Lee-Lean Shu, CEO, GSI Technology.
Is the AI market overhyped? Is it on the verge of a spectacular fall, a drop similar to the dot-com crash nearly 25 years ago?
Some industry watchers certainly believe the bubble is about to burst. They point to the astronomical valuations of AI startups and the frenzied investment in the sector as clear signs of the “irrational exuberance” that Federal Reserve board chair Alan Greenspan pointed out in the dot-com market prior to its implosion.
Indeed, AI investments drove an eye-watering 47% increase in U.S. VC funding in the second quarter of this year, led by “$6 billion raised by Elon Musk’s xAI and $1.1 billion raised by CoreWeave,” an AI hyperscaler.
However, another view suggests that rather than facing an inevitable fall from grace, the AI market may simply be reaching a temporary plateau before its next ascent to widespread adoption and utility.
If AI is taking a breather, it would not be unprecedented. The Internet of Things (IoT) market experienced a period of consolidation before resuming its growth trajectory. What is unprecedented is the speed of AI’s advancement compared to previous paradigm shifts in technology. Unlike IoT, which took time to demonstrate its practical value, AI began addressing real-world needs and providing actionable intelligence much earlier in its development cycle.
Just look at the explosive growth of generative AI since ChatGPT came on the scene two short years ago. The latest annual McKinsey Global Survey on the state of AI, conducted earlier this year, confirms that 65% of organizations are now regularly using GenAI. What’s more, Bloomberg reports that GenAI is set to become a $1.3 trillion market by 2032, up from about $40 billion in 2022.
AI’s Potential
The growth we’ve witnessed so far has largely been in the media and search fields. NFT (non-fungible tokens) and digital art were early adopters that moved capabilities forward. We are already seeing that generational growth in mainstream movie and picture editing software.
ChatGPT was one of the first to take off, and now larger text responses with fewer hallucinations are the norm. Prompt generation—discerning what the operator is asking—is used now in many custom search tools, saving time in data searches.
Companies now use AI to analyze customer feedback. For instance, Amazon generates product summaries from customer reviews. These improvements greatly benefit users but only scratch the surface of AI’s potential.
Looking ahead, AI cannot help but impact virtually all fields in which data can be leveraged to perform, create, improve or act—which is basically every industry under the sun. For this widespread adoption to occur, tools need to catch up and provide the interfaces and actions that various industries will require. This gap between AI’s capabilities and its accessibility is both a challenge and an opportunity.
The Next Breakthrough
The next major breakthrough will come with the release of new applications that build upon and significantly expand the capabilities we’ve seen in early AI implementations and go into other brand-new spaces. This upcoming wave of innovation will combine three key elements:
• Advanced predictive algorithms that can accurately anticipate user needs and intentions.
• Sophisticated, large action models capable of determining the optimal steps to achieve desired outcomes.
• Machine learning systems that can execute these actions autonomously.
For instance, future AI assistants might book a business trip and offer accommodations based on an executive’s preferences. There is market interest for natural customer support AI for online or phone inquiries. AI and AR glasses are being combined with repair manual data to enhance the capabilities of technicians. This has been a vision for AR and IoT, but AI is needed to implement it.
A Critical Challenge
As these more advanced applications emerge and gain traction, we can expect to see another surge of innovation and investment in the AI sector. But a critical challenge in the progression of AI remains: enhancing the accuracy of predictions. This will require the use of more data—which will require more storage and more processing. The good news is that innovative approaches are emerging to optimize AI processing.
One solution is higher-performance GPUs and increasingly complex methods of connecting them to attached DRAM. The Nvidia Blackwell family illustrates this path. A second path is creating large computational arrays that can hold large models. The larger the silicon die, the greater the yield loss, and a complex tradeoff takes place of complexity of compute, density of simpler memory and cost from yield loss of the combination.
A third path is to address the problem with more power and compute efficiency. One programmable solution comes from GSI, which is applying single-bit and ternary processing techniques to its compute-in-memory (CIM) processor architecture. This novel approach maintains flexibility and enables a substantial increase in the size of models that can be addressed while simultaneously reducing power consumption and maintaining accuracy.
The implications of resolving the computational bottlenecks we are currently seeing are far-reaching. With dramatic improvement in the efficiency of AI computations, it becomes possible to run larger models to address accuracy and expanding use cases while controlling energy requirements to manageable levels. This breakthrough has the potential to accelerate AI development across various sectors, from data centers to edge devices.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
Source link
