Michael Clark’s Post

AI models grew 110x in size in 3 years - but their actual performance gains plateaued, with runaway environmental impacts. A preprint from researchers at Inria, Hugging Face, and Signal Messenger challenges the "bigger is better" paradigm in AI development. Their analysis shows that the law of diminishing returns applies to increasing model size, while the environmental and economic costs grow exponentially. For instance, they show that specialized models with 1.3GB can match the performance of 20GB models on specific tasks (reminiscent of Andrej Karpathy). This implies that as computing requirements for state-of-the-art models now exceed the capacity of the world's largest supercomputers, we're creating an unsustainable arms race that benefits only the most resourced players. And, frankly, it looks unstoppable. The paper makes a compelling case for refocusing AI research on efficiency and specialized applications rather than raw scale. As nations appoint AI focus groups (poke the new Crypto/AI czar) and shape policy, or deregulation, around this technology, this research suggests we need to critically examine whether the current path of exponential growth in model size serves society's best interests. Kudos to Gael Varoquaux, Dr. Sasha Luccioni, and Meredith Whittaker, three rockstars in AI for spreading the word.

AI is effectively useless. . . If so, it begs the question. . ."To AI or Not To AI. . .?" James Ferguson, founding partner of MacroStrategy Partnership, puts it bluntly: "AI still remains, I would argue, completely unproven. And fake it till you make it may work in Silicon Valley, but for the rest of us, I think once bitten twice shy may be more appropriate for AI." Ferguson's concerns are echoed by the fact that LLMs often "hallucinate", inventing facts, sources, and more. Moreover, AI's energy consumption and unproven ROI are significant limitations that can't be ignored. But here's the thing: some AI companies getting funding today are just "lipstick on a pig." They're dressing up familiar ideas or incremental improvements on pre-built LLMs as revolutionary breakthroughs. And that means a lot of investor money will be lost when the bubble bursts Ferguson notes, "If AI cannot be trusted... then AI is effectively, in my mind, useless." Let's not forget that the dot-com bubble burst, and many investors were left with significant losses. But will we repeat similar mistakes with AI?                        While AI has made tremendous progress, we need to separate the hype from reality. LLMs and generative AI have a ways to go before they deliver real value to businesses and society. Does that mean AI is useless? Or that we are in a bubble? There will certainly be a lot of money lost due to bad investents – but every innovator knows, technology takes time to take hold. . . Let's take a step back, breathe, and focus on the fundamentals. The next wave of innovation will come from those who can harness AI's potential while addressing its limitations. #innovation #future #futureofai #AI #LLMs #GenerativeAI #DotCom #Hype #RealityCheck

The intersection of AI and decentralization is more than a trend—it’s a necessary evolution. At dFusion AI Labs, we’re tackling one of the AI industry’s biggest challenges: data accountability. AI models often suffer from hidden biases due to low-quality or unchecked data sources. With dFusion AI Labs’s decentralized validation process, we empower communities to curate data, ensuring that trusted inputs shape more reliable and equitable AI systems. Our mission goes beyond technology. By connecting node operators, researchers, and developers, we’re creating a decentralized ecosystem where data ownership and incentives prioritize quality. The future of AI lies in building trustworthy ecosystems where accountability and transparency take center stage.

The MIT Technology Review paints an exciting picture of how Generative AI (GenAI) will transform industries, but let's take a step further into the future. Imagine a world where your Netflix shows are dynamically generated based on your evolving preferences, creating a personalized entertainment experience like never before. Now, consider the financial sector and envision AI-driven financial advising that tailors investment strategies to your genetic profile, predicting risk appetite based on your genes, health, and lifestyle. This technology could use synthetic simulations to analyze market trends and personal circumstances, ensuring that your investments grow and secure a comfortable retirement despite the challenges of longer lifespans and retirement savings shortages. With GenAI, not only could your financial health be optimized in real-time, but it could also make personalized adjustments to keep you financially secure and prepared for the future. As Elon Musk puts it, “Generative AI is the most powerful tool for creativity that has ever been created. It has the potential to unleash a new era of human innovation.” This isn’t just a vision—it’s a roadmap to intelligent financial management.

"Acemoglu envisions three ways the AI story could play out in coming years. The first — and by far most benign — scenario calls for the hype to slowly cool and investments in “modest” uses of the technology to take hold. In the second scenario, the frenzy builds for another year or so, leading to a tech stock crash that leaves investors, executives and students disillusioned with the technology. “AI spring followed by AI winter,” he calls this one. The third — and scariest — scenario is that the mania goes unchecked for years, leading companies to cut scores of jobs and pump hundreds of billions of dollars into AI “without understanding what they’re going to do with it,” only to be left scrambling to try to rehire workers when the technology doesn’t pan out. “Now there are widespread negative outcomes for the whole economy.” The most likely? https://lnkd.in/gmiMWZy6

AI - is it really going to live up to the hype? My views are the industry and the money men are going to catch a cold and so this extract from a recent Telegraph article caught my eye. “MIT economics Professor Daron Acemoglu, a contributor to the Goldman analysis and co-author of the best-seller Why Nations Fail, thinks that the hype obscures the reality of what is really a quite limited technology. “It was always a pipe dream to reach anything resembling complex human cognition on the basis of predicting words,” he says. Far fewer jobs are exposed to automation, he thinks – just 4.6pc of tasks can be reliably automated. Over a decade, Acemoglu envisages a mere 0.53pc improvement in total factor productivity. So instead of a “fourth industrial revolution”, generative AI’s impact more closely resembles that of the Excel macro. Useful but not exactly epoch-defining. “There is pretty much nothing that humans do as a meaningful occupation that generative AI can now do,” Acemoglu warns. And he suspects the vast social cost of fraud enabled by AI will bury any advantage in the public’s mind.”

My latest piece for CIO Online on why the current chill in enthusiasm for AI is not a bad thing and why businesses should continue to experiment and invest in the technology https://lnkd.in/eqH8gEhD

Daron Acemoglu is doubling down and expanding a bit on what is on the GS paper on Generative AI: "Only about a quarter of tasks that could use artificial intelligence will likely be cost-effective in the next decade, according to Daron Acemoglu, Institute Professor at MIT. Speaking on the Goldman Sachs Exchanges podcast, Acemoglu says that even with big breakthroughs in AI, the impact will not be seen for several years." "The current architecture of the large language models has proven to be more impressive than many people would have predicted, but I think it still takes a big leap of faith to say that just on this architecture of predicting the next word, we're going to get something that's as smart as, you know, Hal in 2001: A Space Odyssey," he said. There "could be very severe limits on where we can go with the current (LLM) architecture," Acemoglu said. He is also skeptical that AI can achieve its goals more quickly by simply using more GPU capacity. He added that higher and higher-quality data, rather than capacity, will be needed, and where that data comes from needs to be clarified. It's worth a listen. https://lnkd.in/eMjnGVcy #AI #generativeAI

The Return on Investment for AI is in the use cases