Curious why that feels condescending? Like, the average person who is seeking a recipe needs a photo to know what to shoot for
The menu: Rosemary and garlic roast lamb, Extra-crispy roast potatoes, Honey-roasted carrots and parsnips, broccoli.
The photo: [1] Roast turkey, mashed potatoes, baby carrots, broccoli, brussels sprouts. No lamb, no parsnips.
If you shoot for what's in the photo, you're going to have a bad time.
You're also going to have a bad time when you try to make an apple crumble with no flour, no sugar, and no butter, because they're not on the shopping list.
[1] https://images.openai.com/static-rsc-4/gyzrX8zp3O2KLEPm0FAqh...
I feel like it can only be successful by luck. Either it's reproducing a recipe verbatim that was tried and validated and tasted by a human, in which case, we didn't need AI for that, just a searchable cookbook. Or, it's making one up. I understand that using RLHF has helped improve the quality of questions about history, programming, or TV show recommendations, but I do not believe that there has been some kind of training regime where people prompt the models for "a recipe that uses X, Y, Z random ingredients," follow the recipe, and then score it. And even if they do, I don't see how a model can learn enough from that besides "this exact recipe is good/bad." '1/2 tsp cumin' may be a great addition to one recipe and not enough for another, so improving the output based on a bunch of scored recipes... I just don't believe cooking is an LLM job.
Maybe some other kind of model that I don't know about.
I feel like the technique recommendations are probably the most valuable element, but I'm getting rave reviews like 85% of the time now?
There have been a couple times where I was out of an ingredient and it proposed an adjustment that sounded a little wild to me, but it almost always is well received.
Unfortunately, one does not exist.
The Internet, however, is full of garbage cooking advice, and while AI is quite happy to parrot that garbage, it's not significantly worse than it's sources.
The physical ones have indices, and the digital ones usually do as well (plus string search).
I often find their contents to be both incredibly broad, and incredibly shallow, and often lead to dishes that don't quite suit my tastes.
I have generally had more success with the internet than I have with them.
I usually just find recipes that are close to something I like and then modify them until I like them
I feel like cooking recipes should always be full of little notes from past attempts, to tweak them to your personal taste
And we aren't talking about something esoteric. How To Cook Everything. (4.5 rating on Amazon, 4.0 on Goodreads), one of the most popular cookbooks in the world...
Has its first recipe for chicken cutlets produce undercooked chicken (15 minutes at 325F has not resulted in 165F internally).
Worse yet, its chile recipe for some weird reason involves boiling and simmering an onion with the beans before throwing it out (WTF). It then has you drain the beans only to add that drained water right back (WTF?), unless you optionally replace it with tap water (WTF was the point of boiling the onion then?), so that you could bring them to boil again. The recipe barely has any spices besides actual chile peppers.
This... Is not good or helpful, but is incredibly opinionated with a bunch of bullshit steps.
Fuck. Does nothing good survive on the internet anymore.
I mostly don’t even need a recipe from AI, just the title and a 1-2 sentence description.
For me, the text-only recipes I already get out of ChatGPT today are good enough
With an image of sour grapes
The user is probably in a grocery store. They need to buy the stuff first. Showing them a picture of a finished product is an irrelevant distraction.
Then, if there is some arguably relevant content you can tack it on afterwards.
“My made up scenario is definitely more realistic than your made up scenario.”
Who goes shopping for the meal before they even have a number of guests?
Kinda just feels like google search results in AI, which imo is a step down from distilled information. Chatgpt is already able to generate charts and visuals upon request.
That's never, ever the case with the AI generated ones and makes them useless.
The rest is basically unlimited unless you're doing something very weird.
Meanwhile I cant ask a simple question to Claude, not even to Haiku, because I used my claude code 5h limit.
Regular chat, even on extra high thinking mode, is sort of unlimited. Idk at what point you hit the abuse gaurdrail but its really high whatever it is.
E.g. codex reviews come from codex quota but also has its separate quota they don’t tell you about. Same with security reviews which has another quota still.
So the chat messages can be sent at several effort levels and from various models including 5.6 and 6 pro. The 6 pro messages are limited and they don’t tell you how many you’ve sent or how many are left. But the 5.6 messages seem to be unlimited, however there could just be a much higher unstated quota.
Prepare to lose essentially unlimited chat mode.
I imagine many will move to claude, as I will (return), unless anthropic makes more blunders.
Random link I found looking for the twitter post
https://pasqualepillitteri.it/en/news/21024/openai-merge-cha...
The only things that OAI still have over A/ are better coding agent GUI, no 5hr limit on >$100, much more reasonable cybersec guardrails (that allow most RE work) and... that's it.
I'm more worried about picture accuracy: usually the benefit of pictures in recipes is to see what you're aiming at (eg, how finely chopped something is), but I don't know if the image model is up to that level of detail.
As a highly literate person, it is easy to overestimate the share of the population that is highly literate.
Yes, pictures OF THE FOOD. Not unrelated pictures of another lamb-roast made with another recipe.
Good cookbooks do proper food photography, of the food made with the recipe. That's what sets them apart from slop, be it human or AI slop, cook books.
For almost everything else we can create them - just give a reasonable cook a kitchen and the ingredients and they will do it. For a picture we then have an artist do preparation to make it look good, but they should always start with the real dish.
My most used prompt in the last week is probably "Explain this concisely and simply, like I am a child".
I spent years dumbing things down and creating visuals to support it for decision makers. I am frequently asking ChatGPT to do the same for me.
Explain it to me like I'm a golden retriever.
My go to phrase that I’ve found to work well is “please restate in high school English with direct declarative sentences and no parentheticals or references that require recalling previous turns of this conversation”
Which I have as a hotkey.
All those complaints about Claude language use make me think we're finally seeing the consequences of a generation growing up on instant messaging.
If I say Claude talks like a meringue it's patently obvious what I'm trying to convey, sweet but full of empty space. If you hear that same metaphor day-in day-out then after a while it doesn't call up a thought at all, it's just an annoying verbal habit.
mine is “explain this 2 me in simple terms”
And that is a problem. I am not guiltless ( I am working one such abomination now! ). The dumbing down of everything has to stop somewhere. Just because my boss thinks hiss bosses are retards does not mean I have to endure simplified muse where I can't change much.. I get that normal people like it. Fuck em. They are making their choice.
For me, I try to spend as little time in QuickBooks as possible.
We make little abstractions of the world around us to help us navigate it. Nothing wrong with that.
There is plenty wrong as we use more and more complex machines with less and less access under the hood without crowbar or a super secret handshake. The amusing part is that it is bound to change now -- just not for the reasons I like.
So now the generation that grew up on IMs, whose standards of communication is "lol ok thx" and "i cannot read sentence with commas in it", are having difficulty with reading Claude, and in response, Anthropic is reducing their models' communications skills to match.
I am not exactly the person to normally to argue along the lines of 'build your own ISP/network/platform/model/we', but it will still be easier than fighting against collective nature of society.
What is absolutely crazy to me is that llms allow an insane way to accumulate knowledge and maybe even some skills - though ai doomers/boosters will argue over what that actually means -- and people's apparent response is:
can you make me think even less?
> ASD-STE100 Simplified Technical English (STE) is a controlled natural language that is designed to simplify and clarify technical documentation. It was originally developed in the 1980s by the European Association of Aerospace Industries (AECMA) at the request of the European airline industry, which wanted a standardized form of English for aircraft maintenance documentation that could be easily understood by non-native English-speakers.
I've taken a liking to pointing agents at the Simple English Wikipedia editorial guidelines...but so much of that is for making up for shortcomings of modern Anthropic models. Codex running 6.1-sol explains things so much more clearly that I don't feel I need to assert a style guide on its output.
GPT stays too much in text imho.
- Gemini 3.8 has a beautiful answer. Typography, use of lists, brevity, style, even the font choice on the web harness -- all of it is A tier or even S tier. Notice I didn't say answer quality. The answer is always extremely mid, and the harder the question (or more effort needed to answer it), Google just quits. I give it a "C" on answer quality
- ChatGPT (Chat, pre 6, I assume 5.6 although they commonly hide the model picker): Ugly answer. Runs on printing page after page of unnecessary side notes. Mostly just walls of texts in paragraphs with little thought to information design. However, inside of that mess of text is almost always the answer I'm looking for, and it's almost always incredibly better than the truncated, desgined Gemini answer.
For a while I ran all of my question queries on Gemini and ChatGPT, but I noticed that I basically always picked ChatGPT's answer head to head.
I even found it was capable of stepping back from code implementation and bug hunting to “hang on, it’s not the algorithm that is incorrectly implemented, it is the wrong algorithm for the goal.” GPT Sol (5.6,6 and 6.1) just kept hammering in the problem.
While I do rarely use the web harnesses for technical research, nearly all of my technical questions are going through coding harnesses and layers of my local context, skills, etc, so what I produce from codex/claude code is not comparable to what others produce.
So my comparison of the web harnesses is me asking questions about caring for my baby, or product comparisons of new televisions, or visualizing a room organized differently (google sucked at this so bad, chatgpt did great), or asking for help in a video game, or other normie AI uses lol
I will say, Gemini is the best for asking about google services, such as Flights or Youtube.
That's a very high bar.
As I said I can see the value in this similar to Pinterest boards. It is directionally correct and serves most of the value that a picture does. No perfect but probably close enough for the general audience.
And you see this in the other examples too. The teach the CLT piece was funny. It's and error-ridden mess that isn't even an explanation at all!
The best part is, someone looked at that and thought yes, that's right. Shows you clearly why programmers will be needed in the future. The LLM might be able to run circles around that person in math, but that person also has no idea when they're being bullshitted.
Sometimes one needs a top comment like this to remind them how out of touch the average HN voter is.
The fact we've come so far with textual models & the image stuff is still producing stuff that harks back to early-stage tripophobic AI produce (the garlic with roast potatoes here) is interesting.
Don't get me wrong, I'm certainly starting to see some AI-produced imagery today that I can't tell isn't real, but that's largely because a lot of real photography is overproduced & ugly. I have yet to see anything that's aesthetically good.
Given that OpenAI is making noises about merging work with chat (a horrible idea imo), and Work is very similar to Codex... I dearly hope things like these won't have any meaningful cross-polination into the actual work tools.
Seeing that the chat is based on 6.0 and not 6.1 is disappointing. The "Visual and interactive explanations" seems genuinely useful, but 6.1 is just so much better. I wouldn't truly trust the 6.1 with the explanations, but I'd trust them a fair bit more than 6.0. I understand that compute isn't infinite, but tons of people only interact with the chat and having your "things-explainer" be as good as it can be is important when people increasingly treat AI models as the source of truth, or even use them for academic learning and whatnot.
Still, the models will improve, so the visual explainer seems pretty good as an idea / mvp.