Technology thesis · Artificial Intelligence
high conviction matureDeep learning
Deep learning is consolidating on the transformer while its growth axis shifts from pre-training scale to test-time compute, moving the capital from training clusters to inference silicon.
Position maintained continuously · last reviewed Jun 24, 2026
The thesis
Core thesis
Every major AI capability — LLMs, computer vision, drug discovery, protein folding — runs on deep learning. The transformer architecture (2017) dominates. But state-space models (Mamba), mixture-of-experts, and neuro-symbolic approaches are gaining ground. The 'bitter lesson' holds: scale and compute beat hand-engineering. The question is whether we're approaching diminishing returns on the current paradigm.
State of the art (2026)
The transformer still anchors every frontier system in mid-2026 - Anthropic's Claude Opus 4.8, OpenAI's GPT-5.5, Google's Gemini 3.1 Pro and xAI's Grok 4 - but the scaling story has moved. Pre-training returns on dense models have flattened, so the capability lever is now test-time compute: chain-of-thought reasoning, search and verification at inference, the axis DeepSeek's R1 opened on the cheap. Architecture is hedged rather than replaced - Mamba-style state-space and mixture-of-experts hybrids run in production for long-context and efficiency, yet none has matched transformer quality at frontier scale. The economic centre of gravity is shifting from training clusters toward inference, custom silicon and the compute glut that ramping Blackwell and Rubin may create.
Everything below is live inside CanaryIQ
The full analysis behind the verdict — the structure is real; the content unlocks when you log in.
Signal stack
Evidence stacked leading → lagging
Technology-native KPIs
Metrics that predict trajectory, tracked over time
Landscape map
Who builds what — and who depends on whom
Catalyst calendar
Dated events that will move the position
Technology roadmap
Milestones on the path to maturity
Watchlists
Companies, people and papers — each with a remove-by condition
Decision frameworks
The same call, framed for your desk
Thesis changelog
When our view changed, and why
Change our mind
2 disconfirming conditions
The rest is inside
You've read the verdict. The file is much deeper.
The full signal stack, technology-native KPIs tracked over time, the landscape of who depends on whom, the dated catalyst calendar, decision frameworks for every desk, live watchlists and the changelog of every time our call on Deep learning has changed — all live inside CanaryIQ.