Anthropic Files IPO

This paper from Hugging Face and Data & Society argues that the “Human-in-the-Loop” (HITL) approach – put forward as a means for humans to retain control over AI – is ineffective, and that developments in AI are making it more difficult to implement HITL.

• One major challenge for HITL is the cognitive burden. Powerful AI systems generate a lot of information. Repeated requests for approval lead to approval fatigue – where humans are more likely to accept the AI’s answer, as well as automation bias – where humans gradually begin to believe in the correctness of AI outputs.
• A second major challenge is that HITL requires human skills that AI is negatively impacting in the long run. Notably, HITL requires critical thinking and analytical reasoning, domain expertise, practical skills, and independent judgment. Human dependence on AI tools is leading to a degradation of these skills.
• The paper calls for changes in the design approaches to AI systems where models are not only measures in technical metrics like the number of parameters, but also by HITL metrics such as the number of human interventions in training.

Cybersecurity experts are warning of a “vulnerability tsunami” in the coming months as AI is used to discover more flaws in commonly used software. The warning follows the discovery of a serious vulnerabilities in the PHP programming language runtime.

• The PHP vulnerability affects millions of websites around the world. The sheer number of websites vulnerable has led researchers to refer to the current month as “Black October” with some experts estimating that 70% of websites are vulnerable to hacking.
• A CloudFlare post on X said the number of attacks on websites following the PHP vulnerability discovery “jumped tens of thousands in the 45 minutes post announcement, hundreds of thousands within 2 hours, and millions now”.
• Several experts now believe that it is not possible to patch vulnerabilities at the current rate of discovery of vulnerabilities by AI.
• Several websites across the world were taken down for patching in the aftermath of the patching.

Character.ai, the AI chatbot company founded by former Google employees is filing for bankruptcy after five years of operation. TechCrunch marks the passing by enumerating several much publicized initiatives that failed in the long term.

• OpenAI’s ChatGPT Atlas was announced as a stand-alone AI-powered web browser, but was discontinued in August of this year. Most of its features have been incorporated into the ChatGPT application.
• Another abandoned OpenAI initiative is the Sora video-sharing platform. Its user-retention challenge and high operating costs led it to be shutdown after six months.
• Microsoft’s Recall is an AI “photographic memory” for the desktop that allowed a user to periodically take snapshots of work being done. The initiative was abandoned due to user fears over privacy.
• Notion Mail was an AI-enhanced email client that used AI to help users organize and automate inboxes. This product was discontinued as users looked to other AI agents to do this task.
• Character.ai was founded in 2021 and allowed users to create customizable AI chatbots that they could interact with. The company raised 150 million USD at a 1 billion USD valuation in 2023. The company has faced severe problems such as lawsuits from families claiming the chatbots contributed to self-harm and even suicide. Also, Disney sent a cease-and-desist letter for chatbot characters that infringed on Disney copyright.
• S&P Global Market Intelligence reports that 42% of AI initiatives are abandoned by corporate parents. The reasons evoked are insufficient funding, technical challenges, competition, difficulty scaling, and weak user demand.

Anthropic publicly filed its initial public (IPO) offering this week, and in doing so gave deeper insight into its finances.

• The company’s revenue for the first six months of 2026 was 16.3 billion USD, which is up from 1.3 billion USD one year ago.
• Operating expenses were 17 billion USD – leading to an operating loss of 693 million USD, which is down from 2.14 billion USD one year ago. Expenses are attributed to computing and infrastructure (8.95 billion USD), personnel (3.15 billion USD), as well as sales, marketing and partnerships (3.17 billion USD).
• Anthropic’s rise in revenue over the past year is attributed to enterprise adoption. Over 1900 enterprise customers generated 1 million USD each in revenue, with 250 generating at least 10 million USD.
• These customer contracts are also a source of concern for Anthropic, because these contracts do not require customer commitments, and Anthropic has contracts for data centers that go beyond the duration of the customer contracts. Indeed, the company requires more computing capacity than it is currently able to provide.
• Anthropic also wants to distance itself contractually from damages that its agent may create. It also warns that the insurance industry may not be able or willing to cover damages.

This MIT Technology Review opinion article by Sam Hiner, executive director of the Young People’s Alliance in the US, argues that the current focus on extreme AI harms like the cyberattacks by agents is taking attention away from the more common and insidious harms.

• He cites examples such as students who become reliant on AI to write homework assignments and young people who measure their lives against social media feeds that are engineered to provoke envy and to keep watching. In these cases, the products meet current legal safety standards but leave us worse off culturally.
• Big Tech has avoided addressing these concerns because it would create friction to the adoption of their technologies.
• The companies are accused of “cultural engineering”. People now spend an average of over two hours per day on social media. This is “time that has been quietly taken from ordinary life and handed over to companies whose profits depend on keeping us there”.
• Cultural engineering on social media platforms has contributed to violence: the Rohingya genocide in Myanmar, anti-Muslim riots in Sri Lanka, and alleged calls for violence during Ethiopia’s Tigray war.
• An MIT study found that young people using ChatGPT to write essays showed weaker brain connectivity from the task, and that “cognitive debt” persisted even after the task was complete.

Researchers at Microsoft have developed an AI system called the Microsoft AI Diagnostic Orchestrator (MAI-DxO), that they claim has higher diagnostic accuracy than human generalist physicians. The tests were run on complex medical cases and emergency department patient records.

• Results published suggest that the AI system correctly diagnosed 83.2% of complex cases, compared with 36.1% for human physicians.
• MAI-DxO uses a chained reasoning approach which evaluates the patient through a series of questions. The tool maintains several possible diagnoses and finally eliminates possible candidates based on the results of the tests it orders.
• In the tests, the AI system asked an average of 25.5 questions before ordering its first diagnostic test, compared with only 5.6 questions for the participating physicians.
• On the other hand, consultation costs were higher for the AI system than for the physicians: 4979 USD compared to 3400 USD for emergency room cases.
• Another caveat is that only diagnostic accuracy was measured. Treatment decisions, patient safety and clinical outcomes are not measured. The results therefore do not provide evidence about whether an AI physician could improve patient care, or outperform a human physician in clinical practice.

Reuters reports how investors are evolving in their attitudes to spending by AI companies. Those lacking significant cash flows are being punished by investors. Microsoft is seen as faring well, Google is faring poorly.

• Google’s search revenue grew by 12% in the third quarter but was expected to grow by 14.7%. One reason for the fall is that people are increasingly looking to AI chatbots for information on the Web.
• Google’s Cloud revenue more than doubled to 30.71 billion USD but the company’s overall cashflow was negative with a 9.07 billion USD cash burn.
• Microsoft generated 18.6 billion USD in cash flow in its fiscal first quarter, which was 10 billion USD more than expected. At the same time, it spent 45.2 billion USD on its IT infrastructure. From an AI perspective, Microsoft has a large number of customers to help fund its AI development.
• For one market analyst, AI has become a “sport of kings, which is that if you don't have 150 billion USD to spend next year on capital expenditure, don't even bother showing up”.

The chipmaker Intel is still facing issues as it lost 28% on its stock price in one day of trading after the announcement that its expected revenue for Q4 of 2026 is 15 billion USD instead of an anticipated 16 billion USD.

• The US government took a 10% stake in Intel last year, and the fall in share price corresponds to a loss of 12 billion USD. On the other hand, the stock is still up 70% compared to this time last year.
• The main challenge for Intel is that PC sales are down, largely due to increase in PC prices which in turn is due to supply chain shortages as AI companies buy up chips for their data center construction.
• On the bright side for Intel, its data center business has grown 78% since this time last year. However, it is currently lacking the capacity to keep growing in this area.

As AI coding agents are now seen to be more prolific and efficient at coding that humans, this InfoWorld article gives reasons why humans are better than agents in software projects.

• Humans understand context: only humans can have customer empathy, which is where real understanding of user requirements comes from.
• Only humans can be held accountable for bad software, so there has to be a human in charge of the project.
• Only humans can say “no” to a feature requests: AI agents tend to follow instructions without critical reflection. Further, only humans can initiate projects or come up with new features.
• Only humans have taste. In the case of building a Web site for instance, web sites created by AI agents tend to resemble each other.