How to build your own AI model and trap it in your house
PLUS We are implementing AI strategies in newsrooms all over the world
For the longest time I have told people, “I’m not a coder. I use ChatGPT.” But that has begun to change. I still use ChatGPT for help, but I am getting a tighter grip every day on how software is built and what it is capable of.
This week I upped my tech skills and downloaded my own AI model. I opted for American. Thanks Zuck. And I loaded it on a PC in my lounge (a machine that is traditionally used for storing and playing 90s movies).
The first step was to download the latest Llama model (that’s Meta’s version of ChatGPT). You can use Zuck’s AI on the web and even (in certain countries) straight through WhatsApp. However, on a dedicated PC in my house I have loaded a considerably trimmed down version. To put it in perspective, I am running Llama 3.2 1B (this has 1 billion parameters while the full model has 90 billion). This means that as an AI it struggles to do much more than basic text, but the data stays entirely in my house. I don’t yet have the hardware capacity to go all in.
The key to doing this is I want MediaMap (our app that is mapping the world media’s implementation of AI) to not constantly call the OpenAI API to generate insights and chat responses, because every time it does I have to throw Sam Altman a few coins. My plan is for MediaMap to call the Llama model in my home. Of course, at the moment this model of Llama operates like a broken version of ChatGPT, but you have to start somewhere. And this brings us to the necessary hardware…
There are electricity costs (and loadshedding concerns), but the big expense is the hardware. You are looking at thousands of USD for a system that can run a large version of Llama (with the sweet 90 billion parameters rather than the 1 billion I have now) and run it at a speed which is tolerable. So, I am going to need to upgrade the hardware. But then I’ll have a functional LLM that I can “speak” to (and so can anyone else using MediaMap) with theoretically no ongoing cost.
And then the next step will be to “fine-tune” the model for specific tasks. There are a host of libraries on Hugging Face and I am imagining this set-up being perfect for our Justice AI project that is designed to help journalists choose legal cases that can deliver impactful stories more effectively.
After China dropped the DeepSeek bomb last week, Open Source AI has edged closer to being the norm. I wrote about Meta’s move towards Open Source early last year and it still seems like their strategy is to give away Llama in order to stay relevant and encourage developers to work within the Meta ecosystem. It’s assumed the larger play is to have the world using Llama and then increase user engagement with Meta’s platforms (though I have felt no pressure to use more Facebook or Instagram during my own operation).
I will keep you updated on how this develops (and to what extent the electricity bill balloons).
This week’s AI tool to use…
I have been deep in the OSINT (Open Source Intelligence) mines. The data that is publicly available and the ways you can analyse it with AI is truly outstanding. I’ll be writing about how we can all become mini-investigators in an upcoming letter. I honestly believe crowd encouraged investigating could be the future of media in general, but that’s another story.
In the meantime, a few fun tools have leapt out at me (though be warned, even though these tools use publicly available data, a few are proprietary, so not Open Source, and could lock you into a pricing plan). For example, a tool like Hunchly captures all the websites you visit and documents them so all that juicy research is saved and easily retrievable. And Shodan feels futuristic and frightening as it allows you to search for anything connected to the Internet from power plants to refrigerators to open servers (and helps you secure your own Internet enabled objects).
What AI was used in creating this newsletter?
I have an unhealthy obsession with AI product subscriptions. They are “for work” but like any addiction, they need to be managed. In that spirit, I cancelled Midjourney last week and I am passing this hardship on to you, the reader, because it means the images on this newsletter will now be created with ChatGPT. Sorry.
If you are enjoying what you’re reading please consider paying $5 a month to help support this newsletter.
In the news
The BBC published this report about the inaccuracies of AI models this week. And on its face this report makes sense: LLMs are not reliable for truthful information. They tested ChatGPT, Copilot, Gemini and Perplexity. The report says: 51% of all AI answers to questions about the news had issues. 19% of AI answers which cited BBC content created factual errors. And 13% of the quotes sourced from BBC articles were either altered or made up. I wrote the other week about how you don’t need to even use an LLM to get BBC misinformation, now your phone can add errors to their news directly. I imagine the BBC would love it if you just went directly to their app or website, but I would like to know how these stats compare to the accuracy of an average Google search or Wikipedia page.
The DeepSeek China AI hyper cycle seems to have slowed, for now. Of course, OpenAI was upset and it was delicious when a company that has built its premium product by stealing the images, audio, text and video created by others then starts to moan about another company stealing from them.
Burn of the week: everyone’s rich dad Elon Musk offered robot face Sam Altman $97.4 billion for OpenAI. And Altman hit back (on Musk’s own platform) with, “no thank you but we will buy twitter for $9.74 billion if you want”.
What is happening at Develop AI?
I am training journalists from all around Africa in AI, starting this Thursday. Our course is online and for six weeks. My co-trainer is the formidable Peter Deselaers. A huge thanks to him and DW Akademie for the opportunity to help a bunch of talented journalists use AI more effectively.
I have taken on consulting several newsrooms in The Balkans to help them implement AI successfully. I am looking forward to seeing a host of exciting outcomes after they use AI in productive ways.
We are pushing forward with funding proposals for our Justice AI project to help enhance criminal justice reporting around the world with AI.
I am writing an article on “AI Pioneers” and if you would like to nominate someone, please get in touch.
See you next week. All the best,
Develop Al is an innovative consulting and training company that creates AI strategies for businesses & newsrooms so they can implement AI effectively.
Check out Develop AI’s consulting services and conference appearances.
Ask about our training workshops and read about MediaMap, our project that is using AI to map the world media’s implementation of AI.
Email me directly on paul@developai.co.za. Or find me on our WhatsApp Community.
Watch us on YouTube. Follow us on LinkedIn / Facebook / Instagram / TikTok / Threads and X. Or visit the website.
Listen to Develop Audio’s investigative podcasts, including the award-winning Alibi, Asylum and The Last Afternoon In The Garden. Get in touch and we will make you a world-class podcast of your own.
Thank you to all our paid subscribers. For $5 a month you can help support this newsletter (and keep it free for everyone).