Hacker News Reader

Comment by josefresco | Hacker News Reader

Tried using Bing ChatGTP to create a logo for me because I saw a cool demo on Instagram. It couldn't write text correctly and I tried numerous times. Not only that, the logos were terrible and didn't "mimic" the image I uploaded as a guide (which was in the demo) Today I needed to transcribe an MP3. Nope, can't do that either.

However on Monday I had it write some PHP and that was neat!

50/50 for me.

View on HN

For logos you'd be better off with DALL-E 2, no? How do you even get a useful graphical logo out of a text-based language model?

Bing chat uses DALL-E for image generation. The results were complete gibberish. Characters that looked like an alien alphabet and the graphic was also so abstract to not represent anything.

Generative image systems are notoriously poor at producing legible text without significant rerolls. You'd be far better off generating a textless logo and then layering text afterwards using PS, Photopea, Krita, etc,

why are they so bad at text?

These models work on a "latent" representation of the image, instead of working on the pixels directly. The model essentially learns a compression algorithm which allows it to process images most efficiently according to its loss function. The latent space might not represent certain types of shapes very well, as is the case with text characters.

Models like DeepFloyd IF work in pixel space, without learning a latent space. This means they can capture much more detail, but require more computational power. This is why these types of models usually rely on creating a smaller image and then using a separate model to upscale the image.

Transcribing an MP3 is a solved problem, in some contexts. What were you doing to make this a fail?

Gave Bing Chat an MP3 URL and it just refused. I know I can do this with software on my PC, or maybe there's an online service where I can upload but I was hoping this would be something Bing Chat/ChatGTP would excel at. I tried first giving it the URL of the article where the MP3 was embedded, and then tried a direct link.

I'm still lost on how you reached the conclusion Bing Chat should do be able that? Does MS imply it can do that somewhere?

Did you assume Whisper is from OpenAI => OpenAI is related to Microsoft => Bing Chat can transcribe MP3s?

I'm really just lost on the thought process here.

People hear “ai” and think it’s a near-human that cannot merely answer to text but also see, hear and do shopping.

You can upload images to Bing Chat for analysis (see), you can talk to it (hear) and yes, it will also shop for you lol.

Bing Chat is marketed as an AI assistant with a long list of capabilities beyond simply generating copy. It's not far fetched to think it could also transcribe audio. How is a user supposed to know that their AI assistant can and can't do without asking?

Not exactly the longest list: https://www.microsoft.com/en-us/edge/features/bing-chat

Are you assuming people should know what their AI assistant can do without asking? Because I'm saying that a tool not doing some arbitrary task it never implied it could isn't surprising, I don't think anyone has proposed what you just said.

Can you provide any resources that show transcribing arbitrary audio? In my experience using VAMP plugins and librosa, I could get about 80% there, but I'm very interested in any projects that get this right.

EDIT: I see now that you mean speech to text, and not music to sheet music, so maybe disregard this. Unless you do happen to know about what I'm talking about.

Reply on news.ycombinator.com