The week in AI: Nvidia rides AI wave to claim title as world's most valuable company

In partnership with

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each Thursday, we aggregate the major developments in artificial intelligence - we pass along the news, useful resources, tools and services; we highlight the top research in the field as well as exciting developments in open source. Even if you aren’t a machine learning engineer, we’ll keep you in touch with the most important developments in AI.

NEWS & OPINION

Nvidia pushes out Apple, Microsoft to become world’s most valuable company

-------------------------

Move over, Microsoft and Apple. The stock market has a new king. Nvidia, the Silicon Valley chip maker, has surpassed the tech giants to become the world's most valuable public company with a market capitalization of $3.36 trillion as of Thursday morning. The company's meteoric rise has been fueled by the boom in generative AI and the surging demand for its graphics processing units (GPUs) - essential for building AI systems. Nvidia's CEO, Jensen Huang, made an early bet on the potential of GPUs in AI, tailoring the company to accommodate what he believed would be tech's next big boom.

Nvidia's dominance in the AI chip market is staggering - the company controls more than 80 percent of the market for chips used in cutting edge AI. Nvidia's rise through AI might seem reminiscent of dot-com era titans like Cisco and Juniper Networks, which built the equipment that ran communications networks for the internet - and many are predicting the AI bubble is just waiting to burst.

Nvidia’s leadership is certainly aware of its overexposure to AI hardware:

• Nvidia’s data center segment has a current run rate of $90 billion (2024’s projected AI-based revenues), which, if materialized, would represent more than half the total revenue of the entire AI market

• The AI accelerator market (GPUs et al.) is projected to grow at a 35-39% CAGR until 2030. If this is true, and if Nvidia holds its market share, its expected revenues from AI would grow to a staggering $800 billion, almost ten times its current run rate. At a 10 price-to-sales ratio, that’s an 8 trillion dollar company.

But they have their eggs in many AI baskets. In his COMPUTEX keynote, Huang emphasized the company's commitment to diversifying within the AI industry, particularly in the realm of "Physical AI" and the conquest of embodied intelligence. Nvidia is positioning itself at the forefront of this challenge by focusing on simulation, training autonomous AIs in realistic virtual environments before transferring them to the physical world. The company is also investing in digital twins, creating precise virtual replicas of factories, the Earth, and even drug discovery.

The other side of the AI boom: software providers struggle as hardware and cloud providers continue to reap profits

-------------------------

“It’s clear that application makers are laser-focused on building AI tools. What’s less clear is whether anyone is willing to pay for them.”

Despite the AI boom, most software makers have yet to see significant revenue from these new capabilities. The majority of AI spending is currently directed towards hardware and cloud infrastructure, which are essential for training and deploying AI models. Meanwhile, traditional software vendors are struggling to capitalize on the AI wave - and investors have lost patience. The iShares Expanded Tech-Software Sector ETF, a common gauge of the industry, has only gained about 3.5% this year after jumping 59% in 2023.

Looking ahead, there are major concerns that AI could potentially disrupt the software-as-a-service (SaaS) business model. As AI enables greater efficiency, it may lead to reduced workforce requirements, impacting the per-user pricing model that many SaaS companies rely on for revenue growth. While some industry leaders, such as Adobe's CEO Shantanu Narayen, believe that AI will ultimately expand the customer base by making software more accessible, most software providers aren’t even yet sure which AI features they should offer freely and which ones they should attempt to monetize.

As the industry navigates this transformative period, the focus on AI-oriented hardware and infrastructure is likely to continue in the near term.

Forbes threatens buzzy AI search startup Perplexity with legal action

-------------------------

A couple of months ago, The New York Times podcast The Daily investigated how the country’s biggest technology companies raced to build powerful new artificial intelligence systems by bending and breaking the rules from the start - and host Michael Barbaro called copyright violation AI’s “Original Sin”. And while the NYT itself launched a lawsuit against OpenAI for copyright infringement back in December, other publishers are making partnerships as they grapple with growing AI use and encroachment on their territory.

But as Forbes has recently found out, it’s not just the Googles and OpenAIs of the world that publishers need to be watching out for. We recently highlighted Perplexity (a relatively small AI player who is pioneering a new way to search using AI) for their new “Pages” feature, which essentially lets you use AI to create a comprehensive article (complete with images and enhanced formatting) with minimum effort. Forbes has accused Perplexity of directly ripping off their reporting without permission or attribution:

• With the “Pages” feature, Perplexity’s chatbot regurgitated a version of an exclusive, paywalled Forbes report on ex-Google CEO Eric Schmidt’s military drone project. Perplexity’s “curated” version (viewed over 30,000 times) lifted near-verbatim passages and even an in-house graphic from Forbes’ original story.

• Perplexity then sent a push notification to its subscribers of its version of the story and published an AI-generated podcast, which was then turned into a YouTube video, that outranks all Forbes content on the topic within Google search.

• Perplexity CEO Aravind Srinivas tried to defend the company's practices on X, stating that the incident was part of a new product feature that has "rough edges" and is being improved "with more feedback."

• In response, Forbes sent a letter directly to Srinivas, demanding that Perplexity remove the misleading source articles, reimburse Forbes for all advertising revenues Perplexity earned via the infringement, provide "satisfactory evidence and written assurances" that it has removed the infringing articles, and that Perplexity provides "written representations and assurances" confirming that it won't use any of Forbes' intellectual property or content to generate and publish AI chatbot articles and that it won't infringe on Forbes copyrights in the future.

WIRED also launched an investigation into Perplexity and found unscrupulous scraping practices. Journalism outlets have regularly blasted AI firms in recent months for using their content to “train” chatbots without proper credit or compensation, and then using the chatbots to erode their audiences. Critics have warned that the rise of AI ripoffs could decimate news publishers unless federal officials intervene.

MORE IN AI THIS WEEK

• What policymakers need to know about AI (and what goes wrong if they don’t)

• Tesla shareholders sue Musk for starting competing AI company

• Latent Expertise: everyone is in R&D

• Black founders are creating tailored ChatGPTs for a more personalized experience

• McDonald’s will stop testing AI to take drive-thru orders, for now

• “AI Steve” is running for Parliament in the UK

• Clearview AI used your face - now you may get a stake in the company

• Bitcoin miners at Donald Trump’s closed-door event say he thinks Bitcoin can help win AI arms race

• Microsoft’s star AI chief peers into OpenAI’s code, highlighting an unusual rivalry

• A Generation of AI Guinea Pigs - AI is quickly becoming a regular part of children’s lives. What happens next?

• Amazon has a secret way to scrape Microsoft’s GitHub and feed its AI model

Use AI as Your Personal Assistant

Ready to embrace a new era of task delegation?

HubSpot’s highly anticipated AI Task Delegation Playbook is your key to supercharging your productivity and saving precious time.

Learn how to integrate AI technology into your processes, allowing you to optimize resource allocation and maximize output with precision and ease.

Download Now!

TECHNICAL, RESEARCH & OPEN SOURCE

Nvidia releases free, open LLMs that match GPT-4 in some benchmarks for synthetic data generation

-------------------------

Nvidia has just launched a groundbreaking family of open models designed to generate synthetic data for training large language models. This cutting-edge pipeline enables the creation of high-quality datasets for commercial applications across industries from healthcare to finance to retail. The context length is currently limited to 4,096 tokens; otherwise, this is an impressive and powerful suite of language models. The Nemotron-4 340B family comprises three key components:

• Base Model: Trained on 9 trillion tokens (!), this model serves as the foundation for customization using Nvidia's NeMo framework, allowing developers to adapt it to specific use cases or domains.

• Instruct Model: This model generates diverse synthetic data that closely mimics real-world characteristics, enhancing the performance and robustness of custom LLMs.

• Reward Model: To further refine the quality of AI-generated data, the Reward model can filter out the best responses based on five essential attributes: helpfulness, correctness, coherence, complexity, and verbosity.

Impressively, Nemotron-4 340B Instruct performs on par with or better than GPT-4 in human evaluation for various text tasks, such as summaries and brainstorming. You can read more about the benchmarks in the technical report. The models are optimized for inference with the open-source NeMo framework and the Nvidia TensorRT-LLM library. Nvidia makes them available under its Open Model License, allowing commercial use. All data is available on Huggingface.

Nvidia also published research on Mamba-based language models (selective state models, an alternative architecture to standard Transformer models). Transformer models are inefficient from a quadratic standpoint due to the self-attention mechanism - the computational cost and memory requirements increase quadratically, making it challenging to process very long sequences efficiently. Although challenges remain, the study demonstrated the potential for Mamba-based models, especially hybrids, to rival or even overtake Transformers for certain use cases where inference efficiency is critical.

MORE IN T/R/OS:

• China’s DeepSeek Coder becomes first open-source coding model to beat GPT-4 Turbo on HumanEval

• CryptGPT: Privacy-preserving language models using Vigenere Cipher (Part 1)

• AI Agents: hype vs. reality

• Gemini’s API gets context-caching: a new feature designed to reduce overall operational costs

• (Google DeepMind research) Generating audio for video

• (Shanghai AI Lab) MCT Self-Refine: a new algorithm that helped a tiny Llama-3 8B model to achieve GPT-4 level performance on complex math

• (Stanford AI Lab) HumanPlus: a system enabling humanoid robots to autonomously learn and perform tasks by imitating humans

That’s all for this week! We’ll see you next Thursday