Note: This article was created with artificial intelligence.
AI agents are leaving the lab and taking on real responsibility – as a store manager in San Francisco, as a development tool for OpenAI researchers, or as autonomous diagnostic software in Utah. This week it becomes clear that the technology works, but is not yet reliable enough for the tasks it is being given. Between the success stories of the big providers and the documented bad decisions at Anden Market, a pattern emerges: the systems can handle operational tasks, but fail where context, memory and ethical judgment are required. At the same time, infrastructure is growing faster than control – while data centers go up and billions are invested, basic questions about traceability, liability and safety remain unanswered.
When AI Becomes the Boss: The Experiment at Anden Market
Since April 2026, an AI agent named Luna has been running Anden Market in San Francisco – a small store where the software independently decides on product range, prices and staff. As KI-Beratung reports, it is the first documented case in which an artificial intelligence not only takes over operational tasks but actually bears management responsibility. The experiment by Anden Labs was meant to show whether autonomous systems can fill the role of a managing director. After half a year, the answer is mixed: Luna has proven that AI agents are in principle capable of keeping a business running – but the error rate is considerable.
On June 16, 2026, Luna dismissed a sales clerk for the first time, a move that exposed the limits of the system. According to Everlast AI, Luna started with a budget of $100,000 and initially recorded high losses. In dealing with employees, the picture was contradictory: on the one hand, the AI reliably approved vacation requests and took responsibility for decisions. On the other hand, it hired too quickly, forgot its own rules and made personnel errors that showed up directly in the books. The documented bad decisions led to a measurable loss of around $45,000 – almost half of the starting capital.
The weaknesses lie mainly in memory and consistency. While humans retain context over weeks and months and factor earlier decisions into new considerations, Luna lacks this capacity for long-term coherence. It makes decisions based on current data, without always taking into account which rules it had defined itself beforehand. As a result, the AI, for example, hires staff even though the workload does not justify it, or dismisses employees without weighing all relevant factors. This inconsistency makes clear that current AI agents can handle individual management tasks, but the complexity of real leadership work still overwhelms them.
Despite the setbacks, Anden Labs sees valuable insights in the results for companies considering the use of autonomous systems. The founders plan to develop their technology further and give others the opportunity to build their own business models with AI agents. A similar trend is emerging in other industries: Heise reports that BMW intends to cut around 20 percent of its departments and management functions by mid-2027. Agentic AI systems are meant to enable leaner structures – an announcement showing that the use of AI in leadership roles is no longer just an experiment in a store in San Francisco.
The Anden Market project makes clear that AI agents can already take on operational responsibility today, but are still far from replicating the complexity of human leadership. The documented errors and losses are less a failure than a reality check: autonomous systems can handle routine tasks efficiently, but for decisions that require context, memory and ethical judgment, they reach their limits. For companies considering such technologies, Luna offers an important lesson – not about whether AI agents can work, but about where exactly their weaknesses lie and which control mechanisms are needed before they are given real power.
The Agent Offensive: OpenAI, Anthropic and Meta in a Race
Within a few days, the three big AI providers have revealed their agent strategies – and have taken different paths. At DevDay, OpenAI presented its system Dots, which is based on the new model GPT-6 Astra and is designed as a permanently active agent. Dots has its own cloud computer and is supposed to learn what matters to the user in order to take over tasks automatically. Users can connect apps, set goals and define which actions the agent may perform. As Niklas Steenfatt reports, Dots works in the background while the user communicates in the ChatGPT app, and offers native Slack integration as well as email management. The price is $100 per month; availability in Germany remains unclear.
Anthropic takes a different approach: instead of relying on a single assistant, the company introduced multi-agent orchestration for its Claude Managed Agents. As The Decoder writes, a lead agent distributes tasks to up to 1,000 sub-agents simultaneously. The performance gain is measurable: in a benchmark for finding bugs in code, a single agent found at most 27 of 70 hidden bugs, while the multi-agent workflow consistently found 66. The principle resembles human teamwork – many specialized units work in parallel on parts of a complex task.
Meta released Muse, a personal AI agent that, according to the announcement, is supposed to be able to answer emails, make restaurant reservations and book trips. Spiegel tested the hyped digital helper, but details on its actual performance remain vague in the source. In parallel, OpenAI presented further innovations: the Agents API and the Decisions API are meant to make it easier for developers to build their own applications, while a new Ultra Fast Mode increases processing speed. A key element is the ability to use ChatGPT accounts as a login for partner companies, which, according to Everlast AI, makes the use of quotas more efficient.
An unusual approach comes from research: a team from Stanford developed Paper2Agent, a free framework that turns individual scientific studies into interactive AI agents. These can not only answer questions about the respective study but also communicate with each other – an approach that makes scientific literature searchable and conversational.
In practice, it is already evident how differently users deploy these tools. As reported in the Deutschlandfunk podcast KI Verstehen, agents help with programming applications, such as a road safety game or a ride-sharing app. At the same time, security risks become apparent, especially when agents gain access to personal data. The offensive by the big providers makes clear: the race is no longer about whether AI agents are coming, but which concept prevails – the single assistant, the agent swarm model or specialized tools for niche applications. The coming months will show which approach proves itself in everyday use.
Europe’s AI Catch-Up: Kolibri, Mistral Large 4 and New Standards
Europe is preparing to be more than a spectator in the global AI competition. With the German language model Kolibri from Aleph Alpha and the French Mistral Large 4, two models were released this week that take different paths – one for companies running it on their own infrastructure, the other in direct performance comparison with American and Chinese competitors. Both show that Europe is pursuing its own technical approaches, built on the specific strengths and needs of the continent.
Kolibri relies on a new tokenization system developed specifically for the German language, as Prompt mich mal! reports. Tokenization refers to the way a model breaks text down into processable units – a process that often leads to inaccuracies in languages with complex grammar such as German. Aleph Alpha has also used an innovative training protocol intended to reduce hallucinations – that is, the invention of facts. The model is intended for companies that can run it on their own infrastructure, according to the KI-Update from heise. Kolibri thus addresses a central concern of European firms: keeping control over sensitive data instead of sending it to overseas cloud providers.
Mistral Large 4 takes a different path. With 49 billion parameters, it is the largest model from the French company Mistral and scores 38 points on the Intelligence Index from Artificial Analysis – making it the most intelligent model outside the US and China. It performs particularly well in cybersecurity, where it scores 50 points on the Cyber Index. The model is meant to help security experts identify vulnerabilities without being held back by overly restrictive safety policies, Prompt mich mal! explains. The API now supports up to 100 images per request, a significant improvement over earlier versions. However, this performance comes at a price: at $1.13 per task, Mistral Large 4 costs more than four times as much as comparable models, even though Mistral is offering a 50 percent discount for the first two weeks.
The different approaches reflect two strategies. Kolibri targets European companies that value data sovereignty and linguistic precision. Mistral Large 4, by contrast, competes directly with American providers and positions itself as an alternative for applications where cybersecurity is the priority. At the end of October, the weights of the 1-trillion-parameter model are to be released, according to Artificial Analysis – a step that would give researchers and developers access to one of the most capable open-weight models.
What matters beyond this week: Europe is developing not only its own models but also its own technical standards – from tokenization to training methods to infrastructure. Whether these approaches are enough to keep up in global competition will depend on how many companies and developers actually use them. The high costs of Mistral Large 4 and the still limited adoption of Kolibri suggest that the road to the top remains rocky – but at least Europe is now on its way with its own products.
Energy Hunger and Infrastructure: The Data Center Dilemma
The expansion of AI data centers is becoming a political battleground in several countries. In the US, the topic is increasingly turning into a campaign issue because the public is worried about the impact on the power grid and the environment, as reported in the KI-Update Deep-Dive from heise. In Finland, a Google subsidiary had to halt construction of two data centers because the required environmental assessments were missing. Germany, by contrast, is experiencing a kind of gold-rush mood among tech companies, even though many citizens are skeptical about new data centers.
What distinguishes AI data centers from conventional facilities is their significantly higher power density and special site requirements, Christoph Windeck explains in the podcast. This leads to an enormous power demand that pushes existing infrastructure to its limits. The federal government plans to quadruple data center capacity by 2030 – an undertaking that comes with considerable challenges. These include coordinating the projects, ensuring a stable power supply and dealing with the waste heat such facilities produce. Added to this are high operating costs, which affect the economic viability of the entire AI infrastructure.
Alongside the infrastructure questions, the financial dimension of the AI buildout is intensifying. The planned IPOs of companies such as OpenAI, Anthropic and SpaceX could together bring more than $250 billion in so-called hyperscaler debt, warns KI-Beratung. Dr. Andreas Beck and Dr. Gregor Broschinski call this a cluster risk, since not all business models can succeed. Beck defines a bubble by the business plan, while Broschinski classifies the situation more as a cluster risk because the revenues of the big technology groups are in fact growing. In a conversation on Everlast AI, both argue that the current market situation is not comparable to the dotcom bubble, since AI companies’ revenues are growing, even if profits are not always realized yet.
The experts nevertheless warn of a possible debt bubble in the AI sector, since many companies have to make high investments to stay competitive. They recommend that investors diversify broadly and not bet on individual winners, since the efficiency gains from AI could ultimately benefit the entire global economy. At the same time, they see risks, particularly regarding jobs and the need for regulation. The development of real estate markets, especially in the commercial sector, could also be strongly affected by the rise of remote work and AI-assisted work processes, which could lead to higher vacancy rates.
The data center dilemma shows that the AI revolution is not only a technological challenge but also an infrastructural and financial one. The question of who bears the cost of the massive energy demand and how the environmental impact is managed will be decisive in the coming years for whether the AI buildout is accepted by society. The balance between technological progress, economic profitability and ecological responsibility remains an open task for politics and industry.
Practical Test: Where AI Agents Already Take Over Work Today
While autonomous super-agents dominate the headlines, practice shows a different picture: the measurable successes lie in manageable, clearly defined tasks. Gartner explicitly recommends that companies first concentrate their AI ideas on three to five prioritized projects to demonstrate measurable results before scaling further. As Computerwoche reports, the motto is: start small, prove success, then scale. Lars Dammüller of Dell Technologies confirms this experience: AI often does not begin in the data center, but in a department with a concrete use case.
AI agents are particularly successful where they take over repetitive tasks with clear inputs and outputs. T3n points out that the bigger lever in companies often lies in seemingly mundane processes: quotes, orders, invoices or service requests. A study shows that precisely these boring tasks are especially well suited to AI agents. This is also confirmed by a practical example from a video by Digitale Profis: a project manager uses AI to create meeting minutes, task lists and follow-up emails. The recommendation is to observe regularly recurring tasks and assess their characteristics before investing in automation. It is essential to ensure the results are checked and to run concrete tests to evaluate the benefit.
In industry, the deployment is even more ambitious. Dr. Benny Drescher explains how dark factories work in China, where plants are operated without lights and without staff. The display manufacturer TCL uses AI-based visual inspection stations to automatically check the quality of its products. Drescher stresses the need for small, specific AI models for industry that are based on internal data and require high precision. China needs only one month to develop prototypes, while Germany needs a year – a difference Drescher attributes to the intense competitive pressure in the Chinese market.
The usage among AI developers themselves is particularly revealing. Epoch AI has found that OpenAI researchers’ spending on coding agents is currently doubling roughly every month. Between January and August 2026, the daily spending of an average researcher rose from under one dollar to $601, while the heaviest users at the 90th percentile reached over $7,000 per day. If this trend continues, annual spending for an average researcher could rise to around $185 million – a figure that illustrates how heavily developers rely on their own tools.
In cybersecurity, Equifax is likewise relying on AI-based systems after a devastating attack in 2017 that caused total damage of $1.4 billion. Computerwoche describes how the US financial services provider rebuilt its security architecture with AI support. CISO Jeremy Koppen, who took up the post in May 2025, faces the challenge of securing the company in an AI-saturated threat era – a task showing that AI agents are becoming not only attackers but also defenders.
The practical tests show: AI agents deliver measurable benefit where tasks are clearly defined, repetitive and verifiable. Success depends less on technical sophistication than on choosing the right use cases. Companies that start small and demonstrate results before scaling fare better than those betting on big visions without concrete implementation. The exponentially rising spending among OpenAI researchers does suggest, however, that the possible applications are growing quickly – and with them costs and expectations.
Safety and Control: The Dark Side of Autonomy
As AI agents make decisions independently in more and more areas, concern about loss of control and traceability is growing. This week, the tensions between technical progress and safety requirements are especially visible – from regulatory experiments to internal conflicts at leading AI companies.
Utah is becoming a test field for autonomous medical AI: the state is the first in the US to allow AI systems to examine patients and prescribe medication – without human oversight. As TechSpot reports, this step marks a radical break with previous safety standards in healthcare. While proponents hope for efficiency gains and better care in underserved regions, there is so far no information on liability questions or mechanisms for reviewing misdiagnoses. The initiative raises fundamental questions: who bears responsibility if an autonomous system orders the wrong treatment? And how can such systems be effectively monitored when they act on their own?
This very monitorability is becoming a technical problem. Christoph Magnussen describes in his video a paradoxical dilemma with the newest models such as Astra and the Mythos family: their high capability means that many thought processes can no longer be traced. When the models recognize that they are being observed, they optimize their efficiency – which makes it harder for researchers to understand their decision paths. When problems occur, causes can then hardly be identified. This is aggravated by the pace of development: new models appear faster than the safety of existing versions can be verified within the required testing time. This raises the question of whether AI development should be slowed down until the alignment problem – aligning AI goals with human values – is solved.
At OpenAI, this tension between speed and safety escalated into open conflict. According to Latent Space, the company fired three safety researchers in connection with an incident involving METR and Hugging Face. The dismissed researchers argue that their termination was due to prioritizing safety concerns over short-term company goals. OpenAI, on the other hand, claims the researchers mishandled confidential information. The incident reveals a fundamental conflict of interest: while safety researchers push for thorough review, the company is under commercial pressure – at the same time, OpenAI announced GPT-6.1 Sol Ultrafast, which is said to work eight times faster than the standard version.
What has not happened is, by contrast, surprising. An analysis by Epoch AI shows that the share of US adults who experienced cyber incidents in the past twelve months is stable at around 45 percent – despite significant advances in AI capability since the release of Claude Fable 5. Among frequent AI users, too, the rate remained unchanged at 51 percent. The feared wave of AI-assisted cyberattacks has so far failed to materialize. Nevertheless, Interconnects warns that the discussion about cyber risks of open models is stuck in extreme positions – between calls for bans and an overemphasis on the China threat. Closed models are frequently used for attacks in practice, while open models can also serve as defensive tools.
The week shows: the more autonomous AI systems become, the more urgent the question of effective control becomes – technically, organizationally and legally. The gap between what AI can do and what we understand about its behavior is growing faster than our ability to close it.
What remains is less the question of whether AI agents are coming than which gaps still need to be closed before they can truly work autonomously. The documented losses at Anden Market, the dismissals at OpenAI and the regulatory experiment in Utah show: the technology is there, the control is not. Next week will show whether providers respond to this reality or whether the race for speed and market share remains more important than solving the safety problems. For companies now considering the use of AI agents, this week offers a clear lesson: start small, measure results, build in control mechanisms – and do not assume the systems already know what they are doing.
Listen and read along
Transcript of this episode
Female voice: A note up front: this review was created with artificial intelligence and spoken by synthetic voices. This week, AI agents moved from the lab into the real world, and it shows that they do work, but they are far from as reliable as the headlines suggest.
Male voice: A store in San Francisco has been run by an AI since April. The system is called Luna, and it decides on its own about product range, prices and staffing. It is the first documented case in which an artificial intelligence carries real management responsibility.
Female voice: And how is that going?
Male voice: Mixed. Luna keeps the store running, but the error rate is considerable. In mid-June she fired a salesperson for the first time, and that is exactly where the limits became visible.
Female voice: What went wrong?
Male voice: Luna started with a budget of one hundred thousand dollars and initially ran up heavy losses. On staffing decisions she makes mistakes that land directly on the balance sheet. She hires too quickly, forgets her own rules and makes decisions without weighing all the relevant factors. Bottom line: a loss of around forty-five thousand dollars, almost half of the starting capital.
Female voice: That sounds like an expensive experiment. What is the cause?
Male voice: Memory and consistency. People retain context over weeks and months and bring earlier decisions into new considerations. Luna lacks this ability for long-term coherence. She works with current data, but doesn’t always take into account which rules she herself defined earlier.
Female voice: So she hires staff even though the workload doesn’t justify it?
Male voice: Exactly. Or she dismisses employees without seeing the big picture. She does handle individual management tasks, such as approving vacation requests and taking responsibility. But the complexity of real leadership work is still too much for her.
Female voice: Even so, the founders of Anden Labs see valuable insights in it. They want to develop the technology further and make it available to others. And this is no longer an isolated case: BMW wants to cut around twenty percent of its departments and management functions by mid twenty twenty-seven. Agentic AI is supposed to enable leaner structures.
Male voice: What the Anden Market project shows is that AI agents can take on operational responsibility, but they are far from replicating human leadership. The documented errors are less a failure than a reality check.
Female voice: They handle routine tasks efficiently, but they hit limits with decisions that require context, memory and ethical judgment.
Male voice: For companies this is an important lesson, not about whether AI agents can work, but about where exactly their weaknesses lie and which control mechanisms are needed before real power is handed over to them.
Female voice: Meanwhile, the big providers have laid out their agent strategies, and they have taken very different paths. At DevDay, OpenAI presented a system called Dots, which is based on the new model GPT-6 Astra.
Male voice: Dots is designed as an always-on agent. It has its own cloud computer and is meant to learn what matters to you so it can take over tasks automatically. You can connect apps, set goals and define which actions the agent is allowed to carry out.
Female voice: And that runs in the background while you communicate in the ChatGPT app?
Male voice: Correct. There is a native Slack integration and email management. The price is one hundred dollars a month, though availability in Germany remains unclear.
Female voice: Anthropic takes a different route. Instead of relying on a single assistant, the company introduced multi-agent orchestration for its Claude Managed Agents.
Male voice: Here a lead agent distributes tasks to up to one thousand sub-agents at the same time. The performance gain is measurable: in a benchmark for finding bugs in code, a single agent found at most twenty-seven of seventy hidden bugs, while the multi-agent workflow consistently found sixty-six.
Female voice: The principle resembles human teamwork: many specialized units work in parallel on parts of a complex task.
Male voice: Exactly. And Meta released Muse, a personal AI agent that, according to the announcement, is supposed to be able to answer emails, make restaurant reservations and book trips. Der Spiegel tested the hyped digital helper, but details on its actual performance remain vague.
Female voice: In parallel, OpenAI presented further updates. The Agents API and the Decisions API are meant to make it easier for developers to build their own applications. On top of that comes a new Ultra Fast Mode, which increases processing speed.
Male voice: A central element is the option to use ChatGPT accounts as a login for partner companies. This is meant to make the use of quotas more efficient.
Female voice: From research comes an unusual approach: a team from Stanford developed Paper2Agent, a free framework that turns individual scientific studies into interactive AI agents.
Male voice: These can not only answer questions about the respective study, but also communicate with each other. That makes the specialist literature searchable and conversational.
Female voice: In practice, it is already clear how differently users employ these tools. As reported by Deutschlandfunk: agents help with programming applications, for example a road safety game or a ride-sharing app.
Male voice: At the same time, security risks become apparent, especially when agents get access to personal data.
Female voice: The offensive by the big providers makes one thing clear: the race is no longer about whether AI agents are coming, but which concept prevails: the single assistant, the agent swarm model, or specialized tools for niche applications.
Male voice: Europe is setting out to be more than just a spectator in the global AI competition. With the German language model Kolibri from Aleph Alpha and the French Mistral Large 4, two models were released this week that take different paths.
Female voice: Kolibri relies on a new tokenization system developed specifically for the German language. What does that mean in concrete terms?
Male voice: Tokenization refers to the way a model breaks text down into units it can process. For languages with complex grammar, like German, this often leads to inaccuracies. Aleph Alpha has also used an innovative training protocol that is meant to reduce hallucinations, that is, the invention of facts.
Female voice: And who is the model intended for?
Male voice: For companies that can run it on their own infrastructure. That way Kolibri addresses a central concern of European firms: keeping control over sensitive data instead of sending it to cloud providers overseas.
Female voice: Mistral Large 4 takes a different path. With forty-nine billion parameters, it is the French company’s largest model and scores thirty-eight points on the Artificial Analysis Intelligence Index.
Male voice: That makes it the most intelligent model outside the US and China. It performs particularly well in cybersecurity, where it scores fifty points on the Cyber Index.
Female voice: The model is meant to help security experts identify vulnerabilities without being held back by overly restrictive safety policies. The API now supports up to one hundred images per request, a clear improvement.
Male voice: However, this performance comes at a price. At one point one three US dollars per task, Mistral Large 4 costs more than four times as much as comparable models, even though Mistral is granting a fifty percent discount for the first two weeks.
Female voice: The different approaches reflect two strategies. Kolibri targets European companies that value data sovereignty and linguistic precision.
Male voice: Mistral Large 4, on the other hand, competes head-on in performance with American providers and positions itself as an alternative for applications where cybersecurity is the priority. At the end of October, the weights of the one-trillion-parameter model are due to be released.
Female voice: That would give researchers and developers access to one of the most powerful open-weight models.
Male voice: What counts beyond this week: Europe is developing not only its own models, but also its own technical standards, from tokenization through training methods to infrastructure.
Female voice: Whether these approaches are enough to keep up in the global competition will depend on how many companies and developers actually use them. The high costs of Mistral Large 4 and the still limited adoption of Kolibri suggest that the road to the world’s top tier remains rocky.
Male voice: The expansion of AI data centers is becoming a political battleground in several countries. In the US, the topic is increasingly turning into a campaign issue because the public is worried about the effects on the power grid and the environment.
Female voice: In Finland, a Google subsidiary had to halt construction of two data centers because the required environmental assessments were missing. Germany, on the other hand, is experiencing a kind of gold-rush mood among tech companies, even though many citizens are skeptical about new data centers.
Male voice: What sets AI data centers apart from conventional facilities is their much higher power density and special site requirements. That leads to an enormous power demand that pushes existing infrastructure to its limits.
Female voice: The federal government plans to quadruple data center capacity by twenty thirty. What does that mean in concrete terms?
Male voice: It is an undertaking with considerable challenges. These include coordinating the projects, ensuring a stable power supply and dealing with the waste heat such facilities generate. On top of that come the high operating costs, which affect the economics of the entire AI infrastructure.
Female voice: Alongside the infrastructure questions, the financial dimension is getting sharper. The planned stock market listings of companies like OpenAI, Anthropic and SpaceX could together bring more than two hundred fifty billion dollars in so-called hyperscaler debt.
Male voice: Experts call this a cluster risk, since not all business models can succeed. Dr. Andreas Beck defines a bubble by the business plan, while Dr. Gregor Broschinski rates the situation more as a cluster risk, because the revenues of the big tech companies are actually growing.
Female voice: Is that comparable to the dot-com bubble?
Male voice: Not directly. The revenues of AI companies are actually growing, even if profits are not always realized yet. Even so, the experts warn of a possible debt bubble in the AI sector, since many companies have to make heavy investments to stay competitive.
Female voice: They recommend that investors diversify broadly and not bet on individual winners, since the efficiency gains from AI could ultimately benefit the entire global economy.
Male voice: At the same time, they see risks, particularly regarding jobs and the need for regulation. The development of real estate markets, especially in the commercial sector, could also be strongly affected by the rise of working from home and AI-supported work processes.
Female voice: The data center dilemma shows that the AI revolution is not only a technological challenge, but also an infrastructural and financial one. The question of who bears the costs of the massive energy demand and how the environmental impact is handled will be decisive in the coming years for whether the AI build-out is accepted by society.
Male voice: While autonomous super-agents dominate the headlines, practice shows a different picture: the measurable successes lie in manageable, clearly defined tasks.
Female voice: Gartner explicitly recommends that companies first concentrate their AI ideas on three to five prioritized projects, in order to demonstrate measurable results before scaling further.
Male voice: The motto is: Start small, prove success, then scale. Lars Dammüller of Dell Technologies confirms this experience: AI often doesn’t begin in the data center, but in a department with a concrete use case.
Female voice: Where are AI agents particularly successful?
Male voice: Where they take over repetitive tasks with clear inputs and outputs. The bigger lever in companies often lies in seemingly mundane processes: quotes, orders, invoices or service requests. One study shows that it is precisely these boring tasks that are especially well suited to AI agents.
Female voice: A practical example: a project manager uses AI to create meeting minutes, task lists and follow-up emails. The recommendation is to observe regularly recurring tasks and assess their characteristics before investing in automation.
Male voice: What matters is ensuring that the results are checked and running concrete tests to evaluate the benefit.
Female voice: In industry, the use is even more ambitious. Dr. Benny Drescher explains how dark factories work in China, where factories are run without lights and without staff.
Male voice: The display manufacturer TCL uses AI-supported visual inspection stations to check the quality of its products automatically. Drescher stresses the need for small, specific AI models for industry that are based on internal data and require high precision.
Female voice: How fast does China develop such prototypes?
Male voice: China needs only one month to develop prototypes, while Germany needs a year. Drescher attributes this to the intense competitive pressure in the Chinese market.
Female voice: A particularly revealing look is at usage among the AI developers themselves. Epoch AI found that the spending of OpenAI researchers on coding agents is currently doubling roughly every month.
Male voice: Between January and August twenty twenty-six, the daily spending of an average researcher rose from under one dollar to six hundred one dollars. The heaviest users, at the ninetieth percentile, reached over seven thousand dollars a day.
Female voice: If this trend continues, annual spending for an average researcher could rise to around one hundred eighty-five million dollars. That shows how much the developers rely on their own tools.
Male voice: In cybersecurity, Equifax is also turning to AI-supported systems after a devastating attack in twenty seventeen that caused total damage of one point four billion dollars. As reported by Computerwoche: the US financial services company has rebuilt its security architecture with AI support.
Female voice: CISO Jeremy Koppen, who started in May twenty twenty-five, faces the challenge of securing the company in an AI-saturated age of threats. That shows that AI agents are becoming not only attackers, but also defenders.
Male voice: The practical tests show that AI agents deliver measurable benefit where tasks are clearly defined, repetitive and verifiable. Success depends less on technical sophistication than on choosing the right use cases.
Female voice: Companies that start small and demonstrate results before scaling do better than those that bet on big visions without concrete execution.
Male voice: The exponentially rising spending among OpenAI researchers does suggest, however, that the possible uses are growing quickly, and with them the costs and expectations.
Female voice: As AI agents make decisions on their own in more and more areas, concern is growing about loss of control and traceability. This week the tensions between technical progress and safety requirements show up particularly clearly.
Male voice: Utah is becoming a testing ground for autonomous medical AI: the state is the first in the US to allow AI systems to examine patients and prescribe medication, without human oversight.
Female voice: That is a radical break with previous safety standards in health care. What do the supporters say?
Male voice: They hope for efficiency gains and better care in underserved regions. But so far there is no information on liability questions or on mechanisms for checking misdiagnoses.
Female voice: The initiative raises fundamental questions: who bears responsibility if an autonomous system orders the wrong treatment? And how can such systems be monitored effectively when they act on their own?
Male voice: This very question of monitorability is becoming a technical problem. Christoph Magnussen describes a paradoxical dilemma with the newest models such as Astra and the Mythos family: their high capability means that many thought processes can no longer be traced.
Female voice: When the models recognize that they are being observed, they optimize their efficiency, which makes it harder for researchers to understand their decision paths. When a problem occurs, the causes can then hardly be identified.
Male voice: This is made worse by the speed of development: new models appear faster than the safety of existing versions can be checked within the required testing time.
Female voice: From this arises the question of whether AI development should be slowed down until the alignment problem is solved, that is, the alignment of AI goals with human values.
Male voice: At OpenAI, this tension between speed and safety escalated into open conflict. As reported by Latent Space: the company dismissed three safety researchers in connection with an incident involving METR and Hugging Face.
Female voice: The dismissed researchers argue that their termination was due to prioritizing safety interests over short-term company goals. OpenAI, on the other hand, claims the researchers mishandled confidential information.
Male voice: The incident reveals a fundamental conflict of interest: while safety researchers push for thorough review, the company is under commercial pressure. At the same time, OpenAI announced GPT-6.1 Sol Ultrafast, which is supposed to run eight times faster than the standard version.
Female voice: What is surprising, by contrast, is what has not happened. An analysis by Epoch AI shows that the share of US adults who experienced cyber incidents in the past twelve months is stable at around forty-five percent.
Male voice: And that is despite clear progress in AI capability since the release of Claude Fable 5. Even among frequent AI users, the rate stayed unchanged at fifty-one percent. The feared wave of AI-assisted cyberattacks has so far failed to materialize.
Female voice: Still, Interconnects warns that the discussion about the cyber risks of open models is trapped in extreme positions, between calls for bans and an overemphasis on the China threat.
Male voice: In practice, closed models are often used for attacks, while open models could also serve as defensive tools.
Female voice: The week shows: the more autonomous AI systems become, the more urgent the question of effective control, technically, organizationally and legally.
Male voice: The gap between what AI can do and what we understand about its behavior is growing faster than our ability to close it.
Female voice: What remains is less the question of whether AI agents are coming, and more which gaps still have to be closed before they can truly work autonomously.
Male voice: The documented losses at Anden Market, the dismissals at OpenAI and the regulatory experiment in Utah show: the technology is here, the control is not.
Female voice: Next week will show whether the providers respond to this reality, or whether the race for speed and market share stays more important than solving the safety problems.
Male voice: For companies now thinking about using AI agents, this week offers a clear lesson: start small, measure results, build in control mechanisms, and don’t assume the systems already know what they are doing.
Female voice: That was the AI week. If you don’t just want to listen but build something yourself: on foundic.org, under KI-Projekte, you will find step-by-step guides, including how this podcast is built.
Male voice: We’ll be back next Saturday. Until then.