Over its remarkably short public life, about 18 months, generative AI or large language models (LLMs), have have grown from fatassy to a semblance of reality, but its rapid emergence has been tempered by formidable challenges that are just now coming into focus.
Generative AI cannot exist without huge amounts of data, especially user information, powerful, leading-edge micro-processors originally designed for 3-D gaming and lots of investment capital. Users are the fuel that helps LLM’s to get smarter, but that kind of training requires loads of capital.
Intellectual property issues abound, especially as they relate to copyright and generative AI, as well as those that accompany trademark and rights of publicity and image.
While much of what the leaders like OpenAI and Microsoft are supporting is protected under IP rights, AI open source strategies are being taken more seriously by smaller players, as well as Meta. It is difficult to be proprietary, they believe, about products that are evolving so fast and need widespread use and constant refinement to perfect.
The facts?
Generative AI platforms are very expensive. Not unlike the Internet and social media, it is currently dominated by a few big players with the enormous capital necessary to lead. That will not change quickly. Microsoft owns $13 billion of the public side of OpenAI, which is currently valued at $86 billion.
Anthropic, another foundational AI model, raised $4 billion from Amazon and $2 billion more from Google, both of which also provide it cloud infrastructure. Apple recently revealed a slew of AI developments backed by a new relationship with OpenAI.
In April, Meta announced Llama 3, the next generation of its state-of-the-art open source large language model.
Graphics Processing Unit (GPU) chips supplied by Nvidia for $30K to $40K each, are in short supply, so are the cloud servers. Nvidia’s market value is $3 trillion dollars (that’s right, trillion). It is one of three companies so highly valued. You can guess who the others are.
With global semiconductor revenue is projected to double over the next six years (see graph), continued AI growth is more than likely.
GPUs are a specialized electronic circuit initially designed to accelerate computer graphics and image processing. After their initial design, GPUs were found to be useful for non-graphic calculations and training of neural networks.
Training large learning models or LLMs like ChatGPT and and Gemini (formerly Bard) is time-consuming and expensive. Almost no one except a handful of tech leaders can afford to.
The Power of Chips
Photolithography to design and manufacture advanced chips has gotten enormously complex over the years. Back in the 1970s top-line chips had some 29,000 transistors. Intel’s Pentium had 3.1 million transistors in 1993. Xeon chips which support Gmail, Netflix and other current cloud-related computing functions have billions of transistors on a single waver.
Today’s Nvidia GSU chips have 208 billion. Trillion chip wafers are not that far in the future – and they will be pricey.
The current cost of each Nvidia”entry-level” Blackwell GPU, the B100, has an average selling price between $30,000 and $35,000 per chip. Nvidia’s next generation AI superchips could cost $70K each.
The average full-scale data center (1Q 2023) is 100,000 square feet in size and runs around 100,000 servers, which are essentially powerful computers.
GPT-4 used and estimated $78 million worth of computer power for its AI training, indicates Stanford’s 2024 AI Index Report. Gemini’s Ultra cost $191 million to train.
Anthropic’s co-founder, Dario Amodei, told the Bloomberg Tech Summit on May 9 that he expects “the cost to train an AI model will eventually rise to $100 billion as models get bigger and require more computing power.”
The leading chipmakers are dominated almost exclusively by U.S., Taiwanese and South Korean companies. China is getting stronger in a few areas but has a lot of catching up to do. Europe has been a laggard.
The inventions and products associated with advanced chips are protected by patents, and the software is covered by either patents or copyrights. The data that the systems require is likely protectable under trade secrets. Trademarked names, symbols and brands also play a significant role.
Follow the Money
Not unlike the Internet, a key to how generative AI and machine learning evolve is size and capital. The cost of entry is incredibly high for both AI and the chips that power it. Microsoft’s $13 billion investment in OpenAI is a good example. Alphabet and Amazon can afford to play catchup; Apple, too
Smaller players who rely on ChatGPT and other products are focusing on proprietary industry or company-based data sets and the ability to use them with the most meaningful prompts for the relevant responses. The data falls more under the heading of trade secrets than copyrights or patents, but it is a meaningful fuel as the GPU chips. They will use the big LLM engines to power their part of the AI world.
Brian Chen’s harsh analysis of OpenAI’s latest release, ChatGPT-4o (o as in Omni), which attempted to use a semblance of Scarlett Johansson’s voice for responses which sounded suspiciously like her voice in the film “Hers,” and for which she had refused to grant permission. She threatened to sue. The encounter goes a long way to illustrating just how much and little the leading LLMs have evolved.
“To my chagrin, the [ChaptGPT-4o] demo turned out to be essentially a bait and switch,” wrote Chen in the New York Times last week. “The new ChatGPT was released without most of its new features, including the improved voice (which the company told me it postponed to make fixes)…”
“This tactic, in which A.I. companies promise wild new features and deliver a half-baked product, is becoming a trend that is bound to confuse and frustrate people.
“Companies are releasing A.I. products in a premature state partly because they want people to use the technology to help them learn how to improve it. In the past, when companies unveiled new tech products like phones, what we were shown — features like new cameras and brighter screens — was what we were getting.”
With artificial intelligence, companies are giving a preview of a potential future, demonstrating technologies that are being developed and working only in limited, controlled conditions. A mature, reliable product might arrive — or might not.
Resist the Hype
“The lesson to learn from all this,” concludes Chen, an experienced tech reporter, “is that we, as consumers, should resist the hype and take a slow, cautious approach to A.I. We shouldn’t be spending much cash on any underbaked tech until we see proof that the tools work as advertised.”
Image source: clickworker.com via Adobe Stock; InternationalBusinessStrategies via wsj.com
ipcloseup.com 