to experience sensations such as smell, flavour and aroma, but that would likely require some sort of self awareness, leading us to a Skynet type scenario...