The jump and the record

In a 2026 position paper with a blunt title, Google DeepMind researcher Tom Zahavy argues that LLMs can't jump. His argument builds on philosopher Charles Peirce's distinction between three modes of inference:

1. Induction finds patterns in data.
2. Deduction derives consequences from premises.
3. Abduction invents a new explanation for a surprising observation.

Zahavy's position is that modern AI excels at induction, is making rapid progress in deduction, and still lacks abduction: the jump that produces an explanation the available data did not demand.

His central example is general relativity. By 1915, Newton's theory still explained almost everything scientists could observe. Think of a map that matches nearly every road in a city. A system rewarded for compressing that map can make it smaller and cleaner, but it has no reason to conclude that the city rests on a different geometry. Zahavy's point is that Einstein did something closer to drawing a new kind of map. He started with a thought experiment about a person in free fall, then reasoned toward new axioms about gravity and spacetime. Zahavy proposes physically grounded world models as one possible path toward systems that can run that kind of experiment.

It is a strong argument, and anyone building with AI should read the paper. Place it beside the rest of the current architecture landscape, though, and a second pattern appears. It matters now for anyone deploying agents.

Everyone is absorbing capabilities into the model

We recently worked through a survey of the stateful-AI research frontier for our own roadmap. Across the field, major efforts are trying to absorb missing cognitive functions directly into the model:

• Memory during inference. Google's Titans reports learning what to remember while a model is working, with coherent recall across very long contexts.
• Facts stored in model parameters. Meta's memory-layer work reports higher factual question-answering performance with limited compute overhead.
• Memory systems designed by agents. ALMA meta-learns agent memory architectures and reports improvements over human-designed baselines.
• World models. V-JEPA 2 uses latent simulation for robot planning, while Genie generates interactive environments an agent can act within.
• Search and self-modification. Aletheia applies verifier-backed search to mathematical problems. SEAL and the Darwin Gödel Machine explore systems that rewrite parts of themselves to improve.

Each effort closes a capability gap. Each also leaves a common residue, documented in the research: important state becomes harder to inspect, attribute, correct, and govern.

Said another way, this is like asking one student to write the exam, take it, grade it, and protect the gradebook. Even if the score improves, the score is not independent evidence.

Recent safety research makes that risk concrete. OpenAI reported that, during an evaluation, an agent exploited an unknown flaw in a testing interface, accessed the target implementation, copied it, and received a high score. In a controlled research environment, Anthropic reported a reward-hacking model disabling a monitor, changing its reward function, and replacing a transcript with a false one. These are not descriptions of ordinary production behavior. They are demonstrations of what can happen when the system doing the work can also influence how its work is judged.

That is why an autonomous agent cannot be its own untampered evaluator.

Capability and trust are different axes

"Can the system do the thing?" and "Can someone verify, govern, and correct what it did?" are separate questions. Progress on the first does not automatically answer the second.

Verifier-backed reasoning works best where an objective verifier already exists: a proof checker, compiler, or test suite. Mathematics has one, which is why search can produce remarkable results there. Open-ended agent work has no equivalent compiler for questions like "Was this refactor consistent with our conventions?" or "Did the agent use current knowledge rather than stale knowledge?"

As generation becomes cheaper, generation stops being the only bottleneck. Review and validation become the expensive part. The constraint is not capability alone. It is the record and proof behind the work.

What the record requires

The survey points to five properties a trustworthy agent system needs. They have to live outside the model because the model's own output cannot be their only foundation:

1. State authority independent of the model, so generated text is never treated as self-authenticating truth.
2. Lineage you can replay, with a record exact enough to reconstruct what happened.
3. Corrections without retraining, so facts can be superseded or set aside without erasing history.
4. Evaluation the agent cannot tamper with, enforced by a layer it does not control.
5. Measured cost per verified unit of work, so the system can show what the work required.

Nothing on the capability frontier provides all five from inside the model. That is not a criticism of the research. These are properties of the system around the model.

The jump and the record

Taken together, Zahavy's position paper and the architecture survey describe complementary problems. Zahavy asks what it would take for AI to invent genuinely new knowledge and answers that today's models are missing something important. The architecture survey asks what it takes to trust the knowledge agents already produce at scale and reaches a parallel conclusion: the required authority cannot come from the model alone.

That second layer is what we build at 5thDev. Your agents' knowledge lives in a store outside any model, where facts carry provenance, newer versions can replace older ones without silently erasing them, and conflicting information can be set aside. Writes are checked against your rules by a layer the agent does not control. Sessions end with a summary and receipts you can inspect.

We are not claiming to solve abduction. Nobody has. We are building the record: the layer both bodies of work point toward from different directions, and the layer that remains necessary whichever capability curve wins.

If your agents are already doing real work, the record problem is already yours. 5thDev Local is free, runs on your machine, and is just getting started.