Show HN: I fine-tuned Flan-T5. Can it cook? | Hacker News Reader

Show HN: I fine-tuned Flan-T5. Can it cook?

abuqader.substack.com

55 pointsaqader43 comments

Checkout the app at https://www.lechef.fyi

For some reason, I find it really grating how most deep learning stuff are pretty much impossible to do without paying quite a bit to a 3rd party service by the hour. Its one of the rare things in computing that you can't get started with unless you have some significant resources - most other things you can theoretically do with any old laptop and internet connection (ok - you may need to pay once you get users, but not just to get started)

You can get a lot of success by buying used gpus of the right sort (usually pro gpus, not gaming cards).

There is still cost in terms of power - gpus run hot and loud.

I bought a Tesla K80 for ~$200 off eBay, and now it's going for less than $150, and you can essentially get the same but just 1 GPU for $100. It has 12 GB of VRAM, and works great for "medium-duty" machine learning and AIs like Stable Diffusion.

It does run hot (150W per GPU), and I've had to make a [custom solution to cool it](https://github.com/askiiart/k80-linux-cooling), but it works great considering the price.

By the way, you can essentially never have an application use the 2 GPUs in parallel in any useful way; I'd suggest just getting a K20X instead.

At certain scales that’s true but there’s plenty you can experiment with with just Colaboratory or Kaggle… at least to get some intuition before jumping to spending money.

I think that model he trained can be trained on consumer RTX 3060, you just maybe would need to wait 4 times longer comparing to A100 he used.

I’ve found Google’s Colab to be a pretty great deal for getting started and playing around.

Runpod and Colab are the easiest to get started. Google cloud has a crazy sales funnel that will make you go crazy.

$20 should get you pretty far I would think.

Google colab is $49/mo to get an A100. Trust me when I say you can build multimillion-dollar ML companies with just that (and maybe $100 extra dollars per month of spot credits).

Handled my puerile test input pretty well, including photo: https://i.imgur.com/x60ZuDx.png

This doesn't seem useful beyond serving as a tech demo and shill post for OP's employer's subscription service. A lot of the nuance in recipes is lost when you programmatically clobber them into a format uniform enough to train a model with, and it seems like in this case it can either regurgitate existing recipes, or invent dodgy unappetizing ones.

My Rustic Omelette seems reasonable:

rustic omelette

by le chef

Ingredients

2 tablespoons butter

1 cup onion, chopped

1/2 cup green pepper, chopped

4 eggs

1/4 teaspoon salt

1/4 teaspoon black pepper

1/8 teaspoon garlic powder

1/8 teaspoon onion powder

Directions

Step 1: Melt butter in a skillet over medium heat.

Step 2: Add onion and green pepper; saute until tender.

Step 3: In a large bowl, whisk together the eggs, salt, pepper, garlic powder and onion powder.

Step 4: Pour egg mixture into the skillet.

Step 5: Cook for 3 to 5 minutes or until set.

Step 6: Flip omelette over and cook for another 3 to 5 minutes or until done.

That's the trick with the recipe ones. It's really easy to make a recipe that "seems reasonable" because most people don't have that good an intuition for the size/weight of ingredients, but those ratios are some of the most important things in the quality of the finished dish.

For example my experienced eye is worried about the veg:egg ratio here. By volume that's like 1.5 : 1 or more, depending on how you interpret "until tender." There are egg dishes like a spanish tortilla or some frittatas with that much vegetable but they use different technique to account for it.

It would be edible but depending on the cultural expectations of the person you put it in front of it might not register as omelette. Which could be fine! But if you've ever cooked an omelette more than twice you'd probably do a better job just eyeballing it.

I requested an octopus omelette and was told to boil the egg, fwiw.

Looks like a much tastier version of what the times did here https://www.nytimes.com/2022/11/04/dining/ai-thanksgiving-me..., they should write a followup

Thinking out loud here: Generating recipes for novel, yet to be discovered dishes, sounds really exciting — like that time I added peanut butter to warm rice.

It would need a deep understanding of human brain chemistry to mix and match different flavors.

BTW, tried peanut butter rice with “le chef” and the recipe didn’t contain any liquid.

Not got a lot of Thai restaurants near you, eh?

If you’re referring to peanut butter rice, I put like 7-8 tablespoons of creamy peanut butter into the rice.

I do love Thai food though. Unfortunately there aren’t any good Thai restaurants in Guayaquil.

Ah that's a shame. Thai and Indonesian cuisine has the most delicious peanut sauces. eg http://www.thaifoodandtravel.com/recipes/peanuts.html

Sorry, but 'searching google' for recipes is a surefire way to cook mediocre food. Most recipes on Google are optimized for convenience over quality. Even if that's what you want, it's a lot more useful to build a stable of recipe websites that you like and have had good results with and go from there.

Whenever I want to cook a new dish, I usually search Serious Eats, Chefsteps, Smitten Kitchen, and maybe then I'll look at Bon Appetit or a few cookbooks if I'm really struggling. Life is too short to bother cooking crap recipes, but that's basically what you'll get if you just do a Google search.

Awesome! checkout https://text-generator.io which is a language model as a service similar to OpenAi that can cook but much cheaper

I would recommend for vast majority of people to experiment around with these pretrained models first before getting the dedicated GPU/training etc, but cool that your company helps with it

Sort of minor note but when I cook, I want the directions to contain the quantities.

Rather than “heat oil in pan” I want it to say “heat 1tbsp of oil in a pan”.

When I shop I need the ingredients but when I’m actually doing the prep I want the directions and quantities at the same time because it’s faster to grab the ingredients and prep them as I go than it is to prep everything and then cook

EDIT: no one does this by the way, they all make me scroll between the directions and the ingredients list.

I totally agree. I have started working on a recipe website and for my personal recipes this is exactly what I do.

You can see an example of what it looks like for now here: https://www.hedacuisine.com/r/776.

I suspect once general NLP models become cheap and widely accessible, we'll see a lot of niche tools that can transform data from one type into another. E.g. you could just paste the recipe in and the tool would already be prompted with "instructions" to output it in the desired format.

Here's ChatGPT's take on the problem, which I think it did pretty well: https://i.ibb.co/dfwVH9G/image.png

Edit - bonus transformation: https://i.ibb.co/7CrKWX7/image.png

I mean, ya, that's what you generally want. But oil is a poor example. Are there people out there actually measuring how much oil they put in the pan? It's just a general guideline for if you need a lightly oiled pan, or a pan with a lot of oil. And really the amount of oil you need will vary a lot by the type of pan you are using. A lot of measurements are like this. Like salt, not only are personal tastes very different, but a half teaspoon of one kind of salt can be a lot more salt than a half teaspoon of another type.

Anyway, A lot of people cook by generally knowing good ratios. It takes time though and you can't really get started without explicit measurements.

Wasabi pancakes sounds good

Ingredients

1 cup all-purpose flour 2 teaspoons baking powder 1/2 teaspoon salt 3/4 cup milk 1/4 cup butter, melted 3 tablespoons wasabi paste

Directions

Step 1 In a large bowl, combine the flour, baking powder and salt. Combine the milk, butter and wasabi paste; stir into dry ingredients just until moistened.

Step 2 Drop by heaping tablespoonfuls onto a greased griddle or skillet. Cook until golden brown on both sides.

It can't. Here is "green pasta"

Only problem is that there is no pasta and I'm quite sure this is not very tasty :-D

The green pasta by le chef

Ingredients

- 2 cups fresh spinach leaves, chopped

- 1 tablespoon olive oil

- 1/2 teaspoon salt

- 1/4 teaspoon black pepper

- 1 garlic clove, minced

Directions

Step 1

In a large bowl, combine the spinach, olive oil, salt, and pepper. Toss well to coat.

Step 2

Divide the spinach mixture evenly among 4 plates. Top each serving with 3 tablespoons of the spinach mixture.

Yes it definitely can't cook even a simple steak.

Perfect steak by le chef

Ingredients

1 (3 lb.) round steak

2 Tbsp. salt

1/2 tsp. pepper

1/4 tsp. garlic powder

1/4 tsp. onion powder

1/4 tsp. celery salt

1/4 tsp. oregano

1/4 tsp. basil

1/4 tsp. thyme

1/4 tsp. rosemary

1/4 tsp. marjoram

1/4 tsp. paprika

Directions

Step 1

Cut steak into serving size pieces.

Step 2

Sprinkle with salt, pepper and garlic powder.

Step 3

Place in shallow baking dish.

Step 4

Combine remaining ingredients; pour over steak.

Step 5

Cover and refrigerate overnight.

Step 6

Bake at 350° for 1 hour.

Not pasta but wilted spinach is an excellent side dish or addition to pasta.

Note that this recipe is raw spinach (with raw garlic)

So how do those cookies taste? Did you make them? They have around double the sugar I’d expect, and dropping the dough from a teaspoon seems like it might not work with a dough this consistency, but they might be good.

I'm not too picky when it comes to chocolate chip cookies, but if someone messes up and uses baking soda instead of baking powder, that makes me sad. That's the kind of mistake a human might decide to put into big bold letters when writing up the recipe, but you are not going to get that from GPT. At least, not yet.

This "general gao's tofu lo mein" recipe looks tasty, but has no noodles. :)

I asked for Oregano Cookies and recipe did not contain oregano.

Was it hard to fine-tune FLAN-T5? What setup did you use?

OP does not mention that he's employed by BaseTen, the company behind "Blueprint" (which is mentioned repeatedly in the blog post). I noticed the repeated mentions and did some digging to find this out.

Not cool, OP, not cool.

See: https://www.baseten.co/blog/deploying-stable-diffusion

Not sure if it was added since, but 7 mins after, it has this,

> Quick thanks to Baseten, where I work. As an employee, I had early access to Blueprint where I used their fine-tuning APIs and serverless GPUs, which made this process much faster. You can sign up for the waitlist at: blueprint.baseten.co

Wow, even the disclosure sounds like an ad.

A disclosure ad is fine as long as there is disclosure imo

The history of adtech tells me I can pentest "fine" by squashing so many ads into the disclosure that your device starts to get hot.

relax, it’s just a recipe experiment writeup

no, it's chav in chav's clothing. this is promoting the company product using something that could pass as an experiment to the lay person. to those with spidey senses, this just reeks of "hey look what we did with our product!!!" as they pull some muscles reaching around to pat themselves on the back. probably going to be sore in the morning.

You're inflating your own ego with this comment; probably gonna be bloated in the morning.

lol. i'm inflating my ego for calling out someone for pushing marketing blogspam and you think that's worse than the ego of the blogspam creator? we do live in silly times.

There is an HN term for this :)

http://www.paulgraham.com/submarine.html

Reply on news.ycombinator.com