ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-11-11104 posts

Refresh

Krafton Q3 Operating Profit Falls 35.5% on Event Costs, but Cumulative Results Hit Record High with Additional Buyback

Topic · Krafton三季度业绩Material
2026-11-10 15:03 GMT+8

South Korean game developer Krafton released its preliminary third-quarter results on November 10, following an investor briefing. The company reported Q3 revenue of 971.4 billion KRW, up 11.6% year-over-year, but operating profit fell 35.5% to 225 billion KRW. Krafton attributed this decline primarily to concentrated marketing and participation costs associated with the Gamescom exhibition in Germany held in August.

Despite the single-quarter dip, Krafton’s cumulative performance for the first three quarters reached all-time highs for the period. Total revenue from January to September amounted to 3.633 trillion KRW, a 50.9% increase year-over-year, while cumulative operating profit rose 13.8% to 1.1975 trillion KRW. This divergence suggests that core business growth remains robust, with short-term expense fluctuations not altering the underlying profitability trend.

In terms of capital allocation, Krafton’s board approved an additional share buyback plan on November 3, committing 300 billion KRW to repurchase and cancel its own shares. Combined with approximately 400 billion KRW already spent on buybacks and 99.6 billion KRW in dividends paid through the third quarter, the total shareholder return for the year stands at 917.2 billion KRW. This figure represents the highest annual shareholder return since the company initiated its return policy in 2023, signaling management’s confidence in cash flow generation and a commitment to enhancing per-share value through equity reduction.

Sources:스타트업 생태계의 모든 것 \'플래텀(Platum)\'

Research

Hidden Costs of AI Infrastructure: Google Inventory Surges 5x, Intel Commits Billions to Substrate Supply

Topic · AI基建库存与供应链瓶颈Material
2026-11-11 22:00 GMT+8

As AI infrastructure investment expands, the time lag between equipment procurement and actual compute availability, along with the associated working capital burden, has emerged as a critical industry bottleneck. According to EE Times Japan, Alphabet's inventory levels surged from $2.4 billion at the end of 2025 to $13.3 billion by September 2026. This increase is primarily driven by a shift in Google's business model: beyond providing computing power via Google Cloud, the company has begun selling complete TPU (Tensor Processing Unit) systems directly to customer facilities. This hardware sales approach ties up significant capital in manufacturing and logistics before revenue recognition.

Alphabet CFO Anat Ashkenazi explained during the third-quarter earnings call that once customers sign contracts, orders enter the backlog, and Google initiates system manufacturing ahead of delivery and billing. While this helps stabilize demand visibility, the company has not disclosed the proportion of TPU systems within total inventory or specific delivery timelines. This case highlights that the key metric for measuring AI infrastructure maturity is shifting from simple capital expenditure (Capex) to the 'time-to-deployment' interval.

Meanwhile, Intel faces similar capacity conversion and supply chain constraints. To address surging demand for server CPUs, Intel CFO David Zinsner revealed that the company has reallocated some 'Intel 18A' process capacity originally planned for Panther Lake client chips to produce Granite Rapids server CPUs. Additionally, CEO Lip-Bu Tan noted that shortages of organic substrates required for advanced EMIB packaging have become a major constraint on expansion. Consequently, Intel signed two substrate supply agreements after Q3 2026, with expected funding totaling $5.3 to $6.1 billion, including approximately $3.3 billion payable by the end of 2027.

These dynamics indicate that AI compute supply is constrained not only by wafer fabrication capabilities but also by backend packaging, substrate availability, and the resulting working capital pressure. For investors, attention must shift from aggregate spending figures to specific supply chain bottlenecks and cash flow efficiency.

Sources:EE Times Japan

Research

Aotter Vice President Su Peng-han: AI Can Read Enterprise Data Without Replacing ERP—Standardizing Data and Introducing MCP Logic

Material
2026-11-11 13:09 GMT+8

When East Asia Lighting first approached Aotter, the problem they wanted to solve was not "which model to use," but who could help them replace their old ERP. Replacing an ERP is a massive undertaking for traditional industries—many companies' ERPs were highly customized years ago by system integrators who cannot provide APIs or cooperate with development. Even if a new vendor creates an AI chatbot, integrating it would hit roadblocks. Actually replacing the system could paralyze the company for half a year, and the people and processes bound to the system would be difficult to change alongside it.

Su Peng-han, Aotter's co-founder and vice president, offered a solution: do not ask clients to replace their ERP. His positioning judgment is clear: build tools for AI to use, rather than trying to replace or compete with AI agents, because agents are too powerful. The specific approach is that after the client explains their workflow, Aotter's Forward Deployment Engineer (FDE) team, accompanied by product managers and business development personnel, moves in to pull scattered data from various systems into AotterClam AI's data middleware. This middleware sits on top of the client's original ERP, standardizes the data so AI can understand it, and leaves subsequent analysis and applications to the AI.

Founded in 2012, Aotter began as a software R&D company and owns the YouTube channel Aotter Girls, which has millions of subscribers. Launched in April this year, the enterprise AI data brain AotterClam AI focuses its efforts on organizing enterprise data scattered across different locations. Once organized, the data is handed over to the AI already in use by clients via the MCP protocol. Executives do not need to understand what MCP is; they can simply ask questions in the Claude or GPT interface to receive analysis.

How effective is it in practice? Xie Lun, Aotter's co-founder and CEO, said that mid-to-senior level executives at East Asia Lighting now directly open Claude or GPT. To prepare board reports, they just tell the AI to pull up the gross margin for this quarter, how much it grew, the reasons why, which business metrics are lagging, and which sales figures are high but profits are low—the answers come out instantly.

Xie Lun added a deeper observation: the real pain point in manufacturing often lies in transportation. Choosing between sea freight and air freight involves significant cost differences, yet there is no API for sea freight, though inventory data shows the status. In the past, with specialized division of labor, everyone only looked at their own systems, leaving many gaps in company operations; once data is connected, one can determine that a batch of goods does not need air freight and is still on time, allowing schedules to be optimized immediately.

However, the most difficult step in data standardization often gets stuck on people. Su Peng-han mentioned a phenomenon: every company has a very mysterious person who knows how to fill out the form. This person might be an auntie or an uncle who knows what to write in which field under what circumstances. Having operated the system tens of thousands of times, things that look difficult to others are natural to him; he even knows what to do during the 30 seconds it takes for a submission to go through. Many ERPs are not deeply customized for specific industries, so each industry develops its own special usage methods. 'We don't have orders, so why must the title on that entry form say Order?' Everyone calls it an order, but in reality, it is not an order.

Sources:INSIDE

Research

Gartner Recommends Establishing an AI Central Bank and Value-Risk Flywheel to Bridge the Enterprise AI Governance Gap

Topic · GartnerMaterial
2026-11-11 12:11 GMT+8

At the Gartner IT Symposium/Xpo conference held in Florida, USA, in October 2026, Gartner presented Chief Information Officers (CIOs) with a specific contradiction: 71% of enterprises plan to deploy AI agents within the next 12 months, yet only 21% have established governance capabilities for AI agents. Kristin Moyer, Distinguished Research Vice President at Gartner, pointed out that this gap makes it difficult for the speed of value creation to exceed the accumulation of risks. Meanwhile, a Gartner Q3 survey of 303 CIOs and IT risk management executives showed that only 25% of CIOs can effectively coordinate technological risks with business risks.

The first mechanism proposed by Gartner is the value-risk flywheel model. The flywheel features two core forces: momentum, which accelerates the realization of results, and drag, which may slow down progress. Daryl Plummer, Distinguished Research Vice President at Gartner, emphasized that momentum is not entirely beneficial; appropriate drag can also yield positive effects. Excellent CIOs should know how to create benign momentum, strategically introduce benign drag, while reducing negative momentum and negative drag. A governor located at the center of the flywheel is needed to decide when value acquisition justifies taking risks, which values should take precedence, and how to ensure organizational safety.

The second mechanism is the establishment of an AI Central Bank. This organization serves as a virtual control platform for internal AI decision-making within the enterprise, capable of establishing accountability systems, implementing technology governance, managing economic activities, and forming shared wisdom. Moyer explained that its purpose is not centralized management, but rather achieving common goals, integrating coordination, and fostering cross-departmental collaboration to ensure transparency, coordination, and accountability mechanisms between humans and machines. Plummer simultaneously reminded that the AI Central Bank itself cannot eliminate risks or directly create value; it is still people who truly accomplish these things.

The third mechanism is the cultivation of AI entrepreneurs. Gartner believes that high-level AI users, now referred to as AI Artisans, are gradually becoming the ultimate creators of value. They utilize AI to propose innovative solutions and redefine work models. To expand the value end of the flywheel, enterprises need to build a culture of innovative talent with AI literacy, provide space for personnel to redesign their work methods through serendipitous opportunities, and offer rewards when results are achieved.

It must be clarified that these three mechanisms are a coordination framework proposed by Gartner based on its survey data, rather than standardized schemes verified through implementation. Their value lies in providing CIOs facing governance gaps with a structured thinking path, but specific effectiveness still depends on each enterprise's own strategies and risk tolerance.

Sources:iThome

Research

SK hynix LPDDR5X Earns IEC 61508 SIL 3 and ISO/SAE 21434 Certifications, Expanding Physical AI Safety Standards

Topic · SK海力士物理AI认证Material
2026-11-11 07:59 GMT+8

South Korean semiconductor maker SK hynix announced on November 11, 2026, that its LPDDR5X dynamic random-access memory (DRAM) has obtained Safety Integrity Level 3 (SIL 3) certification under the IEC 61508 standard for industrial functional safety. Additionally, its product development process has earned Cybersecurity Management System (CSMS) certification based on the ISO/SAE 21434 standard. Both certifications were granted following rigorous assessments by the global certification body SGS-TÜV Saar (involving both SGS Korea and SGS Japan).

IEC 61508 is the foundational international standard for industrial functional safety, classified into four levels from SIL 1 to SIL 4, with SIL 3 being the highest level attainable for a single memory semiconductor. The standard requires maintaining safety in real time without interrupting system operation in continuous mode, providing a foundation for applications such as real-time robotics control. Previously, SK hynix had obtained ISO 26262 ASIL-D automotive functional safety certification in January 2026; this new certification extends its functional safety credentials beyond automotive applications to a broader range of industrial scenarios including robotics and smart factories.

Regarding cybersecurity, ISO/SAE 21434 is an international standard for automotive cybersecurity that requires managing cybersecurity activities throughout the entire product lifecycle. As vehicles evolve toward the software-defined vehicle (SDV) model and become increasingly connected to external communication networks, the United Nations Economic Commission for Europe (UNECE) has mandated the establishment of a CSMS as a prerequisite for vehicle type approval. The CSMS certification validates SK hynix's capability to safeguard against malicious attacks, complementing its earlier TISAX certification (a global information security standard for the automotive industry) to form a comprehensive safety and security portfolio required by the physical AI market.

Notably, SK hynix secured certification not only for its LPDDR5X product but also for the entire development process behind it. This demonstrates that the company has established a framework for consistently incorporating safety requirements into both current and future products. Leveraging the low-power, high-performance characteristics of LPDDR5X, SK hynix plans to actively address the functional safety requirements of customers developing robotic platforms and System-on-Chip (SoC) solutions.

Sources:SK hynix Newsroom

Research

Meta Engineering Blog Details "Your Algorithm": How LLMs of Different Sizes Collaborate to Turn a User's Sentence into Recommendation Results

Material
2026-11-11 01:00 GMT+8

The engineering team at the US technology company Meta published a technical blog post on November 10, 2026, detailing the internal architecture of the "Your algorithm" system. This tool was previously launched to users in October, allowing people to write in natural language what they want or do not want to see in their Facebook Feed and Reels. The blog's focus is not on feature introduction, but rather on explaining an engineering challenge: how to transform a casual sentence spoken by a user into a sorting signal that actually takes effect within the recommendation system.

The Meta team points out that traditional recommendation systems rely on implicit behavioral signals (such as clicks and dwell time). While these can run at scale, they adapt slowly and fail to capture subtle intentions. "Your algorithm" attempts to fill this gap: when a user directly states "I want to see more content about urban gardening" or "I do not want to see AI-generated images anymore," the system must immediately understand this statement, store it, and execute it in subsequent recommendations. This presents three specific challenges—new interests must appear quickly without overwhelming other content; exclusion items must filter reliably even when topics are vague; and previously cached old recommendations must be updated.

On the "Interested In" side, each interest added by a user drives four independent retrieval paths: two direct matches (via search and via topic clustering) and two extended coverage paths (via the content graph and item-to-item similarity). The purpose is to ensure that content for brand-new interests does not completely miss the recall stage. Candidates that pass the relevance check receive a boost in the late-stage ranking, after which the control mechanism of the reranking layer determines where they appear. Meta emphasizes "balance over quantity": the control layer spaces out recommendations for the same interest and sets a daily cap for each interest; during the retrieval stage, each request fetches candidates for only one interest, allowing all interests to take turns getting opportunities.

On the "Not Interested In" side, filtering is distributed across multiple stages. During the ranking phase, embedding comparisons identical to those used in the relevance check are employed to remove candidates that are too close to the excluded interest, while demotion rules push down near matches. After ranking, an LLM reranker reviews the final list. On the device end, cache invalidation mechanisms discard content that has been queued for display but matches new exclusions. Notably, not all exclusions are topic-based—requests to reduce AI-generated content or sexually suggestive content map to dedicated demotion signals rather than topic matching; if a user specifically names a creator, that creator's content is filtered out directly.

LLMs appear in three different positions within this system, each undertaking different tasks and corresponding to models of different scales. Upon a user's first visit, an LLM is responsible for generating the initial profile: it converts continuously refreshed interest signals into readable topic names, populates the "Interested In" list, and writes a "Lately" summary. Whenever a user edits their preferences, another smaller LLM handles validation—checking whether the phrase constitutes a coherent topic, detecting duplicates, and mapping them to standardized names. Meta explains that this step uses a small model to serve efficiently at scale. After ranking is completed, a fine-tuned LLM reranker performs a final check.

Sources:Engineering at Meta

Research

Microsoft Patches Actively Exploited Outlook Zero-Day CVE-2026-124194 Allowing NTLMv2 Hash Theft

Material
2026-11-11 11:02 GMT+8

In its November Patch Tuesday release, Microsoft addressed over 1,000 security vulnerabilities, with particular attention drawn to CVE-2026-124194 in Outlook due to detected exploitation activity. This flaw allows unauthenticated attackers to send specially crafted message files that, when opened by a user in Outlook, cause the application to connect to an attacker-controlled network location without displaying the usual warning for accessing external content.

Successful exploitation enables attackers to obtain NTLMv2 hash data, potentially allowing them to impersonate the user. The vulnerability has a CVSS severity score of 6.5. Microsoft urges users to install the patch promptly to mitigate these spoofing attacks.

Sources:iThome

Research

Joint Study by University of Manchester and Universidad del Pais Vasco Reveals Lecturers Overestimate Student AI Usage by 2.12 Points

Topic · AI使用感知差距Quick take
2026-11-11 10:05 GMT+8

A new joint study by the University of Manchester in the UK and the Universidad del Pais Vasco in Spain reveals that university lecturers significantly overestimate the extent to which students use generative artificial intelligence in their studies. Conducted between April and June 2024, the research surveyed 45 academic staff and 86 students from a school of social science disciplines at a research-intensive university.

The data shows that when rating their perception of AI use, responding academic staff recorded an average score of 5.23 out of 7, whereas students self-reported an average of only 3.11. This notable perception gap suggests that educators may be formulating strategies based on assumptions rather than actual experience. Researchers noted that while students might underreport usage, it remains difficult for lecturers to spot AI use, adding complexity to policy development.

Dr. Mark Carrigan, Senior Lecturer in Education at The University of Manchester and one of the study’s authors, stated: “We need to move beyond a climate of suspicion and focus on how these tools can be used appropriately to support learning, while rethinking teaching and assessment for an AI-enabled world.” He emphasized that because staff and students have quite different perceptions of how AI is being used, it is difficult for universities to develop effective responses based on assumptions rather than a clear understanding of students’ experiences.

Sources:TechRadar

Research

US Regulator Approves Construction Permit for X-Energy's First Industrial-Scale Small Modular Reactors

Topic · X-energyMaterial
2026-11-11 04:43 GMT+8

Nuclear startup X-energy announced that the U.S. Nuclear Regulatory Commission (NRC) has granted a construction permit for its first four commercial Xe-100 small modular reactors (SMRs). The reactors will be built at a Dow chemical plant in Texas, with each unit capable of generating 80 megawatts of electricity and 200 megawatts of heat. X-energy stated that this approval represents the first time an advanced nuclear reactor has been permitted for construction at an industrial site in the United States.

Small modular reactors are seen as a key solution to the surging energy demands of AI data centers. Proponents argue that shrinking the reactor size will speed up construction and reduce costs, though these advantages remain unproven. No SMR is currently operational in the U.S., and X-energy expects the Dow reactors to become operational in the early 2030s.

The approval has drawn criticism from safety advocates. Edwin Lyman, director of nuclear power safety at the Union of Concerned Scientists, argued that the NRC achieved this rapid approval by pushing critical safety questions to a later date, rather than resolving them before greenlighting construction.

Despite these concerns, hyperscalers have shown strong interest in nuclear power. Amazon’s Climate Pledge Fund led X-energy’s $500 million Series C-1 round, and the tech giant has committed to buying 5 gigawatts of capacity from the nuclear startup. X-energy went public in April 2026, raising $1 billion, but investor optimism has tempered due to the high costs and long timelines associated with nuclear projects. The company's shares are currently trading at less than half their initial offering price. The recent permit approval provided a slight boost to the stock, recovering some of last week's losses.

Sources:TechCrunch

Research

US Solar Struggles to Decouple from China as Section 232 Tariffs Loom

Topic · 美国太阳能去中国化Material
2026-11-10 19:53 GMT+8

Washington aims to sever US solar firms' reliance on China, but the process proves more complex than anticipated. In December, new Commerce Department rules under Section 232 will require panel manufacturers to pay a fixed minimum price and a tariff of up to 15% for imported cells and certain specialized components. These measures are designed to end years of cat-and-mouse games where Chinese makers shifted factories across Asia and Africa to dodge US tariffs.

Texas-based T1 Energy is navigating this shift. CEO Daniel Barcelo stated that the rules "will really level the playing field" against rivals. However, the resulting price hikes have caused panic among installers. According to research firm Anza, at least 1 gigawatt worth of large-scale solar projects has been cancelled in recent months due to Section 232-related price adjustments.

T1 itself was born from these trade wars. After abandoning an EV battery plan due to Chinese competition and US incentives, Barcelo acquired a Dallas factory from Chinese giant Trina Solar for $621 million in December 2024. Despite Trina dropping its stake, T1 remains deeply tied to Chinese technology: Trina is its anchor customer, buying 1 GW of panels annually; specialized glass comes from Zhejiang; and equipment maintenance relies on Chinese contractors.

Although the US imports almost no core components directly from China now, a full divorce remains elusive. T1 is building its own cell factory based on technology first developed for Chinese firms and awaits Commerce waivers to avoid penalties during its onshoring transition. Wood Mackenzie analyst Elissa Pierce noted that over 40% of the current 66 GW of operational US module capacity originated from facilities owned by Chinese manufacturers or joint ventures.

Trade barriers have not yet sparked a domestic manufacturing renaissance. Clean Investment Monitor data shows new solar facility investment remained flat at about $2.9 billion year-to-date, while $2.8 billion in other investments were cancelled. Pierce attributes this standstill to the combined impact of higher costs from tariffs and phasing-out tax credits, making financiers cautious even as policymakers push for supply chain independence.

Sources:semafor.com

Research

Salesforce Security Team: Replacing Email Inbox with Agentforce Conversational Triage

Topic · SalesforceQuick take
2026-11-11 06:22 GMT+8

Over the past 18 months, the internal security team at Salesforce received approximately 110,000 emails related to security concerns. Roughly half of these originated from employees, and 45% of those employee emails required manual triage. The Corporate Security Operations Center (CSOC) team spent significant time re-routing requests unrelated to security, chasing down missing log details, and deciphering ambiguous short descriptions. Genuine security signals were buried under sheer volume.

In November 2026, the Salesforce security team published an internal practice record detailing how they replaced this email inbox with a Slack-native assistant powered by Agentforce—the Security Reporting Agent. Authors Julissa Marcenaro, Matt Robinson, and Jingqi Wu summarized the core architectural decision as shifting from an invisible backend email parser to an active conversational frontline experience within Slack. This approach guides employees to write comprehensive reports in real time during the conversation, rather than having to fill in blanks after submission through follow-up requests.

This transition involved the coordination of three technical components: Slack served as the conversational entry point; Agentforce executed real-time guided triage and dynamically prompted for missing context; and Service Cloud acted as the underlying engine to generate pre-structured cases and route them directly to CSOC responders. The product team shifted the model from a backend parser to a conversational guide; the engineering team migrated the agent to Agentforce Script, consolidating multiple overlapping topics into a streamlined execution framework; and the operations team analyzed five years of historical data to determine which requests truly belonged to CSOC and which merely created friction—for instance, lost badge requests should be directed to Facilities, and phishing inquiries should be routed directly to Gmail’s native reporting tool.

Regarding quality assessment, Salesforce built an automated evaluation framework. The scorecard comprises six primary categories and twelve graded subcategories, with target success thresholds ranging between 85% and 98%. A continuous simulation testing framework runs simulated conversations constantly to detect subtle regressions in factual accuracy, technical completeness, tone, and empathy, intercepting issues before they reach production.

Results from the beta phase showed that across 378 qualifying conversations involving 261 employees, 72% were handled entirely on the spot by the agent—either answering directly or routing to the correct department without opening a ticket. The remaining 28% represented genuine security concerns converted into complete structured cases, requiring zero preliminary intake work from human analysts. Extrapolating these early patterns across Salesforce’s workforce of over 83,000 employees, the trajectory points toward fewer back-and-forth exchanges, complete context on first touch, and responder bandwidth concentrated on critical threats.

It must be noted that all the above data consists of internal beta results self-reported by Salesforce, with a limited sample size and no disclosed confidence intervals or absolute triage accuracy rates. The 72% instant handling rate may partially reflect the correct diversion of non-security requests rather than an improvement in the quality of security triage itself. The article also notes that the agent is intended solely for internal use by Salesforce employees; customers and partners should continue to report security concerns to .

Sources:Salesforce

Research

Google DeepMind and Artificial Analysis Launch Multimodal Knowledge Agent Benchmark to Evaluate AI on Complex Visual Workflows

Topic · MKA基准测试Material
2026-11-11 00:00 GMT+8

Google's DeepMind team, in partnership with international AI evaluation firm Artificial Analysis, launched the "Multimodal Knowledge Agent" (MKA) benchmark on November 10, 2026. This new standard aims to evaluate how well AI agents perform real-world knowledge work tasks that require reasoning across visually complex files.

Unlike previous benchmarks that relied primarily on text-only inputs, MKA targets the visual nature of everyday professional work, such as cross-referencing architectural blueprints, tracing engineering schematics, and synthesizing findings into polished decks or reports. To ground the benchmark in reality, Google collaborated with domain experts across 10 verticals and 83 occupations to create a 200-task evaluation set. Each task simulates a multi-step project that takes a seasoned professional three to eight hours to complete, paired with an expert-written rubric of roughly 70 yes/no checks, weighted 75% on factual accuracy and 25% on visual design.

To ensure integrity, Artificial Analysis runs the 200-task evaluation set privately, preventing task leakage into training data or discovery via search. A separate set of 10 public tasks has been released on Hugging Face for developers to inspect the format and test their own agents. All models run within the same open-source agent harness, Stirrup, using identical tools, sandboxes, and turn budgets.

Google claims that its latest model, Gemini 4 Carbon, ranks second in quality among tested models on this benchmark, while completing tasks at six times lower average cost compared to the top-scoring model. These results place Gemini 4 Carbon on the Pareto frontier of quality and price for MKA evaluations. However, these performance metrics are currently based on vendor self-testing and have not yet been independently reproduced by third parties.

Sources:Google

Research
Next →
Sign in / Sign up
ReadingResearchRadarInvestment framework

This also explains why the best counterpart for promoting AI adoption is mid-to-senior management rather than the boss. Su Peng-han used the analogy of switching from kerosene lamps to electric lights: 'The masters and mistresses feel nothing because it is the servants who are making the switch.' The servants find it very convenient, but the boss just sees that the lights turn on and doesn't understand why a switch is needed. What truly gives the boss a sense of crisis is population aging—some manufacturing firms originally had 8 to 10 finance and accounting staff, and those familiar with traditional ERP workflows will retire one by one within five to ten years.

Regarding information security strategy, Su Peng-han offers a concise criterion: 'If you are willing to upload this data to Google Sheets, then it should be fine.' At least then you can use cloud models; without overly vertical applications, eighty percent of needs can be handled. For enterprises still wary of the cloud, Aotter can also provide on-premises solutions, but will warn that on-premises models have limited capabilities. The recommended approach is to categorize data by sensitivity: truly sensitive data should be processed slowly on-premises; data that has been aggregated into quarterly reports with names and vendor contact information removed can use the cloud to gain broad insights.

It should be noted that all the above effectiveness descriptions—including Xie Lun's claim that 'a fully customized ERP can be online in 15 minutes'—are self-reported by the vendor and have not been independently verified. Aotter serves mostly older enterprises; these companies are not short of money and have stable niche markets. Once they switch to new systems and processes, they can extend their lifespan, but this condition may not apply to all industrial stages. The data middleware itself becomes a new dependency: if the standardization process misses key business semantics, the 'clean data' read by AI may actually be incorrect data.

Su Peng-han believes that current concerns about AI are similar to concerns about the cloud ten years ago. Back then, enterprises worried that putting hosts on the cloud meant Google would know everything; even earlier, people dared not even swipe cards online. He described this as a paradigm shift; in ten years, the next generation will have been chatting with AI since birth. This analogy itself is also an assumption—it presupposes that trust building for AI will follow the same trajectory as the cloud, but AI handles reasoning rather than storage, so the risk structure may not be the same.

Meta shared a key engineering discovery in the blog: in the reranking filtering stage, a small model fine-tuned to "only return items to be deleted" is both faster and more reliable than having a general-purpose model judge each candidate individually for relevance. This conclusion directly influenced model selection—the approach is not to use a larger general-purpose model for more granular individual reasoning, but rather to use a smaller specialized model for batch exclusion.

How to evaluate these LLM calls themselves is also a question. Meta notes that standard offline metrics cannot determine whether a "Lately" summary is accurate, nor can they judge whether the reranker removed the correct content. Their solution involves using another independent LLM as a judge to score the output of each call against expected behaviors, thereby providing a repeatable comparative baseline before launch.

Meta acknowledges unresolved issues at the end of the article: the ability to interpret ambiguous inputs, making recommendations respond more precisely to each specific request, and collecting user preferences from more entry points. The value of this blog lies in demonstrating a complete engineering pipeline from natural language to sorting signals, and highlighting the differentiated requirements for model scale, output format, and evaluation methods across different tasks within this pipeline.

security@salesforce.com