Janusz Maretski

AI industry analyst and researcher with a focus on next-generation AI architectures.

Claims

As speaker

Mentioned in

Key positions and views

The current paradigm of scaling large language models is a technological dead end due to insurmountable issues like data exhaustion, model collapse, catastrophic forgetting, and hallucination.

The hundreds of billions of dollars being invested in data centers for current AI models represents a 'catastrophic misallocation of capital' based on a flawed and soon-to-be-obsolete technology.

A new generation of AI, inspired by neuroscience and capable of continual learning with vastly superior power efficiency, is already emerging and outperforming existing systems.

The future of powerful AI is decentralized and localized, with the most capable models running on consumer devices like laptops within the next one to three years, disrupting the current cloud-centric model.

The exodus of top-tier talent from major AI labs is the most significant leading indicator that a fundamental paradigm shift away from current LLM architectures is underway.

Podcast consensus on Maretski

Points of consensus

▶Maretski consistently argues that the current AI paradigm of scaling large language models has reached a fundamental limit, citing the exhaustion of high-quality training data, a three-year plateau in benchmark performance, and the problem of 'model collapse' from AI-generated data.Apr 2026

▶He repeatedly describes the hundreds of billions of dollars being invested in data centers for current-generation AI as a 'catastrophic misallocation of capital,' believing the underlying technology is flawed and will soon be obsolete.Apr 2026

▶A core tenet of his view is the imminent rise of a new generation of brain-inspired AI systems (from companies like Fractal Brain and Innate AI) that are capable of continual learning and are orders of magnitude more power-efficient.Apr 2026

▶He points to the exodus of top AI researchers (like Ilya Sutskever, Dave Silver, Jan LeCun) from major labs as a key indicator that the industry's intellectual leaders are abandoning the current LLM approach to pursue fundamentally new ideas.Apr 2026

Points of debate

▶While the prevailing market narrative fuels massive investment in AI data centers, Maretski argues this is a 'false start' and that companies avoiding this capex, like Apple, are positioned to be the long-term winners.Apr 2026

▶In contrast to the common perception of rapid, ongoing AI progress, Maretski asserts that performance on standard benchmarks has actually been stagnant for the past three years.Apr 2026

▶While the industry is focused on large, cloud-based models, Maretski's contrarian view is that the future of powerful AI is decentralized, with the most potent models running locally on consumer laptops within one to three years.Apr 2026

▶Maretski challenges the idea that AI can be improved by training on its own output, instead claiming this practice leads to 'model collapse,' where models progressively degrade in quality and become 'dumber.'Apr 2026

Key themes

▶The End of the Scaling EraApr 2026

Maretski argues that the dominant AI strategy of scaling up models with more data and compute has hit a wall. He cites the exhaustion of quality training data three and a half years ago, the degrading effect of 'model collapse' from AI-generated content, and a three-year performance plateau on benchmarks as definitive evidence.

This theme suggests that competitive advantage in AI will shift away from access to massive capital and compute, and toward fundamental innovation in model architecture and learning algorithms.

▶Catastrophic Misallocation of CapitalApr 2026

He contends that the hundreds of billions being invested in data centers for the current generation of LLMs is a monumental financial mistake. This view is rooted in his belief that these models are architecturally flawed, unable to learn continuously, and will soon be superseded by far more efficient, decentralized alternatives.

Investors should critically evaluate companies with heavy capex commitments to AI infrastructure, as this strategy carries significant risk if Maretski's predicted paradigm shift materializes.

▶The Rise of Brain-Inspired AIApr 2026

Maretski highlights an emerging generation of AI systems from companies like Fractal Brain, Innate AI, and Pathway AI. These models are inspired by neuroscience, using techniques like Hebbian learning and architectures based on the prefrontal cortex to achieve continual learning and 3-4 orders of magnitude greater power efficiency.

This signals a potential disruption where smaller, agile companies with novel architectures could leapfrog the current market leaders, making early-stage tech scouting in this area critical.

▶The Great Talent MigrationApr 2026

Maretski interprets the departure of top-tier AI researchers like Jan LeCun, Ilya Sutskever, and Dave Silver from major labs as a leading indicator of a paradigm shift. He believes this exodus shows that the field's brightest minds recognize the fundamental limitations of current LLMs and are moving to build the next generation of AI.

Tracking the movement of this key talent provides a valuable signal for identifying which new companies and research directions are most likely to define the future of the AI industry.

Source episodes

Timeline

~3.5 years ago

Claims the AI industry ran out of high-quality, diverse training data, marking a critical turning point for the scaling paradigm. [19]

~3 years ago

Asserts that the performance of large language models on standard benchmarks plateaued and has not shown significant improvement since. [18]

End of 2022

Notes the completion of OpenAI's GPT-4 training, which he claims utilized all publicly available internet data at the time. [12]

Recent past

Observes the departure of prominent AI researchers like Dave Silver, Ilya Sutskever, and Jan LeCun from major labs, which he interprets as a signal of a paradigm shift. [17]

Present

Characterizes the current period as one of 'catastrophic misallocation of capital' into data centers for a fundamentally limited generation of AI technology. [4]

Suggested prompts

What specific technical evidence supports Maretski's claim that next-generation AI systems are already outperforming current models on common benchmarks? &nearr;How does Maretski's thesis on 'catastrophic misallocation of capital' in data centers account for the persistent and growing demand for cloud-based AI services from enterprises? &nearr;What are the primary technical and commercial barriers to the widespread adoption of the brain-inspired AI architectures that Maretski champions, such as those from Fractal Brain or Innate AI? &nearr;If Maretski's prediction of powerful, locally-run LLMs comes true, what are the second-order effects on the business models of current AI leaders like OpenAI, Google, and Microsoft? &nearr;

Key concepts

Model Collapse 1 ep Capital Misallocation 1 ep Continual Learning 1 ep Catastrophic Forgetting 1 ep Brain-Inspired AI 1 ep Data Center Investment 1 ep Local LLMs 1 ep Benchmark Plateau 1 ep AI Talent Migration 1 ep Power Efficiency 1 ep

Notable quotes

“the enormous capex band on vast data centers, the hundreds of billions of dollars that have already gone into this and are still projected to go into this, this is a catastrophic misallocation of capital.”

Janusz Maretski · Why You Should Wait Out AI’s Super-Spending False Start | Merryn Talks Money

“if you look at the data that is currently being created on the open internet, it is data created by those LLMs... that, using a technical term, slowly ends up leading to something called model collapse where the models themselves are actually getting dumber.”

Janusz Maretski · Why You Should Wait Out AI’s Super-Spending False Start | Merryn Talks Money

“what's going to happen in a year, two, three years from now, when the majority of laptops out there will be able to run general purpose, the most potent language models with access to the Internet.”

Janusz Maretski · Why You Should Wait Out AI’s Super-Spending False Start | Merryn Talks Money

“I worry more about potential misuse of those next generation systems, less worry about their performance because they're already beating existing systems on common benchmarks.”

Janusz Maretski · Why You Should Wait Out AI’s Super-Spending False Start | Merryn Talks Money

Report last updated: Apr 21, 2026

Get started free

Back to Entities Intelligence Report