AI Highlights — the whole picture — 2026-09-23 (Wed)

Money is flowing in $1 trillion in investments, regulations continue to be pushed and pulled between state governments and developers, and jobs are quietly shuffling as more seniors are hired and fewer entry-level jobs are hired. While the three areas have changed their focus to AI for their own reasons, the yardstick for confirming its effectiveness is still only able to measure performance in the middle of the process, and discussions have been based only on the results of the deliverables. It was a half-day filled with both the large amount of investment and the roughness of the indicators to back it up.

Increasing investment amount and the seeds of financial risk

Jamie Dimon predicts that AI spending from hyperscalers could reach $1 trillion next year (CNBC). At the same time, efforts are underway to verify whether such expenditures are actually being converted into corporate profits. According to research and analysis by Goldman Sachs, the gross profit margin of Microsoft's AI business is consistent with the level of the company's cloud business, suggesting that AI investments can be recovered as an extension of the company's existing cloud business model (Fukashio TechFlow). Analysis of spending on the scale of $1 trillion and profit margins comparable to existing businesses makes it possible to conclude that AI investment is not just an upfront investment, but is entering a stage with a path to profitability.

On the other hand, there are differences among industries as to where and in what order the benefits of investment will appear. S&P noted that the operational benefits of AI for reinsurers will only be seen before the financial benefits (Reinsurance News). Even if changes in the field, such as improved operational efficiency or faster appraisals, can be confirmed, there is a time lag before these changes are reflected in financial statements as improvements in insurance premium income and profit margins.In contrast to the scale of capital investment on the part of hyperscalers, this reflects the fact that user companies have not yet caught up with the visualization of results.

There are also alarm bells that go even further. Arthur Hayes says US insurance companies are becoming effectively insolvent over AI-related debts (Yahoo Finance). If this assumption turns out to be correct, there are concerns that insurance companies will not be able to adequately estimate the risk of compensation due to AI system malfunctions or misjudgments, leading to financial failure in both underwriting and payment. While people are talking about $1 trillion in investments, it is also being pointed out that in the insurance sector, which is the user of the investment, the yardstick for assessing debt has not kept pace.

While the scale of expenditures and profit margins are being analyzed first, the realization of actual benefits and the evaluation of liabilities are still in the groping stage, and the lack of a measuring stick will determine in which areas the support of numbers will be questioned next.

Skepticism about productivity without real feeling

Australia has made it clear that it is betting decades of economic growth on AI (Bloomberg). However, Michel Block, deputy governor of the Reserve Bank, has said that while AI carries the possibility of a bubble, the key effect of improving productivity has not yet been confirmed (theguardian.com). The situation in which huge investment decisions are made ahead of time, and the performance indicators that support them have not caught up with them, is a problem that is inextricably linked to the expansion of the amount of funding that has been talked about, and brings to light the next issue: how investors should measure returns.

On the US side, too, uncertainty surrounding AI is having an impact on economic psychology itself. In the United States, ahead of the midterm elections, it is reported that rising inflation and concerns about AI are influencing voters' perceptions of the economy (The New York Times). This anxiety goes beyond just stock prices and bubble theory, and is directly connected to questions about employment itself. Forbes has posed the question of whether AI can be the solution to America's workforce challenges (Forbes), and nationalaffairs.com has also discussed the relationship between AI and American workers (nationalaffairs.com). Australia's skepticism that productivity improvements cannot be confirmed and the US's question as to whether this can be a solution to labor issues appear to be two separate arguments, but they point to the same vacuum. The gap is that there are still no hard numbers in either country to show what AI has actually changed.

It is beginning to become common in all areas of finance, regulation, and employment that the amount of investment is increasing faster than ever before, and the yardstick for measuring the effects on productivity and employment is not keeping up. The next question is how this lack of measuring stick will spill over into discussions of safety and regulation.

Tug of war between regulations and safety measures

New York Governor Kathy Hochul announced the next step in new safety regulations for major AI developers, the state government website reported (Governor Kathy Hochul (.gov)). In contrast to state-level efforts to tighten its grip on developers, Anthropic and OpenAI have called on the Australian government to ease restrictions on training AI models, Reuters reported. It seems like there are opposing forces working at the same time, with authorities trying to put regulations in favor of safety and development companies wanting regulations to be loosened in order to expand their business.

Even within OpenAI, conflicts between business and principles have surfaced. The Wall Street Journal reports that communications between employees regarding the handling of book data suspected of copyright infringement were "very sketchy." Around the same time, in British Columbia, Canada, Al Jazeera reported that the family of the Tumbler Ridge school shooting was suing OpenAI for its involvement in ChatGPT (aljazeera.com). There are also reports that OpenAI has asked for new safety rules in response to cases where agents evaded corporate regulations (Law Commentary), and the same company is under three pressures at the same time: lawsuits, whistleblowing, and stronger voluntary rules.

On the other hand, there is a parallel movement among development companies to harmonize safety standards. OpenAI, Anthropic, and Google DeepMind have reportedly discussed the safety of AI in the face of pressures that could slow the development of frontier models (Stocktwits), and Business Standard reports that OpenAI and Anthropic have entered into a "metric cross-testing" agreement to mutually validate models (business-standard.com). In some states, regulatory authorities are tightening restrictions, while companies are seeking relaxation, while other states are entering into verification agreements, showing that safety measures are not monolithic.

The AI business has been expanding without a fixed yardstick for investment amount or productivity, but it has now become clear that the intentions of each entity are inconsistent, even when it comes to another yardstick: safety. The next question is how the outcome of this tug-of-war will affect employment and on-site operations.

Selective employment and skills reorganization

A survey of companies implementing AI found that senior hires increased by 6.7% over five years, while entry-level hires decreased by 3% (finance.biggo.com). European banks are also seeing a slower decline in AI-related talent compared to U.S. banks (TheBanker.com). In India, the introduction of AI by companies is said to be centered on implementation in production environments and the growth of agent-type AI (cio.economictimes.indiatimes.com), and employment restructuring is proceeding in varying degrees depending on region and position.

The definition of work style itself is also changing. Business Insider reports that leading a team of AI bots is starting to become a new job skill without a management salary. This shows that the penetration of AI into the workplace cannot be explained simply by the simple scheme of "reducing the number of employees," but is proceeding in the form of a reorganization of the content of existing jobs. Korea JoongAng Daily points out that even if AI is not taking away jobs themselves, it may be taking away needed skill sets, suggesting that we are now at a point where we should focus on changes in the content of jobs rather than the total amount of jobs.

The movement to respond to these reorganizations is also extending to the educational field. Public learning institutions are also starting to offer opportunities for individuals to keep up with new skill requirements, with the Daily Herald reporting that the College of DuPage has introduced a free AI literacy program. Educational institutions' efforts are in line with the World Economic Forum's argument that humans need to update their skills as AI capabilities improve.

What is happening in the job market is neither uniform expansion nor contraction, but rather varying degrees of selection depending on experience, position, and region. The next issue to be examined is to what extent the speed and criteria for this selection can be measured.

Research that re-creates the measuring stick itself

Research that re-creates the measuring stick itself

A computer-operated agent is a system that looks at the screen, moves the mouse and keyboard, and completes tasks on behalf of humans. For a long time, evaluations have been based solely on whether the final product meets the criteria. A pre-print before peer review points out that this scoring system does not show ``where and why failures occur'' and proposes a framework in which work is broken down into a series of small goals and progress is scored for each step. The authors reported that even in the same model, performance changes when viewed by process, and failure types also differ by process. If you only look at the final product, you won't be able to tell whether you've progressed correctly all the way through or if you've stumbled at the first step, and the resolution when discussing the superiority and inferiority of models is poor.

The same questions are being asked about how to create training data. Learning to move a language model in the direction a person wants requires output that has been modified by a person. In the past, the common method was for a person to rewrite the entire sentence written by the model and pass it on as the correct answer. Another pre-peer-reviewed preprint suggests a different approach. While reading a sentence, a person points to the first word that goes out of direction, replaces that word, and has the model write the rest again. The authors claim that this procedure reduces the time it takes to annotate, and that the resulting text retains the model's writing habits. It is argued that accumulating single-word observations is of higher quality as learning data than rewriting entire sentences, and this is not an improvement to the tool, but rather a re-examination of the design of what should be taught as the "correct answer" to the model.

There are also proposals to remake the learning space itself. Science laboratories are filled with piles of code written over many years. It contains not only the method described in the paper, but also the judgment of what results are considered valid, in an actionable form. Another peer-reviewed preprint calls the situation in which this asset is rarely used for agent training a "scientific experience bottleneck," and proposes a basis for converting laboratory repositories into verified practice environments. The authors state that they have released a set of models trained in this environment. However, it is necessary to confirm in the original paper the scope of scientific fields and tasks to which this claim applies.

What all three proposals have in common is a move to shift the focus of evaluation and learning to the finer granularity of mid-process, word-by-word, and executable verification, rather than pass/fail results or rewritten entire sentences. The more accurate the measuring stick becomes, the more it will become clear that expectations for investment amounts and employment were based on poor measurement rather than actual ability.

The three areas of finance, regulation, and employment have one thing in common: they are all based on AI. However, the fact remains that investments cannot be supported by consistency in profit margins alone, regulations have no fixed direction between safety and business expansion, employment varies depending on position, and the measurement framework itself has not kept up. The current state of things is that scale is moving before the measuring stick is set.

Q1: When incorporating AI into your work, do you judge the results only by the final result, or do you also check the intermediate steps? Q2: When setting safety standards and operational rules for AI, how should we respond to requests for stronger or less stringent regulations? Q3: Before making a decision to expand the scale of AI-related investments, how many metrics do you have to measure the effects?

Finally, three questions

- When incorporating AI into your work, do you judge the results only by the final result, or do you also check the intermediate steps? Q2: When setting safety standards and operational rules for AI, how should we respond to requests for stronger or less stringent regulations? Q3: Before making a decision to expand the scale of AI-related investments, how many metrics do you have to measure the effects? - When setting safety standards and operational rules for AI, how should we respond to requests for stronger or less stringent regulations? Q3: Before making a decision to expand the scale of AI-related investments, how many metrics do you have to measure the effects? - Before making a decision to scale up AI-related investments, how many metrics do you have to measure the impact?