- AI Emergence
- Posts
- Greg Brockman (OpenAI Co-Founder) Returns
Greg Brockman (OpenAI Co-Founder) Returns
Along with: Google’s Jarvis AI assistant leaked???
Analytics Vidhya (Curated by Kunal Jain)
November 14, 2024
Hey there,
Drama at OpenAI doesn’t stop!
On the team front, Greg Brockman returns after a 3-month hiatus. On the other hand, Mira Murati (ex-CTO) poaches from OpenAI!
More importantly - there are rumors that OpenAI’s next model, “Orion,” only shows marginal improvements over GPT-4.
If that happens to be true, we would all get a breather to absorb what has happened in the last 12 months, but the valuations and investments in AI would come crashing down!
On a different front, all the big tech players are racing to launch agents that could control computers and browsers to enhance productivity.
Anthropic has already launched one, Google Jarvis was leaked, and OpenAI is expected to launch one in January. It looks like another frontier where a battle will be waged in 2025.
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
Greg Brockman, co-founder of OpenAI, has returned to the company as president after a three-month break, during this time, OpenAI faced challenges, including executive departures and controversies surrounding its shift to a for-profit model.
His return follows OpenAI’s recent funding round which valued the company at $157 billion. OpenAI has seen several key figures, such as CTO Mira Murati and others, leave amid organizational shifts. Brockman’s longstanding alliance with CEO Sam Altman remains central to OpenAI's leadership dynamics. (source)
According to The Information, OpenAI’s latest model, “Orion,” offers only modest improvements over GPT-4. It shows limited progress in areas like coding while incurring higher operational costs. Surprisingly, Orion doesn’t consistently outperform GPT-4 in programming, with notable gains primarily in language processing.
Why? OpenAI is running out of new data sources. Having tapped into all available natural data, they’re now left with synthetic data as the only viable option.
In response, OpenAI has launched a new "Foundations Team" focused on enhancing data quality and accelerating model development.
This challenge isn't unique to OpenAI. Several other AI labs are confronting similar limitations, slowing progress in the field of large language models. (source)
Ilya also highlighted that the era of scaling has plateaued, emphasizing that future breakthroughs depend on "scaling the right thing" rather than just more data. (source)
Google's experimental AI assistant, Jarvis, was briefly leaked on the Chrome Web Store, showcasing its potential to assist with web browsing, task automation, and more advanced actions like booking flights or buying groceries.
Though the leaked version was incomplete, it hinted at Jarvis's capabilities to streamline tasks by interacting with web pages and automating actions like typing and clicking.
Jarvis is set to officially launch in December. It is expected to compete with other AI assistants like Anthropic’s Claude AI Agent, marking a significant step in AI-driven productivity tools. (source)
OpenAI is expected to launch a similar Agent in January next year (source)
KLING, a Chinese AI video generator, has launched a new "Custom Models" feature that allows users to train their own video characters using 10 to 30 video clips, enhancing character consistency across different scenes.
The system utilizes video footage for training, enabling the generation of consistent character representation, even with complex camera angles and time-shifted facial shots.
KLING's international version, released in July 2023, offers 1080p HD resolution and a variety of video formats, while its "Motion Brush" tool helps users control the movement of elements in their videos. (source)
Amazon is set to reveal its new AI chip, Trainium 2, next month as part of its strategy to reduce reliance on NVIDIA and improve AWS efficiency. Developed by Annapurna Labs, Trainium 2 aims to enhance AI model training within Amazon’s cloud infrastructure, helping reduce operational costs and offering AWS clients more affordable services.
Major companies like Anthropic and Databricks have already trialed the chip. This push aligns with Amazon’s increasing tech investments, expected to reach USD 75 billion in 2024, as the company seeks greater control over AI infrastructure. (source)
At the Conference for Robot Learning (CoRL) in Munich, NVIDIA unveiled several new tools to accelerate AI and robotics development.
The company introduced Isaac Lab, an open-source framework for robot learning, and Project GR00T, a set of workflows designed to speed up humanoid robot development.
New tools like the Cosmos tokenizer and NeMo Curator enhance video and image data processing, enabling faster and more efficient creation of world models.
These advancements, alongside the introduction of generative AI-powered environments and robot motion generation, are aimed at accelerating progress in humanoid robotics and general-purpose robots.(source)
XPeng's 2024 AI Day showcased significant advancements in robotics, AI chips, and transportation.
A highlight of the event was Iron, a humanoid robot powered by XPeng's Turing AI chip.
Iron is actively working on the company's production line, and the robot is intended to play a future role in retail and office environments.
Additionally, XPeng introduced their Kunpeng Super Electric System, which aims to improve EV efficiency and enable ultra-fast charging, along with exciting plans for urban flying cars.
The event signified XPeng's ambitions to lead in AI-driven technology, particularly in robotics, electric vehicles (EVs), and autonomous urban mobility.(source)
Magentic-One: AI Agents that go beyond conversations to autonomous actions
AI is advancing from simple conversations to performing complex tasks autonomously, exemplified by systems like Magentic-One.
This new agentic system uses a lead Orchestrator to coordinate specialized agents in handling tasks such as coding, web navigation, and file management.
Magentic-One’s modular structure supports flexibility and efficient problem-solving across various scenarios, from scientific research to software engineering. While its capabilities mark a leap forward, they also bring new risks, like unintended actions or security vulnerabilities.(source)
At Analytics Vidhya, we are launching a program on Agentic AI to help professionals and enterprises understand, implement, and use this cutting-edge technology in their products.
At Baidu's 2024 conference in Shanghai, the company launched new AI products, including smart glasses, a text-to-image generator, and a coding tool to enhance accessibility.
The smart eyewear, competing with Meta’s Ray-Ban glasses, uses Baidu's Ernie model for interactive features that handle a wide range of AI-powered applications, from text generation and question answering to more complex tasks.
Users can interact with the glasses using voice commands, allowing them to take pictures, record videos, and perform other tasks in real time.(source)
Google DeepMind has released AlphaFold 3’s source code and model weights, advancing AI’s ability to predict protein interactions with DNA, RNA, and small molecules. This breakthrough can accelerate drug discovery and disease research. While the open-source release balances science and commercial needs, AlphaFold 3 still faces some accuracy challenges.(source)
Palantir Technologies' stock soared following its Q3 2024 earnings report, as the company posted better-than-expected revenue and earnings, boosted by a 30% revenue growth and a 43% increase in earnings per share.
Palantir's high valuation, trading at 46 times sales, raises questions about its ability to sustain this growth. Despite these concerns, its leading position in AI software platforms and a solid forecast for AI market growth suggest that Palantir may continue its upward trend into 2025. (source)
Alibaba's Qwen AI unit has launched a new series of models, Qwen-2.5-Coder, aimed at helping developers write, analyze, and understand code.
The models, ranging from 0.5 to 32 billion parameters, were tested in applications like AI-powered code editor Cursor and a chatbot similar to ChatGPT.
The largest model, Qwen-2.5-Coder-32B-Instruct, outperforms other open-source systems in code generation and logical reasoning. Trained on over 20 trillion tokens, the models support 40+ programming languages and are available on GitHub for developers to explore.(source)
Despite excelling in text generation and basic problem-solving, AI struggles with complex mathematical reasoning.
Epoch AI's new benchmark, FrontierMath, reveals that advanced models like GPT-4o and Gemini 1.5 Pro can only solve around 2% of the high-level math problems posed.
These problems, developed in collaboration with expert mathematicians, require deep creativity and reasoning, unlike traditional benchmarks where AI models score over 90%.
With this unique test, FrontierMath showcases the significant gap between AI capabilities and human expertise in advanced mathematics, underscoring the long journey ahead for AI in achieving genuine understanding.(source)
Interactive blogs by Analytics Vidhya
Introducing Analytics Vidhya’s Interactive Blogs, a game-changer for engaging, bite-sized learning. With flashcards, quizzes, and real-time interaction, these blogs transform traditional reading into an active, knowledge-enhancing experience. Perfect for learners aiming to solidify their understanding while having fun.
How to Access:
Log in to Analytics Vidhya.
Go to the Blogs section.
Select “Go Interactive.”
Dive into each blog, answer questions, and test your knowledge.
OpenAI CEO Sam Altman recently in a conversation expressed in confidence that artificial general intelligence (AGI) is closer than expected, predicting a breakthrough as early as next year. Altman believes the path to AGI is "basically clear" and achievable through diligent engineering rather than new scientific discoveries.
In the recent episode of Leading with Data, I had an engaging conversation with Eleni Verteouri, Director and AI Tech Lead at UBS, about her experience in AI model development and innovation in finance. We discussed her journey, insights on responsible AI practices, strategies for handling data challenges, and her initiatives to increase women’s representation in tech.
This podcast features a conversation between Lex Fridman and Dario Amodei, CEO of Anthropic, discussing a range of AI-related topics including the technical and ethical dimensions of AI scaling, the competition among leading tech companies, safety frameworks, and the development and future implications of advanced AI models like Claude.
How does it feel to see so much action in a week? It sometimes feels a bit overwhelming, but I feel more excited than ever. I feel like jumping out of bed every morning to get my hands on the new tools and developments happening.
If there is such a thing as living a dream life, I think I am already in one. What about you?
How do you rate this issue of AI Emergence?Would love to hear your thoughts |
Reply