Artificial Intelligence inMedicine, Cybersecurity, Automotive & Society

New proposals

Request a research critique

Requests are considered by the editorial team.

R0:1908369e19bd02a491f28cff06766ff7-TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving

TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving

How much of a theoretical physicist's calculational work can an AI agent perform without guidance? TheoPhysEval puts five agents to work on 50 research-level problems spanning 11 areas of physics, from classical dynamics to quantum field theory and gravity. The strongest tested configuration, GPT-6-Astra with Codex, obtains accepted final answers on 49 problems with Internet access and 46 without it. The study offers a detailed view of what these systems can calculate, where they fail and how much their use costs. Its results also raise a deeper question: what distinguishes a correct answer from a trustworthy scientific contribution? Each configuration was evaluated once, the principal scores concern final answers, and the experiment does not include a matched comparison with human specialists. Those boundaries make the paper particularly useful for understanding both the promise of AI-assisted theoretical research and the further evidence needed before calculation can be equated with autonomous discovery.
R0:184d5f85eb7a86194894ef30d0f8bf57-Organizing Intelligence

Organizing Intelligence

A powerful AI model does not automatically make a powerful organization. In Organizing Intelligence, Alex Imas, Andrew Koh and Shane Legg ask how human and artificial capabilities combine inside firms, and why coordination may matter as much as individual expertise. Their theoretical model separates improvements in task competence from improvements in working together. That distinction changes the possible paths of firm size, market concentration, wages and employment. Complementary human skills can sustain productive teams even when AI outperforms the average individual; stronger AI coordination can, under specified conditions, reorganize production around much larger firms. The paper also distinguishes gradual replacement within mixed teams from abrupt displacement when AI firms outcompete human firms. These are conditional economic mechanisms, illustrated by mathematical examples rather than forecasts of when jobs will disappear. The study offers a demanding question for AI evaluation: how should we measure the capabilities of an organization, rather than those of a model working alone?
R0:e81127dd3f66450ac34b82915cf63584-An ECG biomarker for sudden cardiac death discovered with deep learning

An ECG biomarker for sudden cardiac death discovered with deep learning

A routine electrocardiogram can reveal a disturbance in rhythm while it is happening. Could the shape of an apparently ordinary beat also help identify someone at risk of sudden cardiac death a year later? Ziad Obermeyer and colleagues linked Swedish ECGs to death certificates, then examined what their model had learned. When they altered a beat to raise its predicted risk, the descending edge of one wave became less abrupt in lead aVL. That observation gave them a feature they could measure in actual recordings. The Swedish results show both the promise and the limits of the approach. A small group had a markedly elevated event rate, yet many ECGs preceding a death fell below the chosen threshold. Records from the United States and Taiwan add evidence about ventricular arrhythmias and cardiac arrest. Whether offering defibrillators to people identified through this signal would prolong their lives remains an open clinical question.
R0:0fca8348a9f334da8f8573081f02e921-Autonomous robotic operation controlled by wave-based neuromorphic hardware

Autonomous robotic operation controlled by wave-based neuromorphic hardware

A small robotic car approaches an obstacle. Its distance readings travel by Bluetooth to a dish of water, where six actuators set waves in motion. A camera captures the interference pattern, a classifier reads it and the car receives an instruction to turn. Jacob Zohar and colleagues used this arrangement to demonstrate that a physical wave field can participate in the control of a moving vehicle. Their experiment raises a question that matters whenever a machine acts in the world. Could an independent observer explain why it turned at a particular moment? The published route shows what the car did, and the archived five-class examples show which obstacle configurations that separate classification test included and which it left out. Reconstructing an individual decision requires the readings, images and motor command from the same moment. The distinction also matters for the proposed magnetic successor to the water reservoir, whose obstacle recognition the authors studied through micromagnetic simulation.

IntelligenceNeurosymbolic

21 Nov 2026

An AI finds an equation. What makes it a physical law?

An equation can fit a set of numbers and still fail to describe the physical phenomenon behind them. This article follows an ideal pendulum through two separate questions: whether a proposed expression has compatible units and whether its predictions agree with the evidence. An original example shows why dimensional rules can reject some formulas while admitting others that predict different periods. Research on symbolic regression then explains how learned proposals and explicit constraints can cooperate during the search itself. The…

Read the article →

20 Nov 2026

From a medical report to a checked record: what can rules verify?

A report can mention a finding without saying that it is present now. It may describe an earlier examination, deny an observation or leave a question unresolved. Turning those sentences into a database requires more than finding familiar words. This article follows an original, fictional example through a proposed extraction, explicit checks and a decision to retain uncertainty. A clinical research study provides a bounded comparison, while standards for data validation and provenance help distinguish a well-formed record from a…

Read the article →

Browse the complete collection →

R0:aa77bca17589f025d763d708bbb99028-Can LLMs Improve Their Own Harness? A Controlled Study on ARC-AGI-3

Can LLMs Improve Their Own Harness? A Controlled Study on ARC-AGI-3

If a language model can rewrite the software it uses to play a game, how can it tell whether an edit was worth keeping? Jens Lehmann and Sahar Vahdati let each model build a program, play five practice games and revise that program twice. They then compared all three versions on twenty games whose results had been withheld during development. Across 27 independent sequences, the practice scores rose by 8.42 points on average. On the reserved games, the gain shrank to 2.80 points, and twelve sequences finished below their starting score. Cost tells a similarly mixed story. Eighteen final programs were cheaper to run, but only nine of those also played better. By preserving the original programs and their descendants, the study reveals something a final leaderboard cannot show on its own: when an agent already plays well, how much of that ability did it acquire through revision, and how much was present from the start?
R0:abbc9fad0e8a3823e93438ffbc99b607-FrontierPhysics: Evaluating Agents for End-to-End Physics Research

FrontierPhysics: Evaluating Agents for End-to-End Physics Research

Can a research agent tell when its own calculation has led it astray? In FrontierPhysics, physicists turned investigations from their own laboratories and research groups into 56 problems. An agent receives a brief, explores the literature, chooses a method and submits a manuscript alongside results that an examiner can check. The strongest configuration passed 46 of 168 trials. It solved 25 problems at least once in three attempts, but only eight on every attempt. The most revealing findings emerged when physicists revisited the experimental records. Some agents had checked a result using the very assumption that produced it. In other cases, a sound answer had failed the examiner’s test. Nine such judgements changed in the detailed audit of one configuration. Together, the repeated trials and the expert review illuminate a practical question for any physicist considering an agent’s work. What evidence would persuade another researcher that the result survives a genuinely independent check?
R0:6c8a4ce33496b8be6ac636aa4dfd0fef-Proving and Improving the Platonic Representation Hypothesis with JEPAs

Proving and Improving the Platonic Representation Hypothesis with JEPAs

Can a compact text encoder learn representations that agree with vision more closely than those of a much larger language model? This preprint investigates that question through a conditional mathematical theorem and experiments comparing Joint-Embedding Predictive Architectures (JEPAs) with next-token prediction. Its central proposal is to train on different descriptions of the same scene, encouraging models to preserve shared information. A 109-million-parameter text JEPA achieves stronger visual alignment than the evaluated language models, including models with up to 72 billion parameters. The reported gains concern representation geometry, retrieval and selected classification tasks. The paper also combines a JEPA with an LLM decoder to generate image descriptions. This publication examines those contributions alongside the assumptions of the theorem, the meaning of the reported sample-efficiency advantage, the role of generated training captions and the limits of comparisons with general-purpose language models. The findings suggest a practical direction for representation learning while leaving broader claims about reasoning, planning and autonomous world models open.
R0:36cbb300eb0fca3fc0ffa717d93011b5-Explainable deep learning improves human mental models of self-driving cars

Explainable deep learning improves human mental models of self-driving cars

How can a passenger understand why a self-driving car behaves unexpectedly? This study introduces the Concept-Wrapper Network (CW-Net), which connects a learned planner to human-readable concepts and examines whether explanations improve understanding of vehicle behaviour. Its evaluation combines private-track experiments with a real autonomous vehicle, expert and non-expert studies, and an online study using public-road recordings collected while a human drove. Participants receiving concept explanations performed better on selected questions about surprising situations. The accompanying critique distinguishes the architecture in which concepts determine the planning reward from an alternative that provides diagnostic explanations without determining trajectory selection. It also examines concept-recognition errors, the inclusion of caution questions in a prediction measure, and uncertainty about effects during ordinary situations. The publication offers evidence about interpretable planning and human understanding, while explaining why these results do not establish safer autonomous operation on public roads or unrestricted replication of the complete vehicle system.

Browse all proposals, including earlier publications →

Deterministic Governance of Artificial Intelligence

Research on visual reasoning, evidence from model evaluations, measured hardware behaviour and controlled cybersecurity trials: these four resources document how claims are examined within the Sistema Vectorial (SV). Readers can consult methods, results and limitations, with ongoing work distinguished from demonstrated findings.

These panels provide context for research into the development of artificial intelligence and the companies supplying its models, chips and computing infrastructure. Read alongside the journal’s commentary, market movements can help frame questions about expectations, financing and industrial concentration. They do not measure technical progress or society’s acceptance of AI: prices also reflect other businesses and wider economic conditions. Our purpose is to support critical investigation, not investment decisions.

US listings · USD · Quotes may be delayed

AI & Hardware — Stock Market

NVIDIA, Alphabet, Intel, AMD, Microsoft, Broadcom, Arm, TSMC, Oracle, Amazon, SpaceX, ASML and Anthropic: share prices, daily changes and charts for the AI and computing industry. Select a company to explore its market information.

Market data and charts by TradingView. Data loads directly from the provider when this section comes into view. Quotes may be delayed; when markets are closed, the latest available session remains visible. Refresh the page to reconnect.

The US panel above shows dollar quotations. TSM represents TSMC’s US-listed depositary shares (ADR). These market indicators provide economic context, not a ranking of AI capabilities or investment advice.

SpaceX includes xAI; its share price reflects the wider group, including its space and connectivity businesses. Alphabet is shown as GOOGL, providing exposure to Google and Gemini within another diversified group.

Why SpaceX matters here — Musk, Terafab and global reach

One listed group, several businesses. SpaceX began trading as NASDAQ: SPCX on 12 June 2026. Its acquisition of xAI was completed on 2 February 2026. The quotation is therefore for SpaceX shares, not a separate xAI security. Listing information · Quarterly filing, 30 September 2026.

Elon Musk is the central decision-maker. As founder, chief executive, chief technical officer and chairman, his role combines corporate direction with engineering leadership. The offering documents describe a dual-class share structure concentrating voting control in him, with limited influence for ordinary public shareholders. Governance and risks in the offering documents.

Compute, models and connectivity. Colossus supplies large-scale AI-training infrastructure; Grok is the group’s frontier language-model family. Starlink adds a global communications network: the September filing reports service across 174 countries, territories and other markets. These are distinct capabilities within the group, rather than proof that every service already operates as one integrated product. Colossus · Group segments and reach.

Terafab and the supply of AI chips. Colossus uses chips to train models. With Terafab, SpaceX and Tesla aim to manufacture chips for AI on Earth and for future data centres in orbit. Intel joined the project in April 2026. The partners plan to bring logic chips, memory and advanced packaging together at a single site. Terafab project · SpaceX filing on the partnership.

Tesla describes its ambition as building the largest chip factory in history. In August 2026, Texas announced Grimes County as the site and put planned investment for the first phase at more than $16.8 billion. The long-term target is to make enough new chips each year to equip one terawatt of computing hardware. Neither the size record nor that output has been demonstrated. Tesla’s description · Texas announcement.

The strategic question. Combining computing capacity, AI models, applications, satellite connectivity and launch capability could reduce some dependencies and support coordinated deployment. It could also concentrate access to infrastructure and bargaining power. This is an interpretation of the group’s structure, not a forecast of market dominance; energy, spectrum, regulation, capital requirements and competing networks constrain what can be achieved.

A historical comparison, with limits. Rockefeller’s Standard Oil combined cost reduction and operational integration with acquisitions and favourable railway rates. The Library of Congress reports that it controlled about 90% of American refineries by the late 1880s. The useful parallel is the relationship between scale, distribution and concentration. It does not establish that SpaceX follows the same pricing strategy, holds an equivalent market share or has committed the same antitrust violations. Standard Oil: historical context.

Frontier AI is not confined to this stock. Alphabet also develops frontier models through Google DeepMind’s Gemini programme. Neither SPCX nor GOOGL isolates the value or performance of its AI models. Google DeepMind: Gemini.

Corporate context reviewed 13 November 2026. Terafab sources checked 14 November 2026. This date concerns the explanation above; quotations in the interactive panel are supplied separately by TradingView.

AI Markets and Models in Asia

Alibaba and Baidu develop the Qwen and ERNIE models in China. India’s Tech Mahindra works on models for Indian languages. On the hardware side, follow Cambricon’s accelerators, Sony’s vision sensors in Japan, ASE’s chip packaging in Taiwan and SK hynix’s high-bandwidth memory in South Korea.

Market data by TradingView. Shanghai (CNY) and Bombay (INR) show end-of-day prices. The US listings are depositary receipts in USD and may be delayed. Each price reflects the entire company.

Company guide, sources and Hong Kong listings
Company and market of originRole in AI · company sourceChart listing
AlibabaChinaQwen language modelsNYSE:BABAUSD
BaiduChinaERNIE foundation modelsNASDAQ:BIDUUSD
Tech MahindraIndiaProject Indus for Indian languagesBSE:TECHMINR
CambriconChinaMLU accelerators for AI computingSSE:688256CNY
Sony GroupJapanIMX500 image sensors that run AI on the sensorNYSE:SONYUSD
ASE Technology HoldingTaiwanAdvanced packaging that connects chips in AI systemsNYSE:ASXUSD
SK hynixSouth KoreaHigh-bandwidth memory for AI processorsNASDAQ:SKHYUSD

Two more model developers trade in Hong Kong dollars. Zhipu (2513) develops the GLM family. MiniMax (0100) develops multimodal models. Their charts open on TradingView.

Semiconductor sector · NASDAQ: SMH · USD

Semiconductor market

VanEck Semiconductor ETF (SMH) brings together US-listed semiconductor and semiconductor-equipment companies, including international firms. An ETF is an exchange-traded fund: its share price offers a view of a basket of companies rather than a single business.

Chart by TradingView · Official fund information — VanEck. This is the US fund, quoted in USD, not its European UCITS counterpart.

The chart loads from the provider when this row comes into view. Read its time axis and market status: quotes may be delayed, and the latest session remains visible while the market is closed. A price chart does not include all distributions or represent the whole semiconductor industry. For editorial context, not investment advice.

Power Companies in the AI Industry

Vertiv supplies power and cooling systems for data centers. Vistra, headquartered in Irving, Texas, generates and sells electricity. Forgent Power Solutions manufactures electrical distribution equipment for data centers and power grids.

Quotes are in US dollars. TradingView data may be delayed. Vertiv (VRT) · Vistra (VST) · Forgent Power Solutions (FPS).

Corporate debt · USD · Fixed rate · Senior unsecured

AI infrastructure — bond market

CoreWeave specialises in GPU cloud infrastructure for training and running AI models. Oracle combines GPU computing with a broader enterprise business in databases, cloud services and applications. Amazon (AWS), Alphabet (Google Cloud) and Microsoft (Azure) also operate extensive cloud platforms within diversified groups. The bonds below represent debt of the named issuer as a whole; they do not isolate the performance of its AI activities.

How this updates: the table records fixed coupons and contractual maturity dates. The interactive chart below requests market data directly from TradingView / Bonds.com when it loads. Updates depend on the source; quotes may be delayed or indicative, and a bond with no new observations can retain an older value. Use Refresh chart to request the chart again.

Fixed issue terms · not market quotations
IssuerAnnual couponMaturity
Microsoft1.350%15 Sep 2030
CoreWeave9.625%15 Jul 2032
Oracle4.800%26 Sep 2032
Alphabet4.375%15 Nov 2032
Amazon4.700%1 Dec 2032

Common basis: USD, fixed-rate, non-convertible senior unsecured notes; the same market-data source and daily chart interval. Compare YTM with YTM, or price with price, over the same dates. The four 2032 issues have closer maturities; Microsoft’s 2030 issue provides a shorter horizon. Credit quality, liquidity and early-redemption terms still differ. This is a comparison panel, not a ranking of investment quality.

CoreWeave · USD senior unsecured notes · Fixed annual coupon: 9.625% · Maturity: 15 July 2032 · ISIN: USU2069EAE06 (Reg S).

Read the measure: YTM is the annualised yield to maturity implied by price and scheduled payments, not the coupon or a guaranteed return. Price is expressed as a percentage of face value. Use the cursor and time axis to check observation dates; the viewer’s clock is not a last-trade timestamp.

Data loads directly from TradingView / Bonds.com when this section comes into view. Quotes may be delayed or indicative; illiquid bonds can retain older observations. Reload to reconnect. Market data is fetched separately from the site’s CDN.

Issue documents and scope

Contractual terms last checked on 11 November 2026. This date records the editorial check of the issue documents; it is not a market-data timestamp. The table and the choice of issuers are editorially maintained, not automatically synchronised with corporate actions.

SEC filings: CoreWeave · Oracle · Alphabet · Amazon · Microsoft.

The selection covers verifiable, embeddable issues and is not restricted to one maturity year. Matured securities belong in historical analysis, not the current comparison. Convertibles and floating-rate debt require separate treatment. These are corporate obligations, not dedicated AI bonds or an AI-sector index. CoreWeave’s displayed issue is the Reg S tranche; distribution restrictions may apply. A common currency and debt category do not make the risks identical.

Editorial market information, not investment advice.

The AI Oligarchy and Its Suppliers

AI and Society

  • In a Reuters Institute and YouGov survey across nine countries, 67% of respondents felt comfortable with news produced by a journalist alone. That share fell to 19% when AI did most of the work under human oversight.In Pew Research Center’s 2025 survey, 89% of adults in India trusted their country to regulate AI effectively, compared with 22% in Greece. The midpoint among the 25 countries surveyed was 55%.In two US surveys in 2026, the share who saw data centres as mostly bad for household energy costs rose from 38% to 50%. The bars trace a change in public opinion about electricity bills.

    When AI answers the helpline: who gets help, and who gets left waiting?

    AI can help an adviser resolve a routine query, yet leave a caller waiting when the question takes an unexpected turn. A parcel delivery and a hospital appointment show what human judgement adds, who answers for errors and how automation can change the jobs of those who take the calls.

    The recall offers a revealing detail. Subaru used the welding robots’ records to find 293 vehicles that needed checking. All 293 passed inspection. For the driver waiting for an answer, what mattered was that somebody could find the car, inspect the welds and report the result.

    In the United States, a court order barred one pharmacy chain from using facial recognition for surveillance for five years. In Australia, an inquiry into automated welfare debts asked how people could challenge the calculation behind a demand for repayment. Follow either case from the citizen’s side and a practical test emerges. Who can reopen the file, how long will it take, and who can correct the record before the mistake follows someone elsewhere?

    Read the article →: When AI answers the helpline: who gets help, and who gets left waiting?
  • In a Reuters Institute and YouGov survey across nine countries, 67% of respondents felt comfortable with news produced by a journalist alone. That share fell to 19% when AI did most of the work under human oversight.In Pew Research Center’s 2025 survey, 89% of adults in India trusted their country to regulate AI effectively, compared with 22% in Greece. The midpoint among the 25 countries surveyed was 55%.In two US surveys in 2026, the share who saw data centres as mostly bad for household energy costs rose from 38% to 50%. The bars trace a change in public opinion about electricity bills.

    One citizen, one identity, one political leader and an AI that keeps watch

    Who accounts for a grant worth millions when a family must justify every euro of childcare support? An independent AI auditor could help citizens trace the decision, challenge an abuse of power and see who bears the cost when a public authority causes harm.

    The recall offers a revealing detail. Subaru used the welding robots’ records to find 293 vehicles that needed checking. All 293 passed inspection. For the driver waiting for an answer, what mattered was that somebody could find the car, inspect the welds and report the result.

    In the United States, a court order barred one pharmacy chain from using facial recognition for surveillance for five years. In Australia, an inquiry into automated welfare debts asked how people could challenge the calculation behind a demand for repayment. Follow either case from the citizen’s side and a practical test emerges. Who can reopen the file, how long will it take, and who can correct the record before the mistake follows someone elsewhere?

    Read the article →: One citizen, one identity, one political leader and an AI that keeps watch
Share this on: