- AI Emergence
- Posts
- Less than 2 weeks to OpenAI’s Strawberry?
Less than 2 weeks to OpenAI’s Strawberry?
Along with: First multimodal model from Mistral
Analytics Vidhya (Curated by Kunal Jain)
September 12, 2024
Hey there,
I will be honest - I am having butterflies in my stomach! It is being rumored that Strawberry will be out in less than 2 weeks. As you might already know - I have been waiting for what OpenAI reveals as its next Frontier model!
Assuming this is true, this would be the second time OpenAI has launched an incremental model with better UI before launching a big frontier model. They did this with chatGPT 3.5 and launched GPT 4 within 6 months. This time they launched GPT-4o and it looks like “GPT 5” is on the way!
Let’s go through 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
Project “Strawberry” has been a mystery in the AI world for a few months now - supposed to be the next frontier that takes AI closer to “human intelligence”. Well, it’s not too far away.
As per this report by The Information, Open AI is planning to launch "strawberry" with advanced, in the next two weeks as part of its ChatGPT service.
Project Strawberry is not your usual chatbot. It’s said that it will be much better at "think" before responding. Therefore, it will require more computation and time for response. We may see a ChatGPT where there may be an option to toggle between normal reasoning and the new “High Reasoning” mode.
Although Strawberry will initially focus only on text-based input and output, without multimodal features like image processing, it is expected to enhance how users interact with AI by improving the quality of responses. (source)
Mistral has launched its first multimodal model, Pixtral 12B, which processes both images and text.
With 12 billion parameters, Pixtral 12B can answer questions about images using URLs or base64-encoded data.
It is available for download under an Apache 2.0 license via GitHub and Hugging Face, with potential applications like image captioning and object counting.
However, it's unclear what data was used for training. Pixtral will soon be accessible on Mistral's platforms for further testing.(source)
Apple has finally announced its new "Apple Intelligence" alongside the iPhone 16 series at its Glow Time event on Monday.
So, what does it offer?
Siri Enhancements: Siri will now better understand context, handle speech errors more effectively, and answer a wider range of questions.
New Writing Tools: These tools will proofread your text, suggest tones, and summarize content, including in mails.
Image Search: You can find old images by simply describing them.
Audio Transcription: It can transcribe audio into notes in just a few minutes.
Apple plans to roll it out to the market by December, with the beta version expected in October.(source)
Replit has launched "Replit Agent," an AI tool that enables users to create apps from scratch in minutes, with early access available only to Replit subscribers.
Recently, Masad, founder and CEO of Replit, said, "AI is excellent at writing code, but creating software involves much more, including setting up environments, configuring databases, and deploying apps
In a demonstration video, Masad showed the agent responding to a natural language prompt by planning and coding automatically, while narrating its steps, which users can edit.
After seeking feedback, the agent adds API passkeys and deploys the app. Notable industry figures, like OpenAI co-founder Andrei Karpathy, praised this innovation, calling it a move towards AGI. (source)
HyperWrite's new AI model, Reflection 70B, which was launched as the top open-source AI model, quickly faced scrutiny after third-party evaluators failed to reproduce its impressive benchmark results.
Initially hailed for its "Reflection Tuning" technique, the model's performance discrepancies were attributed to an upload error on Hugging Face by HyperWrite's CEO, Matt Shumer.
Despite some defending the model, others have raised concerns about the transparency of its development, including suspicions that it might be built on proprietary models like Anthropic’s Claude 3. Shumer has since apologized for prematurely announcing the model and promised more transparency as investigations continue.(source)
Safe Superintelligence (SSI), co-founded by Ilya Sutskever, has secured $1 billion in funding to develop AI systems that surpass human capabilities while ensuring safety. SSI, based in Palo Alto and Tel Aviv, aims to create a small, trusted team of researchers and engineers. (source)
AlphaProteo, a new AI system, successfully designs novel protein binders that can target specific molecules, advancing drug development, disease research, and more.
Unlike existing tools like AlphaFold, which predict protein structures, AlphaProteo can create new proteins tailored to bind with specific targets such as those linked to cancer and viral infections.
The AI system has shown higher binding success rates and stronger binding affinities than traditional methods, with some binders proving effective in preventing SARS-CoV-2 infection. While AlphaProteo's potential is vast, it faces challenges with difficult targets and continues to evolve with improvements in bioengineering and responsible development.(source)
If you're looking for real-time code suggestions, want to discuss complex concepts, or need help writing entire code sections, Blackbox AI Code Chat has you covered. It integrates seamlessly with your coding environment, provides access to a vast library of community-shared code, and even interprets code from images. Essentially, Blackbox AI offers various AI-powered tools for developers, focusing on code generation, conversation, and search.
Problem Statement:
Take any website link from Google, paste it into the chat box, and identify the bugs or errors in the website. The tool will generate code solutions to enhance the website's functionality.
How to Access:
Open Blackbox.ai
Enter the website link and your prompt in the chatbox
It will generate solutions and provide the relevant code
This is an example from my trial with the tool.
To learn more, check out this blog by Analytics Vidhya.
In this video, Andrej Karpathy, the former Tesla Autopilot leader and a founding team member of OpenAI, has shared his views on the evolution of self-driving cars and how he thinks that AI capabilities could be further integrated with human cognition.
If you are a beginner in Generative AI and Large Language Models (LLMs), these free courses from Analytics Vidhya can be a great starting point to build a solid foundation in these exciting fields.
Generative AI- A Way of Life- to learn AI-powered text and image generation, use top AI tools, and explore industry applications.
Getting Started with Large Language Models- to delve into the intricacies of natural language processing and model training.
Building LLM Applications Using Prompt Engineering- for mastering prompt engineering, implementing best practices, and developing chatbots using enterprise data using advanced techniques like few-shot or one-shot prompting.
For those who are interested in exploring graph-based approaches for organizing and querying language models, this course by LangGraph Academy focuses on how to effectively use LangGraph. It covers key techniques and practical applications to integrate graphs with language models for more efficient knowledge representation and retrieval.
I am looking forward to reading Nexus, the latest book from Yuval Noah Harari. If you have already read it or read it - would love to hear your reviews on it.
This latest tutorial from Sebastian Raschka is a brilliant workshop to build LLMs from scratch. If you want to hear why he built them from scratch, listen to the podcast I did with him a couple of months back.
I am feeling more excited than I have been for any Apple or Google event - it feels so lucky to be living through this era of AI Emergence!
How do you feel about the fact that we could see a model which would define the new AI frontier in less than 2 weeks? Would love to hear from you.
Keep learning, keep emerging!
Regards,
Kunal
How do you rate this issue of AI Emergence?Would love to hear your thoughts |
Reply