I stumbled across an idea while driving to work a few months ago. I had been testing several automation tools to streamline some projects I was testing, and through this, I discovered a way to create a fully automated news website. This vision was simple, to create a website that pulls information from reliable sources, write an article about the information it pulled, and then publish it without any interaction on my part.
After much deliberation, I decided on the name “Innovation Era”, a play on the growing AI trend and the threat of the results being something aligning with a new ‘era’ or ‘error’. I managed to build out pretty much everything to get it running in a weekend and thought it could be interesting to talk through how I approached this, as well as some of my considerations and learnings.
Ethical Challenges
There are a number of ethical and philosophical challenges with that vision. Firstly, how to source the information without stealing from the hard work that others have written and published unless they consented to their article being used. The objective was not to pass off other people’s work as my own, but if I could get the information from the initial source, then that should appease this ethical gray area. The decision, therefore, was to pull blog articles directly from company blogs that are used for their latest announcements.
AI is known to ‘hallucinate’ some things when prompted. Given the focus was to remove human intervention as much as possible, this meant there was a chance that AI could take the source information and write it in any number of ways, including ‘hallucinating’/making new information up. To address this, the original article would need to be easily accessible for people to read to fact-check if needed. An additional layer was included at the end of every article with a link to the source material.
It was also crucial to ensure that the output produced articles that were well written, abided by journalistic structure, and did not push any agenda. It simply needed to report the facts. Given the source, however, was coming from companies hyping their latest innovation, there was some AI training that needed to take place. Through modeling, an AI agent was trained on the ideal writing style I was looking for, and to ensure that the focus was on the details and not necessarily an opinion. After some fine-tuning and adjustments, the output became a lot tidier and easy to read.
Philosophical Debate & The Best Way To Use AI
Philosophically, it is scary to give complete control over to AI, without human intervention. My approach to using AI is what I call the 20/60/20 Rule. In short, 20% of your focus should be put into prompting and giving AI the information it needs to succeed. The more information you give AI, the better the results will be. Then, AI will do the heavy lifting, equating to roughly 60% of the work required. The final 20% is for you to make the result yours. Adjust and add your unique style and touches to the results.
With Innovation Era, I would need to give up full control of the final 20%. A lot of work was done to set everything up at the start, but little to no oversight with the results. Obviously, I would be reading every article, and can adjust, but it would already be published and available by the time I identified any issues. The ideal state is someone reviews every article before anything is published, but finding time to review things in relatively real time was not going to be possible or scalable.
The First Week
In the first week, Innovation Era published 20 articles. It wasn’t perfect, and got stuck publishing a couple of articles multiple times. However, the overall quality of the articles was solid and it was able to identify a new article across key sites, and within about an hour had written and published an update on Innovation Era.
Minor adjustments were made in the backend to fix some bugs and improve article quality. But overall I was very happy with the results. The only issue I found was in the quality of the images being produced. DALLE-3 was creating the images to align with the article, and it was an inconsistent producer of images. There was a possibility to replace these with Midjourney images, but was going to be an arduous task to change to this. At least at that time.
The First Two Months & Cost
After the initial run of testing, I let the automation take full control of the site. In all, it created close to 90 articles from 6-8 AI companies. Overall, the results were great and proved that with enough setup, AI could run a website with little to no oversight.
Each morning I would wake up and check the site to see what news had come through overnight. Traffic to the site was limited as I did nothing to drive visitors to the site, but there was some organic interest coming through. At this stage, it was less about growing an audience and more about making sure the tech worked, which it did. After the initial trial across the two months, I paused the AI so I could rebuild several backend hurdles to streamline and reduce some of the steps required.
Excluding the cost of hosting the website and buying the domain, the setup cost was relatively low. The automation platforms to source the news and trigger the AI Agents to start writing cost roughly $90 over the two months, or around $1 per article. The AI Agents cost around $0.04 to do all the writing, and each image costs roughly $0.20 to create. Overall, excluding site hosting, each article costs around $1.24 to create. This reduces with each article published, and the image cost has reduced with a recent change.
Latest Improvements
After the latest improvements made in recent days, images are now created by Flux, rather than DALLE, a huge improvement. In addition, the backend setup has been hugely streamlined, now able to do everything it was doing before, but with one-eighth of the number of steps required. Through this reduction, it is now possible to create an article in 15 minutes from the time something is published on a company’s site. This improvement also unlocks so many more opportunities. It is now incredibly easy to include any new sites to pull information from without any additional work. It is possible to add 30 more company sites to the automation in 30 seconds, and the automation will continue as usual.
The long-term ambition was to extend the content into a podcast format, something that I had tested but was still a very manual process. However, with the advancement of some of the AI audio tools, this is now incredibly easy. There is, in fact, now a (near) fully automated Innovation Era podcast that will discuss some of the latest news as it comes out. All are written and presented by AI hosts.
What’s Next?
This started as an idea and is already close to the ideal dream state I wanted to eventually build it towards. I anticipated this would be a fun little project that would take a year to get close to what I wanted to achieve. It took 3 months. The evolution of AI capabilities is rapid, and the challenges I was looking to work on, like the podcast, solved themselves.
The next hurdle is to make it a true news destination for people to frequently visit. With the improved backend adjustments, it is now so much easier to begin publishing content across social media, like Meta, X, and LinkedIn as a new article comes through.
The automations are all seemingly working, and there will likely be continuous tweaks done to improve this. But it’s now time to start trying to build site visits. The proof of concept is reality, and if people find it interesting, I can scale this up in a matter of minutes. The first site was built for AI news, but with everything now created, I can almost instantly create 20 more websites about any topic imaginable. The possibilities are endless.