The Logic Upgrade

Stacking Trades — Thu, 13 Nov 2025 22:02:38 +0000

For the past decade, AI progress has focused on scale. This means bigger models, more tokens, and larger GPU clusters. However, in 2025, researchers are moving toward a different kind of progress. They aim for systems that can reason, not just predict.

This shift centers on neuro-symbolic AI, a hybrid approach that combines deep learning with clear reasoning frameworks. Unlike earlier waves of machine learning, this one isn’t driven by the number of parameters. Instead, it relies on structure.

Why Pure Neural Nets Hit a Limit

Neural networks excel at perception and pattern matching, but they struggle with logic, abstraction, and consistency over long chains of thought.

This is a well-known limitation documented across the field:

• The Stanford HAI 2024 AI Index Report found that large language models still underperform on symbolic reasoning tasks compared to specialized systems.
(Source: Stanford HAI 2024 AI Index, Chapter: Technical Performance)

• The Allen Institute for AI reported that LLMs systematically fail on benchmarks requiring multi-step deductive reasoning.
(Source: AI2 Aristo Reasoning Benchmark, 2024)

• Meta AI researchers published work showing that LLMs diverge on tasks requiring strict logical operators or relational consistency.
(Source: Meta AI “Neural Networks and the Limits of Logical Generalization,” 2023)

These findings create what researchers refer to as the reasoning gap. This is a major weakness of purely neural models.

“Pattern recognition built the last decade. Reasoning will build the next.”

What Neuro-Symbolic AI Actually Combines

Neuro-symbolic AI merges two approaches:

1.Neural components → learn from examples and handle perception

2.Symbolic components → apply rules, logic, constraints, and explicit knowledge

This hybrid design addresses exactly where neural nets fail.

The idea is not new, but major institutions have pushed it forward with real, documented progress:

• IBM’s Neuro-Symbolic AI work has shown dramatic improvements in tasks requiring explainability and rule-following.
IBM’s 2022–2024 papers in AAAI and NeurIPS established practical neuro-symbolic architectures for visual question answering.

• MIT CSAIL research on “compositionality” continues to demonstrate that hybrid models generalize better from fewer examples.
Source: MIT CSAIL, “Compositional Abstractions in Neural Models,” 2023–2024.

• Google DeepMind’s AlphaGeometry system (2023) solved thousands of Olympic-level geometry problems using a hybrid neural + symbolic approach.
Source: Nature article, January 2024.

AlphaGeometry is one of the strongest real-world proofs that combining learning with logic can exceed pure neural nets.

Why It Matters Now

Three forces are pushing neuro-symbolic systems into mainstream use:

1. Regulation and Compliance
Finance, healthcare, and government now need explainable AI. Symbolic components provide clear, auditable reasoning chains. Deep nets alone cannot offer this.

2. Efficiency Pressure
As model training becomes much more expensive, hybrid systems can reach similar reasoning ability with significantly less computing power. This matches findings from the Stanford HAI Index, which show that energy use is increasing among frontier models.

3. Reliability
Symbolic systems enforce constraints that reduce hallucination — a documented weakness of LLMs across academic benchmarks.

The Shift From Scale to Structure

The AI field is no longer united behind the idea that bigger is always better.

Across research labs and industry groups, a new consensus is forming:

• Deep learning provides intuition.

• Symbolic reasoning provides structure.

• Together, they form systems that can both learn and understand.

Neuro-symbolic AI represents the logic upgrade — the next layer above pattern recognition.

Trading FLOPs

Stacking Trades — Fri, 07 Nov 2025 22:45:38 +0000

The Invisible Commodity

Training the latest generative-AI models costs billions, not millions. By mid-2025, leading labs reported that compute budgets had surged, driven by rising demand for high-performance GPUs and data-center scale-up. For example, CoreWeave signed a five-year contract worth US $11.9 billion with OpenAI in March 2025.

In September, CoreWeave strengthened its partnership with OpenAI for up to $6.5 billion. This brings their total deal to about $22.4 billion.

In essence: compute has become a bottleneck, and bottlenecks trade like commodities.

From Cloud to Market

What started off as dealing with cloud compute as a cost center is now turning into a way of looking at it as a tradable asset. Nvidia announced the new software platform Lepton, which will create a marketplace for cloud-based AI chip capacity, thus marking the formalization of compute-markets.

Meanwhile, the global “GPU as a Service” market is projected to grow from US $6.54 billion in 2024 to US $8.21 billion in 2025, and further to US $26.62 billion by 2030 (CAGR ~26.5%).

Running a training job is no longer just a tech operation—it’s a market signal.

“We used to buy capacity; now we’re trading it.”

Liquidity of Intelligence

Investors aren’t just backing software anymore; they’re backing compute infrastructure. Firms treat GPU clusters, training pipelines, and model-hosting farms as utility assets.

The global data-centre GPU market is predicted to grow from about US$119.97 billion in 2025 to US$228.04 billion by 2030.

In other words, compute throughput is becoming the next lever of market advantage.

The Speculation Layer

Compute markets are still thinly regulated but rapidly gaining complexity. For example, Nvidia’s deal to invest up to US $100 billion in OpenAI (and supply chips) shows how compute commitments now resemble infrastructure finance.

The question becomes, if you’re buying access to AI compute, what asset class do you own?

It sits at the intersection of technology, finance, and commodity markets.

The Human Parallel

Commodity trading desks once monitored oil tankers and futures curves. Today some firms are monitoring compute pipelines and GPU lease contracts.

The analogies are real: just as power plants once provided scale for industrial growth, model-labs now seek scale via compute liquidity.

Value is shifting upstream: from software features to underlying infrastructure.

What It Signals

If compute becomes a core asset class:

• Infrastructure providers (not just hardware vendors) gain strategic importance: marketplaces, leasing platforms, and secondary GPU markets become key nodes.

• Software companies face pressure: increasingly the differentiator is access to compute + data, not just features.

• Investors need new metrics: compute utilisation, petaflops per dollar, latency arbitrage become valuable indicators.

• Markets may shift: competition becomes about compute liquidity as much as talent or data.

AI Model – Stacking Trades