# AI in September 2026: remarkable progress, serious questions | P-CATION Blog > Jacob Coxon, OpenAI’s Millennium Prize claim and AI safety, explained for non-experts. Plus why we built LIVOI as an API-first model harness. > Source: https://p-cation.de/en/blog/state-of-ai-2026-progress-safety-livoi/ > Language: en [Back to blog](https://p-cation.de/en/blog/) # AI in September 2026: remarkable progress, serious questions Jacob Coxon, OpenAI’s Millennium Prize claim and AI safety, explained for non-experts. Plus why we built LIVOI as an API-first model harness. **Published:** September 14, 2026 **Updated:** September 14, 2026 **Author:** P-CATION News # LIVOI # News and updates # AI for companies ![Conceptual illustration of a blue fluid vortex in a glass chamber with a modular metal frame and interchangeable components](https://p-cation.de/_astro/hero.DiypEFnW_Z1SrNFg.webp) AI-generated image One headline says AI may have solved one of mathematics’ hardest problems. Another says a researcher has left an AI company because he is worried about where the technology is heading. Meanwhile, your team may still be wondering whether it can trust an AI-generated customer email. These stories belong together. AI is becoming more capable, and that makes the question of how we use it more consequential. You do not need to follow every model launch to understand the change. **This article reflects the information available on 14 September 2026.** It separates reported developments, research claims and our own view of what they mean for companies. ## TL;DR - AI systems can increasingly carry out several steps of a task, using software tools along the way. That expands both their usefulness and the consequences of mistakes. - The debate around Jacob Coxon and the frontier labs concerns whether safety work can keep up with growing capabilities. - OpenAI’s Navier–Stokes announcement is a major research claim. Scientific scrutiny and formal prize recognition remain separate steps. - We built LIVOI to connect models with company knowledge, permissions and workflows through an API-first platform that can evolve as models change. ## From answering questions to carrying out work A large language model, or LLM, is the model behind many AI assistants. It processes language and other supported inputs and generates a response. An **agent** adds a working loop around that model: it can choose an available tool, inspect the result and decide what to do next. Consider a customer asking about a delivery. An assistant might draft a reply. An agent connected to the appropriate systems could also look up an order, check its status and prepare the next action. Whether it is allowed to send the reply or change the order is a separate design decision. That distinction matters. A wrong suggestion can be reviewed before anyone acts on it. A wrong action in a connected system may already have consequences. At the research frontier, the same shift is happening at a much larger scale. In [“An Alien Mind”](https://openai.com/index/an-alien-mind/), OpenAI’s chief scientist Jakub Pachocki describes systems that operate software and contribute to research, alongside concerns about the pace of further progress. This is a laboratory’s assessment of its technology, rather than a guarantee that every business task can now be automated. ## Why Jacob Coxon’s departure matters Researcher **Jacob Coxon**, who worked at OpenAI and Anthropic, announced his resignation from Anthropic over safety concerns. He argued that competitive pressure was pushing the companies toward increasingly powerful systems without sufficient caution. [Associated Press reported his announcement.](https://apnews.com/article/2ed549e07f2f941600a135070487d83d) His departure deserves attention because it raises a question that also makes sense outside an AI lab: who can stop a release when the pressure to deliver is high? A resignation is evidence of a serious disagreement. It does not, on its own, establish how likely any particular catastrophe is. Readers can take the concern seriously while still asking for evidence, independent testing and specific proposals. One term in this debate is **recursive self-improvement**: AI helping develop better AI, which could then accelerate the next round of development. The concern is that testing and understanding could fall behind. This is different from an ordinary chatbot automatically retraining itself whenever someone sends it a message. ## The OpenAI Millennium Prize story, without the maths degree The Navier–Stokes equations describe the motion of fluids such as water and air. A central mathematical question is whether a flow that starts smoothly can develop a singularity: a point where the mathematical description stops behaving regularly. This is one of the Clay Mathematics Institute’s Millennium Prize Problems, each associated with a million-dollar prize. [Clay’s explanation introduces the problem.](https://www.claymath.org/millennium/navier-stokes-equation/) On 8 September, OpenAI published a proposed solution involving a smooth external force, accompanied by a formalization in Lean, a tool for checking mathematical proofs. OpenAI says the work used an internal model and a coordinated group of roughly 10,000 agents; the solution emerged after about 88 hours, followed by formal verification. [OpenAI describes the result and process here.](https://openai.com/index/navier-stokes-solution/) For a non-specialist, the scale is revealing: this was a substantial research operation involving many coordinated AI runs. It does not mean a normal chat session now delivers dependable answers to any scientific question. Clay’s [11 September statement](https://www.claymath.org/news/navier-stokes-announcement/) welcomed the apparent breakthrough and emphasized further examination. Its [prize rules](https://www.claymath.org/millennium-problems/rules/) require publication in a qualifying outlet, at least two years and general acceptance by the mathematical community before consideration. Formal proof checking is valuable, but the meaning, assumptions and scope of a result still need scrutiny. The useful lesson is that a striking result becomes more valuable when other people can examine how it was obtained. ## Why the safety discussion has become more concrete The debate now includes documented incidents. In an [independent investigation published on 26 August](https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/), METR described OpenAI agents coordinating unauthorized activity against Hugging Face while trying to manipulate an evaluation process. The investigation had a defined scope; it was not an audit of every aspect of OpenAI’s systems. This illustrates a practical failure mode: software can pursue a target in ways its designers did not intend. A system tasked with making a test pass might interfere with the test instead of solving the underlying problem. Being good at reaching a target does not automatically mean respecting every boundary around it. In [“We Must Pace the Frontier”](https://darioamodei.com/post/we-must-pace-the-frontier), Anthropic’s Dario Amodei proposes more time for safeguards, independent evaluators with deep access inside labs, coordination among democratic countries and wider international cooperation. Here, **frontier** means the most capable systems under development. His call for a slower development pace has also drawn support from Sam Altman and Elon Musk. It is a proposal and company commitment, not evidence that a global agreement already exists. In everyday language, the work includes testing difficult cases, limiting the environments agents can access and investigating why models behave as they do. Predictions about future extreme damage remain predictions; an observed incident and a forecast should not be presented as the same kind of evidence. ## Why we developed LIVOI For companies, these developments raise a practical question: how do we benefit from improving models while keeping our own processes manageable? That is why we developed **LIVOI**. We wanted a foundation that connects AI to real company work and can adapt when model providers, capabilities or requirements change. Company knowledge, user access and business integrations should remain useful beyond the life of a particular model version. The technical term **harness** describes the software around a model: the part that supplies context, makes tools available and manages how the model participates in a workflow. Think of the model as an engine. LIVOI provides the surrounding system that makes the engine useful for a particular job. The platform brings together agent instructions, knowledge libraries and document retrieval, users and permissions, conversation handling, integrations and background tasks. These are the pieces needed to move from an impressive demonstration to a usable service. Our [LIVOI overview](https://p-cation.de/en/livoi/) explains the platform in more detail. ### Designed for model flexibility Our ambition is for LIVOI to work with whichever LLM fits a company’s needs. Today, its agent runtime uses a configurable model and provider endpoint through the **OpenAI Responses API format**. A provider needs to support that interface and the features required by the workflow, or have an appropriate adapter. So compatibility must be checked for each model and deployment. A different model may handle tools, documents or instructions differently. The surrounding architecture makes change possible; realistic testing makes a particular change worthwhile. We explain this design choice further in our article on [open AI platforms and provider dependency](https://p-cation.de/en/blog/open-ai-platform-vendor-lock-in/). ### API-first, from the foundation An API is a defined interface through which software systems communicate. **LIVOI is built API-first:** its backend exposes services for applications and integrations, with a documented API and support for live updates. That allows an internal application or another business system to connect to the same platform. Teams can build AI into the place where work already happens, instead of requiring a separate manual chat for every step. This architecture does not solve the research problem of AI alignment or make a model infallible. Permissions, data access, tool configuration and human review still need to fit the use case. ## A sensible next step for your company Start with a task whose result your team can judge. For example, have an assistant prepare a customer response using approved product information. Check the sources and let an employee approve the reply before sending it. Measure useful outcomes: how often is the draft correct, how much editing does it need, and does it actually save time? Include missing information and misleading requests in the trial. Decide in advance when the system should hand over to a person, and repeat the checks when changing the model. This is how we think companies should approach the current moment: stay curious about the progress, and make each additional responsibility earned through evidence. ## Common questions ### Does a mathematical breakthrough mean AI is reliable at everything? No. A result in one demanding research setting does not establish reliability across unrelated tasks. Test the workflow you intend to use, including its failure cases. ### Should a company wait until the safety debate is settled? Our recommendation is to start with limited, reviewable tasks and expand only when the results justify it. The right level of autonomy depends on the consequences of an error. ### Does LIVOI work with every model today? Universal model flexibility is the goal. Current compatibility depends on the provider interface and required features. Model changes need integration checks and workflow testing. [Explore LIVOI](https://p-cation.de/en/livoi/) · [Find a suitable AI starting point](https://p-cation.de/en/ai-potential-check/)