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

An Exploration into Automated Information Era 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 day by day — typically hourly — foundation, a lot information to devour, so many tweets to learn, how do I make all of it work?

Wouldn’t or not 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 actually matter, and summarize them, so I can get all of the information with morning espresso?

So I’ve determined to perform a little experiment — a social experiment with no people concerned — and easily let the generative fashions learn, combination, filter and summarize the vital information for me. The whole lot will probably be performed mechanically, with none human intervention. How good will the outcome be? Will it is smart? 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 day by day — this isn’t a trivial name. I noticed I’ve two major necessities from the LLM I’ll select:

1. It wants a protracted context window. The mannequin will scan by means of and browse a number of totally different articles earlier than serving me with one thing, so it wants the power to retailer quite a lot of knowledge in its reminiscence.
2. It must work properly with exterior instruments. Clearly, the mannequin will probably be required to look the net and entry web sites on my behalf, so working with exterior instruments in an efficient method is essential.

With these two necessities in thoughts, I got here to the conclusion that GPT-4 Turbo is the mannequin to go along 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 (despite the fact that I haven’t used it on this venture), I’ve determined to go along with the second possibility — I’ll have a number of brokers, every with their very own position, interacting with each other to create a day by day concern for my AI information journal. After some trial-and-error, I’ve converged to 4 varieties of brokers, working collectively:

1. Editor-in-Chief. That’s the agent that governs every thing, and ultimately has the final phrase. The Editor doesn’t write any article — they solely edit the reporters articles. The Editor can also be the one to transient the reporters about what to search for, and likewise has the ultimate determination in what will probably be featured within the day by day concern.
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 special system immediate, which ought to ideally lead to totally different web-searches and totally different article choice.
3. Tutorial Reporter. One of many issues I shortly realized is rather like people, giving brokers to many choices yields confusion. As a substitute of asking the identical reporters to do analysis each on-line and on Arxiv, I break up the duties, and gave the academic-research activity to a separate reporter, dealing solely with this.
4. Twitter Analyst. Within the area of AI, information and traits typically begin off as tweets earlier than getting headlines on extra conventional media. Realizing that, I created an agent specializing in looking knowledge on Twitter, which then notifies the editor what everybody’s is speaking about.

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

Instruments

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

1. Internet Search. The standard of the journal will instantly correlated to the brokers search skill. Due to this fact, I gave them entry to Google Search. Getting began with entails organising a Google Console account with an lively Search API, and organising a Customized Search Engine. As soon as performed, the official Python bundle might 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 possibility 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 might be performed with a couple of traces of code utilizing requests and BeautifulSoup.
3. Accessing Arxiv. Somewhat documentation-lacking too, however Arxiv makes it very straightforward to look 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 information. I used PyPDF.
4. Accessing Twitter. This one is a bit of difficult. Twitter below Elon Musk costs $100/month to entry Twitter API. As a workaround, I used Google search whereas limiting it with web site:twitter.com. This appears to be working fairly properly for public tweets, that are the overwhelming majority.
5. Journal Archive. Information can someday be duplicated, and a subject mentioned on one web site at this time may 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 performed, I created embeddings of every article within the journal, and permit the the Editor to look in an identical 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 net search capabilities to Twitter knowledge entry, I used to be able to outline the day by day operations of my AI-driven newsroom. This setup dictated how the brokers would work together and the way the complete course of would unfold every day.

The Routine

Now we’ve the decided the brokers and arrange their instruments, it’s time to find out how the day by day routine will appear to be. I had two conflicting tips 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 with a view to cut back prices. Finally, the next routine was the one which labored greatest for me:

It goes like this:

1. The routine begins with the Editor getting a common overview of what I’m anticipating of the journal to be — what’s the fields and particular subjects I’m it.
2. Within the meantime, the Twitter Analyst comes up with an inventory of individuals to comply with on Twitter, and checks what they’re speaking about. It compiles an inventory of traits, and sends them to the Editor.
3. The Editor takes under 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 net and Arxiv, and ship an inventory of one of the best objects they discovered again to the Editor. Who decides what are the highest objects? The reporters themselves, in fact.
5. The Editor appears in any respect the ideas and does a number of issues:
   – It decides what are the objects to be featured within the concern, and asks the reporters to write down
   – It combines a number of ideas about the identical matter from totally different sources, to keep away from duplications
   – It appears up the articles subjects within the Journal Archive, verifying this matter wasn’t lined 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 day by day routine that effectively makes use of our brokers and instruments, I targeted subsequent on the individuality 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 gained’t be the one writing them.

Because the Editor is the one in cost, the primary activity 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 instantly, 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 exploit one other agent, with just one activity — to explain to me a number of totally 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 totally different points. Each concern is exclusive.

Creating content material is nice, however it must be served someway. I made a decision to go along with a easy and environment friendly resolution — GiHub Pages. All I wanted to do is to ensure 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 acquired an internet site. I additionally built-in GitHub Actions to set off the routine each morning, so when wake there’s a brand new recent concern prepared for me.

However then I noticed that I truly wish to get my information after 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 part 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 concern is added, creating an RSS feed is simple. Now, for those who didn’t know, apparently organising a podcast solely requires one factor — an RSS feed. So, in a matter of minutes, my information narration grew to become obtainable on Spotify, and now I get me information each morning whereas I’m out for a stroll.

Whereas the day by 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 quite a bit, 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 part — the one the place the reporters search on-line for articles — was the most costly 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 usually 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 often very lengthy — I noticed a single URL would possibly require even 27 tokens. Take into account the quantity of URLs which are being processed — that turns into quite a lot 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 motive — all numbers which have as much as three digits (0–999) are transformed to a single token. That easy change within the URLs illustration dropped the prices of the analysis part by greater than 50%!

There are most likely extra methods to scale back prices. I’m nonetheless taking part in round with this, studying find out how to optimize it higher 💪.