Here’s what’s on our plate today:

• 🧪 Decoding the monoculture problem in AI.

• 📰 Zuck's tough-love memo, SpaceX's IPO reveal, HSBC bets on AI honesty.

• 💡 Roko's Pro Tip: diversify your AI stack before a single paradigm shift wipes you out.

• 🗳️ Poll: What finally breaks the AI monoculture?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• One architecture won, not on merit: The transformer model became dominant less through scientific superiority and more through economic feedback loops: GPUs happened to fit, capital followed, and everything else got crowded out.

• The cracks are real: Models hallucinate 15-50% of the time on enterprise tasks, fail when math problems are slightly rephrased, and, per ICML 2024 research, cannot genuinely plan or self-verify.

• Alternatives exist, starved not dead: Symbolic, Bayesian, neuromorphic, and neuro-symbolic approaches never failed scientifically. They failed to attract GPU-scale infrastructure, funding, and tooling.

• Capital is hedging, barely: $5B flowed into non-GPU architectures between 2024 and early 2026, but neuromorphic revenue was only $50M in 2025. The pivot is cautious at best.

• The stakes: If transformer limitations become unavoidable before alternatives mature, there may be no Cavendish ready to replace the crop.

Decoding the monoculture problem in AI

Until the 1950s, the banana most people ate was not the one on supermarket shelves today. The Gros Michel, sweeter and richer than its successor, dominated global commercial trade until a soil fungus called Panama disease swept through Central and South American plantations, wiping out what had taken decades to build. The reason one fungus caused the collapse of a thriving industry was that it had grown almost all the bananas on Earth from a single genetic clone. When a single pathogen found the right key, every lock opened at once. The replacement, the Cavendish, now accounts for 99% of bananas exported worldwide and faces a different strain of the same fungus that already threatens its survival.

The banana industry of the 1950s did not collapse because the Gros Michel was bad. It collapsed because everyone depended on just one approach to growing bananas, and no one maintained the infrastructure to grow anything else.

As of 2026, something structurally similar is unfolding in artificial intelligence, except that the monoculture in AI was not created by an agricultural accident but by economic logic, and what happens when it finally reaches its limits remains an open question.

How one architecture won

The current generation of AI tools, the large language models (LLMs) behind ChatGPT, Claude, and Gemini, are built on a single architectural idea called the transformer (a system that processes language by learning which words and concepts are most relevant to each other, rather than reading text sequentially). It was introduced by Google researchers in 2017.

Like AlexNet before it, the transformer model represented a genuine scientific breakthrough, effectively launching the deep learning era. These were real discoveries, not market constructs.

What followed, however, was driven far more by economics than by science. Graphics processing units, or GPUs, were originally built for video games but turned out to be unusually well-suited for the parallel computations required by neural networks. NVIDIA did not design them for AI; they simply happened to fit the problem almost perfectly.

That accidental alignment created a powerful feedback loop. As demand for AI systems surged, cloud providers poured hundreds of billions of dollars into GPU infrastructure. The growing scale of that infrastructure made the transformer-plus-GPU model cheaper and easier to access, which attracted more researchers and companies, which in turn justified even larger investments in data centers and compute capacity. According to the International Energy Agency, roughly $580B was spent globally on AI-focused data centers in 2025 alone. OpenAI, meanwhile, has projected $50B in compute spending for 2026, a figure disclosed under oath by the company's president, Greg Brockman.

When venture capital entered the picture, the field narrowed even further. Large language models were easy to fund because their progress was clearly measurable: benchmark scores kept improving, model sizes kept growing, and each new release could be presented as a visible leap forward. The story translated neatly into investor decks, growth projections, and commercial narratives, which drew even more capital into the transformer ecosystem.

And while this dynamic accelerated the growth of the AI industry, it also discouraged investment in alternative approaches, including methods grounded in decades of scientific research. Those approaches increasingly struggled to compete in an environment shaped by benchmark scores, scaling curves, and investor expectations. Their progress was often slower, less linear, or harder to translate into the kind of clean performance metrics venture capital tends to reward. As Gary Marcus has argued in his published work, the enormous concentration of capital around deep learning has made it structurally difficult for alternative hypotheses to gain serious attention, regardless of their scientific merit.

The current state of AI, then, has reached a point where the commercial success of one dominant approach has increasingly crowded out investment in alternatives, creating a dynamic not unlike putting all the eggs in one basket, a strategy that rarely ends well in the long run.

What the current crop can’t do

In October 2024, researchers at Apple introduced a benchmark called GSM-Symbolic to test whether leading AI models were actually reasoning through problems or simply reproducing patterns from training data. They modified math problems that models had previously solved correctly by making small changes, such as altering numbers or adding irrelevant details, and found that accuracy dropped sharply, in some cases by as much as 65%. The results suggested that the systems relied less on logical reasoning than on familiar statistical patterns that could easily break down under slight variations.

Subbarao Kambhampati reached a similar conclusion through separate experiments presented at ICML 2024, arguing that auto-regressive LLMs “cannot, by themselves, do planning or self-verification” and should instead be treated as “universal approximate knowledge sources” that require external verification systems.

The consequences of those limitations are already visible. Independent benchmarking has placed hallucination rates, instances where models confidently present false information as fact, between 15% and 50% across enterprise tasks. In legal research, a study from the Stanford Institute for Human-Centered Artificial Intelligence found that specialized legal AI tools hallucinated between 17% and 34% of the time on difficult queries. A peer-reviewed 2025 study titled “Trust Me, I’m Wrong” found that these systems also routinely produce what researchers called “high-certainty hallucinations,” confidently delivering incorrect answers even when the correct information is available.

None of this has meaningfully slowed deployment, but it has revived a question the industry largely sidelined during the scaling boom: what kind of intelligence were these systems actually building toward?

The fields that didn’t get farmed

Before transformers consolidated the field, AI research was far more pluralistic. Symbolic AI relied on explicit logical rules that allowed machines to reason step by step. Bayesian systems treated knowledge as a probability distribution, continuously updating conclusions as new evidence arrived. Evolutionary algorithms searched for solutions through processes resembling natural selection. Neuromorphic computing, an area attracting renewed attention today, aims to mimic the architecture of biological brains, in which neurons fire only when necessary rather than constantly performing large-scale matrix calculations. The contrast in efficiency is stark: the human brain consumes roughly 20 watts, while training a frontier language model requires vastly more energy.

None of these approaches failed because they were scientifically invalid. They lost because the economic infrastructure around transformers and GPUs became self-reinforcing, while competing paradigms lacked comparable scale in investment, tooling, and compute ecosystems. A peer-reviewed paper published in Communications Psychology in 2026 described the result as a “scientific monoculture,” arguing that funding structures, publishing incentives, and career pressures increasingly reward work aligned with transformer-based AI while marginalizing research that does not fit neatly into that framework. The paper drew a parallel between biodiversity in ecosystems and intellectual diversity: both protect systems from collapse when dominant paradigms reach their limits.

Even some of the researchers most responsible for modern deep learning have begun expressing doubts about its trajectory. Yann LeCun, one of the three recipients of the 2018 Turing Award for foundational work in deep learning, has repeatedly described current large language models as a “fundamental dead end,” while fellow laureate Yoshua Bengio has taken a more cautious position focused on managing risk rather than abandoning the paradigm altogether. The disagreement itself is revealing: even the architects of the current system do not fully agree on where it leads.

Money moving slowly

And there are signs that capital is beginning to hedge, though not yet pivot decisively. Unconventional, founded by former Databricks AI head Naveen Rao, raised $475M in seed funding with backing from Andreessen Horowitz, Lightspeed, and Jeff Bezos, while Intel launched its Hala Point neuromorphic system in 2024. More broadly, over $5B flowed into non-GPU AI architectures between 2024 and early 2026, although the neuromorphic sector generated only around $50M in revenue in 2025, highlighting how early the field still is.

At the same time, research interest is also slowly broadening. Work in neuro-symbolic AI, which combines neural networks with symbolic reasoning systems, has accelerated in recent years. Google DeepMind’s AlphaGeometry demonstrated the promise of this hybrid approach by solving Olympiad-level geometry problems using a combination of neural networks and symbolic deduction rather than relying purely on scaled language modeling.

Waiting for the next Cavendish

The banana monoculture eventually produced a replacement crop, but only after the industry had become dangerously dependent on a single vulnerable system. The lesson is not that monocultures inevitably collapse; it is that their risks often become visible only after alternative infrastructures have withered away.

AI may still be in the phase where the dominant crop continues to produce impressive results, but the research ecosystem surrounding alternatives remains thin compared to the immense concentration of money and compute flowing into transformer-based systems. Whether those alternatives mature before the limitations of the current paradigm become unavoidable is a question no benchmark can yet answer.

Roko Pro Tip

Monday Poll

Bite-Sized Brains

• Zuck's tough-love memo: Mark Zuckerberg is reportedly taunting Meta employees ahead of fresh layoffs, framing performance cuts as a culture reset rather than cost-cutting.

• SpaceX's IPO reveal: The filing exposes the company's full AI ambitions, Starship economics, and the extent to which Elon Musk's broader empire is intertwined with them.

• HSBC bets on AI honesty: HSBC's CEO is openly telling staff that AI will destroy jobs and create new ones, urging employees to adapt rather than fight the shift.

Meme Of The Day

The Toolkit

• Mirage: AI video generator that turns prompts or photos into cinematic clips with multimodal foundation models.

• Speechmatics: Speech-to-text engine built for accuracy across accents, noisy audio, and real-time use.

• Superhuman: AI-powered email client that drafts replies, summarizes threads, and gets you to inbox zero faster.

Rate This Edition