AI Models compete, 2024 Elections, and more

Hi everyone, welcome back! After a brief hiatus we are back delivering some of the most interesting news to you in AI and as it applies to politics, international relations and social entities.

Thanks to your feedback, were editing our format to include more deep dives into a smaller set of topics. Hope you enjoy this week’s edition!

Apple and OpenAI LLM Models Compete: ReALM-3, Apple’s LLM, outperformed GPT-4 and why that might be a bit misleading. First, let's learn a bit about the differences between a ReALM (Retrieval-enhanced Language Model) vs an LLM (Large Language Model). ReALM: Unlike traditional models that primarily rely on the vast amounts of data they were trained on, ReALM incorporates a retrieval component that allows it to access an external knowledge base in real-time to answer queries. This mechanism is designed to help the model remain current with the latest information, a notable limitation in models that solely depend on their training data.

ReALM therefore is really helpful in identifying ambiguity like words “they”, “it”, “that one” with a context, such as prior statements or a visual on a user’s phone, to send to an LLM. This is especially important with conversational assistants like Siri and Alexa who can understand the way a human naturally talks. Think of you saying “Can you direct me to the one on 5th Ave” while on your phone looking at a list of restaurants, a ReALM can see what you’re looking at on-screen to inform the LLM on the most appropriate response.??

ReALM does so by converting visual components such as an image or what's on a screen into text-format, which a powerful language understanding model can reason. This makes it much more practical for devices like smart-phones. Pretty soon we’ll be able to ask our phones “buy the ice cream maker I was researching yesterday” or “book the flight I just looked at” without having to provide a great amount of detail. Read the original Arxiv paper: here ?

AI Impact on 2024 Elections: Every week there seems to be new updates on ways AI is being used in the election.?

So lets quickly break down some of the ways AI is and might be used during the 2024 elections.

• Deep fakes: Examples of false images, audio and video are likely to be seen during this election cycle. One of the biggest questions is how, or even if, social media platforms will attempt to stop and screen this content. We’ve already seen examples of a Joe Biden robo-call in New Hampshire. Even with citizens becoming aware of these possibilities, there isn’t any evidence?
• AI systems influencing constituent feedback: We’ve already seen a study that points to this possibility and expect it to become more prevalent. Government officials and hopefuls will need to find ways to comb through correspondence to search for genuine responses.
• Curated Messaging: With the wealth of consumer data available, it seems like custom and highly curated messages to each individual person landing in our inboxes and social media ads are unavoidable.?
• Misinformation and Stealing PII: With AI becoming increasingly prevalent across the digital ecosystem, many organizations are turning to chatbots or similar AI systems to first engage and interact with a user. There is a chance that unless the implementors of such tools are aware of what they’re doing, some of the interactions between users could be send to malicious actors and some responses could be corrupted.?

Big tech’s race for training data: by now it has become clear that one of the largest factors in model improvement is the quality of the data it is trained on. By quality we mean the quantity of data (which is chief among them), the diversity of the data and actual quality for a given set of tasks.

Having access to a large and differentiated data set has become the hot resource in the quest to create increasingly superior LLMs. We’ve seen OpenAI, Google and Meta ignore corporate policies, side skirting copyright laws, and altering their rules as they seek more data. OpenAI created Whisper, a speech recognition tool, which it used to scrape over a million hours of YouTube videos. Here’s a viral video of OpenAI’s CTO Mira Murati’s response when asked if they trained on videos from YouTube, Facebook and Instagram.

Google also used data from YouTube which is owned by the creators in addition to reviews from Google Maps and public Google Docs.?

Meta also considered purchasing a publishing house to obtain long form data, in addition to choosing to scrape data from various sites knowing copyright laws would be broken since negotiations would take too long.?

One notable omission from this list was Amazon who has been busy doubling down on AI with its additional $2.75 Billion investment into Anthropic. In this symbiotic relationship in the quest for AI dominance has Anthropic using AWS service centers and chips, in addition to Alphabets. It’s expected that as this relationship grows, Amazon’s commercial product and review data will be a powerful repository for Anthropic to potentially pull from. As the quest for dominance in the AI industry accelerates, the ethical and legal ramifications of data acquisition continue to loom large. These tech giants, in their pursuit of data to train more capable models, must navigate the fine line between innovation and infringement. It remains to be seen how the industry will reconcile the push for technological advancement with the imperative to uphold ethical standards and respect intellectual property rights. That’s it for this week, see you on the next one!