CONVO OVER COFFEE ☕

• Meta has taken a major stride forward with the launch of their Open Language Model, Llama v2. Unlike its predecessor, Llama v1, the license for v2 permits the use and distribution of its weights as part of a commercial product, though some restrictions apply. The v2 release features both the base model and a fine-tuned version designed for chat. Presently, 7b, 13b, and a whopping 70b parameter models are available. Notably, in tests, the 13b model proved to be on par with GPT-3.5-turbo in most areas. The true strength of open models like Llama lies in the community that produces diverse derivatives of the base model, tailor-made for different functions. This is a power move by Meta, and step further into commoditization of LLMs.

• Naturally, all major partners new about above step, and Meta and Qualcomm straight after announced that they are teaming up to run Meta's Llama 2 on Qualcomm chips in phones and PCs from 2024 onwards. This shift suggests a move from cloud to device-based AI processing, which could enhance voice assistants and other apps, while potentially reducing running costs. After PCs and Laptops, mobile was perceived as a challenge due to the computing power required for models to operate - now we are seeing models way friendlier towards low-power processing architectures.

• The big surprise came from Microsoft and Meta when both behemoths have announced support for Llama 2 on Azure and Windows. This partnership aims to democratize AI (this sounds like copy paste from every announcement of this type…), enabling developers and organizations to create AI-powered tools and experiences with greater ease and safety. This move underscores a significant stride in AI innovation, bringing large language models to a broader range of platforms. But what it actually means is trouble for OpenAI.

• Just yesterday, OpenAI has introduced Custom Instructions, allowing Plus users to give ChatGPT custom requests or context to apply to every conversation. Early testers have identified a multitude of use cases, such as expertise calibration, language learning, localization, novel writing, response formatting, and writing style personalization.

• However, it appears that the ChatGPT fever might be cooling down, with traffic to OpenAI's ChatGPT website falling by 9.7% in June, according to web analytics firm Similarweb. Despite the drop, ChatGPT still leads as the most visited AI-powered chatbot site, surpassing Microsoft's Bing and Google's Bard.

• Rumors that OpenAI's GPT-4 is experiencing a reduction in performance over time have been given some scientific backing. A recent Stanford study suggests that GPT-4's proficiency in coding and math has declined significantly over the past few months. This underscores the fragility of large language models and the potential challenges in relying on them.

• A relief to OpenAI can still come from Microsoft side, as they announced exciting developments: the expansion of Bing (powered by OpenAI models) to new audiences with Bing Chat Enterprise—an AI-powered chat for work; together with the rollout of Visual Search in Chat on Bing Chat. Microsoft's AI tools are designed with business-grade data protection, ensuring sensitive data remains within the organization and is not used to train models. This move paves the way for a future where employees can harness AI to enhance creativity and productivity while maintaining data privacy. Or at least this is what Microsoft says.

• On a note of smaller companies, Reworkd AI, a YC S23 company, is revolutionizing workflow automation with AI Agents. After identifying limitations in AI Agents handling business-critical processes, the team has developed a solution that pairs AI Agents with a structured workflow system. The result is a more reliable and efficient AI-driven workflow automation. These are the guys behind AgentGPT, so quite interesting what they can achieve with backing from Y Combinator.

• Elon Musk has launched a new AI company, xAI, aiming "to understand the true nature of the universe". The company will function alongside Musk's X Corp. brand, which includes Twitter and Tesla. Despite being an initial supporter of OpenAI, Musk has criticized it for "training the AI to lie", making his new venture a potential competitor. The staff of xAI include veterans from Tesla, DeepMind, OpenAI, Google Research, and Microsoft Research.

• Lastly, Apple takes a leap forward in the AI race, unveiling plans to create its own AI chatbot, the Apple GPT, powered by an innovative AI framework, Ajax. Boasting greater efficiency and scalability, Ajax, based on the groundbreaking "sparse transformers" architecture, might disrupt the existing AI landscape. Currently being tested internally, Ajax could soon empower features like chatbots, text generation, and enhanced services across Apple's product range, including a revamped Siri and Messages app. With an enormous consumer base of over 2 billion devices, Apple could significantly influence the LLM landscape, potentially revolutionizing AI-powered applications for the Apple ecosystem. Funny to hear, that with all of this promise, they struggle to envision how to use it…