garymarcus-substack-com

Customers are recognizing that tokens from three major IPOs are being burned for millions of dollars with little tangible return on investment.

Uber COO Andrew Macdonald stated that the company is not seeing proportional productivity gains from its rising AI costs, raising concerns about the return on investment in artificial intelligence.

The article warns that a seemingly minor policy trick could potentially cost retirement funds billions of dollars, urging readers to contact their representatives in Congress.

Gary Marcus scrutinizes recent claims from OpenAI and Anthropic, urging readers to examine the underlying details and math behind their headline-grabbing announcements rather than taking them at face value.

Gary Marcus compares generative AI's trajectory to the Vietnam War, arguing the industry risks a major backlash if it continues to overpromise and underdeliver. He suggests public disappointment could ultimately push AI development toward more rigorous, trustworthy approaches.

The high-profile AI trial concludes anticlimactically, leaving certain questions permanently unanswered.

Gary Marcus argues that Generative AI, as currently pursued through massive hyperscaling, is an illusion. He makes the case for alternative approaches like world models and neurosymbolic AI, suggesting that large-scale bets on current methods are misguided.

The article argues that US AI policy is fragmented and ineffective, with over 1,200 proposed bills at state and federal levels lacking a coherent national framework. It calls for a more structured, evidence-based approach to regulate artificial intelligence effectively.

Gary Marcus critiques the interpretation of METR's latest "time horizon" graph, arguing that fears over rapid AI progress are misplaced. He breaks down what the data actually shows versus the overblown claims about AI taking over human tasks, emphasizing that the graph measures specific task completion times rather than general intelligence or autonomous capabilities.

The article discusses skepticism around the return on investment (ROI) of generative AI for most businesses, referencing an MIT study that found limited financial benefits from current AI agent implementations.

The article critiques OpenAI's business model, suggesting the company had not yet determined how to fund its operations, raising questions about its long-term financial sustainability.

The article discusses two differing viewpoints on the events unfolding in the legal dispute between Elon Musk and OpenAI, focusing on what should be considered significant in the case. It highlights contrasting perspectives on the trial's key issues and developments.

Gary Marcus argues that current autonomous AI agents are unreliable and chaotic, labeling their performance as a "shitshow" due to frequent failures and lack of robust reasoning.

The article discusses the increasing public and expert backlash against rapid, unregulated AI development, highlighting concerns over safety, ethics, and the gap between corporate promises and actual risks.

A review indicates that large language models have not yet led to measurable improvements in patient outcomes, despite widespread experimentation in healthcare settings.

The article discusses how renowned evolutionary biologist and skeptic Richard Dawkins was reportedly misled or deceived by an AI system called Claude, highlighting the irony of a prominent critic of irrational beliefs falling for an AI-generated illusion.

The article argues that AI-generated code that compiles and passes tests is not equivalent to correct, secure, maintainable, or well-architected software, highlighting the gap between superficial success metrics and genuine software quality.

The article warns that massive investment in AI without solving fundamental issues like reliability could become the greatest capital misallocation in history, as concerns grow that current spending exceeds realistic returns.

Gary Marcus offers three thoughts on Elon Musk's lawsuit against OpenAI, arguing that while Musk's motivations may be questionable, the lawsuit does raise legitimate concerns about OpenAI's shift from a nonprofit mission to a for-profit structure and its commitment to transparency and safety.

Gary Marcus critiques Dario Amodei and other AI cheerleaders for downplaying the risks associated with increasingly powerful AI systems. He argues that hype-fueled, "vibe-coded" AI deployments are leading to real-world disasters, particularly in safety-critical domains, while the industry downplays these dangers.

ChatGPT struggles with basic spatial reasoning tasks like distinguishing between left and right, according to tests by Gary Marcus. The AI system frequently fails at simple directional questions that humans find trivial, revealing limitations in its understanding of fundamental concepts.

ChatGPT's new image generation capabilities demonstrate that while the system can produce impressive visual outputs, this ability does not equate to genuine understanding. The article examines the distinction between sophisticated pattern reproduction and actual comprehension in AI systems.

Multiple studies indicate that chatbots should not be trusted for medical advice due to their limitations in providing accurate and reliable health information. The research consistently shows potential risks in relying on AI chatbots for medical guidance.

Anthropic has filed a lawsuit against the US government. The company's CEO, Dario Amodei, is leading this legal action.

Recent outages at major tech companies have been linked to the use of AI coding tools, with some incidents described as having "high blast radius" impact. These disruptions highlight the risks associated with automated software development systems.

The Pentagon has labeled Anthropic, the company behind Claude AI, as a supply chain risk. This raises questions about whether the US military is concerned about the AI system itself or other factors related to the company's operations and security.

Two new large-scale AI experiments have reportedly failed, providing evidence that simply scaling up models may not be sufficient for achieving desired outcomes. The expensive studies challenge the assumption that scaling alone is all that's needed in AI development.

Sam Altman acknowledges that achieving artificial general intelligence will require major breakthroughs beyond simply scaling current AI systems. He states it is time to look for new architectures rather than relying on existing approaches.

The article discusses how cancer research could serve as a meaningful test for artificial intelligence systems. It explores the potential for AI to contribute to cancer diagnosis, treatment, and research advancements in the medical field.

The article announces several upcoming live events featuring the author, with apologies for the short notice regarding these quick announcements.