Video models are impressive but still clearly have a ways to go for that "last mile". I'm excited to test Seedance 2.0 when it's available. Also, the scores are completely vibes based (this is a fun weekend thing, not a serious benchmark) so if you complain enough I might change them if you make good arguments.
Hi all, I just wanted to share a quick utility for converting a PDF to Anki cards. I saw some paid options, but this is a simple enough process that it seemed an open source version should exist. You just have to add your API key in a file called "key.txt" in the same directory.
If anyone wants to improve upon it by refining prompts or doing some post-processing of the cards, please feel free! Also, if anyone knows of a better developed open-source version that's already out there, I'd be happy to start using that instead.
I actually think Dwarkesh is usually pretty good - this interview wasn’t his best (maybe he was a bit nervous because it’s Zuck?) but his show has had a lot of good conversations that get more into the weeds than other shows in my experience
Seconding this opinion: Dwarkesh's podcast is really good. I haven't watched all of the Zuck interview but I recommend others to check out a couple extra episodes to get a more representative sample. He is one of the few postcasters who does his homework.
The demo was incredible, and this seems perfect for my current project. I am going to try to integrate this as soon as I can to see if it works for me. How responsive can I expect the support to be?
I’m trying to make a smart speaker where you push a button to talk and when you release it sends the audio to a server for transcription and response. It also has to do some basic logic with LEDs. I also want it to be always on and available as long as it’s plugged in. Do you have any advice on what might be a better alternative to a Pi? This is basically my first foray into hardware so I’m trying to learn as much as possible!
I like the idea, but when I started practicing some Spanish, I noticed it gave me incorrect grammar advice (it claimed that the subjunctive form of “creer” is “creen” - it was right that the subjunctive was needed, but that would be “crean”). That makes me nervous to use it for a language that I know less well like Korean because I won’t be able to tell if the grammar feedback I’m getting is accurate.
If there’s a way to make it more robust on that front I could definitely see myself using it!
We're using LLMs and good ol'ML. These systems are never going to be 100% accurate. Then again humans are also not 100% correct and we're working hard on ironing out the kinks like the one you just discovered.
That is true of LLMs, but if you’re looking to replace iTalki, the standard to meet or beat on grammar correctness and explanation is that of a human instructor - this is not the type of mistake that my native tutor would make, and if he somehow did, I would find a new tutor that I could be confident is teaching me correct grammar.
I certainly don’t point it out to be discouraging - on the contrary, I feel like I am the exact target audience for a product like this and would happily pay for it if it can reach or be near the trustworthiness of a human teacher.
How are you guys assessing the accuracy? Sure, LLM's aren't going to be 100% accurate. But if I pick up a Spanish textbook, there is a right answer 100% of the time (okay, aside from publishing errors)
If you can't make any promises as to the quality of what you're offering, then how can you reasonably expect people to pay for it? At the end of the day you need to be able to tell consumers what they can reasonably expect from your product in terms of its capabilities and accuracy.
It could be somewhat embarassing for someone who trusts the app and then says the wrong thing to another human. As an off-the-cuff idea (not because you haven't already thought deeply but because I hate to find a problem without offering a solution), maybe use videos of real speakers for basic instruction, then the AI for responsive lessons, etc.?
As you no doubt have considered, but if you define you mission by the tool (AI) and not the outcome (accurate, faster language acquisition) you'll have a great tool and lesser outcomes. Whatever you prioritize, that's what you'll get! :)
