Reading

A software engineer at NAVER in South Korea, who has been programming since 1989, published a long-form article on the D2 blog on October 27, 2026, proposing a core judgment: the underlying executors within organizations are now AI, and every developer has become a manager of an AI team. This is not rhetorical exaggeration but a description of his daily work—assigning tasks to multiple AI agents, providing background information and evaluation criteria, and quickly providing feedback to correct deviations when results occur.

He maps the responsibilities of traditional managers one-to-one with AI agent management: setting goals and values for multiple individuals (AIs), reducing information retrieval costs, assigning work based on respective strengths while controlling the scope of context, and establishing efficient feedback loops between decision-makers and executors. He also points out a specific engineering consequence: when module boundaries are clear, AI agents narrow their scope of modification; when boundaries are vague, AI replicates similar code in multiple places, thereby accumulating technical debt. This means architectural decisions no longer only affect the collaboration efficiency of human developers but directly determine the work quality of AI agents.

However, the author believes that what AI agents cannot replace is human accountability. Even if AI completes all coding, testing, and deployment operations, it is still humans who decide to let users use a certain software; therefore, responsibility for problems falls on the final approver. From this, he infers that if the title 'developer' continues to exist in the future, its most likely meaning would be 'the person who approves the deployment of code for sensitive systems to production environments.' To gain this trust, methods to verify and ensure code behavior—code reviews, automated testing, multi-layer security checks—become even more important, not less so.

He provides a specific risk mechanism to support this judgment: LLM outputs follow a probability distribution, and the same instruction may produce different code; AI feedback loops primarily run in development environments, whereas the development environment and runtime environment of large-scale systems differ significantly in traffic, data scale, and diversity of execution paths. Therefore, code passed by AI in the development environment cannot guarantee the absence of performance latency, memory overflow, or data pollution in the runtime environment. He cites his own experience: participating in system architecture transformations involving annual transaction volumes of tens of trillions of won, multiple production incidents made him realize this gap is '운영 환경의 중압감' (the pressure of the runtime environment).

Regarding what capabilities developers need to cultivate, the author unfolds two lines of thought. The first is human-to-human communication and the collection of implicit information: AI can amplify the direction determined by humans, but the direction itself needs to be determined through communication; implicit premises within organizations, temperature differences among stakeholders, and inspiration emerging from face-to-face interactions—information that AI cannot obtain from training data—needs to be actively mined by humans. The second is establishing standards higher than the general level in specific fields and combining standards from multiple domains to make judgments: he gives an example where AI considers N+1 queries acceptable in the runtime environment, but given the actual deployment scale, he cannot accept this standard; caching is good from a performance perspective but potentially dangerous from a data consistency perspective—such comprehensive judgments require humans to have sufficient depth across multiple dimensions.

Deployed large language models face a fundamental contradiction: directly updating parameters destroys old capabilities (catastrophic forgetting), while relying solely on the context window means all new information is lost once the session ends. In a blog post published on October 27, 2026, Ali Behrouz and Vahab Mirrokni of Google Research propose a path around this dilemma—allowing models to have 'sleep' cycles like the human brain.

The sleep framework divides model operation into two alternating phases. The 'wake' phase normally receives and processes inputs, storing temporary information in fast short-term memory components; the 'sleep' phase pauses new inputs and redirects internal computation toward memory consolidation and self-improvement. The authors emphasize that although it borrows neuroscience terminology, the framework operates entirely within a fixed model architecture and human-defined training rules, rather than simulating biological sleep.

The sleep phase contains two sub-processes. The first is memory consolidation: the model uses Mixture-of-Experts (MoE) and Hope architectures to dynamically expand parameter capacity, writing newly unlocked low-rank experts into stable, low-frequency memory blocks to accommodate new knowledge without disturbing older parameters. Simultaneously, fragile short-term memories in high-frequency layers are 'seeded' into newly expanded long-term parameters through Generalized Knowledge Distillation (GKD) combined with reinforcement learning-based imitation learning. After consolidation is complete, high-frequency temporary parameters are reset to free up capacity for the next wake phase.

The second sub-process is 'dreaming'—an autonomous self-improvement loop. The MoE router randomly activates non-standard experts, forcing the model to mix different knowledge domains to discover latent patterns; based on this, the model generates synthetic training samples ('dreams'), filters them using gradient-based data selection to retain those with the highest improvement potential, and finally performs supervised fine-tuning on these high-scoring dream samples while evaluating downstream task performance.

The authors report a set of self-test results: in continual translation for unseen languages such as Manchu and Kalamang, the sleep framework maintained near-zero performance degradation across multi-stage evaluations, whereas ICL, Cartridges, and SFT all suffered severe catastrophic forgetting; on the BABILong benchmark, the method claims to achieve near-perfect context scaling up to 10 million tokens, surpassing GPT-4 and Titans; on SQuAD factual knowledge integration, it reached 48.9% (single-task) and 46.2% (continued pre-training), outperforming baseline models, SFT, and SEAL; on AIME 24/25 and HMMT 25 mathematical reasoning, it achieved accuracy rates of 79.2%, 69.0%, and 46.1% respectively, exceeding SFT, GRPO, and OPSD.

All these figures come from author self-tests; the blog does not disclose hardware configurations, absolute scores for each baseline, or statistical confidence intervals, and the full paper (arXiv:2606.03979) has not undergone independent reproduction. Therefore, these results should be viewed as preliminary evidence attributed to the Google Research team, rather than verified industry benchmarks. The actual deployment costs of the framework—the computational overhead of each sleep cycle, storage growth from parameter expansion, and controllability of synthetic sample quality during the 'dreaming' phase—are not quantified in the blog.

According to a report by Tom's Hardware, the American Civil Liberties Union (ACLU) published findings on October 26, 2026, alleging that US police are increasingly using artificial intelligence (AI) algorithms and automated license plate reader (ALPR) surveillance data to flag motorists whose travel histories appear "suspicious," even without evidence of wrongdoing.

Jay Stanley, a senior policy analyst with the ACLU Speech, Privacy, and Technology Project, explained that this practice relies on networks of ALPR cameras installed along roads and highways. These cameras photograph passing vehicles and record their registration numbers, locations, and timestamps. By combining sightings from different cameras, law enforcement can determine where a vehicle has traveled, which routes it used, and how frequently it visited particular locations. An AI system with access to this surveillance data can quickly analyze huge amounts of information and identify patterns it has been trained to deem as "suspicious," triggering an alert to law enforcement. Stanley implies officers may receive the alert and look for a minor traffic violation to justify pulling the driver over. In some documented cases, officers have concealed the actual reason for the stop.

The report details several incidents, including one in which a woman claims an officer stopped her during a cross-country road trip because a surveillance system he was using had flagged her movements as suspicious. The system had accessed ALPR data on where she had been driving.

This finding comes amid growing concerns about the impact of massive mass surveillance camera networks, involving issues such as privacy, abuse, and limited regulation. For instance, security researchers disclosed a network of over 335,000 Flock ALPR cameras across the United States. Companies such as Flock Safety have been developing AI-enhanced surveillance systems that can identify "unusual" travel behavior and associations between vehicles, marketing them as effective crime-prevention tools. However, while license plate readers have traditionally helped police track known suspects or locate vehicles already associated with a crime, the current scale and blanket surveillance are unprecedented.

Condemning the approach and the use of mass surveillance in general, Stanley expressed concerns that this method would lead to an uptick in controversial "pretextual stops." He warned, "More and more people will be subject to police encounters as these AI algorithms inevitably sweep in far more innocent people than actual wrongdoers."

Solar equipment company Exowatt provides dispatchable solar power systems for AI data centers. Its energy management system operates across an expanding fleet of sites, each continuously generating hardware alarms. The team needed to automatically classify alarms and provide next-step action recommendations on edge controllers without GPUs and often without network connectivity. Exowatt described this engineering practice in an Amazon Web Services (AWS) machine learning blog: fine-tuning the Qwen3-1.7B model for the alarm classification task achieved a self-test macro-F1 of 0.911, surpassing Claude Sonnet 4.6's 0.698, with zero inference costs (running on existing site CPUs).

Why are frontier models unfeasible in this scenario? Exowatt cites three reasons: the task is narrow and proprietary, as alarm classification relies on their specific hardware and operational logic; it must run at the edge with no GPU and often no network; and costs are unsustainable at the fleet scale, making it economically unfeasible to call cloud-based frontier models for every decision. The team initially attempted Qwen3-8B with prompting and guardrails, but macro-F1 plateaued at 0.735, missing the target. They then discovered a counterintuitive result: none of thirteen fine-tuned Qwen3-8B schemes across twelve dataset versions exceeded the prompted 8B baseline. Conversely, Qwen3-1.7B, with only one-fifth the parameters, surpassed the 8B model when paired with more diverse training data.

Data strategy was the core finding. The training set consisted of only 2,522 items: 684 verified site-scenario decisions plus 1,838 rewrites generated by Qwen3-VL-235B (preserving decisions, labels, evidence citations, and all numbers, altering only phrasing). The paraphrase generation cost approximately $3.72, with 1,838 of 2,218 candidates passing deterministic checks. Key control experiments showed that using only the 684 original data points for 11 epochs (matching the step count of the expanded dataset) yielded a macro-F1 of 0.879; using the expanded, diverse 2,522-item dataset for just one epoch yielded 0.911. Exowatt concluded that for tightly bounded tasks, data quality and diversity outweigh model scale and volume; simply increasing training epochs performs worse than increasing data diversity.

In a cloud-based energy market analysis task, the same pipeline lifted Qwen3-14B from a baseline macro-F1 of 0.415 to 0.975, surpassing Claude Opus 4.7's 0.625, with a cost of $0.60 per thousand responses compared to $29.26 (a roughly 49-fold difference). Exowatt emphasized that these two models were merely the first outputs of the pipeline; the more enduring outcome was the pipeline itself: a repeatable path from raw operational data to certified training sets, fine-tuned models, and evaluated deployment artifacts, compressing iteration cycles from months to days.

Boundaries to note: All performance data are self-tests by Exowatt and AWS, without independent reproduction. Both tasks involved single-turn strict JSON output classification, where every claim returned by the model had to cite specific rows from the input data; unsupported claims were discarded. Test sets comprised 320 and 2,000 samples respectively. Cost comparisons used Amazon Bedrock list prices multiplied by measured tokens; the edge model's zero cost refers to the absence of API call fees, excluding hardware depreciation and operations/maintenance. These figures illustrate relative advantages on specific narrow tasks and cannot be extrapolated to multi-step reasoning, open-ended generation, or scenarios requiring world knowledge.

Avante, a Seattle-based startup building an AI platform for managing employee health care and benefits, announced a $30 million Series A funding round on October 27, 2026. The round was led by Frist Cressey Ventures, with participation from Sorenson Capital, Hearst Ventures, Morgan Stanley Next Level Fund, and existing investor Fuse. This brings the company’s total capital raised to $40 million, following a $10 million seed round in 2024.

The company’s technology uses AI agents to automate tasks such as comparing vendor contracts against actual claims, reviewing plan documents for errors, and preparing for renewals. Employees can also interact with the platform to ask questions about their coverage. Avante plans to use the new funds to expand internationally and develop additional AI agents, building on the recent beta release of its “Ava Agents” product.

According to GeekWire, Avante now serves more than 70 customers, including Samsung Semiconductor, Real Chemistry, Zscaler, and News Corp, and reported that bookings grew nearly fivefold over the past year. The platform currently reaches more than 385,000 employees and family members. CEO Rohan D’Souza noted that large employers are projecting a median 9.2% increase in health care costs for 2027, driving demand for tools that help control these expenses.

U.S. cloud provider Oracle announced on October 27, 2026, at the Oracle AI World conference in Las Vegas, new capabilities in Oracle Acceleron, a suite of networking and infrastructure technologies within Oracle Cloud Infrastructure (OCI). These updates are designed to help customers run demanding AI workloads reliably, securely, and at scale.

Key additions include the Oracle Acceleron High-Performance Fabric, which uses RDMA connectivity on front-end compute to boost performance for inference, storage, and HPC workloads; expanded NVMe block-storage attachment options, claiming up to 3x performance improvements; and the Oracle Collective Communications Library. The latter spreads traffic across OCI's backend Multiplanar Network to help distributed AI workloads instantaneously recover from link, port, or NIC failures, significantly improving network availability.

Pradeep Vincent, Executive Vice President of Compute and AI Infrastructure at Oracle, stated that when thousands of accelerators work together, one congested or failed path can have consequences far beyond a single connection. The new architecture aims to respond by rerouting traffic and recovering from failures.

Oracle included a "Future Product Disclaimer" in the release, noting that these features represent general product direction, do not constitute contractual commitments, and that development, release timing, and pricing may change.

Dongwon Parts, a South Korean manufacturer of semiconductor equipment components, held an IPO press conference in Seoul on October 27. Vice President Jo Jun stated that the company's motto is "not competing with customers like TSMC," emphasizing a strategy focused on aftermarket parts for semiconductor processes to avoid direct competition and patent disputes with equipment manufacturers.

Financial data revealed that Dongwon Parts' first-half sales reached 63.3 billion KRW, up 44.9% year-over-year. Operating profit surged from 3 billion KRW in the same period last year to 14.5 billion KRW. The company attributed this improvement to a higher proportion of high-value-added products. A key driver was the "showerhead" component, which evenly distributes gas onto wafers; its average selling price (ASP) rose from $1,646 in 2022 to $2,140 in the first half of this year.

The share of aftermarket component sales has increased from 39.3% in 2023 to 45.1% in 2025. Although it temporarily dipped to 39.7% in the first half of this year due to increased demand for new equipment, management projects that the aftermarket share will exceed 60% in the long term. Current major clients include Lam Research (45.9%), SEMES (33.4%), and Exelis (12.6%). The company is expanding its customer base to include Applied Materials (AMAT) and plans to introduce metal 3D printing technology to reduce costs and develop lightweight aerospace components.

The IPO involves issuing 2.03 million new shares at a desired price range of 23,000 to 26,500 KRW, aiming to raise between 46.7 billion and 53.8 billion KRW. The expected post-listing market capitalization is approximately 213.4 billion to 245.8 billion KRW. Samsung Securities serves as the lead underwriter.

A focused equity fund delivered an annualized return of 18.5% between 2000 and 2010. During the same period, the S&P 500 Index declined slightly; this decade is known on Wall Street as the 'lost decade,' making the 18.5% performance exceptional by any standard. However, if you examine the actual returns of the average investor in that fund, the answer is a loss. The fund made money while its investors lost it; the discrepancy stems from the timing of buy and sell decisions.

Daniel Crosby, Chief Behavioral Officer at Orion Advisor Solutions and a psychologist, discussed this case while promoting the new paperback edition of his book The Behavioral Investor on the Excess Returns podcast. He used the example to introduce his core argument: there is a fundamental distinction between a 'mathematically optimal' portfolio and a 'behaviorally optimal' one. A mathematically optimal portfolio offers the highest theoretical return within a mean-variance framework; a behaviorally optimal portfolio is one that investors can actually stick with despite real-world emotional volatility. If an investor cannot resist selling during a sharp decline, even the highest theoretical return is irrelevant to them.

Crosby argues that traditional risk assessment measures only two dimensions, missing a third. The first is risk capacity: how much risk you can afford to take based on your goals, timeline, and existing wealth. The second is risk tolerance: the trade-off you are willing to make between risk and reward when in a calm emotional state. Crosby notes that research indicates risk tolerance remains relatively stable throughout a person's life and does not fluctuate significantly.

The third dimension he proposes is risk composure: the likelihood that strong emotions such as fear or greed will alter your perception of preferences established in a calm state. In other words, instead of asking 'how much risk are you willing to bear?', the question becomes 'what is the probability that you will overturn yesterday’s answer when the market crashes?' Crosby links this dimension to the personality trait of 'neuroticism'—individual differences in anxiety proneness—and observes that this variation is substantial across the population.

In practice, this means portfolio construction must address not only 'what should be held' but also 'under what emotional conditions one might abandon holding.' Crosby suggests providing additional behavioral coaching and early warning support for investors with lower composure rather than simply reducing their risk exposure, as lowering risk might prevent them from achieving their financial goals. He simultaneously warns that over-intervening with investors who have high composure can be counterproductive. He cites A/B testing results from Betterment, a robo-advisor, which showed that sending 'don't panic' alerts during volatile periods actually triggered trading among clients who had previously shown no anxiety, effectively planting seeds of worry where none existed.

Programming and customer support were the first two product-market fits in AI; Tomasz Tunguz, a partner at Theory Ventures, considers security to be the third. His core argument is not that AI makes security better, but that attackers' parallelization capabilities have overwhelmed human defenders' response speeds, creating a structural demand for token consumption.

The attack data cited by Tunguz comes from multiple sources. Black Kite reported that ransomware incidents against the financial services sector rose from 156 in 2024 to 202 in 2025, reaching 65 in the first quarter of 2026 alone, a 76% year-over-year increase. Cisco Talos discovered that eight intermediate certificates from eMudhra were compromised, with private keys evaluating to single- or double-digit integers, allowing attackers to impersonate any website or service. StepSecurity tracked 56 supply chain attacks over twelve months, one of which was a five-day campaign that stole 78,330 secrets from 2,186 organizations, including 92 publicly traded companies. These figures are self-reported by various institutions, and Tunguz did not independently verify them.

A key step in the economic mechanism is the collapse of response times. Tunguz wrote that intrusion response times have plummeted from months to minutes, making it impossible for human teams to read every alert at that speed. This forces Security Operations Centers (SOCs) to introduce AI for automated triage. He surveyed 75 enterprise security practitioners, asking how they would allocate their security budgets excluding endpoint security, resulting in five categories: Security Operations Center, SIEM, penetration testing, threat intelligence, and vulnerability management. Each category involves significant token consumption. In penetration testing, models continuously probe every endpoint at a cost far lower than human red teams; in vulnerability management, AI ranks massive scanner findings by reachability and exploitability; in SIEM, analysts query databases in natural language, receiving answers in seconds rather than filing tickets to wait for query writing.

Tunguz quotes OpenAI President Greg Brockman from the October 2026 No Priors podcast:

"Security spend should actually vastly outstrip software spend. Every company needs AI security because attackers can parallelize their efforts." He added that the global cybersecurity market was $175 billion last year, growing 13% annually, noting this figure does not yet include the incremental impact of security entering the token path.

Tunguz himself notes limitations in his survey: the sample of 75 people is largely qualitative, and error bars on charts are significant. His argumentative chain is: attack parallelization → response time collapse → five budget categories must adopt AI → each generates token consumption → security market enters the token path → this constitutes the third PMF. A weak link in this chain is that inertia in existing security budget allocation may not automatically translate into token expenditure simply due to technical feasibility; vendors (including OpenAI) have an incentive to promote this narrative; and the $175 billion figure is a Forrester forecast rather than observed token consumption data. As a General Partner at Theory Ventures, Tunguz has a vested interest in the AI infrastructure narrative.

American database and cloud services giant Oracle announced "Oracle Media Fabric" on October 27, 2026, at the Oracle AI World conference in Las Vegas. This integrated platform is designed to help media companies, streaming providers, and brands reduce content production time and cost, distribute high-quality video at scale, and create more personalized experiences for viewers.

The platform combines Oracle Creative Studio (an AI-powered production suite), Oracle AI Media Models (for video and image generation), and Oracle Video Edge (infrastructure for low-latency large-scale live delivery). Oracle stated that Media Fabric brings sophisticated content creation and streaming capabilities, previously available only to the largest proprietary streaming platforms, to content creators of all sizes.

Oracle CEO Clay Magouyrk said the initiative uses AI across the entire production and distribution workflow to help creators better utilize existing content and tailor viewing experiences in real-time. Notably, Oracle included a "Future Product Disclaimer" in the announcement, stating that the described features represent general product direction, do not constitute contractual commitments, and that development, release timing, and pricing may change. The platform is planned to be available within the next 12 months.

According to WIRED, the Independent Chief Inspector of Borders and Immigration (ICIBI) is considering an investigation into the UK Home Office's use of generative AI tools to process asylum applications. The potential probe follows public questioning by lawmakers and researchers regarding the tools' accuracy and the government's secrecy surrounding their use.

The ICIBI monitors the Home Office's management of borders, immigration, and customs. It has asked the digital rights nonprofit Open Rights Group (ORG) to share more information on the tools as it draws up plans for topics to inspect next year. ORG's findings come from a six-month investigation involving 11 parliamentary questions, two ministerial letters, and 10 public records requests, which the Home Office has largely withheld details against.

Documents obtained by ORG reveal that the Home Office uses the Asylum Case Summarisation (ACS) and Asylum Policy Search (APS) tools, built on OpenAI's 2023 GPT-4 large language model and Microsoft Foundry, to summarize interview transcripts and country origin information. The Labour government began piloting these tools in late 2024 to clear a backlog of around 100,000 asylum cases. While the number of people waiting for an initial decision halved in the year through June, appeals against decisions have ballooned.

Researchers are concerned about the accuracy of these AI tools in high-stakes decisions determining whether to send someone back to a country where they fear persecution. ORG found that the percentage of ACS-generated summaries deemed not fit for use—often due to factual errors or misrepresentations—is significantly higher than the published 9 percent error rate from the pilot evaluation. There is no quality control for APS summaries. In July, a UK judge accused the Home Office of citing an "AI hallucinated" document to refuse an asylum claim.

A Home Office spokesperson stated, "Asylum decisions are made by trained caseworkers, with AI tools used only to assist in navigating interview records and policy information. They cannot be used in isolation and do not make decisions on claims." However, critics note that under the UK's "AI Playbook," departments must be transparent about algorithmic tools used in decision-making and list them on a public repository, yet the Home Office's tools are absent from this list. Furthermore, asylum seekers are not informed when these tools are used, potentially breaching UK data rights laws.

According to SDx Central, the European Union unveiled five new 6G research projects at the Tech Territory conference held in Latvia on October 20-21. The initiative establishes global partnerships with Taiwan and India, aiming to build a 6G experimental platform across six EU member states and set up a coordination hub between the EU and India.

The Smart Networks and Services Joint Undertaking (SNS JU), an EU-funded public-private partnership body, will provide $25 million (approximately 3.3 billion KRW) in support for these projects.

The program drafts include HEMES-6G, which features a plan centered on Artificial Intelligence Radio Access Network (AI-RAN) cooperation with Taiwan. AI-RAN integrates AI computing directly into base station hardware and is considered a key architectural element of future 6G networks. These projects also serve as successors to the Hexa-X-II project, which concluded last year.

Elzbieta Pitory, Executive Director of SNS JU, stated: "As AI and 6G technologies mature, we are placing significant emphasis on large-scale demonstrations. Through these projects, we aim to pave the path from laboratory to market."