Weekly Findings In Data Science and AI
A New AI Search Engine And A Mobile App For Stable Diffusion — Weekly Findings
An iPhone app for Stable Diffusion, GitHub Codespaces and a new AI search engine for the web in this week's findings.
Metaphor — Search The Web Using Generative AI
Metaphor is a new search engine that uses generative AI similar to GPT-3, but instead of predicting text, it predicts links.
I’ve tried using GPT-3 to do this before by asking it to suggest me links for certain things, but the results have been poor. Often the provided links return “not found”. Metaphor seems to be doing much better.
On Metaphor’s Twitter page they gave an example where they prompted for persons similar to Elon Musk with specific characteristics, like a person living during the 19th century. I tried to do the same with Sherlock Holmes with the following prompt:
Examples of clever fictional characters that are similar to Sherlock Holmes are.
A couple of the results were different pages describing Sherlock Holmes. Others retrieved what was searched for with Wikipedia pages of similar detectives: Hercule Poirot, Father Brown, Lord Peter Wimsey, etc.
If you ask Google the same question, most results are written about the question. For instance, one of the results from Google had the title“Is there any other fictional character like Sherlock Holmes?”. But not links to the actual answer (i.e a site describing a character that answers the question). I think this is a good example of a discrepancy between Google and Metaphor.
There are clearly some weaknesses in the search engine though. For instance, it seems to be returning multiple identical pages such as the following:
This is especially annoying if you search for some library and there are multiple copies of it for different versions of the library, i.e. version 1.2, version 1.1, etc.
Overall, I think the idea has potential. It would be interesting to see a comprehensive comparison between it and other search engines to see if it’s more functional or not. The link to the website can be found below:
Mobile Stable Diffusion
Stable Diffusion has found its way to the iPhone in a free app called Draw Things: AI Generation. In this app you do not send prompts to a server, the models are actually running directly on your phone! Here are some of the features included in the app:
- Edit existing images
- Use different models
- Inpainting (repainting parts of an image)
- Outpainting (extending an image by generating outside of it)
- Negative prompts (specify what you don’t want in the image)
- Able to change different parameters for the image generation
Be aware that the models will be downloaded and consume a couple of GBs of memory. Testing it on an iPhone 13 Pro, it took about a minute to generate an image using the default settings.
The app could be a fun way to pass time when you’re bored, commuting, etc. Also, since it’s run locally, you don’t need internet access and you don’t have to share the images.
I can only imagine more and more mobile apps will start to utilize these models, and it will only become easier as they become smaller and more efficient. Great opportunity to create new innovative applications.
The link to the app can be found below:
GitHub Codespaces
GitHub Codespaces is a development environment running in the cloud that can be accessed from the browser or IDEs and was recently released for everyone. It is similar to Google Colab but for entire projects with options such as configurable computing power, choice of editor (VS Code, JetBrains Gateway, JupyterLab), and integrations with GitHub repositories.
You can use it for free (for individual use) for 60 hours each month using your GitHub account.
I see several advantages with GitHub Codespaces:
- Fewer problems with people developing on different operating systems (i.e. Linux vs. Windows vs. Mac). Becomes a sort of “Docker for the development environment”.
- Less need for a powerful computer
- Potential to work from a variety of devices, even an iPad
I also see some disadvantages though:
- Seems a bit wasteful at times, why run something on a remote computer when you don’t have to? Sometimes it’s useful, e.g. running machine learning tasks on a powerful computer, but in other circumstances?
- Not sure if it will always work, e.g. if you have a large local database you’re testing against
- Will not work without an internet connection
- Less privacy. Employers could potentially obtain extensive analytics of everything you do.
- Another cost of development
Perhaps the usage of online tools such as GitHub Codespaces will become more frequent in the future.
Link below:
That was it for this week, see you next week!
Check out the next article in the series:
If you’re interested in reading more articles about data science or AI, check out my reading lists below:
If you’d like to get a Medium membership you can use my referral link if you wish. Have a nice day.
