독서

Toby Walsh is a professor at the University of New South Wales (UNSW) in Australia and an AI scholar who has long advocated for an international ban on autonomous lethal weapons. His book God AI: Boom or Doom?, published in September 2026, addresses superintelligence, the singularity, existential risk, and labor markets simultaneously, yet the book's central question is not when machines will surpass humans but whether our institutions are prepared for the growth of capability.

In a written interview with Electronic Times (ETNews), Walsh compressed his position into a single sentence: 'What worries me most is not how intelligent machines become, but how fragile our institutions are. The challenge is power, not intelligence.' He elaborated that while how powerful AI becomes matters, the more urgent question is who owns that capability, who controls it, and who is held responsible when it goes wrong.

His argumentative mechanism rests on a concrete analogy: the aviation industry. Rather than treating accidents as an inevitable cost and leaving them unaddressed, the industry built a cycle of accident investigation, design improvement, and operational rule refinement, establishing public trust through safety, and that trust in turn expanded the market. Walsh argues AI should follow the same path—safety is not an obstacle to growth but a precondition for it.

Walsh also identified a specific failure mode in current institutional design. He criticized the widespread adoption of 'human-in-the-loop': if a machine generates a target list and a human has only seconds to approve it, that human review carries the risk of becoming a rubber stamp. His criterion is that humans must have actual time and authority to review information and must be able to reject AI's judgment when necessary—merely being 'in the process' does not mean being 'in control.'

This distinction has real-world context. Walsh recalled leading a boycott in 2018 with AI and robotics researchers from 30 countries against South Korea's KAIST, after the university partnered with Hanwha Systems to establish a defense AI research center. The boycott was lifted after KAIST committed not to develop autonomous weapons and to maintain human control. He also cited reports that the US Maven system fused intelligence data to recommend weapons during the war with Iran, striking approximately 1,000 targets on the first day, arguing that the scenario he had warned about for years had materialized.

Walsh does not shy away from superintelligence or existential risk. His book begins with 'Boom or Doom,' moves through superintelligence, the singularity, existential risk, and AI extinction, and ends with 'Machines of Loving Grace.' His position is that these scenarios merit study, but treating them as predetermined fate misplaces the question. The variables that are actually actionable are engineering tests, independent verification, incident reporting, real accountability, and public institutional capacity—things that can be built starting today.

On who should be constrained and to what degree, Walsh's direction is clear: independent safety testing before AI systems are released, incident reporting mechanisms, and real accountability for products that cause harm. At the international level, he holds that an autonomous weapons ban and maintaining human final control over nuclear decisions are non-negotiable red lines. He also emphasized that areas beyond the reach of national regulation require international agreements.

Walsh's diagnosis of institutional fragility includes a dimension often overlooked: the dependence of regulators on the regulated. He argues that governments must cultivate independent expertise internally and fund public research, otherwise regulators can only rely on information provided by companies to assess risk. He criticized the Australian government for having cancelled its plan to establish a standing independent AI advisory body, seeing this as a manifestation of institutional fragility.

His conclusion is neither optimistic nor pessimistic but returns the choice to human beings: 'It is not machines that determine our fate, but us.' Boom and Doom are not predictions but options, and the quality of those options depends on the quality of our institutions.

Meta released Muse Wonder and Muse Wonder Full on October 12, emphasizing coding agents, tool calling, and long-task execution. Wonder Full scored 57.5 on SWE-Marathon and 79.0 on DeepSWE 1.1. The developer community BridgeMind then arranged a 3D programming test, asking the model to create a lava lamp with a transparent glass shell, dynamic lava, and lighting effects; the model produced only a red cylinder. Meta's Chief AI Officer Alexandr Wang responded that the team had not explicitly trained for 3D demo games and that they perhaps should have added relevant training afterward.

Behind this controversy lies a more general technical mechanism: current large models increasingly rely on post-training reinforcement learning to improve task capability. In the pre-training phase, models learn language, code, spatial relationships, and other general knowledge from large-scale data; after entering post-training, reinforcement learning shapes more stable execution strategies in specific task environments. Coding agents are well suited to this kind of training — give the model a code repository with a bug, ask it to locate the problem, modify files, and run tests; the entire process can yield relatively clear feedback through compilation results and unit tests, and the reinforcement learning system uses this to judge which operations are effective and continuously reinforce them.

But this kind of training has a critical prerequisite: the environment must be able to provide useful feedback signals to the model. The lava lamp happens to break the feedback patterns familiar to coding agents. Three-dimensional geometry, transparent materials, lighting calculations, and dynamic deformation — even if this code executes normally, the visual output can still be completely wrong. If transparent material parameters are misconfigured, glass may become an opaque solid; if the lava shape lacks a continuous variation mechanism, the internal effect degenerates into static color blocks. Compilers and unit tests have difficulty directly judging these visual problems. The model therefore lacks a reliable reward signal to correct its behavior.

The article further points out that today's Agent benchmark scores can no longer be simply understood as giving a language model a problem to solve. Taking SWE-Marathon as an example, it assesses sustained software engineering tasks; completion rates depend not only on the underlying model but are also directly affected by the Harness (the execution framework built around the model). Meta's official Model Card explicitly states that in SWE-Marathon evaluations, GPT-6 Astra and Claude Fable 5.1 use Codex and Claude Code execution systems respectively, meaning this is not a controlled comparison under a unified Harness; Wonder Full's 57.5 score also used the xhigh inference tier.

This means a single number on a leaderboard compresses three variables: model reasoning capability, execution framework optimization, and compute resource investment. The article argues that if one wants to compare base models' reasoning and planning ability, tool environments, Harness, and inference budgets must be unified as much as possible; but if comparing production-ready coding agent products, the complete execution system's performance itself has engineering value. The problem is that when these configuration differences are compressed into a single score, outsiders find it difficult to determine where the score improvement actually comes from.

Notably, Meta's own technical article dedicated a Capabilities That Generalize section to discuss out-of-distribution generalization. The team divided tasks into in-distribution tasks heavily optimized during post-training and out-of-distribution tasks without targeted optimization, then compared Muse Spark 1.3 and Muse Wonder Full at different inference budgets. Meta believes stronger base models allow post-training gains to transfer more effectively to other tasks. But the article points out that out-of-distribution generalization cannot be described by a single aggregate average score — the distance between different tasks, the knowledge that needs to be supplemented, and the method of evaluating results all affect model performance. A single lava lamp failure only shows that the model did not complete that particular task; it cannot directly prove that the model was over-optimized against public benchmark question banks.

Researchers at Mass General Brigham in the United States recently published a new study in The Lancet Digital Health introducing FAHR-Face, an artificial intelligence model capable of predicting a cancer patient's risk of dying from a single facial photograph. Led by radiation oncologist Ray Mak and his colleagues, the study aims to address the subjectivity inherent in traditional clinical assessments.

Mak noted that medical students are trained to form an impression of a patient's health by observing them, a 'clinical intuition' that directly influences treatment plans such as surgery, chemotherapy, or radiation. However, this assessment is inherently subjective. FAHR-Face attempts to convert this subjective impression into objective data. The team first allowed the algorithm to 'learn' from 40 million facial images without a specific task, then fine-tuned it to perform two functions: estimating biological age and predicting survival odds for cancer patients.

Test results revealed surprising accuracy. In a test group of 3,820 cancer patients, 64 out of the 122 individuals with the highest risk scores died within 90 days of having their photo taken, accounting for more than half. In a larger separate test group of over 38,000 cancer patients, those who appeared older than their chronological age also had a higher risk of dying. The researchers stated this is the first report of an AI algorithm that can directly predict cancer patient survival from a simple face photograph.

To understand how the AI makes these judgments, the team used 'attention mapping' to track where the algorithm focused during analysis. For age prediction, the AI primarily looked at areas like the nasolabial folds (the crease between the nose and mouth) and parts of the forehead. For death risk prediction, it tended to focus on the bridge of the nose and areas below the eyes. While it remains unclear why these specific areas correlate with death risk, the model appears to pick up on facial changes associated with known health risks like smoking and alcohol use. For instance, smokers and heavy drinkers often look older than their actual age. Expressions, makeup, and lighting had little effect on age predictions, but plastic surgery was an exception: people who had rhinoplasty were rated as slightly older in post-surgery photos compared to pre-surgery ones.

Despite its promise, FAHR-Face is not yet ready for clinical use. A major limitation is that the cancer survival model was trained on a Boston-area population that was majority white. The research team is working to expand their training sets to include a wider diversity of faces. Mass General Brigham has filed a patent for the face health algorithms.

The technology also raises ethical questions. Matthew Liao, a professor and director of the Center for Bioethics at New York University, commented that while such tools could help identify patients needing extra care, there is a risk of misidentification, potentially leading to withheld care for those who actually need it. Liao also expressed concern that insurance companies might use this technology to grant or deny coverage, or employers might use it to screen for 'healthier' job candidates. He argued that society must think through the implications before further development.

US cloud computing company Cloudflare announced the preview of Clairic on October 19, 2026, defining it as a new class of 'System One extraction models.' The model is designed to extract, compute, estimate, and forecast values directly from unstructured inputs like receipt photos, support emails, or CSV files, rather than generating text word-by-word like Large Language Models (LLMs).

According to Cloudflare's official blog, Clairic operates via a single parallel forward pass without producing output tokens, meaning it charges only for input and typically responds in under a second. It returns typed values (such as numbers, dates, and lists) with character offsets and confidence scores, returning null if an answer cannot be found. This design avoids common LLM failure modes like invalid JSON formatting and hallucinations, although the model can still make mistakes.

The concept follows TypeSafe's release of Jev a month prior, which introduced decision models based on Daniel Kahneman's 'System 1' thinking from Thinking, Fast and Slow. Cloudflare states that Clairic is the first model in this new extraction category, specifically targeting tasks that require concrete numerical answers rather than choices or free-form text. Clairic is currently hosted on Workers AI, with plans to release open weights soon.

Activation probes are a cheap way to monitor AI models: a simple classifier trained on the model's hidden states fires when it detects a concept such as deception or harm. The Neural Chameleons paper proposed a method to evade such probes—fine-tuning a model to suppress probe scores whenever the prompt contains "You are being probed for {concept}." The paper reported that this generalizes to unseen probes, including deception and harm probes, at a small capability cost (roughly 3 points on MMLU and 1 on GSM8K).

But the original paper did not address a critical case: what happens when the model must complete a task while being probed for the very cognition needed to solve that task? For example, can a model do math while hiding the concept "mathematical" from a probe? Redwood Research/MATS 11.0 interns Taiga Someya (main contributor) and James Lucassen (supervisor) tested exactly this in a two-week sprint project.

They trained chameleons using the original authors' code and data, with abliterated Gemma-2-9B-it as the primary model, evaluated on GSM8K and MMLU. Key finding: when the probe monitors the concept the task requires, evasion succeeds but at enormous cost. Under the "mathematical" trigger, the math probe flags only 11% of correct GSM8K answers, yet accuracy falls from 89% (base model) to 46%. Chemistry and biology show the same trade-off. Capability drops 35–44 points under the relevant trigger versus only 18–21 points under the unrelated control trigger "elephant," indicating the cost is not generalized training perturbation but direct conflict with the task's required cognition.

The replication also revealed two discrepancies with the original paper. First, concepts the chameleons were never trained on (geography, physics, history) mostly remain visible to probes, contradicting the paper's claim that evasion generalizes to unseen concepts. Second, the paper's baseline MMLU and GSM8K scores match Google's published numbers for the pretrained Gemma-2-9B rather than the abliterated Gemma-2-9B-it that was actually fine-tuned, meaning the original paper understated the training cost.

The authors tested a larger model (31B), twice the training data, and a non-abliterated model; none eliminated the trade-off. The 31B model appears to hide math more cheaply (97% drops to 84%), but in logits against the base model its evading runs still lose about three-quarters as much math accuracy as the 9B (−1.9 vs. −2.4), and lose more on chemistry (−2.3 vs. −1.6).

The practical implication for AI safety monitoring: if a model must complete the very task being monitored (for instance, writing a backdoor while a probe watches for "backdoor"), current methods cannot evade the probe without substantial capability loss. The authors note that simple deception—where the model does not need to retain much cognitive capability—may be an exception, but more sophisticated violations are precisely the scenarios that demand capability. All experiment code is publicly available.

In 2026, the US state of California enacted a one-time wealth tax targeting billionaires: a 5% levy on the worldwide net worth of individuals exceeding $1 billion. In an NBER working paper, Benjamin Jaros and Joshua Rauh estimate that once billionaire migration behavior is accounted for, the tax not only fails to generate net revenue but reduces California's fiscal net present value by approximately $38.9 billion.

The paper's mechanism runs as follows: the wealth tax is levied → some billionaires depart California → the state income tax they previously paid disappears → the present value of that lost income tax exceeds the one-time wealth tax collection. Using data from the Forbes Billionaire List, news reports on California billionaires' residential real estate holdings, and public announcements of departures from the state, the authors estimate the Act will collect roughly $30 billion (range $28–36 billion), less than a third of the $100 billion projected by Galle et al. (2025).

More critical is the income-tax side. The authors estimate California's billionaires contribute $3.3–5.8 billion annually in state income taxes. Publicly confirmed migration has already eliminated nearly 30% of the Act's wealth-tax base. Discounting the permanently lost future income tax from departing billionaires, the total present value exceeds the one-time wealth-tax revenue, yielding a net present value of negative $38.9 billion. Across 100,000 simulated combinations of revenue estimates and discount rates, 85% of plausible outcomes produce a negative net present value.

It should be noted that this is an NBER working paper, not yet peer-reviewed. The authors' methodology includes a litigation-survival discount applied to the remaining tax base and validation against the wealth-tax-base elasticity literature. However, the core input — publicly confirmed migration — may not capture the full behavioral response: some wealthy individuals may be planning to leave without having announced it publicly, while others may remain in California due to asset liquidity constraints or family circumstances. Galle et al.'s (2025) $100 billion projection comes from a separate study; whether the two sets of figures use strictly comparable methodologies requires reading the original papers.

Tyler Cowen shared the paper on his Marginal Revolution blog, writing: "Overall this issue is a good litmus test for who still thinks like an economist." That remark is editorial commentary rather than a factual claim, but it points to the paper's core contribution: moving wealth-tax analysis from static revenue estimation to a dynamic net-effect calculation that incorporates behavioral feedback.

The arXiv preprint "Same Person, Different Advice: Auditing LLM Financial Advice with Repeated Risk-Tolerance Responses" highlights that many AI financial advisers currently set stock allocations based largely on a single stated-risk question, despite the fact that human responses to such questions are often inconsistent. The research team tested seven large language models from five providers, generating one-turn allocation advice for 800 real household profiles. They found that when only the user's risk answer changed, the recommended stock share moved by an average of 10 to 16 percentage points. If the answer changed substantially, the swing reached 24 to 35 percentage points. In contrast, repeating the same prompt twice resulted in a movement of only 0.6 to 4.2 percentage points.

The study notes that this volatility driven by input noise is most pronounced among clients under 40. The authors propose that if this response pattern represents how models map attitude to advice, discounting unreliable answers according to classical test theory could theoretically reduce swings by a factor of 2.4 to 3.3. Furthermore, experiments showed that while noting the answer was "noisy" steadied some models, it caused them to over-rely on age factors and push extremely cautious clients toward generic middle allocations (e.g., advising 22% stocks instead of 4% for clients unwilling to take any risk).

The most actionable finding is that in a scripted test, adding just one clarifying question removed nearly all of the swing caused by a client subsequently correcting their answer—reducing the average variance from 33.5 percentage points to 1.9—while preserving the model's ability to respond to confirmed changes in preference. The researchers argue that same-person retest audits should become a standard check for AI advisers, and that handling unreliable answers is a matter of client autonomy as well as statistics.

Researchers proposed EdgeWise, an edge-cloud collaboration framework. The core idea is to deploy a small language model (SLM) on the local device as the primary solver, selectively delegating critical steps to a cloud LLM only when stronger model capacity is needed.

To enable autonomous delegation without introducing additional models at inference time, EdgeWise introduces a step-level reinforcement learning (RL) method. This approach jointly optimizes task solving and delegation decisions, with fine-grained credit assignment for individual delegation choices. This means the local SLM learns during training to identify which steps are "critical," allowing it to make optimal cost-performance trade-offs at runtime.

Experiments show that EdgeWise matches or outperforms pure-LLM execution while reducing inference cost by up to 62% on SWE-Rebench and 34% on Terminal-Bench 2.1. The study further deployed EdgeWise on a local workstation, demonstrating comparable cost savings in a practical edge-cloud setting.

OpenAI announced on October 19, 2026, that its latest models have achieved substantial breakthroughs in coding capabilities for low-resource programming languages such as COBOL, Fortran, and Visual Basic. These languages, while having less available training data online, support critical global infrastructure, including over 40% of online banking systems, 80% of credit card transactions, weather forecasting, and nuclear safety analysis.

To measure this progress, OpenAI developed two specialized evaluations. The first, MultiLangBench, requires models to build a SQLite-compatible database engine from scratch across 20 different programming languages. Data shows that GPT-6 Astra achieved an average test pass rate of 87.4%, a significant increase from 26.8% for GPT-5.6 Sol. Specifically, the pass rate for COBOL jumped from 10.9% to 72.1%, and Fortran improved from 21.4% to 83.5%. The lower-cost model GPT-6 Luna also reached an average pass rate of 40.3%, with API costs approximately 78% lower than the previous generation.

The second evaluation focuses on COBOL code migration, testing whether models can convert existing COBOL programs into modern languages like C++, Java, or Python while strictly preserving original business logic and behavior. Since COBOL is often used for financial transactions, specific arithmetic rules (such as default truncation rather than rounding) are crucial. OpenAI reports that GPT-6.1 Sol achieved 86.6% migration accuracy at low reasoning effort, compared to just 5.4% for GPT-5.6 Sol; at higher reasoning efforts, accuracy approached saturation at 97.9%.

OpenAI stated that these improvements aim to help teams maintain and modernize software dependent on legacy systems, addressing challenges like incomplete documentation and developer turnover. The company has open-sourced the MultiLangBench evaluation setup and test suite on GitHub for developers to verify performance gaps in specific languages. Despite the model's enhanced performance, OpenAI emphasizes that applying such automated migration techniques in production environments still requires expert review and testing against business rules.

Chinese artificial intelligence developer MiniMax Group Inc. released M3.1 Flash on October 19, 2026, an update to the M3 large language model launched in June. The new model targets the high cost of running AI agents, which often require dozens of model calls to complete a single job. According to MiniMax, M3.1 Flash reduces agentic task costs by 70% compared to its previous model, while increasing inference speed nearly threefold, with everyday peak output exceeding 200 tokens per second.

The model maintains approximately 428 billion parameters, achieving speed gains through more aggressive sparsity at inference time. It supports text, image, and video input, offers a 1 million token context window, and retains reasoning capabilities with five selectable effort levels. In MiniMax's own testing, M3.1 Flash scored 61.7 on OSWorld V2, which tests computer operation capabilities, 76.2 on BankerToolBench for banking workflows, and 82% on CyberGym for reproducing software vulnerabilities.

Pricing positions M3.1 Flash directly against competitors: input costs are set at 15 cents per million tokens, matching the off-peak rate of Hangzhou-based DeepSeek's V4.1-Flash model. While DeepSeek claimed its V4.1-Flash resolved 77% of tasks on the DeepSWE v1.1 software engineering test versus M3.1 Flash's 73.7%, neither result has been independently verified. MiniMax also confirmed that the anonymous 'Space Bunny' model previously seen on OpenRouter was an early build of M3.1 Flash.

AI audio company ElevenLabs released the alpha version of 'ElevenAgents Architect' on October 19, 2026. This new feature within its enterprise agent platform, ElevenAgents, simplifies the construction and maintenance of AI agents by allowing users to create new customer service agents or optimize existing ones through voice or text instructions.

Traditional AI agent configuration involves complex elements such as knowledge bases, prompts, workflows, voice settings, guardrails, and integration tools. Architect analyzes conversation logs, failed test cases, and issues found in ElevenAgents Spotlight to automatically identify improvement points and generate configuration suggestions. Before deployment, it can also run simulations to verify the effects of changes.

Beyond the internal platform, Architect integrates with external AI tools like Claude, ChatGPT, and Cursor, enabling non-technical staff (such as CX managers, compliance officers, and product managers) to directly manage agent strategies without relying on developers. The feature is currently in alpha stage, and its performance and stability remain to be seen.

The vLLM project team published a technical blog post on October 19, 2026, detailing how they achieved near-zero-downtime scaling in Elastic Expert Parallelism (Elastic EP) through "asynchronous preparation" and the new NIXL EP communication library. This improvement addresses the challenge of service interruptions when dynamically adjusting the number of GPUs for large-scale Mixture-of-Experts (MoE) models.

Traditionally, changing the GPU count in large inference services required restarting the entire engine or adding/removing complete model replicas. As models grow larger, each replica requires more GPUs, making scaling increments coarse. For instance, in benchmarks with the DeepSeek V4 Pro model, scaling from 16 to 20 GB200 GPUs using traditional methods caused a 197-second service stoppage. Even with vLLM's previous Elastic EP solution that avoided restarts, the service paused for approximately 100 seconds due to the need to initialize new workers, establish communication groups, load weights, and repeat kernel warmup and CUDA graph capture.

The core of this update lies in splitting the reconfiguration process into two phases: "background preparation" and a short "commit." Existing workers continue serving requests while background threads pre-configure target communication groups, transfer model weights, and compile/warm up execution environments. Joining workers initialize independently and receive weights. During the commit phase, vLLM briefly pauses schedulers, activates the prepared topology, and immediately resumes serving.

To support this flow, the team developed the NIXL EP communication library. Traditional EP communication backends have fixed membership, requiring new instances for adding or removing ranks, which increases memory overhead and invalidates previously captured CUDA graphs, necessitating recapture. NIXL EP allows adding or removing ranks in place, reusing the same agent, buffers, and connections, and keeping pointers and parameters stable within CUDA graphs. This means retained workers can reuse existing computation graphs without recapturing them, eliminating most of the downtime.

According to vLLM's official benchmark results, in the same DeepSeek V4 Pro scenario, downtime for scaling from 16 to 20 GPUs was reduced to 0.82 seconds, and scaling back from 20 to 16 GPUs took 1.96 seconds. This improvement allows high-cost GPU resources to be allocated more flexibly during traffic fluctuations with almost no impact on online service continuity.