AI loses focus and forgets what it was doing when it goes off track, reinforcing the investment bubble

 

AI loses focus and forgets what it was doing when it goes off track, reinforcing the investment bubble

 

An analysis suggests that AI's technological limitations, such as forgetting what it was doing and losing focus when it goes off track, could persist and even grow next year.

The New York Times published two columns on its outlook for AI next year on the 25th, focusing on "technological limitations" and "persisting the bubble."

 

Sam Rodriquez, an AI engineer at Edison Scientific, which launched Cosmos, stated in the NYT Dealbook, "One of the key limitations of today's AI systems is that they are limited in the length of their tasks and the sophistication of the tasks they can perform before they go off track. They forget what they were doing and can no longer focus." 

He continued, “In terms of how we build, we build on top of various language models from OpenAI, Google, and Anthropic, but we (Cosmos) use models from various providers. Cosmos also has its own models trained internally for specific tasks, which are much better suited to the specific tasks we train on than models built by frontier providers. The core insight of Cosmos is basically what we call a ‘structured world model.’”

 

Cosmos, he said, “is a way to contribute to a world model that builds over time, and this model essentially describes the entire knowledge of the task they’re working on. You can run hundreds of different agents in parallel, in serial, and have them all work towards a consistent goal, and that’s the real unlock.” The NYT Dealbook host told him that Cosmos costs “$200 per prompt,” and that one person on Kosmos writes an average of 42,000 lines of code and reads 1,500 research papers, so running Cloud might only cost a few hundred lines of code, and if you were a scientist doing experiments and collecting data, it would cost $5,000 to $10,000 to collect that data.

 

He said that Cosmos is the first thing we’ve built that actually feels like an AI scientist. “You set a research goal, and then the idea goes away, and you come back with really deep and interesting insights, sometimes wrong, and about 80% of the time it’s right,” he said. “It’s like asking a human to do something, and a similar percentage of the time they’ll be right.”

Regarding a new contribution to the scientific literature, he said, "The human genome contains millions of genetic variants, and these are differences in DNA from person to person that are associated with disease. We know that most variants are associated with disease, but we have no idea why." 

He continued, "COSMOS provided us with a wealth of raw data on numerous genetic factors. We asked them to identify the variants, which proteins bind near them, and other things. For example, when we asked them to identify the mechanism associated with one of these variants in patients with type 2 diabetes, they identified a mutation that wasn't present in the gene."

He continued, "COSMOS identified this region as a site where other proteins actually bind, and we were able to identify which proteins bind, which genes are expressed, and connect that to the actual mechanism of that gene, SSR1. Ultimately, SSR1 plays a role in insulin secretion in the pancreas."

The New York Times interpreted this as meaning that the COSMOS model made sophisticated inferences based on existing data, uncovering something that other human scientists had long been unable to address, and explained the difference with AI. 

Regarding the New Year bubble that investors are worried about, the New York Times reported that 57% of respondents to Deutsche Bank Research's annual global asset manager survey cited waning enthusiasm for artificial intelligence and declining technology indicators as the biggest threats to the bull market rally, far outpacing concerns about Federal Reserve policy.

"It's misleading to lump everything into one bubble," Adrian Cox, thematic strategist at Deutsche Bank Research, told the Times. "There are at least three bubbles in play: valuation, investment, and technology. In all of them, there's evidence of inflation that could eventually burst. But right now, it feels like we're still in the early stages of that process."

Van Geelen of Citrini Research sees this technology as a revolutionary productivity tool for businesses in the near future, citing AI in robotics and agent AI—autonomous systems that can plan, research, and accomplish goals with limited human supervision—as potential game-changers. "Last year, people were talking about this, but the technology wasn't there yet," he told the Times.

The bubble needed a lot of data centers to reach its goal, sparking a historic construction boom that strained local resources and unleashed a torrent of debt-fueled spending.

The Times reported, "According to the International Energy Agency (IEA), power consumption from U.S. data centers is expected to more than double last year's total by 2030." "This thirst for power is contributing to rising costs for everyday Americans and could become a major issue in next year's midterm elections."

In the latest example of this rapid move to secure power supplies, Google parent company Alphabet agreed on Monday to acquire clean energy developer Intersect Power in an all-cash $4.75 billion deal, aimed at powering Google's data centers.

The sheer scale of investment needed to support this boom has unsettled some on Wall Street. According to Bloomberg, tech giants, including Microsoft and Meta, have agreed to invest a total of $500 billion in data center leases over the next few years.

Oracle alone has committed to $248 billion in leases, a move that sent its stock price plunging earlier this month.

More optimistic observers, who deny the bubble, argue that Mag7 companies, which are heavily funding construction, have cash flow without carrying heavy debt.

OpenAI, which best exemplifies the AI bubble debate, claims its flagship ChatGPT has about 800 million active users, but the New York Times reports that only a small fraction are actually paying customers.

OpenAI CEO Sam Altman recently announced that the company is on track to achieve $20 billion in average annual revenue by the end of this year and has pledged $1.4 trillion in data center construction over the next eight years. “Despite this accounting disconnect, OpenAI, a startup and nonprofit, continues to raise funds at eye-popping valuations, with SoftBank leading a $300 billion funding round in March and a recent funding round in October valuing OpenAI at $500 billion,” the Times reported. “Last week, there were reports that OpenAI was in talks for a $100 billion round that would value the company at $830 billion, and there were also rumors of an IPO in 2027.”

 

Competition is intensifying as OpenAI seeks to monetize its massive user base. Regarding competition, the New York Times reported, “Google’s Gemini 3, which debuted in November, recently outperformed ChatGPT in industry benchmark performance tests.” 

Anthropic, a competing startup founded by OpenAI executives, focuses on enterprise applications, and its latest version of the Claude chatbot reportedly performed “autonomously” for 30 hours with little supervision. However, the Wall Street Journal already lost hundreds of dollars when Claude was put to work running a vending machine in its office.

 

OpenAI faces increased competition from open-source AI models like Chinese startup DeepSec and Alibaba’s Qwen, which are attracting startups to develop on their architectures.

“There’s room for multiple models,” warned Van Halen, founder and CEO of Citrini Research, which specializes in megatrend investment ideas. “What may now seem like slow progress in integrating AI-based tools into everyday business operations could change rapidly.” 

Regarding AI, he said, "Technology advances at an exponential rate, but humans adopt technology at a linear rate." He added, "2026 is the year when AI will start replacing certain jobs, and by this time next year, we may be asking a whole new question: How big can the bubble get?" The NYT reported.

 

When asked, "Should I verify my findings myself? Should I hand them off to a research team and make sure they work? What happens next?" The Cosmos developer replied, "You have to go and verify them yourself. In fact, in the paper, we explain how we verified that particular variant."

 

"Generally, when people use it, you have to get in there," he said. "When you run a Kosmos run, the first thing you do is understand what the game is saying. Once you've read and understood it, you'll want to do a lot of experiments, do your own analysis, and cross-reference to convince yourself that it's true. Then, depending on your research goals, you'll decide what to do next."

In its column, "The 1% Solution to the Coming AI Job Apocalypse," the New York Times stated, "AI will replace workers on a scale many people don't yet realize. In less than a decade, Uber and Lyft have decimated the taxi industry, and self-driving cars could quickly replace human drivers, one of the largest occupations among American men." The paper also stated, "All companies benefiting from automation—most American companies—should follow suit and commit 1% of their profits to retraining."

 

In the Philippines, 80% of the call center workforce, which accounts for 7-10% of GDP, has been replaced by AI agents, prompting companies to decide to invest "1% of their profits" in retraining.

 

The fund, run by an independent nonprofit, partners with companies to ensure the skills being developed are up to the required standards. The paper added, "For 15 years, online learning platforms have demonstrated the applicability of academic learning, and many of the principles apply to technical training."

 

The threat of AI isn't just a job crisis; it's also an educational crisis and challenge. The New York Times stated, "The problem isn't that people can't work. The bigger issue is that we haven't built systems that allow them to continue learning and connect them to new opportunities in a rapidly changing world."