I will admit that it is impressive that Meta allows you to download Llama 3 (their version of ChatGPT) and run it on your home machine without connecting to the Internet. This is a rare opportunity for folks like us. It means we can experiment with AI technologies and build our own applications without being plugged into the cloud.

However, most of what Meta is doing around AI involves their ubiquitous suit of apps. Their assistant Meta AI is entirely free to use and is slowly being included in all places where you keep in touch with your loved ones. So, Instagram now has an AI bot in its search bar. This feels slightly bolted on at the moment, but I imagine it will become more integrated. Even though Meta AI is available to everyone, the bumper features (like image generation) are only for people with a Facebook account.

To hear ‘Zuck’ talk like he is not one of the big evil tech guys he seems to dislike so much, check out the Dwarkesh Patel podcast on YouTube. It feels like Meta has read the room and understands that they can clear the very low bar of morality in this new wave of tech without doing too much. ‘Zuck’ is emerging as the tech world’s tallest dwarf.

‘Zuck’ insists on calling Llama 3 “open source” because it is a term people associate with ye olde Internet of good, free and fair. However, it technically isn’t open source. It is “open weight” which is fundamentally different and not as rosy.

If your model is open source then you make available the full training source code and the dataset required for retraining the model from scratch. This includes the model architecture code, training methodology and the original training dataset.

Read more in Daily Maverick: Meta Debuts Llama 3, Latest AI Model for Training Chatbots

Meta has not included these crucial parts. However, it has published “the weights and the inference code”. In brief: weights are the “knowledge” the model has gained from the training data (so one step away from the “source”) and the inference code is the set of instructions that uses the knowledge to make predictions or classify data.

They are claiming that Llama-3 will be tremendous for the open-weight community. And, to be fair, Llama-2 already has an ecosystem where dozens of decent open-weight models are released each week on HuggingFace. People are building chatbots, translation models and assistants for research. They can customise this model more than others and it’s free. Let’s be clear: at this point, Meta knows it doesn’t have to do this and they want us to be grateful.

What’s fascinating, when it comes to more day-to-day use, is how Meta AI stacks up against its rivals. They are adding real-time AI image generation to WhatsApp — if you were worried about visual disinformation on the platform before, it is about to go stratospheric. And when it comes to coding, the new platform absolutely stinks. When asked to write a WordPress plugin, rewrite a string function, find an annoying bug and write a script it is clear which assistant you need to keep using. ChatGPT nailed all four tasks and Meta AI only succeeded at finding the bug.

The strangest part of this is we are well into year two of the AI explosion and only now Meta is showing its hand. But regardless of its functionality right now they have the user base to accelerate its capacity like no other. Though, I am serious when it comes to disinformation: these integrations mean everything you see on social media is about to be way more dubious than ever before.

This week’s AI tool for people to use…

OpenAI’s Custom GPT Store made a huge splash when it arrived. It promised the chance for you to knock together your own GPT (using their built-in tools) and then sell what you created to other users. The financial side has not entirely materialised, but there are plenty of gems if you have the time to sift through the rubbish. We should all be happy for this post on Medium where they have done extensive work in assessing the tools available for content creation. Here are a few highlights:

• Improve my Writing is great at finding grammatical errors and correcting sentence structure. It is also top at adjusting the tone of a piece.
• Legal Eye is good for writing average legal articles with a dash of political analysis.
• Video Maker by Lucas AI converts your writing into reasonable videos. They are not of a high standard, but with additional editing, you can spruce them up.

In the news…

I’m not sure if this is good or bad news. Profluent just developed OpenCRISPR-1, the world’s first open-source AI-powered gene editor. It says it is capable of editing the human genome and uses LLM tech to produce new proteins. According to the website it promises to: “bypass the random process of evolution and move us towards intentionally designing proteins for a specific purpose”. CEO Ali Madani said there is a future where AI “precisely designs what is needed to create a range of bespoke cures for disease”. But they also cite Botox as an example of a “transformative protein in our society” so who knows what is going on.

What AI was used in creating this article?

Meta AI refused to create an image for the prompt “create an image for the article ‘With AI will Facebook finally improve the world?’” After testing a bunch of variations of the prompt I discovered that it wasn’t “AI” or “Facebook” that it was baulking at, but that I was producing an image for an “article”. Once I removed that word it was completely happy to do what I asked. DM

Subscribe to Develop AI’s newsletter here.

Develop AI is an innovative company that reports on AI, provides training, mentoring and consulting on how to use AI and builds AI tools.