10X AI (Issue #3): DeepFloyd IF, StarCoder, Speech-To-Notes AI, and a Cheating Surfer
Plus more stuff coming to Bing, Stability AI's new open-source LLM, learning negotiation skills with AI, and other tools and tips.
Happy Sunday, friends!
Welcome to 10X AI: a weekly look at beginner-focused AI news, tools, and tips.
Let’s dive right in.
AI news
Here are this week’s AI developments.
1. Text-to-image AI that can spell
Man, Stability AI are on a roll these days!
Two weeks ago, I wrote about their entry into the LLM game with StableLM.
Now they dropped a text-to-image model called DeepFloyd IF. What sets it apart from all other text-to-image models on the market is its ability to spell.
Here’s the pre-upscale grid from DeepFloyd IF for “shop sign that says ‘WELCOME’”:
This might not seem too impressive, but here’s what the latest and greatest Midjourney V5.1 returns for the same prompt:
Midjourney, I love you, but you spell like I drunk-texted back in my student days.
Here’s a Hugging Face demo of DeepFloyd IF if you’d like to give it a shot!
2. StarCoder is a coding star
Hey, what a convenient segue…
The Hugging Face crew aren’t sleeping either. Last week it was Hugging Chat, this week it’s StarCoder.
Hugging Face cooperated with ServiceNow to develop and release this fully open-source LLM model for code. And it does impressively well on commonly used benchmarks when pitted against similar coding LLMs:
Outside of actually downloading and installing the model, there are several ways to give it a try:
In a chat-like interface for complete beginners.
In a more advanced cooperative interface for those who can code.
3. More Stable chatbot goodness
But we’re not quite done with Stability AI just yet, folks!
Two weeks after dropping the StableLM family of LLMs, the company released StableVicuna, “the first large-scale open source chatbot trained via reinforced learning from human feedback.”
Here it is scoring high marks across multiple benchmarks I know nothing about:
There should be a public demo coming out shortly, so stay tuned!
4. Bing is about to get even better
Not to be outdone, Microsoft announced a whole range of new and evolving capabilities coming to AI-powered Bing, including:
No more waitlist. Bing AI is now available to everyone.
Visually rich answers and—soon—multimodal support. (If the “multimodal” part is anything like what we’ve seen in the GPT-4 demo, that could be massive.)
Saving and continuing chats across multiple sessions.
Support for third-party tools within Bing (similar to ChatGPT’s plugins).
Most of this should start rolling out shortly, so let’s see how it works out in practice.
AI tools
Check out this week’s selection of beginner-friendly AI tools.
5. Movie Deep Search
In early March, I showcased two AI-driven sites for movie recommendations.
Well, Movie Deep Search does something similar but supposedly has a deeper, more nuanced understanding of context and your search intent. So you can describe what you’re in the mood for in natural language (and plenty of detail), then get a bunch of movies to check out:
See if it helps you discover something new and relevant.
6. AudioPen
Do you sometimes tend to, like, say, uh, all sorts of just, like, stuff and then maybe you don’t know what the actual point even is but then some of it is good, but you can’t tell which parts, and wouldn’t it be nice to have someone sift through it all, or something, and then kind of make sense of it and summarize it?
That’s what AudioPen does!
It lets you just ramble into the microphone, then helpfully cleans up your stream-of-consciousness babbling, turning it into a coherent summary.
Like so (the original, intentionally rambling transcript is right below the orange summary box):
Quite cool!
AudioPen captured every word correctly and managed to distill the gist of my verbal vomit into a shorter text. It even came up with an appropriate headline for the note.
I also tested a competitor—AudioNotes—which did an equally solid job. But AudioPen has a more generous and unlimited free plan, so it wins this round!
7. DebateDevil
If you want to practice your debate skills or simply expand on your understanding of a topic, DebateDevil is a fun way to do so.
It’s basically a pre-prompted AI chat that has three modes:
I find the “Devil Mode” to be the most frustrating fun. It’s a bit like talking to your know-it-all-friend who plays the Devil’s Advocate in any discussion. (You know who you are, Gary!)
Here’s me debating the most controversial subject imaginable: the use of pineapples as pizza toppings:
As if the debate itself isn’t exasperating enough, there’s also an “AI Judge” that will evaluate the entire chat (and frequently side with its AI debater buddy):
Good times!
AI tips
Here are this week’s tips.
8. Practice negotiation skills with AI
Fed up with your Devil Mode chats?
Time to switch from debates to something more relaxing…like negotiations.
Turns out Bing chat (as well as ChatGPT) are pretty great at simulating different negotiation scenarios, letting you role play and practice in a low-stakes environment.
You could start with the following prompt from this article by
, but it's certainly possible to tweak it to your exact needs:I want to do deliberate practice about how to conduct negotiations. You will be my negotiation teacher. You will simulate a detailed scenario in which I have to engage in a negotiation. You will fill the role of one party, I will fill the role of the other. You will ask for my response in each step of the scenario and wait until you receive it. After getting my response, you will give me details of what the other party does and says. You will grade my response and give me detailed feedback about what to do better using the science of negotiation. You will give me a harder scenario if I do well, and an easier one if I fail.
Thanks to Bing chat’s ongoing tips, I just successfully bought a non-existent virtual car with an asking price of $10,000 for a mere $8,000:
9. Get paragraph-by-paragraph article summaries
I’m sure we’ve all asked Bing to summarize long texts for us before. But there are times when simply getting a bullet list of key takeaways isn’t enough.
For long-form articles and opinion pieces, I like to get a feel for the author’s cadence and structure as well.
In those cases, I explicitly ask Bing to condense each paragraph of a linked article into a sentence:
Read the following article [URL]. Condense each paragraph of text into a single short sentence that captures its essence.
What I get back is a much-condensed version of the article that still reflects its overall presentation and flow:
10. AI fail of the week
“Getting a leg up!”
"Midjourney, I love you, but you spell like I drunk-texted back in my student days." 😆 😆