Don’t be fooled—LLMs don’t reason

Former DeepMind AlphaGo researcher Thore Graepel argues that current large language models lack genuinely auditable reasoning because their chain-of-thought outputs remain extensions of next-token prediction. He advocates AI architectures modeled on AlphaGo's combination of neural intuition and explicit search, maintaining an inspectable epistemic state that updates beliefs only with evidence. The argument is particularly relevant to high-stakes AI deployment in medicine, science, engineering, drug discovery and climate research, where explainable error tracing and trustworthy conclusions are required.

Analysis

The investable implication is not a near-term impairment to GOOG’s model quality, but a potential shift in where AI value accrues over the next 6-18 months: from frontier-model training scale toward verifiable workflow infrastructure. If enterprise buyers increasingly require audit trails, persistent state, evidence provenance and tool-level controls, hyperscalers retain compute demand but application-layer vendors with embedded data and regulated workflows—MSFT, ORCL, NOW, RELX and Thomson Reuters—could capture a larger share of AI software economics. This would challenge the market’s implicit assumption that benchmark gains alone justify expanding model-provider multiples.

For GOOG, the relevant risk is that incremental Gemini capex produces impressive consumer capability without proportionate enterprise monetization where trust, liability and integration determine purchasing. The near-term offset is that more agentic, tool-using systems are compute-intensive and favor Google Cloud’s infrastructure revenues; however, this becomes a margin question if inference and verification costs rise faster than pricing. In healthcare and biotech, the thesis is structurally positive but commercially slower: auditable systems may unlock higher-value deployment, yet validation, data rights and regulatory accountability make a 2027+ revenue pool rather than a 1-3 month catalyst.

Consensus may be underestimating the architecture transition while overestimating its immediacy. Current language models can be packaged with retrieval, deterministic tools and human approval to meet many enterprise use cases without a wholesale redesign. The falsifier for the skepticism is evidence that Gemini enterprise adoption, Cloud AI backlog and AI-related operating-margin performance accelerate simultaneously; that would show buyers are paying for current systems despite imperfect interpretability.

Key Decisions for Investors

• No directional GOOG trade solely on this commentary; maintain a watch item into the next earnings cycle for Google Cloud growth, AI backlog disclosure and incremental capex versus consolidated operating-margin guidance. A combination of rising capex and unchanged Cloud monetization would support a 1-3 month underweight.
• Consider a 6-12 month quality basket long MSFT, ORCL and NOW versus equal-weight short a broad high-beta AI software proxy (IGV) only if enterprise IT surveys show governance/audit requirements delaying standalone generative-AI deployments. Target 10-15% relative upside; exit if IGV earnings revisions reaccelerate relative to the basket.
• Accumulate RELX or TRI on market weakness for a 12-24 month regulated-information exposure rather than treating healthcare AI as an immediate biotech trade. Their proprietary legal, scientific and workflow data can become the evidence layer required for defensible AI outputs; key risk is customers choosing hyperscaler-native tooling, visible through slowing organic subscription growth.
• Avoid adding to pure model-capability beneficiaries on valuation expansion alone. Set an alert for enterprise AI contracts that explicitly price verification, provenance or agent-control features; sustained adoption would favor workflow incumbents and cloud platforms over firms whose differentiation is primarily foundation-model benchmarks.