🤝 Beyond the big three: Why chatbot diversity matters
Dear curious minds,
The majority of humanity has never used a chatbot. Those who have, are most likely using one or multiple of the models from OpenAI, Google or Anthropic. However, it is important that this powerful technology is not in the hands of only a few companies. The recent open release of the first Llama 3 models by Meta is a step in the direction we need to keep our world in balance.
In this issue:
💡 Shared Insight
A Balanced Future: Why We Need Open-Source LLMs📰 AI Update
Llama 3: Meta Releases Top-Performing Open LLMs with 8B and 70B
The Rise of AI Wearables: Pitfall and Promise🌟 Media Recommendation
The Next Wave: Exclusive Interview with Perplexity CEO
💡 Shared Insight
A Balanced Future: Why We Need Open-Source LLMs
So far, a few big tech companies, namely OpenAI, Google and Anthropic, control the state-of-the-art LLMs and share at least some mighty models with the world. No one knows if they already use way more powerful models internally.
There were and still are discussions if open-source models can be competitive. On the one hand, there was the leaked "We Have No Moat, And Neither Does OpenAI" memo from a Google researcher in 2023. On the other hand, everything which is shared as open-source can in theory, even if the license agreement states differently, be used by the companies building proprietary models. Furthermore, training of the currently best performing models is super cost intensive. Happily, Meta is investing a lot of effort and money on creating LLMs, for which they share everything needed to run them. The recent release of the first two models from the Llama 3 family, perform on a very competitive level (covered later in this issue). This is the highly desired balance we need to the powerful state-of-the-art closed models from OpenAI, Google and Anthropic. And yes, Meta does not release everything and also has some limitations in the license they use, but that is nothing I want to explore right now.
Open-source LLMs are ensuring fair access to AI technology. By being openly available, these models allow researchers, developers, and businesses to understand, improve, and adapt the technology without the constraints of proprietary systems. This openness accelerates technological advancements and democratizes AI by enabling a wider community to contribute to and benefit from state-of-the-art models.
Without open-source alternatives, a few large corporations will dominate the LLM market. This would give these companies significant control over the development and deployment of AI technologies.
Especially concerning is the potential centralization of Artificial General Intelligence (AGI) technologies in the hands of a few corporations. Their goals for the technology might not align with the best interests of society. These corporations might prioritize profit or other objectives over the well-being of humanity. This could lead to scenarios where AGI is used for unequal distribution of benefits, favoring the corporations and their interests.
By promoting open-source alternatives and international collaboration, we can encourage a more diverse range of perspectives and values. This will help to maintain a global power balance and ensures that AGI is used for the benefit of humanity.
📰 AI Update
Llama 3: Meta Releases Top-Performing Open LLMs with 8B and 70B
Meta released with the 8B and 70B parameter versions, the first two models of the Llama 3 family.
The shared benchmark results show a stunning performance for both variants. The smaller 8B version, which can be run locally on recent PCs and even smartphones, outperforms all competitors with a similar size. The larger 70B version gets already quite close to the currently best models from the major companies in the LLM space. Both models are already listed in the Chatbot Arena Leaderboard and are ranked really high.
Key innovations include enhanced coding abilities and multimodal training on images. In the coming month Meta will release multiple models with new capabilities, including multimodality which especially will enable the usage of image inputs.
Meta also stated that a model with more than 400B parameters is still being trained, but the current checkpoint already shows a competitive performance.
The biggest downside is the context size of only 8k tokens, which is rather small compared to the state-of-the-art models which allow all more than 100k tokens to be processed. However, already Llama 2 showed that this can be changed by fine-tuned variations of the released models. Furthermore, a recent paper by Meta named “Megalodon” might be the key to an unlimited context size.
My take: The Llama 3 release is the result of a major investment by Meta into generative AI compute infrastructure. To be honest, I did not expect that the company formerly known as Facebook is in the driver seat of releasing the most powerful open LLMs. Really appreciate this and looking forward to the upcoming additions to the Llama 3 family.
The Rise of AI Wearables: Pitfall and Promise
There is another interesting topic besides the battle for the best performing LLMs and the race to create AI agents which help you in doing work and achieving goals. A new market is arising, and it can best be described as new mobile devices which are made possible by the advancements in AI.
Past issues covered already the Rabbit r1, Rewind Pendant and Ray-Ban Meta smart glasses. Recent media coverage was mainly about the bad reviews from the early adopters evaluating their pre-ordered Humane AI pins. On competitor is trying to use the current discussions to get some traction on their rather different approach with their device named Pendant.
Humane AI pin
The AI pin from Humane is a compact, wearable form factor designed to be clipped onto clothing. The device features a monochrome green laser projector for displaying output on a user’s hand. It has a camera for scanning QR codes and potentially other visual input. The device can be controlled by gesture and voice.
There was quite some hype when the announcement video was shared five months ago. Besides the new possibilities, especially the high price tag of $699 or $799 (depending on the color) and a required subscription of $24/month was discussed.
Meanwhile, the first pre-orders were shipped and there are many reviews published which state that the device is not usable in practice. The video from the YouTuber Marques Brownlee with the title “The Worst Product I've Ever Reviewed... For Now” was going viral.
Limitless Pendant (formerly Rewind Pendant)
The company Rewind initially built a Mac application which recorded your screen and audio to create a personal AI assistant.
The company has rebranded to Limitless. They now focus on audio recordings which are stored in an encrypted cloud storage. Limitless offers an app (free with limited features)
The Pendant clip was already announced last October and could be pre-ordered for $59. Now the hardware device can be ordered for $99, and it is stated that the first batch will be shipped in August 2024.
In contrast to other AI wearables, its only feature is the ability to record conversations. The device has a "consent mode" which only records voices from people that agreed to be recorded, but it's unclear if this is forced to be used.
Using AI features like transcription, summaries or extracting tasks requires an active subscription for $19/month if more than 10 h/month are processed.
My take: Hard to judge which AI wearables will be successful. In contrast to the smartwatch, the Google Glass was, despite its large potential, not accepted by society. I see the possibilities for automated note-taking with the 24/7 audio recording devices like offered by the Limitless Pendant. But at the same time, I doubt that it will be accepted by people you interact with that every word they speak will be recorded by a device you wear.
🌟 Media Recommendation
The Next Wave: Exclusive Interview with Perplexity CEO
The first episode of the podcast “The Next Wave” featuring Aravind Srinivas, CEO of Perplexity, was published.
Both hosts are well known in the AI news space and if you are not already following them, you should at least check their publications:
The weekly newsletter “Lore” by Nathan Lands
The YouTube Channel by Matt Wolfe
The episode talks about how Perplexity became a competitor to Google search.
Furthermore, Aravind elaborates on the open-source vs closed-source AI question. He stated that control of AI by a single corporation can lead to biases and highlighted the importance of open-source alternatives. Transparency and accessibility in technology are essential for safety and preventing misuse, involving a larger community in decision-making.
Disclaimer: This newsletter is written with the aid of AI. I use AI as an assistant to generate and optimize the text. However, the amount of AI used varies depending on the topic and the content. I always curate and edit the text myself to ensure quality and accuracy. The opinions and views expressed in this newsletter are my own and do not necessarily reflect those of the sources or the AI models.