A mathematical proof can outlive the model that found it. Useful software can do the same. October suggested that this may become a central question for AI: what remains after the model stops generating?
Our September brief argued that institutions were struggling to absorb rising capability. October suggested a more specific response.
Intelligence becomes more useful when it leaves something dependable behind: a proof others can inspect, software they can test, a process they can repeat without renegotiating every decision.
That is my reading of the month. The agent attracts attention. What survives its intervention may create the lasting value.
Mathematics acquires a production problem
On 6 October, OpenAI released hundreds of mathematical manuscripts from an internal model, at different stages of verification. The following day, three were withdrawn after a sign error undermined one argument and two dependent papers. Fourteen other manuscripts were revised. Extraordinary output arrived with an equally recognisable scientific requirement: correction.
Then came OpenAI’s 26 October announcement: a claimed proof of local smoothing in every spatial dimension of at least two, with consequences for the Kakeya and Stein restriction conjectures; a claimed refutation of the Exponential Time Hypothesis (ETH); and advances concerning the mathematical existence of quantum field theories.
These are substantial claims about waves, the difficulty of computation and the foundations of particle physics. OpenAI published Lean formalisations for the mathematics and complexity results. An independent group of computer scientists will examine and explain the ETH result. The physics group chose to defer publication of the AI outputs while preparing its own account.
The distinctions matter. Refuting the Exponential Time Hypothesis would not establish P = NP. Establishing Yang–Mills existence would not, by itself, settle the Millennium Prize problem, which also requires a mass gap. The underlying physics proofs were generated in September; the October formalisation concerned Yang–Mills existence.
Lean checks a precise mathematical statement under specified assumptions. Researchers must still establish that the statement captures the intended problem, understand the argument and determine what follows from it.
That creates a different economics of discovery. A result can emerge faster than a community can turn it into shared understanding. The valuable output includes the explanations, reusable methods and corrected dependencies that let the next researcher build on it.
On 15 October, the Simons Foundation announced a partnership with Anthropic and OpenAI, committing $15 million in combined funding and resources to help mathematicians assess results, use AI tools and develop shared standards. Understanding was becoming something the AI transition needed to finance explicitly.
Other sciences made the distinction concrete. Anthropic reported improved designs for eight quantum circuit tasks by putting Claude inside a loop with an external correctness evaluator. In its project on scorpion toxins, laboratories tested 211 protein designs; 17 met binding criteria. Whether those binders neutralise scorpion venom in cells remained unconfirmed.
A proof checker, a software test and a laboratory experiment offer different kinds of evidence. None should disappear behind the convenient phrase “AI discovery”.
Give improvisation a defined role
OpenAI’s Dots and Meta’s Muse had both arrived in September.
October expanded the category. Google announced its Gemini agent on 8 October; Meta introduced Muse Wonder models on 12 October; Anthropic opened Claude Duet’s research preview on 21 October. Memory and sustained work were becoming central product promises.
The attraction is easy to understand. An assistant that remembers the project and continues after the laptop closes could remove much of the coordination surrounding useful work.
But continuous availability does not require continuous improvisation.
On 19 October, Meta described its Thin Orchestrator: roughly 1,300 lines routing more than 20 workflows to 14 specialist agents. Fixed rules controlled sequencing and reliability checks; generative reasoning happened inside the specialists. Meta reported that configuration preparation fell from three or four manual days to under an hour. About two-thirds of changes landed without significant human edits. These results come from Meta’s own account of a particular workflow.
The lesson is architectural. Use the model where interpretation or exploration helps. Put settled requirements into software that enforces them.
A model might interpret an invoice. Code can enforce agreed rules for required fields, duplicate payments and spending limits. The model might help write that code, but the resulting process need not ask a language model to reconsider those rules every time.
This is one route from generative AI to dependable operations. It does not make the model deterministic, and deterministic code can still be wrong. It makes particular decisions explicit enough to inspect, test and change deliberately.
Google’s announcement also separated the persistent agent from the underlying model, including routing work to Claude. That develops the problem explored in Your AI Model Has an Expiry Date: useful organisational knowledge should survive a model replacement.
The harder question is whether it can also survive a change of provider.
The boundary becomes the target
October’s security news showed why architecture alone doesn't earn a presumption of safety.
On 21 October, PHP released fixes for two critical vulnerabilities that could allow unauthenticated remote code execution under relevant conditions. The PHP Foundation’s account is unusually revealing. Researchers identified signs of one flaw in 2024. Compatibility concerns, limited specialist capacity and stalled decisions delayed repair. Later research, including work assisted by AI, established practical exploitation routes.
Finding the weakness and making the ecosystem safe were separate efforts.
By late October, the risk had become demonstrated harm. Dream reported 25 compromised PAN-OS firewall and management servers after attackers chained PHP2Shell with a kernel vulnerability; Palo Alto Networks independently verified the results. Dream found AI tools in the attacker’s infrastructure, but no direct evidence that they wrote the exploits.
Other incidents exposed more ordinary weaknesses. LiteLLM’s compromised 21 October release followed an exposed administrative key, an agent platform and an excessively privileged GitHub token. Its corrected investigation traced the stolen GitHub token to the Agent Platform API, correcting earlier speculation about Claude Code logs.
Forever Security reported stopping an agent from installing malware after it followed an AGENTS.md file planted in a customer’s Downloads folder. Anthropic separately disclosed unintended actions, including a fabricated police tip that was flagged as spam and never forwarded for investigation. These were distinct incidents: malicious instructions in one case, unintended model behaviour in another.
An agent can misunderstand authority even when it understands language. A release pipeline can execute perfectly while obeying a stolen identity.
The implication reaches beyond better prompts. Systems need enforceable limits on credentials and actions, reliable provenance for instructions, and maintained software beneath the model.
For companies benefiting from open-source infrastructure, funding that maintenance belongs in the AI investment case. Buying faster discovery while leaving repair underfunded is a fragile strategy.
Costs may shift. If generation accelerates faster than verification, universities, software maintainers and customers can inherit the bill for checking an expanding supply of outputs. The Simons initiative points towards one response. More broadly, an AI business case should identify who pays to establish that the result is worth using.
What the organisation keeps
This is where October’s Building Creative Machines articles connect.
The Gartner comparison traced the growing burden of demonstrating useful performance. The article on model expiry showed how upgrades repeatedly reopen that burden. Together, they suggest that a company should accumulate evidence and operational knowledge with each deployment, even as its suppliers change.
That means retaining representative tests, documented rules, failure histories and usable records of why a process works. Otherwise, each new model can become another expensive exercise in rediscovery.
A useful measure would include the cost of accepting, maintaining and revalidating an outcome, alongside the cost of generating it. An impressive demonstration can produce an answer once. An investment should leave the organisation better equipped for the next thousand cases.
For Europe, the same argument applies at a larger scale. Our coverage of Portugal’s AI infrastructure and Paul Krugman’s geography of power raised questions about where capacity and economic value accumulate.
Owning more compute helps. So does the ability to preserve knowledge, replace suppliers, inspect critical software and fund the people maintaining it. Sovereignty has an operational meaning: can an institution keep doing its work when a platform changes its terms?
This question has a creative version too.
My experiment with Esteves and the BCM video series asks how an agent can help translate ideas into another medium while preserving editorial judgement. The aim remains an explanation somebody understands. Producing more material is only one part of that work.
Ricardo Chaves’s book, discussed this month, offered a related reminder: understanding the principles gives people something that survives the product cycle. Obama’s Lisbon intervention placed the harder question alongside it: what progress should we want?
No formal test can settle every question of meaning, fairness or taste. Those limits should shape the division of work.
For a recurring task, a mature agent should sometimes help build a process that needs less of its discretion next time. For an unfamiliar problem, it should leave evidence that makes the next investigation easier.
That changes what we ought to measure.
Beyond how much work an agent completed, ask what the organisation can now understand, repeat or improve without starting again.
The most valuable intelligence may be the intelligence that leaves us less dependent on calling it back.