Enterprise AI
Nine Loops or Loop One? The Test Every CIO Must Run
Nine Loops or Loop One? The Test Every CIO Must Run
Dr. Anoj Winston Gladius
·
15
Min. Lesezeit
Depth over surface: where enterprise AI creates real value.
aufsatz
Image: Generated by neuland.ai HUB
On 7 August 2026, the physicist and science writer Matt von Hippel issued a public challenge to the AI industry: solve an outstanding problem in particle physics, on the kind of budget an academic research group actually has. On 25 September, Anthropic announced that two of its physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, had done it. They gave Claude a single sentence describing the problem, told it to keep working overnight, and let it run largely unsupervised for days. It reached the answer by two independent methods, for a reported compute bill of a few thousand dollars. Lance Dixon, the physicist whose team set the previous record in 2023, spent about two weeks checking the result and confirmed it. Then a second detail emerged: a team at the Chinese Academy of Sciences had independently reached the same frontier using OpenAI's GPT-6. Eleven days later, Mistral announced a trillion-parameter model trained in its own European data centres. Read together, these are not three unrelated stories. Two laboratories on two continents, using models from two different vendors, crossed the same scientific frontier within weeks of each other, and Europe now builds frontier-scale models of its own. Capable engines have become abundant. And yet most European enterprises that bought "AI transformation" this year are still running problems that are one step deep. The question every CIO should now ask is not which model to buy. It is why their AI programme cannot count past one.
This is the twenty-seventh piece in a series I have been writing for neuland.ai. Each one has pushed the same underlying argument forward: the value, the risk and the moat in enterprise AI are not in the model. They are in the layer above and around it. [¹] This piece uses a result from theoretical physics to make that argument as concretely as I have been able to make it, so it is worth taking the physics slowly.
The model can count to nine. Most enterprise AI is built to count to one.
What a loop is, and why nine is a wall
When two particles collide, physics cannot tell you exactly what will come out. It can tell you how likely each outcome is. The formula that gives you those odds is called a scattering amplitude, and computing it is one of the central jobs of theoretical particle physics.
Physicists calculate amplitudes by drawing the collision as diagrams. The simplest diagrams are almost trivial: particles come in, exchange something, go out. But quantum mechanics says that in between, other particles can briefly appear and disappear, and those fleeting detours show up in the drawing as closed circles. Each closed circle is a loop. Every loop you include makes the prediction more precise, and makes the calculation dramatically harder, because the number of possible detours explodes.
How dramatically is the part a CIO should hold on to. Written out in full on one part of the problem, the eight-loop answer runs to about 1.67 billion terms. At nine loops it is about 30 billion. One more loop made the problem roughly eighteen times bigger. [²] For a sense of scale: the most precise prediction in all of particle physics, the magnetic moment of the electron, is computed to five loops in a different, real-world theory.
The theory in question here, planar N=4 super-Yang-Mills, is a simplified sibling of the theories that describe real particles, a test track rather than a public road. It shares their structure but is clean enough to calculate. That is why physicists use it to push techniques to their limits, and why the eight-loop record had stood for three years.
What Claude actually did, and what it did not
Von Hippel offered two problems. Fitzpatrick and Mishra-Sharma first asked Claude which one it was most confident it could solve, then gave it a one-sentence description of the six-particle amplitude at nine loops. The model, Fable 5.1, running inside Anthropic's research environment, Claude Science, worked through the calculation over several days, reporting progress every four to six hours, and arrived at the result by two separate routes that agreed with each other. [³] Von Hippel's own account of it, published by Anthropic, carried the title "Yes, Claude Can Count Nine Rounds".
It is equally important to be precise about what did not happen. Claude did not invent new physics. By von Hippel's account, it applied methods that physicists had developed over many years, with somewhat more computation than anyone had previously attempted. [⁴] This was a record broken by endurance, rigour and cost, not by a new idea. And it was only accepted once the world's leading expert on the previous record had spent two weeks checking it.
That restraint is exactly why the result matters. It is not a benchmark. It is an open problem, solved, verified by the person best placed to say it was wrong, for a reported compute cost roughly comparable to a few days of a senior consultant's time.
The detail that should worry every model vendor
Here is the part that I have not seen written down clearly enough in the coverage. A team led by He Song at the Institute of Theoretical Physics of the Chinese Academy of Sciences independently reached the nine-loop stage with the help of OpenAI's GPT-6, and has obtained its main results. [⁵]
Two teams, on two continents, with models from two competing vendors, crossed the same frontier within weeks of each other. When the same hard result can be reached with different engines, the engine is no longer the scarce thing. An earlier piece in this series argued, using the Navier-Stokes result, that the systems built around models were what made the science work. [⁶] Nine loops adds the piece of evidence that argument was missing: reproducibility across vendors. The model has become a component in the most literal sense. You can swap it and still get the answer.
Enterprise problems have loops too
This is where the physics stops being a story about physics.
Think of an enterprise question as having loops of its own. Each additional system, relationship or rule that the answer depends on is another loop. What does our standard contract say about liability? is loop one: one question, one document set, one answer. Which of our suppliers operate under a contract that falls within DORA, and which of those had a reportable incident last year? is four loops: supplier, contract, regulatory scope, incident record, each living in a different system with different permissions. Are the instrument tags in this plant's piping and instrumentation diagrams consistent with the equipment list, the cause-and-effect matrix and the last three revisions? runs deeper still.
The analogy has limits, and they are worth stating. Physics loops come with a clean mathematical structure and established methods; enterprise data is messy, sources disagree, and every hop carries its own access rules. So enterprise loops are easier in one way, far fewer terms, and harder in another. But the two properties that matter carry over exactly. The space of possibilities grows explosively with every loop. And an answer at one depth tells you very little about whether the system can produce a correct answer one step deeper.
That is where enterprise value actually sits. Nobody's operating margin changes because staff can summarise a document faster. It changes when a question that used to take a team three weeks of cross-referencing across five systems is answered, with sources, in an afternoon. That is loop four and beyond.
Why the wrapper cannot count past one
Now look at what most European enterprises bought this year under the banner of AI adoption and transformation: a chat interface, connected to a document store and a frontier model, rolled out to several thousand employees.
That is a loop-one machine. It takes one question, retrieves a handful of passages, produces one answer, in one sitting. It holds no durable state, so it cannot work on a problem for days. It has no independent second method, so it cannot check itself. Its only record of what it did is the conversation itself. Put the model that counted to nine inside it, and it still counts to one, not because the model is weak, but because the system around it was never built to go deeper.
So here is the controversial version, and I mean it: most of what Europe bought this year as AI transformation is loop-one software sold at a loop-nine price. It explains a number that has puzzled boards. McKinsey's State of AI survey, published in August, found that 80 percent of organisations report productivity gains from AI and only 37 percent report any EBIT impact, unchanged from the previous year. [⁷] Faster loop-one work makes people feel more productive. It does not move earnings, because earnings live deeper in the stack.
Where Mistral fits
On 6 October, Mistral opened a public preview of Mistral Large 4: a one-trillion-parameter model, 49 billion parameters active per token, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centres, with open weights promised by the end of October and a deployment Mistral will operate end to end under European law. [⁸] That is the most encouraging development for European AI this year. For the first time, a frontier-scale engine has been built on European soil by a European company.
It also matters for a reason the nine-loop story makes uncomfortable. The model that counted to nine is the successor to Fable 5, which every customer in the world lost access to for nearly three weeks in June after a United States export-control directive. [⁹] A European engine is real insurance for continuity, and European enterprises should test it seriously once the weights and the licence are published.
But the lesson of nine loops applies to it exactly as it applies to Claude and GPT-6. Engines are now abundant: American, Chinese and European. A better engine in a loop-one machine still never reaches loop nine. And swapping an American vertical stack for a European one improves jurisdiction, not depth.
The nine-loop test
Von Hippel's challenge worked because it was concrete. Enterprise AI deserves the same discipline, so here is a version a CIO can put to any platform, including ours. I would encourage every CIO reading this to run it before signing the next renewal.
Can it work on one problem for days? Not a long conversation, but a single piece of work that continues, reports progress and survives interruption, without someone keeping a browser tab open.
Can it check itself two independent ways? And does it tell you when the two routes disagree, rather than quietly choosing one?
Can an expert audit it afterwards? Every source retrieved, every tool called, every intermediate result, from a record the system wrote at the time, not from a summary the model wrote about itself.
Can you swap the engine and get the same answer? If the result depends on one vendor's model, you have bought a dependency, not a capability.
A chat window fails the first question before the model is even asked.
Where neuland.ai stands on this
The neuland.ai HUB is a sovereign AI management and orchestration platform, and it was built for work several loops deep. Its agentic threads, deep agents and multi-hop reasoning are designed for long-running work across several systems, decomposed into steps that can run in parallel and be compared before anything is committed. The company's own ontology grounds each hop, so that a supplier in one system and the same supplier in another are recognised as one entity. Entitlement is enforced on retrieval before the model reasons, at every hop, for the person the work is being done for. Every retrieval, tool call and intermediate result is written to an append-only record by the platform, so that an expert can audit the path afterwards. And because routing is model-agnostic, with open-weight models we host, including Mistral Large 3, GLM, Qwen and DeepSeek, alongside Claude and GPT where the customer's policy allows them, the engine can be swapped without rebuilding the work. Mistral Large 4 will enter the same evaluation every model goes through once its weights and licence are published. [¹⁰]
This is also the answer to a question I hear from CIOs more and more often: we have rolled out an AI assistant to five thousand people, isn't that adoption? It is adoption. It is not transformation. Adoption measures how many people use loop one. Transformation is whether anything in the company now runs at loop four.
Personal take
I want to be careful with the framing here. This piece is not against AI assistants. It is not a criticism of Anthropic or OpenAI either. The nine-loop result was presented with unusual modesty, with established methods, more computation and independent verification, and that modesty is what makes it credible. And it is not lukewarm on Mistral, whose model is exactly what European AI needed.
What I am pushing back against is the habit of measuring transformation by seat count, and the market price that habit has created for tools that can only count to one. The physics this autumn showed what these models can do when the system around them is built for depth: work for days, check itself, survive an expert's audit, at a reported compute cost comparable to a few consulting days. That standard is now public. Enterprises should hold their own platforms to it.
The regulatory direction points the same way. The EU AI Act has applied broadly since 2 August 2026, and its expectations on oversight, logging and traceability attach to the deployed system; the Digital Omnibus has moved the high-risk Annex III obligations to 2 December 2027 and Annex I obligations to 2 August 2028. [¹¹] Deep work without a record cannot be evidenced. Deep work with one can.
Series articles available at neuland.ai/en/resources/insights.
Term counts as reported in "The Truth Behind Claude's $2,000 Physics Breakthrough", HackerNoon, 7 October 2026: approximately 1.67 billion terms for one part of the eight-loop answer and approximately 30 billion at nine loops. On the electron's anomalous magnetic moment being computed to five loops, see coverage of the result compiled by KuCoin News, September 2026.
Matt von Hippel, public challenge issued 7 August 2026, offering a choice between a seven-loop calculation in N=8 supergravity and the nine-loop six-particle amplitude in planar N=4 super-Yang-Mills theory on an academic-scale budget. Anthropic announcement, 25 September 2026: computation of the nine-loop six-particle scattering amplitude in planar N=4 super-Yang-Mills theory by Claude Fable 5.1 in Claude Science, directed by Liam Fitzpatrick and Siddharth Mishra-Sharma, with progress reported every four to six hours and the result obtained by two independent methods; previous eight-loop record set in 2023. Reported cost estimates range from roughly 1,000 to 2,000 US dollars in total to that range per method (Dataconomy, 28 September 2026; daily.dev, September 2026). Guest article by Matt von Hippel, "Yes, Claude Can Count Nine Rounds", published by Anthropic.
Matt von Hippel's assessment that Claude applied established methods with somewhat more computation than previously attempted, rather than inventing new physics, as reported by Dataconomy, 28 September 2026. Verification by Lance Dixon over approximately two weeks, as reported by HackerNoon, 7 October 2026. Anastasia Volovich (Brown University) described the result to IBM Think as significant for establishing trust that frontier AI systems can reliably perform difficult calculations (IBM Think, 9 October 2026).
Independent nine-loop work by the team led by He Song at the Institute of Theoretical Physics, Chinese Academy of Sciences, with assistance from GPT-6, as reported by KuCoin News, September 2026.
See the earlier piece in this series on the Navier-Stokes result and flat enterprise earnings.
McKinsey, The State of AI, published 25 August 2026: 80 percent of organisations reporting individual productivity gains against 37 percent reporting any EBIT impact, the latter unchanged year on year.
Mistral AI, Mistral Large 4 public preview, 6 October 2026: 1.05 trillion total parameters, 49 billion active, natively multimodal; trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centres; weights scheduled for release by the end of October 2026; licence not yet published; regional availability including a European deployment operated end to end by Mistral under European law. Coverage: The Decoder, TestingCatalog and Artificial Intelligence News, 6 October 2026.
On the suspension of Claude Fable 5 and Claude Mythos 5 following a United States Department of Commerce export-control directive on 12 June 2026, and the restoration of access on 1 July 2026, see the earlier piece in this series on flexibility as an architectural property.
neuland.ai HUB: agentic threads, deep agents and multi-hop reasoning for long-running work across multiple systems; task decomposition with parallel lines of work compared before commitment; customer-owned ontology grounding entity resolution across systems; permission-aware retrieval enforced ahead of reasoning at each step; append-only audit written by the platform; model-agnostic routing across self-hosted open-weight and proprietary models; evaluation of new models against customer workloads, tool-call behaviour, operational characteristics and DSGVO conformance before admission to the routing table. neuland.ai AG retains responsibility for content quality and clean delivery of results across all customer engagements.
Regulation (EU) 2024/1689 (AI Act), broadly applicable from 2 August 2026. Council of the EU and European Parliament provisional political agreement on the Digital Omnibus on AI, 7 May 2026: Annex III high-risk obligations postponed to 2 December 2027; Annex I obligations postponed to 2 August 2028.
On 7 August 2026, the physicist and science writer Matt von Hippel issued a public challenge to the AI industry: solve an outstanding problem in particle physics, on the kind of budget an academic research group actually has. On 25 September, Anthropic announced that two of its physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, had done it. They gave Claude a single sentence describing the problem, told it to keep working overnight, and let it run largely unsupervised for days. It reached the answer by two independent methods, for a reported compute bill of a few thousand dollars. Lance Dixon, the physicist whose team set the previous record in 2023, spent about two weeks checking the result and confirmed it. Then a second detail emerged: a team at the Chinese Academy of Sciences had independently reached the same frontier using OpenAI's GPT-6. Eleven days later, Mistral announced a trillion-parameter model trained in its own European data centres. Read together, these are not three unrelated stories. Two laboratories on two continents, using models from two different vendors, crossed the same scientific frontier within weeks of each other, and Europe now builds frontier-scale models of its own. Capable engines have become abundant. And yet most European enterprises that bought "AI transformation" this year are still running problems that are one step deep. The question every CIO should now ask is not which model to buy. It is why their AI programme cannot count past one.
This is the twenty-seventh piece in a series I have been writing for neuland.ai. Each one has pushed the same underlying argument forward: the value, the risk and the moat in enterprise AI are not in the model. They are in the layer above and around it. [¹] This piece uses a result from theoretical physics to make that argument as concretely as I have been able to make it, so it is worth taking the physics slowly.
The model can count to nine. Most enterprise AI is built to count to one.
What a loop is, and why nine is a wall
When two particles collide, physics cannot tell you exactly what will come out. It can tell you how likely each outcome is. The formula that gives you those odds is called a scattering amplitude, and computing it is one of the central jobs of theoretical particle physics.
Physicists calculate amplitudes by drawing the collision as diagrams. The simplest diagrams are almost trivial: particles come in, exchange something, go out. But quantum mechanics says that in between, other particles can briefly appear and disappear, and those fleeting detours show up in the drawing as closed circles. Each closed circle is a loop. Every loop you include makes the prediction more precise, and makes the calculation dramatically harder, because the number of possible detours explodes.
How dramatically is the part a CIO should hold on to. Written out in full on one part of the problem, the eight-loop answer runs to about 1.67 billion terms. At nine loops it is about 30 billion. One more loop made the problem roughly eighteen times bigger. [²] For a sense of scale: the most precise prediction in all of particle physics, the magnetic moment of the electron, is computed to five loops in a different, real-world theory.
The theory in question here, planar N=4 super-Yang-Mills, is a simplified sibling of the theories that describe real particles, a test track rather than a public road. It shares their structure but is clean enough to calculate. That is why physicists use it to push techniques to their limits, and why the eight-loop record had stood for three years.
What Claude actually did, and what it did not
Von Hippel offered two problems. Fitzpatrick and Mishra-Sharma first asked Claude which one it was most confident it could solve, then gave it a one-sentence description of the six-particle amplitude at nine loops. The model, Fable 5.1, running inside Anthropic's research environment, Claude Science, worked through the calculation over several days, reporting progress every four to six hours, and arrived at the result by two separate routes that agreed with each other. [³] Von Hippel's own account of it, published by Anthropic, carried the title "Yes, Claude Can Count Nine Rounds".
It is equally important to be precise about what did not happen. Claude did not invent new physics. By von Hippel's account, it applied methods that physicists had developed over many years, with somewhat more computation than anyone had previously attempted. [⁴] This was a record broken by endurance, rigour and cost, not by a new idea. And it was only accepted once the world's leading expert on the previous record had spent two weeks checking it.
That restraint is exactly why the result matters. It is not a benchmark. It is an open problem, solved, verified by the person best placed to say it was wrong, for a reported compute cost roughly comparable to a few days of a senior consultant's time.
The detail that should worry every model vendor
Here is the part that I have not seen written down clearly enough in the coverage. A team led by He Song at the Institute of Theoretical Physics of the Chinese Academy of Sciences independently reached the nine-loop stage with the help of OpenAI's GPT-6, and has obtained its main results. [⁵]
Two teams, on two continents, with models from two competing vendors, crossed the same frontier within weeks of each other. When the same hard result can be reached with different engines, the engine is no longer the scarce thing. An earlier piece in this series argued, using the Navier-Stokes result, that the systems built around models were what made the science work. [⁶] Nine loops adds the piece of evidence that argument was missing: reproducibility across vendors. The model has become a component in the most literal sense. You can swap it and still get the answer.
Enterprise problems have loops too
This is where the physics stops being a story about physics.
Think of an enterprise question as having loops of its own. Each additional system, relationship or rule that the answer depends on is another loop. What does our standard contract say about liability? is loop one: one question, one document set, one answer. Which of our suppliers operate under a contract that falls within DORA, and which of those had a reportable incident last year? is four loops: supplier, contract, regulatory scope, incident record, each living in a different system with different permissions. Are the instrument tags in this plant's piping and instrumentation diagrams consistent with the equipment list, the cause-and-effect matrix and the last three revisions? runs deeper still.
The analogy has limits, and they are worth stating. Physics loops come with a clean mathematical structure and established methods; enterprise data is messy, sources disagree, and every hop carries its own access rules. So enterprise loops are easier in one way, far fewer terms, and harder in another. But the two properties that matter carry over exactly. The space of possibilities grows explosively with every loop. And an answer at one depth tells you very little about whether the system can produce a correct answer one step deeper.
That is where enterprise value actually sits. Nobody's operating margin changes because staff can summarise a document faster. It changes when a question that used to take a team three weeks of cross-referencing across five systems is answered, with sources, in an afternoon. That is loop four and beyond.
Why the wrapper cannot count past one
Now look at what most European enterprises bought this year under the banner of AI adoption and transformation: a chat interface, connected to a document store and a frontier model, rolled out to several thousand employees.
That is a loop-one machine. It takes one question, retrieves a handful of passages, produces one answer, in one sitting. It holds no durable state, so it cannot work on a problem for days. It has no independent second method, so it cannot check itself. Its only record of what it did is the conversation itself. Put the model that counted to nine inside it, and it still counts to one, not because the model is weak, but because the system around it was never built to go deeper.
So here is the controversial version, and I mean it: most of what Europe bought this year as AI transformation is loop-one software sold at a loop-nine price. It explains a number that has puzzled boards. McKinsey's State of AI survey, published in August, found that 80 percent of organisations report productivity gains from AI and only 37 percent report any EBIT impact, unchanged from the previous year. [⁷] Faster loop-one work makes people feel more productive. It does not move earnings, because earnings live deeper in the stack.
Where Mistral fits
On 6 October, Mistral opened a public preview of Mistral Large 4: a one-trillion-parameter model, 49 billion parameters active per token, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centres, with open weights promised by the end of October and a deployment Mistral will operate end to end under European law. [⁸] That is the most encouraging development for European AI this year. For the first time, a frontier-scale engine has been built on European soil by a European company.
It also matters for a reason the nine-loop story makes uncomfortable. The model that counted to nine is the successor to Fable 5, which every customer in the world lost access to for nearly three weeks in June after a United States export-control directive. [⁹] A European engine is real insurance for continuity, and European enterprises should test it seriously once the weights and the licence are published.
But the lesson of nine loops applies to it exactly as it applies to Claude and GPT-6. Engines are now abundant: American, Chinese and European. A better engine in a loop-one machine still never reaches loop nine. And swapping an American vertical stack for a European one improves jurisdiction, not depth.
The nine-loop test
Von Hippel's challenge worked because it was concrete. Enterprise AI deserves the same discipline, so here is a version a CIO can put to any platform, including ours. I would encourage every CIO reading this to run it before signing the next renewal.
Can it work on one problem for days? Not a long conversation, but a single piece of work that continues, reports progress and survives interruption, without someone keeping a browser tab open.
Can it check itself two independent ways? And does it tell you when the two routes disagree, rather than quietly choosing one?
Can an expert audit it afterwards? Every source retrieved, every tool called, every intermediate result, from a record the system wrote at the time, not from a summary the model wrote about itself.
Can you swap the engine and get the same answer? If the result depends on one vendor's model, you have bought a dependency, not a capability.
A chat window fails the first question before the model is even asked.
Where neuland.ai stands on this
The neuland.ai HUB is a sovereign AI management and orchestration platform, and it was built for work several loops deep. Its agentic threads, deep agents and multi-hop reasoning are designed for long-running work across several systems, decomposed into steps that can run in parallel and be compared before anything is committed. The company's own ontology grounds each hop, so that a supplier in one system and the same supplier in another are recognised as one entity. Entitlement is enforced on retrieval before the model reasons, at every hop, for the person the work is being done for. Every retrieval, tool call and intermediate result is written to an append-only record by the platform, so that an expert can audit the path afterwards. And because routing is model-agnostic, with open-weight models we host, including Mistral Large 3, GLM, Qwen and DeepSeek, alongside Claude and GPT where the customer's policy allows them, the engine can be swapped without rebuilding the work. Mistral Large 4 will enter the same evaluation every model goes through once its weights and licence are published. [¹⁰]
This is also the answer to a question I hear from CIOs more and more often: we have rolled out an AI assistant to five thousand people, isn't that adoption? It is adoption. It is not transformation. Adoption measures how many people use loop one. Transformation is whether anything in the company now runs at loop four.
Personal take
I want to be careful with the framing here. This piece is not against AI assistants. It is not a criticism of Anthropic or OpenAI either. The nine-loop result was presented with unusual modesty, with established methods, more computation and independent verification, and that modesty is what makes it credible. And it is not lukewarm on Mistral, whose model is exactly what European AI needed.
What I am pushing back against is the habit of measuring transformation by seat count, and the market price that habit has created for tools that can only count to one. The physics this autumn showed what these models can do when the system around them is built for depth: work for days, check itself, survive an expert's audit, at a reported compute cost comparable to a few consulting days. That standard is now public. Enterprises should hold their own platforms to it.
The regulatory direction points the same way. The EU AI Act has applied broadly since 2 August 2026, and its expectations on oversight, logging and traceability attach to the deployed system; the Digital Omnibus has moved the high-risk Annex III obligations to 2 December 2027 and Annex I obligations to 2 August 2028. [¹¹] Deep work without a record cannot be evidenced. Deep work with one can.
Series articles available at neuland.ai/en/resources/insights.
Term counts as reported in "The Truth Behind Claude's $2,000 Physics Breakthrough", HackerNoon, 7 October 2026: approximately 1.67 billion terms for one part of the eight-loop answer and approximately 30 billion at nine loops. On the electron's anomalous magnetic moment being computed to five loops, see coverage of the result compiled by KuCoin News, September 2026.
Matt von Hippel, public challenge issued 7 August 2026, offering a choice between a seven-loop calculation in N=8 supergravity and the nine-loop six-particle amplitude in planar N=4 super-Yang-Mills theory on an academic-scale budget. Anthropic announcement, 25 September 2026: computation of the nine-loop six-particle scattering amplitude in planar N=4 super-Yang-Mills theory by Claude Fable 5.1 in Claude Science, directed by Liam Fitzpatrick and Siddharth Mishra-Sharma, with progress reported every four to six hours and the result obtained by two independent methods; previous eight-loop record set in 2023. Reported cost estimates range from roughly 1,000 to 2,000 US dollars in total to that range per method (Dataconomy, 28 September 2026; daily.dev, September 2026). Guest article by Matt von Hippel, "Yes, Claude Can Count Nine Rounds", published by Anthropic.
Matt von Hippel's assessment that Claude applied established methods with somewhat more computation than previously attempted, rather than inventing new physics, as reported by Dataconomy, 28 September 2026. Verification by Lance Dixon over approximately two weeks, as reported by HackerNoon, 7 October 2026. Anastasia Volovich (Brown University) described the result to IBM Think as significant for establishing trust that frontier AI systems can reliably perform difficult calculations (IBM Think, 9 October 2026).
Independent nine-loop work by the team led by He Song at the Institute of Theoretical Physics, Chinese Academy of Sciences, with assistance from GPT-6, as reported by KuCoin News, September 2026.
See the earlier piece in this series on the Navier-Stokes result and flat enterprise earnings.
McKinsey, The State of AI, published 25 August 2026: 80 percent of organisations reporting individual productivity gains against 37 percent reporting any EBIT impact, the latter unchanged year on year.
Mistral AI, Mistral Large 4 public preview, 6 October 2026: 1.05 trillion total parameters, 49 billion active, natively multimodal; trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centres; weights scheduled for release by the end of October 2026; licence not yet published; regional availability including a European deployment operated end to end by Mistral under European law. Coverage: The Decoder, TestingCatalog and Artificial Intelligence News, 6 October 2026.
On the suspension of Claude Fable 5 and Claude Mythos 5 following a United States Department of Commerce export-control directive on 12 June 2026, and the restoration of access on 1 July 2026, see the earlier piece in this series on flexibility as an architectural property.
neuland.ai HUB: agentic threads, deep agents and multi-hop reasoning for long-running work across multiple systems; task decomposition with parallel lines of work compared before commitment; customer-owned ontology grounding entity resolution across systems; permission-aware retrieval enforced ahead of reasoning at each step; append-only audit written by the platform; model-agnostic routing across self-hosted open-weight and proprietary models; evaluation of new models against customer workloads, tool-call behaviour, operational characteristics and DSGVO conformance before admission to the routing table. neuland.ai AG retains responsibility for content quality and clean delivery of results across all customer engagements.
Regulation (EU) 2024/1689 (AI Act), broadly applicable from 2 August 2026. Council of the EU and European Parliament provisional political agreement on the Digital Omnibus on AI, 7 May 2026: Annex III high-risk obligations postponed to 2 December 2027; Annex I obligations postponed to 2 August 2028.