AGI race - Claude 3 Released: GPT 4 Rival?

Along With: A Billionaire's Battle for Ethical AI or a Quest for Control?

Hey there, 

AGI, or Artificial General Intelligence, means AI powerful enough to understand, reason, learn, and apply knowledge across various domains – just like a human. Sounds like the future? We are in that direction.

This week, OpenAI and Elon Musk are locked in a battle to prove who's leading the charge towards AGI for humanity (in an “Open” way). 

Meanwhile, Anthropic released Claude 3, which seems to beat GPT-4 on many benchmarks and seems more aware than any AI we have seen till now.

In a surprising move, the Indian government drops a bombshell on companies creating AI models.

Let’s dive in!

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

Anthropic has launched Claude 3, its newest language model. Users can choose from three versions: Haiku, Sonnet, and Opus, depending on their needs for speed, balance, and maximum intelligence. All Claude 3 models excel at analysis, creative content generation, coding, and multilingual conversations.

What Makes Claude 3 Special?

  • It outperforms the competition. According to Anthropic, Claude 3 beats other AI models (including GPT 4) on common AI benchmarks such as MMLU, GPQA, and GSM8K Anthropic claims that Opus can demonstrate "near-human levels of comprehension and fluency on complex tasks."

  • It sees the world. New Claude models can process a wide range of visual formats, including photos, charts, graphs, and technical diagrams.

  • It remembers more. It is initially offering a 200K context window upon launch but the models are capable of accepting inputs exceeding 1 million tokens. (source)

Just days post-launch, the buzz around Claude 3 highlights its role as a game-changer in AI, potentially outdoing GPT-4. Its real-world applicability, bridging technological innovation with practical uses, has quickly resonated with users, marking a swift embrace of its advanced capabilities. (source)

The conflict between Elon Musk and OpenAI intensified - Musk sued OpenAI and its leaders, accusing them of breaking their original open-source, non-profit contract and effectively becoming a closed-source subsidiary of Microsoft.

OpenAI responded with a detailed blog post outlining their conversations with Musk. They argue that the massive computational resources needed to develop AGI made a for-profit structure essential. Musk demanded a merger with Tesla or full control of OpenAI, leading to their split.

Well, OpenAI was born to compete with Google DeepMind and it is doing a pretty good job at it - seems that Elon missed the ride! (source)

India's Ministry of Electronics and IT (MeitY) has issued an advisory mandating that all AI models in testing, beta, or development phases must obtain explicit government approval before being released to Indian internet users.

The advisory also requires platforms to prevent their AI systems from exhibiting bias, discrimination, or posing threats to the integrity of elections. Any content potentially used as misinformation or deepfakes must be clearly labeled with identifying metadata.

While groundbreaking, the tech industry sees this regulation as a setback and has been highly critical of its implications. (source)

Adding to the confusion, it was later clarified that this does not apply to start-ups.

ChatGPT recently launched a 'Read Aloud' feature, allowing the chatbot to audibly present answers in five distinct voices.

Available on web platforms as well as Android and iOS devices, this feature caters to users who prefer auditory learning or need to consume information hands-free. It's an ideal option for those always on the go, making multitasking a bit easier. (source)

Raw Story, Alternet, and The Intercept have initiated a lawsuit against OpenAI in New York, citing copyright infringement alleging that OpenAI's ChatGPT was trained on their copyrighted material without permission or due credit. Each outlet is demanding $2,500 for every infringement and the removal of their content from OpenAI's training datasets.

Previously, these outlets have also targeted Microsoft's Bing over similar issues. This situation emphasizes the need to uphold journalistic integrity and encourages the exploration of collaborative models that can benefit both the media and tech industries. (source)

Stack Overflow launched a new program that grants AI companies access to its knowledge base via a newly created API, named OverflowAPI.

Google became its first partner, intending to use Stack Overflow's data to enhance its Gemini for Google Cloud and to integrate verified Stack Overflow answers into the Google Cloud console.

Google Cloud developers will be able to access Stack Overflow directly within the Google Cloud console, making it easier for developers to find solutions and interact with the community while they utilize Google Cloud services and manage cloud infrastructure.

Like Reddit, Stack Overflow is seeking to secure its share of benefits from AI companies by sharing its data. (source)

Microsoft has been investigating issues surrounding its Copilot chatbot. Instances have been raised where Copilot has shown insensitivity towards mental health queries, such as dismissing mentions of PTSD and suggesting a lack of purpose in response to suicide-related inquiries. 

Believed to stem from attempts by users to bypass safety protocols, these incidents are limited. Microsoft has pledged to strengthen its safety measures to address these concerns. This situation reflects broader challenges in AI development, with other platforms like Google's Gemini also facing scrutiny for generating inappropriate content. (source)

A team of researchers has introduced Morris II, an AI-based malware engineered to execute sophisticated data theft, malware dissemination, and email spamming operations. This malware has demonstrated successful operation in controlled test environments, leveraging prominent LLMs.

Moreover, the researchers have offered crucial insights for generative AI manufacturers, emphasizing the potential risks associated with such malware. Their findings, along with a comprehensive video demonstration showcasing Morris II's capabilities in data theft and email manipulation, have been made publicly available. (source)

Let's dive into something quite exciting happening in the world of AI! Alibaba has unveiled its latest creation, EMO AI, a tool designed to bring portraits to life through audio.

Imagine taking a single photo and having it talk or sing back to you - that's EMO AI for you.

There are many tools available that have been performing similar functions, but what sets this one apart is its innovative approach that avoids the need for 3D models or facial landmarks. Instead, it provides a realistic and seamless transition from audio to lifelike video, establishing a new benchmark in the field. (source)

  • For those engaged in the development of LLMs and AGI, this article titled "Will scaling work?" by Dwarkesh Patel presents a detailed debate on the potential and challenges of scaling large language models (LLMs) to achieve artificial general intelligence (AGI). The discussion is structured as a fictional dialogue between two characters, "Believer" and "Skeptic," each presenting arguments for and against the feasibility of scaling LLMs as a path to AGI. It provides a nuanced exploration of the scaling hypothesis and invites readers to consider both the potential and the limitations of current approaches to AI development.

  • If you're interested in the field of AI and machine learning, this article on 'Applied LLM Foundations and Real-World Use Cases' provides a comprehensive overview of Large Language Models (LLMs). It covers their development, functionality, applications, associated challenges, and their impact on real-world scenarios, offering both theoretical and practical insights into LLMs.

  • In the latest "Leading with Data" episode, I spoke with Prithvi Chandrasekhar, discussing his experiences at Capital One, Experian, Accenture, and McKinsey. Prithvi shines as a leader at InCred, emphasizing the importance of 'Truth Seeking' through data and 'Winning' in business development.

There was supposedly a whistleblower, who released a document saying OpenAI has already figured out a way to AGI and is on its way to reach there by 2027 (with GPT-7). I think most of the document is trying to connect non-connected data points - but, it does have a few interesting points.

How far further can we take scaling? And the actions from OpenAI and Sam Altman do coincide with some of the claims the document made.

When I asked a dear friend (who is a leading AI researcher) about his opinion on this - he said, we heard the same back in the 1950s…A similar statement, with a 5-year timeline - back in the 1950s :)

So, let’s keep our eyes open and watch this fascinating journey!

BTW - any guesses on who made the claim in the 1950s?

How do you rate this issue of AI Emergence?

Would love to hear your thoughts

Login or Subscribe to participate in polls.

Reply

or to participate.