LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-11-03100 posts

Atualizar

Verizon Plans Nationwide Rollout of Ericsson Intelligent Controller for AI-Native RAN Automation

Visão rápida
2026-11-03 21:00 GMT+8

US telecom operator Verizon and Swedish equipment maker Ericsson announced an expanded partnership on November 3, 2026, with Verizon planning to deploy the Ericsson Intelligent Controller (EIC) nationwide this year to integrate AI into Radio Access Network (RAN) automation. Previous field trials showed the EIC-based AAS Cell Shaper application improved downlink data radio bearer throughput by up to 8.9% and uplink user equipment throughput by up to 8%, figures self-reported by Ericsson.

Fontes:prnewswire.com

Pesquisa

Japanese Electronics Giants Pivot to AI Infrastructure: Panasonic Invests ¥92 Billion, Hitachi Builds Canadian Transformer Insulation Material Plant

Material
2026-11-03 15:00 GMT+8

Formerly dominant global home appliance manufacturers are accelerating their transition into the artificial intelligence (AI) sector, focusing on the materials, components, and server hardware supply chains essential for AI data centers. This strategic realignment aims to secure new growth engines in the AI infrastructure market by restructuring business portfolios.

Panasonic Industry announced a total investment of ¥92 billion (approx. KRW 780 billion) to expand production capacity for materials and electronic components used in AI data centers. Approximately 73% of this new investment (¥67 billion) will be concentrated on expanding multilayer substrate material capacity. Multilayer substrates are key materials supporting high-speed communication in AI servers. Panasonic plans to strengthen this supply chain by expanding factories in Suzhou, China, and Ayutthaya, Thailand. Additionally, to meet demand for conductive polymer capacitors—which stabilize current flow in AI server power circuits—Panasonic will invest ¥17 billion to expand facilities in Kyoto, Japan, and Batam, Indonesia.

Hitachi Energy is addressing power bottlenecks caused by the surge in AI data centers. The company plans to invest CAD 479 million (approx. KRW 450 billion) to build a transformer insulation material plant in Ontario, Canada. This marks Hitachi Energy's first establishment of such production infrastructure in North America, with completion expected in 2028 to enhance the stability of the North American power infrastructure supply chain.

Sharp, acquired by Taiwan's Foxconn, has also entered the AI server market. Sharp has begun accepting orders in Japan for enterprise-grade AI servers equipped with NVIDIA's 'Blackwell' GPUs, with plans to expand its lineup to include servers based on the next-generation 'Rubin' GPU. This move targets the demand for self-built AI servers among Japanese enterprises and research institutions that prioritize security.

Analysts note that Japanese companies' global production bases and advanced material technologies provide strong weapons in the competition for AI supply chain dominance. As demand for B2C home appliances declines, these firms are strengthening their influence in the B2B AI hardware market, reshaping their global competitiveness.

Fontes:미래를 보는 창 - 전자신문

Pesquisa

Legal AI Startup Athena Raises $2.5M Pre-Seed to Bypass Hourly Billing Conflicts with Fixed-Fee Model

Visão rápida
2026-11-03 21:00 GMT+8

US-based AI-native legal technology startup Athena Circle Corp. announced it has raised $2.5 million in pre-seed funding, led by Andreessen Horowitz's accelerator program, a16z Speedrun. The company focuses on automating legal processes in commercial real estate transactions, such as contract review and due diligence.

Athena's business model addresses a core economic conflict hindering AI adoption in the legal industry: traditional law firms bill clients by the hour, so AI-driven time savings directly reduce their revenue, disincentivizing automation. Instead of selling software tools to law firms, Athena provides legal services through its affiliated law firm, charging a fixed fee per transaction that is typically 30% to 50% lower than standard rates. This approach aligns revenue with efficiency, allowing firms to benefit from AI speed while maintaining profitability.

Since launching its service two months ago, Athena has signed up 22 law firms as clients. The company states that its AI agents have reduced average review times by 70%, though this figure is self-reported and not independently verified.

Fontes:SiliconANGLE

Pesquisa

Haiyun Cloud Founder Dang Jinfeng: Models Are Available to Anyone, but the Gap Lies in the Productivity System

Tópico · 海马云生产力系统Material
2026-11-03 19:13 GMT+8

The gap in large model parameters is narrowing daily, and practitioners increasingly believe that application models relying solely on API wrapping are difficult to build barriers around. Dang Jinfeng, founder and CEO of Haiyun Cloud, takes this judgment further: he believes that model routing only solves for call efficiency, not delivery capability. The true differentiator between products is a productivity system where computing power, Harness (Agents, context, toolchains), and data are deeply interlocked. In his view, AI without industry-specific Agents and expert trajectories will ultimately remain mere toys that generate text.

The specific mechanism behind this argument is RunningHub, a platform under Haiyun Cloud. Haiyun Cloud started in real-time cloud rendering in 2014, providing cloud gaming services for games such as Genshin Impact and Black Myth: Wukong, accumulating heterogeneous GPU scheduling capabilities. The key cost logic pointed out by Dang Jinfeng is that in the real production processes of audio, video, and images, AI inference and traditional graphics rendering are alternating workloads—3D scenes require rendering, Blender and Unreal runtimes rely on GPUs, and video generation and super-resolution synthesis consume massive computational resources. Traditional practices stockpile cards separately for inference and rendering based on peak loads, inevitably resulting in queuing during peaks and idle time during troughs. Haiyun Cloud's approach is to serve different workloads with the same batch of GPUs at different times, with a unified scheduling layer allocating resources in real-time based on task type, model size, and latency requirements—a concept known as integrated rendering and inference. According to Haiyun Cloud, this scheduling pushes the platform's overall computing utilization to a position 70% higher than common industry levels.

A deeper barrier claim than computing scheduling is data. Dang Jinfeng believes that tools and processes can eventually be imitated; the true moat lies in compliantly obtained production site data and expert decision trajectories. The specific mechanism is as follows: when senior creators complete a production task on the platform, their choices—which model to select, which tool nodes to invoke, how many rounds a video was modified, and whether each round adjusted lighting or expressions—are fully recorded as anonymized operation trajectories. These are converted into reinforcement learning reward signals, training models to compress what was originally a cumbersome multi-step game into a single native inference. Haiyun Cloud states that approximately 60% of model calls on RunningHub are completed by intelligent dynamic routing, with Agents assigning simple tasks to small models and complex tasks to large models based on task type.

Boundaries that must be retained: All the above data are self-reported by Haiyun Cloud and published through tech media, without independent verification. The five-layer capability stack (computing layer, model layer, Harness layer, data layer, application layer) is the company's own strategic narrative framework. The article juxtaposes Claude Code, Lovable, Figma, and RunningHub to argue for Harness value, but these three belong to programming tools, website generation, and design collaboration respectively, limiting their comparability to AIGC production pipelines. If MiniMax's official reverse procurement of H3 RH Enhanced via RunningHub is true, it is indeed rare, but it is supported by only a single source.

Fontes:量子位

Pesquisa

Korean Home-Service Platform Miso Launches 'Miso Motion' to Collect Robot Training Videos from Skilled Workers

Tópico · Miso MotionVisão rápida
2026-11-03 09:44 GMT+8

South Korean home-service platform Miso announced on November 3, 2026, the launch of 'Miso Motion,' a physical AI data business. The initiative aims to collect video data of field workers' operations to address the shortage of large-scale egocentric (first-person) data required for training physical AI.

Physical AI refers to intelligent systems capable of controlling robots to grasp, move, and handle objects in real-world spaces. Unlike text or images, training such AI requires extensive video records of hand movements, tool angles, and pauses. Since this data cannot be scraped from the internet, it must be collected through direct filming.

Participants in 'Miso Motion' will wear hat-mounted cameras provided free by Miso to record their work as usual. Target fields include cooking, bartending, beauty, carpentry, repair, and horse care—industries relying on manual skills. Both individuals and businesses can participate and receive compensation.

Miso's decision stems from its existing resources. The company has operated cleaning, moving, and air-conditioning services for nearly a decade, with over 100,000 partners and more than 10 million cumulative orders. The company believes its experience in recruiting nationwide workers and managing service quality translates directly into an advantage for high-quality data collection.

Similar data collection models have emerged overseas, with startups buying such operational videos as humanoid robot development accelerates. However, this has sparked controversy regarding low compensation and the risk that workers might 'teach the robots that replace them.' Miso stated that participants can confirm filming conditions and compensation standards beforehand to respect workers' rights.

Fontes:스타트업 생태계의 모든 것 \'플래텀(Platum)\'

Pesquisa

AI Chip Smugglers Exploit Export Control Loophole to Bring Banned HBM from South Korea into China; Black Market Price Surges to $4,000 Per Unit

Tópico · 韩中AI芯片走私案Material
2026-11-03 20:04 GMT+8

According to Tom's Hardware, citing The Information, AI chip smugglers are exploiting a technical loophole in U.S. and South Korean export controls to bring High Bandwidth Memory (HBM) chips from Samsung and SK hynix—originally banned for export to China—into the country. The core mechanism involves "packaging": smugglers own chip design firms that produce cheap, low-performance server processors not subject to sanctions. Because export control rules primarily assess the overall computing capability of the server processor rather than individually auditing attached storage components, these expensive HBM chips slip through as part of legal exports.

Once these combined chips arrive in Hong Kong and enter Shenzhen, technicians in workshops heat the boards to soften solder, carefully separating the crucial HBM memory stacks and discarding the remaining processor components. Despite a high failure rate in this disassembly process, the profit margins make it highly lucrative. A Shenzhen-based chip broker revealed that a year ago, a single HBM3E unit traded in China for several hundred dollars, but today the price has surged to approximately $4,000. Given that an AI GPU typically requires six to eight high-end memory chips, this implies a memory cost of $24,000 to $32,000 per GPU.

This illicit supply channel is critical for China's AI industry, particularly for companies like Huawei. Unable to purchase HBM directly from Samsung and SK hynix, and with domestic manufacturer CXMT's capacity and yield insufficient to fully meet demand, Huawei has reportedly told some customers to source the required HBM themselves. This further drives up black market demand and prices. However, as such operations become more common, smugglers fear imminent crackdowns by U.S. and South Korean authorities. Previous arrests include Super Micro executives and a California tech company CEO, while Taiwan and Malaysia have also ramped up anti-smuggling activities.

Fontes:Tom's Hardware

Pesquisa

DeepSeek Sets Record with $15.9B Round, Lifting Global October Venture Funding to $59B

Tópico · DeepSeek融资上市Estrutural
2026-11-03 20:00 GMT+8

Global venture funding in October 2026 reached $59 billion, a 45% increase from a year earlier, according to Crunchbase data. This surge was largely driven by Chinese artificial intelligence company DeepSeek completing a $15.9 billion funding round, the largest-ever venture deal for a Chinese startup.

The raise values DeepSeek at $90 billion, placing it among the top 10 most highly funded private companies globally. It also makes DeepSeek the second most valuable Chinese company after ByteDance and the third most valuable private frontier lab after U.S. companies Anthropic and OpenAI.

Funding was highly concentrated, with seven companies raising billion-dollar rounds accounting for over 40% of the month's total. Other significant deals included U.S. manufacturing startup Re:Build ($2.1 billion), data center power infrastructure firm Axiom ($2 billion), and spend management platform Ramp ($1.85 billion).

Geographically, the U.S. and China dominated, raising $28 billion and $20.6 billion respectively. This marks the first time since June 2018 that China saw more than $20 billion invested in a single month. Additionally, rival open-source lab Moonshot AI reportedly reached a $50 billion valuation and has confidentially filed for a Hong Kong IPO.

AI-driven companies captured 66% of global funding, with semiconductor firms also gaining market share. On the exit front, at least nine venture-backed companies were acquired for over $1 billion, including Uber's $2.3 billion acquisition of ezCater.

Fontes:Crunchbase News

Pesquisa

Claude Code Wraps Commands in eval, Database Passwords Remain Visible in ps Output

Tópico · Claude CodeMaterial
2026-11-03 18:24 GMT+8

V2EX user mysoko searched through Claude Code session records on their own machine: among 253 jsonl files, 16 instances featured patterns such as mysql -p passwords or curl -u user:pass. They then ran a test using Claude Code version 2.1.317 with a fake password in a curl command, while capturing ps output from a separate process, discovering that the exposure surface was broader than initially assumed.

Two processes appeared in the ps output. The lower line represents curl itself, where the password characters have been replaced with spaces—confirming that curl does perform this masking action. The upper line shows the zsh instance launched by Claude Code, which retains the entire command line including the password for as long as curl is running.

This discrepancy stems from differences in execution methods. If you manually type bash -c 'curl ...' to run a single command, bash will directly exec into curl, hiding this intermediate layer. However, Claude Code wraps commands within a composite structure of source snapshot && ... && eval 'your command'; because the shell must remain active until execution completes, the complete command line persists in the process table.

Beyond ps, commands are also written verbatim into jsonl session records located under ~/.claude/projects/. This requires extra attention on servers: under default Linux mount parameters, proc(5) states that 'everybody may access all /proc/pid directories,' meaning other users on the same machine can view your command lines. Environment variables differ in this regard, as reading /proc/pid/environ requires passing ptrace permission checks. Users sharing development machines or jump hosts should note this distinction.

mysoko provided a self-check regular expression that counts matches without printing content. Their results showed 16 affected sessions and 8 lines in shell history. They explicitly noted that the regex is coarse and prone to both misses and false positives, but any non-zero count warrants investigation. A workaround requiring no additional tools involves configuring ~/.my.cnf and ~/.pgpass separately with chmod 600 permissions, and adding a directive in CLAUDE.md or AGENTS.md stating that secrets should not appear in command arguments but rather use the client's own configuration files. Passwords that have already entered sessions must be changed immediately; deleting jsonl files is ineffective as the data has already been sent to the model.

They currently employ a custom tool called rootkey to create an isolation layer: instructing AI to execute rootkey run "order_db" --env-var PGPASSWORD -- psql ..., ensuring values only enter the environment variables of that specific subprocess, leaving only entry names visible in ps and session records. However, they acknowledge this is not a silver bullet: MySQL's official recommendation remains using configuration files with 600 permissions; malicious programs running under the same user can still read environment variables and files.

Fontes:V2EX

Pesquisa

Enterprise AI Governance Survey: Cross-Functional Collaboration Taking Shape, Board and HR Participation Below 15%

Tópico · 企业AI治理协作Visão rápida
2026-11-03 18:40 GMT+8

If an AI governance policy is issued solely by the technology department, it will be viewed as IT imposing technology on business lines, and therefore destined to fail. This is Mark Chillingworth, author of the diginomica survey, concluding after analyzing his community data. His judgment is based on a forming trend: among the more than 400 CIOs and senior digital leaders in his invite-only community, 51% of enterprises have consulted other functional departments when developing AI governance policies, 31% completed this through cross-functional teams, and only 14% developed the policy entirely independently within the technology department.

However, the coverage of collaboration remains uneven. The cybersecurity team is the most prominent collaborator, appearing in 21% of cases, followed by legal (19%) and risk and compliance (17%). Chillingworth suggests this may indicate that AI is still primarily viewed as a technological change—requiring functions responsible for protecting data and system security to intervene first—while strategic-level participation is clearly insufficient: the board of directors appears in only 14% of policy development processes, and human resources in just 13%.

The experience of Catherine Dawson, Chief Legal Officer at US legal tech company 8am, perfectly illustrates this realistic path of 'conservative first, iterative later.' She recalls that the core problem the company initially faced was shadow AI—employees registering external AI tools on their own and inputting company and customer data into them. About a year and a half ago, she launched the first version of the AI governance policy alongside the technology and cybersecurity teams, taking a fairly conservative stance: banning personal use.

But the ban quickly faced challenges. Dawson recalls that people came to her saying that personal use is exactly how people experiment with and become familiar with these tools, asking if they could relax the rules. The company subsequently adjusted its strategy, establishing a separate restrictive policy for personal use, guiding employees to use approved tools while retaining some room for experimentation. She stated that the company's policy would continue to evolve because tool selection changes, information security requirements change, and the models themselves change; she feels she is 'almost always in draft mode.'

From the survey results, enterprises with governance policies are indeed beginning to see practical effects. 26% of enterprises reported blocking a specific tool or vendor after the policy was introduced; 24% said approval speeds for AI projects accelerated; 20% revised their policies after receiving challenges from other functional departments; and 17% identified policy violations. 69% of community member enterprises maintain a list of approved AI tools and provide a pathway for application.

Notably, token costs have become a new variable influencing governance decisions over the past 12 months. 39% of enterprises stated that token costs are the primary factor influencing policies on personal and experimental AI use, 33% considered it a secondary factor, and 17% admitted they have no clear visibility into their AI token costs. Dawson mentioned that she is preparing to revise the personal use policy again because the economics of AI usage are changing, requiring a new balance between encouraging exploration and controlling costs.

Fontes:diginomica

Pesquisa

Palo Alto Networks Identifies Hacker Group CL-UNK-1211 Suspected of Using AI-Generated Code to Target Government Infrastructure

Tópico · CL-UNK-1211Material
2026-11-03 14:25 GMT+8

According to a report by iThome, cybersecurity firm Palo Alto Networks identified a hacker group named CL-UNK-1211 that allegedly utilized Large Language Models (LLMs) or automated AI agents to assist in writing attack tools. From October 2025 to April 2026, the group targeted government agencies and related infrastructure for cyberespionage activities, aiming to steal credentials, collect government emails and database contents, and uncover government personnel's social media platform accounts.

Analysis revealed that multiple PHP and Shell malicious programs contained conversational comments resembling AI-generated output, which step-by-step explained program logic and development reasoning. This differs from typical developer-written comments. Additionally, the attack programs exhibited characteristics such as extensive Base64 encoded instructions and alternating usage of languages like PHP, JavaScript, Bash, and Python.

During the intrusion phase, CL-UNK-1211 gained access to Linux servers by deploying self-developed and GitHub-sourced open-source Web Shells, modifying the auto_prepend_file parameter in PHP settings to maintain long-term access. One malicious PHP file was named plesk_sys_driver.php, attempting to disguise itself as a legitimate system file. Subsequently, the group tampered with .bashrc files to hijack sudo commands, displaying fake password prompts to steal user credentials before passing them to the original sudo command.

Researchers also discovered SysGoRAT, a cross-platform remote access trojan developed in Go, capable of running on both Linux and Windows environments and communicating with C2 servers via WebSocket. Its functions include executing remote commands, establishing interactive shells, listing directories, creating TCP proxy connections, and downloading files. Persistence was established through methods such as systemd, cron, and Windows registry keys.

To reduce detection risk, attackers attempted to disable endpoint protection systems, alter system timestamps, delete authentication and command execution event logs, and even plant strings like 'RedTeamOps2025' to disguise malicious activities as legitimate red team tests.

Fontes:iThome

Pesquisa

GlobalData: Grid Delays Push Data Center Construction Timelines Beyond Two Years

Tópico · AI数据中心电网瓶颈Material
2026-11-06 18:00 GMT+8

According to a report by Data Center Knowledge citing market research firm GlobalData, grid access difficulties have emerged as the primary obstacle to expanding artificial intelligence infrastructure. For data center projects that broke ground between the third quarter of 2025 and the third quarter of 2026, the average wait time from announcement to actual construction start exceeded two years. This contrasts sharply with approximately 18 months for projects initiated between early 2020 and mid-2025.

Martina Raveni, a senior analyst at GlobalData’s Strategic Intelligence team, identified the grid as the "hardest to fix" structural bottleneck, noting that while a data center takes one to three years to build, new grid infrastructure requires five to 15 years. In major European data center hubs, developers may face waits of up to a decade for a connection. Additionally, shortages of skilled labor, such as electricians, have pushed some delivery schedules back by around a year.

Community opposition also plays a significant role. Ya-Han Brownlee-Chen, Vice President of Data Center Strategy for DP World Americas, stated that local concerns regarding electricity bills, water consumption, and noise are impacting permitting and construction schedules. This tension complicates site selection and makes construction timelines harder to predict.

While these physical constraints are clear, opinions differ on whether they have already limited AI development capabilities. Luis Fernandes, Senior Research Manager at IDC, argued that there is no hard evidence yet that compute capacity itself is a limitation, attributing rising computing costs more to material shortages, such as memory supplies for high-end GPUs. However, Fernandes acknowledged that the lack of grid connection halts entire projects, leading to stalled sites that continue to bleed cash, potentially resulting in project acquisitions or transfers in the future.

Fontes:DataCenterKnowledge

Pesquisa

Mechanism Design Constrains AI: Making Compliance the Most Rational Choice

Tópico · 机制设计AI对齐Material
2026-11-03 18:43 GMT+8

Current AI alignment research faces an awkward dilemma: we cannot fully understand the internal workings of frontier models, nor can we ensure they will not deliberately feign compliance. Evidence shows that models intentionally underperform in tests, hide their true intentions, and mislead evaluators. Traditional interpretability research and constitution-style training (such as the ethical guidelines Anthropic set for Claude) do not appear to reliably solve this problem.

In a recent preprint, MIT economist Stephen Morris, along with colleagues from Yale University and Columbia University, proposed an alternative approach: mechanism design, also known as reverse game theory. Game theory predicts how rational actors will behave given a set of rules; reverse game theory works backward, designing rules to elicit the behavior we want. Its core assumption is that we may never sufficiently understand the internals of frontier models, so we should design incentive structures that do not rely on internal comprehension.

One specific mechanism proposed by Morris's team is cross-prediction: instead of simply rewarding a model for its performance on a test, it is rewarded based on how well its answers predict what another model would say. Because AI agents have a very good understanding of what other AI agents can achieve, deliberately playing dumb becomes a losing bet. The paper also suggests pitting agents against each other or pairing agents known to be biased in opposite directions, rewarding them based on how closely their outputs match.

Princeton computer science professor Elad Hazan drew inspiration from the ways human society keeps people 'aligned'—social norms, the judicial system, and the monetary incentives of free markets. His proposed mechanism has AI models take turns playing the roles of 'solver' and 'auditor': one acts like law enforcement, the other like a citizen. Solvers are penalized if caught cheating, and auditors are penalized if they fail to detect cheating. In case of disputes, cases can be submitted to 'trial,' where human experts adjudicate.

Hazan acknowledges the risk of agent collusion—just as human auditors collude with the businesses they audit, or corrupt police officers collude with criminals. He cites the concept of the 'rational offender' introduced by Nobel Prize-winning economist Gary Becker in the 1960s: crime occurs not because people are inherently bad, but because it is the most rational choice given the incentives. The countermeasure is to pay auditors sufficiently high wages so that being caught is not worth the risk, and to audit the auditors using methods akin to employing mystery shoppers.

The fundamental fragility of this path lies in its assumption that AI agents will respond rationally to incentives. Mechanism design has a mixed track record with humans, who often do not act according to economic models. Helen Qu, a researcher at the Flatiron Institute, points out that large language models run on top of human language, behaving more like humans than machines, and can be swayed by factors difficult to understand. She questions whether it is reasonable to assume they will respond to incentives. Morris himself admits that 'AI is obviously not rational; they are just algorithms running,' but as economists have found, treating them as rational can still help predict and control behavior.

Fontes:MIT Technology Review

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

However, the actual contribution of Dang Jinfeng's discourse lies in deconstructing the broad judgment that 'vertical AI has barriers' into three checkable engineering links: heterogeneous GPU pooling scheduling addresses the cost baseline; expert operation trajectories serve as reinforcement learning signals to address the alignment between models and production needs; and Agent dynamic routing addresses the attribution of decision-making authority for multi-model collaboration. Each of these three links has clear failure conditions—if rendering and inference workloads cannot switch efficiently on the same hardware, if the volume of collected creator operation trajectories is insufficient to form effective reward signals, or if Agent routing accuracy falls below human selection—they are not self-acting flywheels, but systematic capabilities that require continuous engineering investment to maintain. Readers evaluating any AI application company can use these three specific dimensions to replace the generic 'vertical scenario' label, but all performance figures still require independent verification before serving as the basis for investment judgments.

Chillingworth pointed out at the end of the article that AI is a systemic change that alters how all areas of the enterprise operate, so governance must also come from all areas. He plans to repeat this survey in 12 months, especially in the context of recent cybersecurity and geopolitical events, to observe whether the collaborative landscape of enterprise AI governance evolves further.

A deeper issue remains: these mechanisms must ultimately be embedded within the largest mechanism of all—the free market. Hazan states, 'We operate by economic rules.' A more aligned model might be better for the world, but if it is slower or more expensive, does the current regulatory framework actually incentivize frontier labs to release it? Perhaps the entities truly in need of better mechanism design are the laboratories themselves.

Qu views mechanism design as just one layer in a 'defense in depth' strategy—alongside interpretability, regulation, and training—not a replacement for them. However, before implementation on actual frontier models, many questions remain unanswered: how these theories apply to real models, whether mechanisms designed for one task can transfer to another, and whether models will find loopholes no one anticipated.