Introducing the First-Ever AI Journal & Podcast Made By AI | by Shaked Zychlinski 🎗️ | The Startup | Apr, 2024

An Exploration into Automated Information Technology with No Human Oversight

Tech by AI is on the market at techbyai.information. Open supply code is on the market on github.com/shakedzy/techbyai

Tech and AI are advancing quick. Actually quick. So quick I can’t sustain with the tempo, and located myself misplaced when making an attempt to. There are new discoveries and fashions on a each day — generally hourly — foundation, a lot information to devour, so many tweets to learn, how do I make all of it work?

Wouldn’t it’s nice if somebody — or let’s say, one thing — would collect all of the information for me, filter out solely the issues that basically matter, and summarize them, so I can get all of the information with morning espresso?

So I’ve determined to do some experiment — a social experiment with no people concerned — and easily let the generative fashions learn, mixture, filter and summarize the necessary information for me. Every little thing will likely be accomplished robotically, with none human intervention. How good will the end result be? Will it is sensible? How a lot will it value? There’s just one solution to discover out.

Selecting a Mannequin

Clearly, essentially the most essential half is which LLM to make use of? There are such a lot of available on the market — with new ones becoming a member of each day — this isn’t a trivial name. I noticed I’ve two foremost necessities from the LLM I’ll select:

1. It wants an extended context window. The mannequin will scan via and browse a number of completely different articles earlier than serving me with one thing, so it wants the power to retailer loads of knowledge in its reminiscence.
2. It must work effectively with exterior instruments. Clearly, the mannequin will likely be required to go looking the online and entry web sites on my behalf, so working with exterior instruments in an efficient means is essential.

With these two necessities in thoughts, I got here to the conclusion that GPT-4 Turbo is the mannequin to go together with. So now that I’ve the mannequin to energy my newsroom, it was time to ask how will the newsroom function? Am I simply going to ask GPT to “summarize information on the internet” for me, or do I would like it to work together with different folks — or fashions — like an actual newsroom?

Brokers

Impressed a lot by Microsoft’s AutoGen (regardless that I haven’t used it on this challenge), I’ve determined to go together with the second choice — I’ll have a number of brokers, every with their very own function, interacting with each other to create a each day problem for my AI information journal. After some trial-and-error, I’ve converged to 4 forms of brokers, working collectively:

1. Editor-in-Chief. That’s the agent that governs every little thing, and ultimately has the final phrase. The Editor doesn’t write any article — they solely edit the reporters articles. The Editor can be the one to temporary the reporters about what to search for, and likewise has the ultimate choice in what will likely be featured within the each day problem.
2. Reporters. Reporters are the brokers which do the analysis on-line, choose the highest articles and write about these chosen by the Editor. There’s a couple of reporter, as the purpose is to have every which a unique system immediate, which ought to ideally lead to completely different web-searches and completely different article choice.
3. Educational Reporter. One of many issues I shortly realized is rather like people, giving brokers to many choices yields confusion. As an alternative of asking the identical reporters to do analysis each on-line and on Arxiv, I cut up the duties, and gave the academic-research job to a separate reporter, dealing solely with this.
4. Twitter Analyst. Within the area of AI, information and traits generally begin off as tweets earlier than getting headlines on extra conventional media. Realizing that, I created an agent specializing in looking out knowledge on Twitter, which then notifies the editor what everybody’s is speaking about.

Having established these roles, it turned clear that I must focus now on offering them with strong instruments to successfully collect and course of info. This requirement led me to discover and arrange the required digital infrastructure.

Instruments

Speaking with the outside-world is crucial factor for my newsroom brokers to efficiently accomplish their assignments. Listed here are the instruments I wanted, and the way I created them:

1. Net Search. The standard of the journal will straight correlated to the brokers search capability. Due to this fact, I gave them entry to Google Search. Getting began with includes establishing a Google Console account with an energetic Search API, and establishing a Customized Search Engine. As soon as accomplished, the official Python bundle may be put in from PyPI: google-api-python-client. The documentation isn’t nice, although.
   (FYI, there’s one other free, out-of-the-box, no-questions-asked choice by DuckDuckGo).
2. Accessing Web sites. As soon as discovered, the articles must be learn. In Python, making a easy too to scrape textual content from web site may be accomplished with just a few traces of code utilizing requests and BeautifulSoup.
3. Accessing Arxiv. A bit documentation-lacking too, however Arxiv makes it very straightforward to go looking and obtain PDFs from it. There’s additionally a fairly straightforward to make use of Python library named arxiv. We’ll want one other library for parsing the PDF recordsdata. I used PyPDF.
4. Accessing Twitter. This one is a bit of difficult. Twitter beneath Elon Musk expenses $100/month to entry Twitter API. As a workaround, I used Google search whereas limiting it with website:twitter.com. This appears to be working fairly effectively for public tweets, that are the overwhelming majority.
5. Journal Archive. Information can someday be duplicated, and a subject mentioned on one website at the moment might need appeared on one other yesterday. I wished to present the Editor an choice to seek for articles within the journal’s archive, and test if there are any related headlines from earlier than. To get this accomplished, I created embeddings of every article within the journal, and permit the the Editor to go looking in an analogous solution to how RAG works. As this little or no knowledge, I used a naive Numpy array and Pandas DataFrame because the vector DB.

With the instruments in place, from internet search capabilities to Twitter knowledge entry, I used to be able to outline the each day operations of my AI-driven newsroom. This setup dictated how the brokers would work together and the way your complete course of would unfold every day.

The Routine

Now we’ve got the decided the brokers and arrange their instruments, it’s time to find out how the each day routine will seem like. I had two conflicting pointers right here — the primary was to let the brokers work together as a lot as wanted with each other, and the second was to restrict their interactions as a way to cut back prices. Ultimately, the next routine was the one which labored greatest for me:

It goes like this:

1. The routine begins with the Editor getting a normal overview of what I’m anticipating of the journal to be — what’s the fields and particular matters I’m it.
2. Within the meantime, the Twitter Analyst comes up with a listing of individuals to comply with on Twitter, and checks what they’re speaking about. It compiles a listing of traits, and sends them to the Editor.
3. The Editor takes into consideration all these inputs, and creates a briefing for the reporters, asking them what to search for and write about.
4. The reporters search around the online and Arxiv, and ship a listing of the most effective objects they discovered again to the Editor. Who decides what are the highest objects? The reporters themselves, after all.
5. The Editor appears to be like in any respect the solutions and does a number of issues:
   – It decides what are the objects to be featured within the problem, and asks the reporters to write down
   – It combines a number of solutions about the identical subject from completely different sources, to keep away from duplications
   – It appears to be like up the articles matters within the Journal Archive, verifying this subject wasn’t coated already
6. Reporters summarize the articles, and hand their drafts to the Editor.
7. The Editor has the ultimate say, and has the choice to edit the texts. The ultimate edit is being served to me

This whole course of takes rather less than 5 minutes, and prices differ from $1 to $5, relying on the size of texts learn by the brokers.

After outlining the each day routine that effectively makes use of our brokers and instruments, I targeted subsequent on the distinctiveness of every publication. This uniqueness is primarily pushed by the system prompts of every agent, curated to inject selection and depth into the content material they generate. Which is why I made a decision I received’t be the one writing them.

Because the Editor is the one in cost, the primary job it will get is to rent the reporters. The Editor is requested to explain the traits of the reporters which would be the greatest match for the newsroom. I ask the Editor to explain them in second physique, as if addressing them straight, telling them who they’re. I then take these descriptions and use them because the reporters system prompts.

And who decides what’s the system immediate of the Editor? For that I take advantage of one other agent, with just one job — to explain to me a number of completely different editors and their traits, once more in second physique. From these I randomly choose one, and assign it because the Editor. Add to that the truth that all brokers temperature is ready to ~0.5, and also you’ll understand that for those who run the identical routine 10 occasions in arow, you’ll get utterly completely different points. Each problem is exclusive.

Creating content material is nice, but it surely must be served by some means. I made a decision to go together with a easy and environment friendly resolution — GiHub Pages. All I wanted to do is to verify the ultimate edit is written in Markdown. I used a clear and MIT-licensed Jekyll theme I discovered on-line, and that’s just about it — I received a web site. I additionally built-in GitHub Actions to set off the routine each morning, so when wake there’s a brand new contemporary problem prepared for me.

However then I noticed that I truly wish to get my information once I stroll my canine within the morning — and it’ll be nice if the information could possibly be narrated for me. So I added one final section to the routine — narration. To maintain it easy, as I’m already utilizing OpenAI API each for GPT and the embeddings, I made a decision to make use of the corporate’s text-to-speech API too. And as Jekyll and GitHub Pages render my web site each time a brand new problem is added, creating an RSS feed is easy. Now, for those who didn’t know, apparently establishing a podcast solely requires one factor — an RSS feed. So, in a matter of minutes, my information narration turned accessible on Spotify, and now I get me information each morning whereas I’m out for a stroll.

Whereas the each day prices had been at all times within the vary of $1 to $5, as days glided by, I observed they stabilized round ~$3.5. Which is isn’t so much, however that’s nonetheless greater than I used to be anticipating, because it provides as much as ~$105 a month. So I took a deeper look into the prices breakdown, and observed that the analysis section — the one the place the reporters search on-line for articles — was the costliest a part of the method, reaching ~$2.7. Is there a solution to cut back the prices with out affecting outcomes? Sure — decreasing tokens.

Whereas English phrases are most often both a single token or two, URLs are a bit extra problematic. As there aren’t any areas, and phrases are both separated by dashes, slashes or by nothing in any respect, and are sometimes blended with numbers — and are additionally normally very lengthy — I noticed a single URL would possibly require even 27 tokens. Contemplate the quantity of URLs which are being processed — that turns into loads of tokens.

The answer was to map URLs to IDs. Behind the scenes I changed all URLs with a numeric ID, and gave that ID to the brokers. My code transformed URLs to IDs and vice-versa. I selected numeric IDs for a cause — all numbers which have as much as three digits (0–999) are transformed to a single token. That straightforward change within the URLs illustration dropped the prices of the analysis section by greater than 50%!

There are in all probability extra methods to cut back prices. I’m nonetheless taking part in round with this, studying learn how to optimize it higher 💪.