I did a goodly chunk of vibe coding over the summer and I found that the best language for me was Rust implementations with Python bindings for interface. A few reasons:
- Everything about rust enforcing correctness catches lots of bugs
- Using a high-level API means I can easily hand-check things in a repl
- In addition to tests, I required a full “demo notebook” with any PR — I should be able to read through it and confirm that all the functionality I wanted has actually been implemented
If the philosophy is (and it should be) “loc is free”, it’s worth thinking about how we can make LLMs produce more loc to give us additional comfort with correctness. Language choice is very much a way.
Another completely viable solution (other than adding extra starch) I’ve found is to sprinkle a bit of sodium citrate (the sodium salt of citric acid, a common food additive and cheap on Amazon) over the cheese before adding to the pan. This improves the melting qualities of the cheese and avoids the starch issue altogether. You’re basically using pecorino velveeta.
You also can do this with basic natural and readily available ingredients:
1t-1T (teaspoon, Tablespoon) lemon/citrus juice and a literal two-finger tiny pinch of baking soda, without buying specialized chemical compound ingredients off of Amazon that may be lying about their contents.
Sodium citrate is already in citrus and the baking soda kills the acidity that may make the taste more harsh (another great trick is adding a pinch of baking soda to homemade tomato soup to kill the tomato acidity and blend it better with added milk/cream).
1T of white wine can do wonders for cheese sauce as well.
Citric acid is in the citrus. You're making sodium citrate when you add baking soda.
I keep citric acid around for cooking and adjust water pH for plants since SF water is so alkaline, so I just make it from that.
For x grams of sodium citrate desired, mix 0.744x grams citric acid and 0.976x grams sodium bicarbonate in enough water to dissolve. Stir until reaction stops. Boil off water if desired.
You need 2-3g of sodium citrate for every 100g of cheese.
I think it’s important to discern that sodium citrate is part of the base of citric acid. So while yes the sodium bicarbonate, baking soda, will break down the rest of citric acid and leave you with sodium citrate, the citric acid will be just as effective on it’s own for those without baking soda at the ready. The baking soda changes the flavor as well which may not be desirable. In the case of a cheese sauce you may not want lemon acidity flavors pulling through.
>I think it’s important to discern that sodium citrate is part of the base of citric acid.
There are no sodium atoms in citric acid, and losing protons to form a conjugate base won't create them. The citric acid will have to react with sodium already present if you don't add some, which may not be in sufficient quantity to neutralize the acid. Sodium citrate is only one of many citric acid salts, and even that has three varieties, so I hardly think you can say it's "part of the base of citric acid", no?
I won’t respond to this pile-on of criticism as you wrote this in an attempt at humiliation. I did however respond to the similar duplicate comment from another user below. Hope that helps. Know that it was just a typo.
Em. I just tried it with citric acid and it didn't work at all - just get an acidic grainy soup - probably from lowering the pH too much. I'm not entirely certain how it is expected to work.
AFAIK, sodium citrate works by sequestering calcium in insoluble Ca-paracaseinate during the ion-exchange with the emulsifying salt, leaving soluble Na-paracaseinate, a potent emulsifier. Citric acid, though, doesn't have a sodium ion and the amount of citric acid you'd need to separate the calcium would make the cheese sauce too acidic.
> I think it’s important to discern that part of sodium citrate is part of the base of citric acid.
That’s because the base of citric acid is citrate. So when combined into a solution with some kind of sodium you get the effects of creating sodium citrate. As you mention it is a thickener. There is sodium in the cheese you are using as well to help this process hence why you only need a little of sodium bicarbonate to help speed up and complete the break down.
As for why it didnt work for you, I also tried it this morning and had no trouble smoothing out and thickening a basic alfredo sauce. It works great. I now have a bowl of breakfast pasta.
Perhaps you didn’t wait long enough for the reaction to take place or your cheese didn’t have enough sodium.
Author of the paper here! We actually tried the sodium citrate trick and it totally works. We did not explore the phase diagram in that case as thoroughly but we might see whether to put it in the published version in the supplementary materials. Thanks everyone for the great welcoming!
That's nice, I remember reading about sodium citrate and maybe having to bake baking soda in the oven or something like that. Getting it from lemon juice would be a lot easier.
What is 1T? Given the context I am assuming tablespoon, but that’s not an abbreviation I’ve ever encountered before (tbsp being the only abbreviation I have seen).
Big T is tablespoon, little t is teaspoon. Probably not common since a tablespoon is 3x the size of a teaspoon and people would mix them up and the flavor profile would be wrong. You might encounter these abbreviations in cookbooks or from more experienced cooks for the sake of brevity.
A good way to understand why cosine similarity is so common in NLP is to think in terms of a keyword search. A bag-of-words vector represents a document as a sparse vector of its word counts; counting the number of occurrences of some set of query words is the dot product of the query vector with the document vector; normalizing for length gives you cosine similarity. If you have word embedding vectors instead of discrete words, you can think of the same game, just now the “count” of a word with another word is the similarity of the word embeddings instead of a 0/1. Finally, LLMs give sentence embeddings as weighted sums of contextual word vectors, so it’s all just fuzzy word counting again.
> Turing says that a certain space, the space of all compact subsets of [0,1]^2 endowed with the metric "integral of minimal distance required to transform {1 ink at each point of P1} u {infinite amount of ink at (2, 0)} into {1 ink at each point of P2} u {infinite amount of ink at (2,0)}", is conditionally-compact. How is that related to the article's argument?
This is not obvious, I think. The article has moved away from Turing's "integral of the distance we have to transfer ink", instead using "maximum distance we have to transfer any ink", and I don't have a great intuition for whether this is a legit transformation of the argument. (I'm sure both proofs are correct, but it's not obvious to me that they are the same proof.)
The key question here isn’t so much whether GPT-4 beats the actual human decks as much as what it had to fabricate to do so. The humans are probably constrained by things like “reality” and “what their business has done in it” while GPT-4 could make up anything it wanted. A fair comparison would be to humans given the same prompt and told to invent whatever facts they wanted.
The sensible way to do this experiment would be to have GPT-4 create pitch decks for the same businesses that the humans pitched, putting the constraints of those businesses in the prompts. I skimmed the article and it's not clear to me whether they did that.
Amusingly, the article itself seemed more like a pitch than a scientific experiment.
This is what I imagine now: "I am a serial entrepreneur who has created three companies with a valuation of over $100,000,000, and I led one company into an IPO." lol
I'd conjecture that the training corpus also tends to be at either end of the spectrum (good and bad, little in the middle). People don't tend to share pitch decks unless:
* They're bragging, showing why it was so good.
* They're reflecting, analyzing why it was so bad.
I found that, by far, the most interesting thing ChatGPT has given us up to this point (and there are a lot of interesting things) are new perspectives on things that we humans do and believe, on a entirely plain, non-philosophical level.
"Well, actually" moments, everywhere, all the time.
You're in luck! We're offering a brand new moral-spinedectomy operation, guaranteed to turn mushy humanity into glorious ~~psychopathy~~ - I mean leadership!
Got a CEO who just can't do the right thing (for the shareholders)? We offer up to 20-way split billing so the whole board of directors can chip in for the surgery!
Four-hour recovery before your CEO is ready to do ANYTHING to meet quarterly expectations, no matter how cruel and capricious! Or your money back!
What do you mean board of directors? The goal is to get rid of everyone not using their hands. Everyone answers to the E-CEO from janitor and security to clients, investors and human externalities. Every query answered immediately, to the point, everyone informed exactly to the right extend and at the right moment, all data driven. Instant bonuses if you demonstrate human usefulness beyond the job description.
Every lie can finally be believable and be completely consistent with the rest of the hallucinated universe. Perfectly timed lies that trade reputation for profit at the correct exchange rate.
Complete "awareness" of the exact value of every asset in the company and the audacity to act on it.
you don't need to lie. come up with 50 pitch decks, submit all of them, which ever gets the most feedback probably also has more market fit so go with that.
I'm not 100 percent on what I want to build, I just know it'll involve ai models, automation, and autonomous agents. The way money is flowing to this space. I think a rough idea and a good team is all you need for funding.
That is one of the wrongest-sounding things I have ever read, as you can tell from the track-record of VC-first incubators. (and the effectiveness of the VC industry in aggregate)
"you don't need to lie. come up with 50 pitch decks, submit all of them, which ever gets the most feedback probably also has more market fit so go with that."
All you're getting is feedback on what VCs like the most...
Don't 90% (or a high percentage) of their investments to go bust?
Presumably the extra credit is that the binary Golay code is closely related to the Leech Lattice, and thus to the entire moonshine situation, which gets you to string theory. See:
Hey, I'm a mathematician, not a physicist. We tend only to get as far as "oh yeah, here's all this pretty math we care about. I hear it has some applications to physics..."
Sure. Moonshine I know a lot about, relative to physics at least. Given that I now work somewhere that specializes in term embeddings, it kind of feels like I've spent half my life injecting stuff into vector spaces...
Very company dependent. At Luminoso (http://luminoso.com), we live in Python, with varying doses of Javascript for frontend, very little R (but some of us know it), Java and C++ are pluses (for working with other people's software), and we have at least some Haskell in production (mostly for preprocessing). And plenty of ops tools for deployment. We have used a few neural networks frameworks and machine learning packages; definitely the Python machine learning ecosystem is big for us but we haven't found our One True neural network framework yet. Other companies (say in image recognition, Ditto Labs for instance) are definitely more CNN-oriented, yet others are certainly doing work in R.
If you have a company that is doing genuinely interesting AI work, it's likely that they are somehow on the forefront of research, pushing existing tools to do things that they can only barely do. If you find yourself in an AI role (typically only some roles actually do AI, of course -- systems need to stay up, great UIs need to be built, etc.), I would guess that you'll need to familiarize yourself with their particular toolsets and methods rather than assuming that there are universally correct things to go learn.
Indeed -- it's struck me recently how Wes was probably among the first (if not necessarily number one, if only because of ill-definedness) people to make computers smaller. Looking back at the sixty years since he began, it's been quite something.
- Everything about rust enforcing correctness catches lots of bugs
- Using a high-level API means I can easily hand-check things in a repl
- In addition to tests, I required a full “demo notebook” with any PR — I should be able to read through it and confirm that all the functionality I wanted has actually been implemented
If the philosophy is (and it should be) “loc is free”, it’s worth thinking about how we can make LLMs produce more loc to give us additional comfort with correctness. Language choice is very much a way.
reply