AI Highlights — the whole picture — 2026-10-30 (Fri) Daily News

Huge sums of money are being made, and at the same time, announcements are being made to reduce the number of workers. For example, the bank and Blackstone provided a $63 billion line of credit for Broadcom's AI technology. On the other hand, it is difficult to tell from statistics how much that investment has increased production and income. The contract amounts and numbers communicated as statements only indicate the movement of funds or someone's point of view, but do not confirm the results. Today's articles cover grid connectivity, diagnostic support, emotional products, and regulations to protect children. The common question in each story is not how quickly it was implemented, but who verified the results and how. I would like readers to look at the companies and systems in each section from this perspective.

A huge amount of money and a decreasing number of workers

Banks and Blackstone have set up a $63 billion line of credit for Broadcom's AI technology, according to WSJ (WSJ). Meanwhile, Bloomberg discusses why billions of dollars in AI investment are hard to see in GDP statistics (Bloomberg.com). The scale of funds can be confirmed by the amount of contracts such as loan lines, but it is difficult to see from statistics how much the investment has increased production and income. Readers should therefore treat the $63 billion figure as separate from evidence of success. The fact that the amount was large can only be seen by the movement of funds.

At the same time, there has been ongoing talk on the labor side that assumes a decline in the number of workers. The Verge reported that Microsoft's AI future will have fewer workers. Qz.com lists 15 companies that will cut staff due to AI in 2026 (qz.com). Silicon UK argues that beyond the organizational chart, the question is how to manage the coexistence of people and AI agents (Silicon UK). Even if AI is cited as a reason for reductions, it is difficult to tell just from a list how much work has been replaced at the company and how the jobs of those who remain have changed. If the number of workers decreases without a management method in place, the impact will be felt first on the ground.

In the pharmaceutical sales field, attempts are being made to change roles rather than reduce the number of people. Inizio's Nick Sindercom said that 27 of the world's largest pharmaceutical companies are piloting STEM AI that turns sales calls into coaching (BigGo Finance). What this statement confirms is that 27 companies have started testing. It is not clear from the material presented whether the conversion to coaching improved the quality or outcomes of the visits. It is appropriate to read this as a statement from Inizio's side, not an employee's.

Funds go first, personnel reductions and role changes proceed in parallel, and results are deferred until later. If things continue in this order, the only way to decide how much to believe each company's announcements will be based on whether or not they have been verified.

Electricity and competition between nations

The UK's energy minister has announced that major data centers will be given priority grid connections (Reuters). Governments are starting to decide the order in which computing resources are connected to the power grid, rather than where they are placed. Around the same time, a former Google CEO warned that there was a six-month difference between China's AI and the US's (Android Headlines). This number was reported as a statement, and no statistics were provided to measure the difference. For Argentina, there is an article (Americas Quarterly) that discusses what is needed to realize the AI vision. Having a vision and actually arranging the equipment and funds are two different stages. The speed of implementation in each country can be determined by the extent to which they can secure the latter.

On the industrial side, competition to secure supplies is also manifesting itself in different forms. Daeong Pharmaceutical will partner with Higashinada City Hospital to build a smart ward at the Seoul AI Expo (BigGo Finance). Catalan biotech company Allox has raised €9.4M to power its AI drug discovery platform (Dealroom). The former controls front-line partners such as hospitals, and the latter controls funding. Although what they are getting is different, both are examples of first setting the conditions for movement. However, the scope of these articles does not show what kind of results were achieved in the hospital wards or what the drug discovery platform produced.

Priority quotas for electricity, alarming differences between countries, partnerships, and procurement are all indicators of who prepared first. The speed of introduction can be explained by the availability of equipment and funds, but this alone does not mean that the results have been confirmed. The next thing to look at is to what extent the introduction is backed up by statistics and on-site verification.

The question of whether you can trust the judge

To what extent can we trust the answers of generated AI used at sites where radio waves do not reach? A preprint (on their own website) that deals with this question proposes a system that divides the connection status into stages and updates them with signed "knowledge packs." The verification was performed by running a completely local RAG on a Raspberry Pi 5. The authors report that properly signed poisoned packs passed all pack checks. The signature guarantees that the pack has not been tampered with and even knows who created it. You cannot tell from the signature whether the contents are true or not. The authors themselves conclude that they do not confirm the truth of the basis for any of the mechanisms they evaluated.

There are other examples where the judge is wrong. The preprint (own site) proposing AgentProofBench measures a frozen verifier that determines whether an agent's work is completed according to requirements. It is designed to measure both testing with changed requirements and execution from the same starting state. The authors report that when a verifier has no errors in judgment, the results are almost always correct, but when a verifier has errors, the results tend to collapse. However, this is an association, not a cause and effect. The FORK reward preprint (own site) deals with questions that require longer answers, unnecessary descriptions, and more errors in order to meet the scoring criteria. In addition to coverage, the authors report that the accuracy of each answer by asking whether it is necessary or correct is included in the reward, and the winning rate over the base model in medical evaluations has significantly increased. All three cases were pre-peer-review reports, showing that unless a grading system is in place, a high score does not necessarily mean a good answer.

The idea that measurement methods determine reliability also comes up in chemistry and pharmaceutical topics. Scale AI has released an index called "DualChem" to measure dual-use risks in the chemical field (Scale AI). Only when we have a measuring stick in place can we compare the magnitude of risks. Meanwhile, Anthropic announced that, as a result of its collaboration with six pharmaceutical companies, it has reduced the time required to review regulatory documents by 89% and detected human oversights (AI Utilization Guide for Small and Medium Enterprises in Kurume and Fukuoka). This 89% figure can only be used as a basis for deciding whether to introduce it or not, only after confirming what the 89% value was calculated against and what criteria it was calculated against.

Therefore, what readers should look at is whether or not the judgment mechanism that produced the numbers has been verified, rather than the results themselves. Are huge amounts of funds and personnel allocated based on unverified numbers? In the next section, we will explore clues to confirm this point.

Accuracy of diagnostic support and information for patients

The U.S. Food and Drug Administration (FDA) has approved AI software (Diagnostic Imaging) that makes CTA images of head and neck arteries easier to see. The product has been reviewed by regulatory authorities and is positioned to help improve images. Meanwhile, in cancer research, reports of models that predict the future based on images and pathology have been published one after another. A study has been published that predicts the cancer risk of patients with Barrett's esophagus by spatially profiling immune and molecular characteristics from histopathological images (Cancer Research). In breast cancer, there is a study comparing a machine learning model that uses MRI radiomics to predict pathological complete response after neoadjuvant chemotherapy (BMC Medical Imaging). In the case of glioblastoma, there has been a report that predicts recurrence on a pixel-by-pixel basis using MRI and deep learning during treatment (Cancers).

There are also studies that predict response to treatment. For high-grade serous ovarian cancer, conventional histopathology slides were analyzed using deep learning to predict response to PARP inhibitors (Apollo). A radio-transcriptomic signature was shown to predict response to atezolizumab plus bevacizumab on a nodule-by-node basis (Liver Cancer). For prostate cancer complicated by type 2 diabetes, multi-omics research has been published that uses machine learning to identify aging-related genes and focuses on SOX4, PRNP, and FOXO4 (Artificial cells, nanomedicine, and biotechnology). These methods have in common that they attempt to extract information related to treatment strategies from existing tests such as images and pathology. However, the title only indicates that the model made a prediction, and it is not clear from this material whether outcomes have improved in actual practice.

There are also reports showing that forecasts do not always lead to upside. Internal evaluation of a random forest model for predicting clinically problematic postoperative pancreatic fistulas after pancreaticoduodenectomy found that the addition of inflammatory and nutritional markers had limited additive effect (Journal of Clinical Medicine). Increasing the number of variables does not necessarily increase accuracy, and this is an internal evaluation. This report alone cannot confirm whether the results would be the same in another facility or patient group.

Evaluation has also begun in providing information to patients. A cross-sectional methodological study has been published that evaluated information provided to patients by an AI chatbot regarding intravesical BCG therapy for bladder cancer (Yeni Üroloji Dergisi). There are real situations in which patients ask questions to chatbots before and after a medical examination, and we are currently in the process of developing a method to measure whether the answers are accurate.

The clues for readers when making a decision are whether there is approval or not, whether there is external verification, and whether the additional effects are measured. Rather than looking at numbers that indicate high accuracy, we need to look at the conditions under which those numbers were obtained. In the next section, we will look at how these confirmation methods are handled by workers.

AI that touches emotions and protecting children

An AI chatbot evaluation tool that measures responses to adolescent mental illness was featured in Nature. Research on AI that aims to read users' hidden emotions and provide safer mental health support has also been reported (Bioengineer.org). The science behind the popularity of AI lovers is also explained (Technology Org). What all three have in common is that AI is moving from providing information to influencing emotions. Products related to emotions should be looked at not so much for convenience as for whether they can respond appropriately in a dangerous situation. The advent of assessment tools indicates that a means to externally measure that response is beginning to take shape. However, there is a difference between having a measuring tool and having each company use it and publish the results. It is the latter that readers want to confirm.

The system also moved. New York Governor signs into law to regulate AI chatbots to protect children (New York Daily News). The decision was made to stipulate the scope of protection by law, rather than leaving it up to companies to self-report. The signature is the starting point, and we would like to see how the effectiveness will be measured after implementation. Regulation of AI that touches on emotions has meaning only when considered in conjunction with evaluation methods.

The user's understanding is also being questioned. The Pennsylvania State University is developing AI literacy in a summer program for freshmen. An article (Her Campus) that explores the gender bias hidden in artificial intelligence shows that it is necessary to use it with the assumption that the output is not necessarily neutral. Author Janet Winterson says no dystopia is inevitable (The Guardian). If the future is not decided, the outcome will be determined by the extent to which users understand it and how the system prepares for it.

There are also situations where AI can produce results. Police find a painting stolen 20 years ago with the help of AI (BBC). I'm sure there are effective uses for it. However, one discovery shows that it was actually usable. It does not go so far as to show that safety has been confirmed in situations involving emotions and children. As the number of examples of convenience increases, the difference between companies and systems will depend on whether or not they have procedures for checking. The same question can be asked in the following fields.

In each of the stories covered in today's section, there were two separate materials: one that showed the scale and speed of introduction, and one that confirmed the results. The amount of the loan facility, the list of layoffs, the priority line of connections, the approved imaging software, and the signing of regulatory legislation are all records of action. On the other hand, as shown in examples of signed knowledge packs passing verification even if they were poisoned, and attempts to measure the reliability of the discriminator itself, the fact that the verification mechanism passed is not the same as the contents being correct. It is only when external verification methods such as statistics, on-site verification, and publication of results are available that results turn from claims to fact. Therefore, when looking at companies and systems, readers should base their criteria on how reliable the verification is and how open they are, rather than the speed of introduction.

Q1: Are there any records left at the site to confirm the results of the introduced AI compared to the state before introduction? Q2: How do you separately check that the verification process has been passed and that the content of the output is correct? Q3: Who verified the results figures used in making investment and personnel decisions and by what method?

Finally, three questions

- Are there any records left at the site of comparing the results of the introduced AI with the state before introduction? Q2: How do you separately check that the verification process has been passed and that the content of the output is correct? Q3: Who verified the results figures used in making investment and personnel decisions and by what method? - How do you separately check that the verification process has passed and that the output is correct? Q3: Who verified the results figures used in making investment and personnel decisions and by what method? - Who verified the results figures used in investment and personnel decisions and by what method?