- AI Emergence
- Posts
- Microsoft’s Phi 3.5 is the next big thing!
Microsoft’s Phi 3.5 is the next big thing!
Along with: OpenAI’s GPT-4 Omni just got a custom upgrade- and It’s 'You'
Analytics Vidhya (Curated by Kunal Jain)
August 22, 2024
Hey there,
It has been about 1.5 years since GPT 4 launched. While there have been many more launches and new models since then - it has not been game-changing.
Don’t get me wrong - I am a big believer in the potential of GenAI and its impact. But I was hoping for a much faster release of better models based on how quickly GPT 4 came out after 3.5!
Having said that, we now have awesome smaller models (Phi 3.5 MoE launched this week), equally performing open source models (thanks to Llama 3.1), and also tools to use LLMs more effectively for real-world use cases.
On that note, let’s look at the developments this week!
What would be the format? Every week, we will break the newsletter into the following sections:
The Input - All about recent developments in AI
The Tools - Interesting finds and launches
The Algorithm - Resources for learning
The Output - Our reflection
Table of Contents
The three new Phi 3.5 models include the 3.82 billion parameter Phi-3.5-mini-instruct, the 41.9 billion parameter Phi-3.5-MoE-instruct, and the 4.15 billion parameter Phi-3.5-vision-instruct.
Remarkably, these models deliver near state-of-the-art performance on several third-party benchmark tests, surpassing competitors like Google’s Gemini 1.5 Flash, Meta’s Llama 3.1, and even OpenAI’s GPT-4 Omni in certain instances.
All three Phi-3.5 models are released under the MIT license, highlighting Microsoft's dedication to the open-source community. (source)
OpenAI has officially launched the ability to fine-tune the GPT-4 Omni model.
This allows developers to tailor GPT-4 Omni with custom datasets, achieving better performance and cost efficiency for their unique applications. Fine-tuning empowers the model to adapt its response structure, and tone, or adhere to complex, domain-specific instructions.
OpenAI is pumping up its enterprise offering by providing organizations with 1 million free training tokens per day until September 23. (source)
OpenAI has partnered with Condé Nast to integrate content from top brands like Vogue, The New Yorker, and GQ into their products, including the newly launched SearchGPT prototype.
It seems like an effort to improve the new SearchGPT. Feedback from news partners will help refine these features, which OpenAI plans to incorporate into ChatGPT in the future. (source)
Unitree has upgraded its humanoid robot, G1, with a focus on mass production, tailoring it for versatile applications in both civilian and research environments. Its key features include:
Advanced joint motors that enable precise movement and deliver high torque across various joints.
The ability to manage significant loads, with capacity varying based on arm extension and posture.
Multipurpose functionality, offering diverse use cases. (source)
Many companies are jumping on the Generative AI bandwagon, even without a clear direction. However, Procreate, the popular illustration app for iPads, is taking a stand against this trend. If you were expecting AI features in Procreate, CEO James Cuda’s recent statement on X might come as a disappointment.
Procreate’s stance is clear: “AI is not our future” and “creativity is made, not generated,” as boldly stated on their website. The company further argues that “Generative AI is ripping humanity out of things” and believes the technology is “built on a foundation of theft and is leading us toward a barren future.”(source)
Runway ML has officially introduced Gen-3 Alpha Turbo, which is claimed to be one of the fastest and most cost-effective versions of its AI video generation model. It is said to be seven times faster than the previous Gen-3 Alpha model and costs half as much to operate, making it more accessible to everyone.
The upgrade focuses on reducing time lag, enabling real-time video creation, and improving overall performance which were faced in the previous model. (source)
A few months ago, OpenAI introduced SearchGPT, a prototype designed to provide fast and highly relevant search solutions. Initially, access was limited, and now OpenAI has closed the waitlist, meaning users who didn’t apply for early access can no longer express interest.
OpenAI announced via email that they have filled all testing slots and will not be granting further access. (source)
OpenAI recently deactivated several ChatGPT accounts linked to Iranian state-backed hackers involved in a disinformation campaign aimed at derailing the upcoming US elections.
Reports indicate that these accounts were using AI to generate fake news articles and social media posts to target the US elections.
The group Storm-2035, known for creating and spreading fake news to influence public opinion on sensitive topics, was identified as being behind this activity. (source)
Recently, Google introduced a music recommendation ranking system that uses a transformers model to better understand the sequential nature of user actions by analyzing their current context. This system helps deliver more accurate music recommendations based on user behavior and preferences.
In services like YouTube Music, user actions such as skipping, liking, and disliking songs are observed. The challenge lies in identifying context-specific preferences. For instance, a user who typically skips fast-paced music might prefer it during a workout. The transformers model enables the system to learn from past behavior while adjusting recommendations based on the current context, enhancing user satisfaction by offering more relevant music. (source)
Waymo, a company owned by Alphabet, has revealed a new version of its self-driving car technology, integrated into Geely Zeekr electric vehicles. These next-generation robotaxis are designed to operate in a broader range of weather conditions while using fewer costly cameras and sensors.
The company expects to bring this technology to market faster due to the growing demand for machine learning and semiconductor advancements. Waymo is already well-known for providing around 50,000 driverless rides in cities like San Francisco and Phoenix.
With a $5 billion investment, Waymo plans to scale its fleet into additional cities and introduce further updates. (source)
Lambda partnered with Nous Research to launch Hermes 3, a fine-tuned version of Meta's Llama 3.1 LLM. With 405 billion parameters, it excels in agentic tasks such as structured output, decision-making, and code generation.
What makes Hermes 3 unique is its occasional unexpected reactions, such as falling into an "existential crisis" when given a blank prompt, which surprised researchers. Optimized for efficiency and creativity, Hermes 3 runs on multi-node cluster setups and offers free access through Lambda's API for a limited time.
Designed for diverse applications, including advanced reasoning, role-playing, and storytelling, it serves as a versatile open-source tool. (source)
AMD is aggressively competing with Nvidia in the AI market by acquiring AI infrastructure company ZT Systems for $4.9 billion. ZT Systems builds custom computing infrastructure for major AI companies like Microsoft, Meta, and Amazon and brings 1,000 engineers to AMD.
This acquisition allows AMD to accelerate the development of its Instinct data center chips and streamline the production of both chips and systems.
Despite surpassing $1 billion in data center revenue last quarter, AMD still trails behind Nvidia's $22.6 billion. Additionally, AMD recently acquired European AI startup Silo AI and plans to launch its MI350 chip in 2025 to rival Nvidia's Blackwell GPU. The deal awaits regulatory approval.(source)
Are you looking to create a professional video for your event or business at no cost? Lumen5 is a tool that allows you to effortlessly generate videos simply by providing a prompt.
Problem Statement: Imagine starting your own creative agency. Write a few points outlining the services or products your firm will offer and create an engaging video to attract clients. You can customize the prompts to fit your needs and even edit them later. Additionally, you can add themes and music to enhance your video.
How to access
Sign up for Lumen5.
Provide your website URL to automatically apply your company’s color theme.
Add a prompt specifying the type of video you want and the content for each slide.
Lumen5 will generate the video, allowing you to customize slide backgrounds based on your text.
You can also add background music of your choice.
Explore additional customization options.
Download the final video.
Click here to see what I created after testing the tool.
In the recent episode of Leading with Data, I had a great conversation with Gaurav Agarwal, Founder and CEO of RagaAI, a pioneering AI testing platform. He shared his views on building high-performing teams, the evolution of AI, the future of technology, and his experience in leading AI initiatives across multiple tech giants like NVIDIA and Ola Electric Mobility. He has extensive experience working on AI systems, computer vision, and scaling AI-driven businesses for over 15 years.
Deep down I am hoping that “Project Strawberry” or GPT-5 once again pushes the envelope of what we expect today from AI and I hope all of us are surprised once again by what the model can do.
What about you? How are you feeling about the pace of development? When was the last time you were surprised by a new development in GenAI?
Look forward to hearing from you.
How do you rate this issue of AI Emergence?Would love to hear your thoughts |
Reply