AI Today and Everyday — April 10, 2024

What’s new in the world of AI

Elon Musk predicts AI surpassing human intelligence next year.

Imaged by Midjourney and Magna Ding

Elon Musk, in a recent interview with Nicolai Tangen, the CEO of Norway’s wealth fund, has provided a forward-looking perspective on the development of AI. Musk predicts a near-future where AI could outsmart the most intelligent human and discusses the broader implications and challenges of rapid AI advancements, including regulation and potential limitations.

The Details:

• AI’s Intellectual Leap: Musk anticipates that AI will surpass the intelligence of any human by the end of 2025.
• Sentient Compute Milestone: He projects the total sentient compute capacity of AI to exceed that of all humans within five years.
• Electricity as a Limiting Factor: Musk identifies electricity availability as a significant constraint to AI’s advancement.

Why It Matters:

• xAI’s Grok Model: Musk’s AI startup, xAI, is set to complete training a new AI model, Grok, which is expected to outperform GPT-4.
• The Power Surge: Following Grok, Musk suggests the development of models exponentially more potent than current iterations.
• Regulatory Challenges: Musk emphasizes the necessity of AI regulation but expresses concern over the rapid pace of development potentially outstripping regulatory efforts. He also touches on the dangers of programming AI with political correctness biases.

Meta confirms Llama 3 coming within the month

Image source: Midjourney

The Rundown: At an event in London, Meta confirmed that it plans an initial release of Llama 3, the next generation of its large language model, within the next month.

The details:

• Meta will release several different versions of Llama 3 with different capabilities over the year — with the initial release within the next month.
• The company did not disclose the size of the parameters used in Llama 3, but it’s expected to have about 140 billion parameters and rival OpenAI’s GPT-4.
• Meta previously released Llama 2 in three sizes, with the largest at 70B parameters.
• Meta has stockpiled 350,000 coveted H100 GPUs over the last year to ramp up AI infrastructure, a supply that dwarfs other competitors.

Why it matters: Meta’s confirmation of Llama 3 shows its commitment to catching up to OpenAI. But the open model simply ‘rivaling’ GPT-4 is suddenly not as impressive of a feat (see today’s AI research section). Regardless, open-source AI is on the rise

Cornell: More Agents, smarter AI systems

Image source: Midjourney

The Rundown: A new study from Cornell researchers just found that increasing the number of AI agents collaborating on a problem can significantly improve performance.

The details:

• The researchers created many agents of an AI model and had them each work on a problem independently.
• The agents’ answers were then combined using a voting system to determine the best overall solution.
• Increasing the number of agents improved accuracy across tasks using various LLMs like Llama and GPT.
• A smaller LLM could match or outperform a larger one by scaling up the number of agents, with a 13B parameter Llama model beating a 70B version on some tasks.

Why it matters: This study shows that there’s a simple and effective way to make AI systems smarter: strength in numbers. As compute power and abilities continue to scale, automating unlimited amounts of agents to complete tasks could lead to mind-blowing increases in capabilities.

AI training data gold rush surges

Image source: Getty Images

The Rundown: A new report from Reuters revealed new details on how tech giants like Google, Meta, OpenAI, and Apple are racing to secure vast quantities of online data to feed their AI models.

The details:

• Reuters reports that Meta, Google, Amazon, and Apple all reached a deal with Shutterstock in 2022 following ChatGPT’s debut.
• The agreement included hundreds of millions of images, videos, and music files for AI training, with deals ranging between $25–50M.
• Prices for training data range from cents per image to hundreds of dollars per hour of video.
• Companies are also paying for access to private content archives, including Photobucket’s 13B photos and videos and other old internet platforms.

Why it matters: While initially scraping the web, a flurry of lawsuits and the wave of AI-generated content flooding the internet now have tech giants paying big bucks for high-quality content for AI models. The gold rush is a boon for data-heavy platforms eager to cash out — but also treads into murky waters regarding privacy and consent.

Google Unleashes Powerful AI Tools and Services at Cloud Next ‘24

At Google Cloud Next ’24, the tech giant revealed a host of powerful AI-driven updates. This includes integrating Gemini into Databases, Looker, and BigQuery to enhance data management and analysis. Additionally, Vertex AI Agent Builder simplifies AI agent creation for developers. Google is also building a comprehensive platform for generative AI development, expanding its Vertex AI Model Garden and supporting AI startups.

The updates extend to Google Workspace as well. Google Vids, a generative AI-powered video service, streamlines content creation. AI-driven features for messaging, meetings, and security further bolster productivity and collaboration. Google also announced AI-powered cybersecurity advancements, including Gemini-assisted threat detection and response.

Showcasing real-world applications, Google highlighted how customers like Bayer, Best Buy, and Discover Financial are innovating with Google’s generative AI services. The company also expanded its Gemma family of models, introducing CodeGemma for coding assistance and RecurrentGemma for research workflows.

Alongside software advancements, Google showcased hardware innovations, including the TPU v5p chip and the Arm-based Axion data center processor. These AI-focused chips promise substantial performance gains, underscoring Google’s commitment to providing accessible AI hardware and meeting diverse workload requirements.

New Gemini feature lets you fine-tune answers

The Rundown: In this tutorial, you will learn how to use Gemini’s latest update, which lets you fine-tune generated answers and get the output you need without rewriting your entire prompt.

Step-by-step:

1. Visit Google Gemini. You will need to sign up or log in with your Gmail account.
2. Enter your prompt and let Gemini generate an initial answer.
3. Highlight the specific part of its output that you want to change and click the “Pencil with a star” button. Choose to either modify it with additional instructions, regenerate it, make it shorter/longer, or remove it entirely.
4. Repeat the fine-tuning process as needed until you have the ideal result.

Sam Altman And Jony Ive Seeking Funding For Personal AI Device (1 minute read)

Sam Altman and Jony Ive, former Apple design chief, have officially joined forces to develop an AI-powered personal device that won’t resemble a smartphone. They are currently seeking funding of up to $1 billion.

Elon Musk said that Grok V2 is currently being trained and expected to be completed in May, also revealing the model “should be better” than GPT-4.

Microsoft announced the opening of a new AI hub in London to advance language models, develop foundation model tooling, and collaborate with AI teams across the company and partners like OpenAI.

Spotify introduced AI playlists in beta, allowing users to generate personalized playlists based on written prompts.

A New York Times report revealed that OpenAI transcribed over 1M hours of YouTube videos for GPT-4 training data, raising legal concerns.

Oracle unveiled a new AI-powered cloud service called Compliance Agent to help banks mitigate anti-money laundering risks.

NTT and Yomiuri Shimbun, two of Japan’s most influential companies, released an AI manifesto warning that democracy and social order could collapse if AI is left unchecked.

Poe introduced a new ‘price per message’ monetization model, allowing bot creators to generate revenue each time a user interacts with their creation.