--- title: "Beyond LLMs: What Reasoning Limits Mean for Clinical AI" slug: "beyond-llms-reasoning-limits-clinical-ai" author: "Miguel Amigot" date: "2026-08-17 14:00:00" category: "Premium" topics: "clinical ai, healthcare ai, llm limitations, reasoning architectures, hipaa, ai governance, self-hosted ai, model-agnostic ai" summary: "A widely-shared DeepMind position paper argues LLMs cannot make the abductive leap that produces new scientific theories. It is a narrower claim than the headlines suggest, and it is not the reason clinical AI fails today — but it does explain why a health system should build for model replacement rather than model selection." banner: "" thumbnail: "" linkedin: | A Google DeepMind paper arguing we need to move beyond LLMs went around widely this week. Worth reading what it actually says before repeating it. The paper is "LLMs can't jump," by Tom Zahavy. Its claim is specific: LLMs handle deduction (proving results from established premises) and induction (finding patterns), but not abduction — the leap that invents the premises in the first place. Einstein's seven-year road to General Relativity is the case study. Two clarifications the headlines dropped. It is a position paper by one researcher, and Zahavy has publicly said it is his personal view, not an institutional DeepMind position. And it is about scientific discovery, not everyday task performance. So it is not the reason a clinical decision-support tool gets a drug interaction wrong. That failure mode is retrieval, grounding, and evaluation — not the absence of Einstein. But there is a real conclusion for health systems in it, and it is architectural rather than philosophical. If serious researchers are arguing the dominant architecture has a ceiling, then any model you deploy today is a temporary tenant. The half-life of a model choice is shorter than the life of a clinical workflow. Which makes the durable question not "which model is best for our health system" but "what happens to our deployment when the answer changes." On rented infrastructure, that is a migration. On a model-agnostic platform you own, it is a configuration change. The model matters. Where it runs matters more. #iblai #HealthcareAI #ClinicalAI #HIPAA #AIGovernance --- ## The Short Answer **A widely-shared DeepMind position paper argues LLMs cannot perform abduction — the leap that generates new scientific premises. It is one researcher's personal view, not an institutional position, and it does not explain why clinical AI errs today. Its real lesson for health systems is architectural: build so the model can be replaced. On ibl.ai you own all the code and the data and run it model-agnostic across any LLM.** The gap between what the paper claims and what it is being cited for is itself worth documenting. ## What did the DeepMind position paper actually argue? That LLMs cannot make an abductive leap. The paper is ["LLMs can't jump" by Tom Zahavy of Google DeepMind](https://philsci-archive.pitt.edu/28024/), dated 27 January 2026. Its argument runs through Peirce's three modes of inference. Modern models do induction — pattern recognition across data — and deduction — deriving conclusions from established premises — well. Abduction, the intuitive jump that proposes a novel explanatory hypothesis, is the one it says they are structurally incapable of. The case study is Einstein's seven-year path to General Relativity. The paper's contention is that a model could execute the deductive phase of proving theorems from given premises, but not formulate those premises when observational data is scarce. Two qualifications matter and are usually dropped in the retelling. Zahavy has publicly clarified this is a personal position paper rather than DeepMind's institutional view. And the scope is scientific invention, not general task performance. The honest summary is narrower than "DeepMind says move beyond LLMs," and considerably narrower than "reasoning architectures instead of next-token prediction." ## Does that limitation matter for a hospital deploying AI today? Not directly, and conflating the two produces bad procurement decisions. A clinical decision-support tool that surfaces a wrong drug interaction is not failing because it cannot originate a new physical theory. It is failing at retrieval, grounding, or evaluation — the answer was not checked against a current, authoritative source the health system controls. Those failures are addressable with today's architecture: grounded retrieval against a formulary the institution maintains, guardrails that refuse rather than guess, human review at the point of action, and audit records that let a pharmacist reconstruct what the system saw. That is why the [95% of enterprise AI pilots that MIT's Project NANDA found delivered no measurable P&L impact](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/) traced to data foundations and workflow gaps rather than model quality. So the paper is not a reason to delay clinical AI. It is a reason to doubt anyone who tells you the current architecture is the final one. ## Why does where a clinical model runs matter more than which model it is? Because model choice has a shorter half-life than clinical workflow. If credible researchers are arguing the dominant architecture has a ceiling, then whichever model a health system standardizes on this year is a temporary occupant. Meanwhile the workflow it is embedded in — triage, documentation, prior authorization, care coordination — will outlive several model generations. That asymmetry decides the architecture. A deployment coupled to one vendor's API converts every model change into a migration: re-integration, re-validation, re-negotiation. There is a second reason specific to healthcare. PHI moving to a third party requires a BAA and a trust relationship for every processor in the chain. A model swap on hosted infrastructure means re-running that review; a model swap inside your own perimeter does not move PHI anywhere new.
When the better model ships Hosted, single-vendor Model-agnostic, self-hosted
Switching effort Re-integration project Configuration change
PHI review New processor, new BAA PHI never left the perimeter
Clinical revalidation Required, on vendor's timing Required, on your timing
Cost at 8,000 clinicians Per seat, scales with headcount Usage-based or flat license
## How should a health system choose an AI architecture that survives model change? Optimize for replaceability rather than for this quarter's benchmark leader: 1. **Require model-agnostic routing.** If swapping the underlying LLM is a code change rather than a configuration change, the platform has made a bet on your behalf. 2. **Keep PHI inside the perimeter.** Inference where the data already lives removes the third-party custodian rather than papering over it with an agreement. 3. **Pin and record versions.** Clinical validation means reproducing behaviour; a silently upgraded model makes that impossible. 4. **Log tool calls, not just conversations.** The action creates the obligation, so the tool call is the record that matters at review. 5. **Price it at full clinician count.** Per-seat licensing prices the org chart rather than the work. ## Where ibl.ai fits ibl.ai is the agentic AI platform where **you own all the code and the data**. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with **no per-seat pricing** — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped. For a health system that means PHI never reaches a third-party processor, the audit trail belongs to the institution, and next year's better model is a configuration change rather than a migration. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. *Related: [Self-Hosted AI Agents for Healthcare](/blog/self-hosted-ai-agents-for-healthcare) · [Healthcare AI Reference Architecture](/blog/healthcare-ai-reference-architecture) · [Model-Agnostic AI: The Real Risk Is Vendor Lock-In](/blog/model-agnostic-ai-the-real-risk-is-vendor-lock-in)* ## Why does owning the AI stack matter? **ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.** - **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours. - **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform. - **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ. - **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.