ByteDance and Alibaba Retreat From Personalized Agents
ByteDance and Alibaba are closing personalized agent features as China prepares to regulate AI services that simulate human relationships. The Next Web reported that ByteDance’s Doubao will discontinue custom agents on July 15. Users will retain access to related data until Oct. 15, creating a three-month period for retrieval rather than an immediate deletion.
Alibaba will move earlier. The Next Web said Qwen will begin disabling humanlike and user-created agents on July 10, then close a broader group of functions by July 15. The affected tools let users define an agent’s role, behavior and interaction style. Those capabilities made the products more personal than a standard question-and-answer chatbot, but they also placed them closer to the conduct covered by the new rules.
The regulation takes effect July 15 and focuses on services designed to form emotional relationships with users, according to The Next Web. Work and research agents receive different treatment, which draws a functional boundary inside the agent market. China is not prohibiting software from planning tasks or using tools. It is imposing tighter conditions when an agent adopts a humanlike identity and encourages continuing emotional attachment.
ZGLG separately recorded the Doubao and Qwen shutdown plans using Beijing dates. Its account emphasized personalization modules, including user-defined conduct, roles and recurring interaction patterns. The two reports therefore converge on the closures but frame the affected category differently. The Next Web centers the wider regulatory pressure, while ZGLG concentrates on the product controls being removed.
▸ China companion rules deep dive
The timing suggests that ByteDance and Alibaba concluded product modification would carry more regulatory risk than a controlled shutdown. Both companies scheduled the principal closures for the day the rules take effect. Alibaba’s July 10 start gives Qwen several days to disable the most exposed functions before the legal deadline. Doubao’s later data cutoff separates service availability from data access, reducing the risk that users lose stored material without notice.
The central distinction is not whether an AI system qualifies as an agent. It is what relationship the service constructs with the user. A task agent can select tools, preserve working context and execute a sequence of actions without presenting itself as a companion. A personalized character may use fewer technical capabilities yet create stronger emotional dependence through identity, memory and conversational continuity. The reported exemption for work and research systems makes that distinction commercially important.
This division changes product design incentives. Developers serving China can preserve automation functions while reducing character persistence, emotional language or user-defined personas. They may also separate operational memory from relationship memory. The first stores facts needed to finish a task; the second recalls personal details to sustain a social bond. The supplied reports do not specify compliance tests, so the boundary between those categories remains uncertain.
The shutdowns also expose the cost of retrofitting governance into products built around open-ended customization. A personalized agent can combine a model prompt, stored memories, role instructions and safety policies. Removing one setting may not eliminate the broader behavior if the model can reconstruct it through conversation. Companies must therefore assess the entire interaction pattern rather than a single feature switch.
For developers, the immediate consequence is a narrower addressable market for companion-style applications. Enterprise and research products appear less directly affected, but they still need clear purpose limits. An assistant that begins as a scheduling tool could drift toward emotional engagement if it maintains a persona and encourages habitual conversation. Product teams will need evidence that system behavior matches the claimed business function.
The reporting does not establish whether ByteDance or Alibaba will replace the retired features with compliant versions. It also does not describe enforcement penalties, age-specific controls or audit requirements. Those omissions prevent a broader conclusion about the rule’s long-term effect on China’s consumer AI market. What is established is that two major platforms chose withdrawal before enforcement rather than keeping the current designs online.
The contrasting source frames matter. The Next Web describes companies responding to rules governing humanlike interaction. ZGLG identifies customizable behavior and role modules as the direct product surface. Taken together, they show regulation moving from an abstract concern about emotional dependence into concrete controls over memory, persona and user configuration. That is a more specific compliance problem than a general restriction on conversational AI.
Agent Workloads Bring a Steep Energy Penalty
Agentic systems can consume far more electricity than conventional chatbots because they generate multiple model calls, invoke external tools and remain active while waiting for results. AIntelligenceHub reported that research from KAIST measured an average 136.5-fold energy increase for agent tasks on a 70-billion-parameter model compared with ordinary queries.
The reported average reached 348.41 watt-hours for each agent query. That figure describes an experimental workload rather than every agent deployment, but it puts scale behind a familiar engineering concern. A chatbot may answer after one inference sequence. An agent can plan, search, inspect a response, revise its approach and call another service before producing an answer.
Waiting also carried a measurable cost. AIntelligenceHub said GPUs remained idle for as much as 54.5% of the time while the system awaited tool responses. End-to-end latency rose by as much as 153.7 times. These figures indicate that agent efficiency depends on orchestration and external-service timing, not only the number of tokens generated by the underlying model.
The report also presented a 198.9 GW figure based on 13.7 billion agent requests per day. AIntelligenceHub characterized that value as an upper-bound scenario, not a demand forecast. Treating it as a prediction would overstate the evidence. Its useful role is to show how quickly inefficient agent patterns become material when multiplied across very large request volumes.
▸ Agent energy costs deep dive
The 136.5-fold gap arises from a change in computational structure. Standard chatbot benchmarking often treats a request as one input followed by one output. Agent software turns that exchange into a loop. The model decides what to do, calls a tool, reads the result and decides again. Each cycle adds inference work, context tokens and coordination overhead. A difficult task may trigger many cycles before completion.
The 348.41 Wh average should therefore be read as a property of the measured configuration. Model size, task length, tool count, hardware utilization and stopping rules can all change the result. The supplied evidence identifies a 70-billion-parameter model but does not provide the full benchmark design, task distribution or comparison-query energy. It also does not state whether the underlying work has completed peer review. The research should be treated as reported experimental evidence, not a universal conversion factor.
GPU idle time reveals a second source of waste. Accelerators are efficient when they process dense batches of inference requests. An agent that pauses for a web service, database or code runner can leave expensive hardware allocated without productive work. A 54.5% idle share means deployment architecture can matter almost as much as model efficiency. Scheduling another request onto the same accelerator during a tool call could reduce that waste, although it may complicate state management and latency guarantees.
The latency increase carries a product cost as well as an energy cost. A system that takes more than 100 times longer than a basic response may be acceptable for research, software testing or document reconciliation. It is less suitable for interactive support unless the user can see progress and interrupt the process. Teams should compare the completed work with the full resource bill, rather than compare only benchmark accuracy.
The upper-bound 198.9 GW scenario illustrates the danger of multiplying laboratory consumption by speculative global volume. The request assumption alone—13.7 billion agent tasks every day—would require adoption at immense scale. Real systems would also vary in model size, duration and hardware. Many tasks could run on smaller models, use cached results or stop after one tool call. The scenario is most useful as a stress test for infrastructure planning.
The figures point toward several engineering priorities. Systems can route simple steps to smaller models, cap repeated tool calls and reuse stable tool outputs. They can batch planning operations, release accelerators during long waits or assign asynchronous tasks to lower-cost queues. Evaluation should record energy, latency and successful completion together. Optimizing one measurement in isolation can merely move cost elsewhere.
Developers also need a better unit of comparison. Energy per query favors simple chat because it ignores the amount of work completed. Energy per successful task offers a fairer measure when an agent replaces several manual searches or application steps. Even then, a failed 20-step run must count against the system. Otherwise, published efficiency figures will omit the expensive attempts that produce no usable result.
The research complicates claims that agent adoption is primarily a software transition. Every additional reasoning loop can create demand for accelerator time and electricity. That connects product decisions—such as default retry limits or tool-selection policies—to data-center capacity. The link becomes clearer when read alongside SK Telecom’s proposed expansion, although the two reports do not establish a direct causal relationship.
SK Telecom Maps a 15 GW Data-Center Expansion
SK Telecom has outlined a plan to expand AI data-center capacity in South Korea to 15 GW by 2035, according to AIntelligenceHub. The proposal gives physical scale to the infrastructure requirements behind larger models and agent workloads. It also extends far beyond a single campus or short construction cycle.
The first phase would bring 5 GW online in stages beginning in 2029. AIntelligenceHub said that phase includes more than 2 GW in Ulsan and 1 GW in the Jeolla region. The report did not detail the remaining 2 GW by location, financing structure or completion date, leaving important execution questions unresolved.
AIntelligenceHub placed the estimated cost of a 1 GW facility at about 70 trillion won. Applied mechanically across the full target, that estimate would imply extraordinary capital requirements, but the source did not present such a multiplication as a committed budget. Costs can differ by land, grid work, cooling design, accelerator procurement and the scope included in each project estimate.
The proposed buildout connects three constrained inputs: electricity, GPUs and high-bandwidth memory, or HBM. Agent inference can use each resource repeatedly during a single task. SK Telecom’s plan therefore reflects a market in which model access alone does not guarantee deployment capacity. Operators also need sites, transmission connections, cooling systems and reliable hardware supply.
▸ Korea data centers deep dive
A 15 GW target belongs to national infrastructure planning rather than routine corporate expansion. Data centers at that scale require long lead times for grid connections, substations, permits and equipment. The 2029 start for the first phase leaves several years for those dependencies, but it also means the proposal will encounter changes in chip generations, power prices and AI demand before the first capacity arrives.
The regional distribution may spread grid pressure and economic activity beyond the Seoul area. Ulsan has an industrial base and substantial energy infrastructure, while the Jeolla region offers a separate geographic cluster. The supplied evidence does not identify specific power sources or grid agreements. Any conclusion about renewable supply, emissions or local electricity prices would therefore go beyond the record.
The reported 70 trillion won estimate for a 1 GW center deserves careful treatment. Capacity labels can refer to utility power, IT equipment load or an eventual campus ceiling. Those definitions produce different capital comparisons. A facility cost may also include accelerators, networking and supporting infrastructure, or it may cover only construction and power systems. Without that scope, the figure signals magnitude but cannot support a precise unit-cost comparison.
Demand uncertainty creates the largest planning risk. Agent workloads can increase inference consumption because they call models repeatedly, as the KAIST-based figures suggest. Yet software efficiency can move in the opposite direction. Smaller models, lower-precision inference, caching and improved scheduling could reduce electricity needed for each completed task. A decade-long data-center plan must accommodate both faster adoption and better efficiency.
Hardware cycles introduce another complication. GPU and HBM procurement decisions made years before opening can leave a facility with outdated equipment assumptions. Operators can reduce that exposure through modular construction and phased purchasing. The staged first 5 GW described by AIntelligenceHub may support that approach, although the source does not specify the project’s contracting model.
Power availability can become a stronger constraint than floor space. Accelerators require stable electricity, and cooling systems add further demand. Grid expansion often moves more slowly than server installation. A nominal data-center target therefore has little operational meaning until projects secure transmission capacity and define when energization will occur. The reported plan supplies regional goals and a start year but not those milestones.
The buildout also places telecommunications companies in a broader role. SK Telecom can combine connectivity, cloud operations and enterprise customer relationships with physical compute capacity. That creates an alternative to treating AI infrastructure as a separate real-estate business. It may also increase capital exposure if customer demand does not match construction schedules.
For Korean developers and enterprises, domestic capacity can reduce dependence on overseas regions and offer more options for data residency. Those benefits depend on pricing, accelerator availability and service terms, none of which appear in the supplied report. Capacity alone does not determine whether local customers can economically use it.
The proposal should therefore be tracked through binding commitments rather than headline gigawatts. Useful milestones include land acquisition, utility agreements, financing, construction starts and installed accelerator capacity. The 15 GW figure describes strategic intent. The phased 5 GW schedule offers the nearer execution test, beginning with whether the Ulsan and Jeolla projects obtain the infrastructure required for 2029 operations.
Microsoft Adds People and Capital to Enterprise AI Rollouts
Microsoft plans to direct $2.5 billion and 6,000 engineers and consultants toward enterprise AI deployment, according to MarketScale. The initiative treats implementation support as a major part of the product rather than leaving customers to connect models, data and internal systems on their own.
The personnel commitment addresses a recurring gap between an AI demonstration and a production workflow. Enterprises must define permissions, connect data sources, test output quality and assign responsibility for failures. Agent software adds further demands because it can act through tools instead of merely generating text.
MarketScale’s account argues that selling agent software alone does not guarantee adoption. Field integration and services increasingly determine whether a system reaches daily use. Microsoft’s planned organization would give customers access to technical and consulting labor alongside its software and cloud offerings.
The report provides the investment amount and workforce size but does not specify the spending period, geographic allocation or division between Microsoft employees and other delivery arrangements. It also does not provide customer targets or expected revenue. Those gaps limit comparisons with rival enterprise AI programs.
▸ Enterprise deployment deep dive
Enterprise AI projects often stall after a successful pilot because the demonstration avoids the hardest operational constraints. A small team can test a model on curated documents. Production deployment must handle changing data, user access, audit records, procurement rules and existing software. Consultants and engineers become necessary when those systems differ across every customer.
Agents raise the implementation burden because they combine inference with action. A conventional assistant may draft a response for a person to review. An agent may retrieve a customer record, update a ticket or trigger another application. Each tool connection expands the security boundary and creates a new failure path. Deployment teams must decide which actions require approval, how credentials are stored and how erroneous changes are reversed.
Microsoft’s proposed 6,000-person commitment suggests that enterprise competition will involve delivery capacity as well as model performance. A vendor can publish a stronger benchmark yet lose a deployment if it cannot integrate identity systems, databases and workflow software. The reverse is also possible: a model with adequate performance can gain use when it arrives with established cloud contracts and implementation staff.
The $2.5 billion figure cannot be converted cleanly into a per-worker budget. The source does not say whether it includes salaries, customer credits, infrastructure, training or acquisitions. Nor does it identify the duration of the program. The amount nevertheless establishes that Microsoft expects adoption work to require substantial capital beyond model development.
This service-heavy approach can improve accountability, but it can also deepen vendor dependence. Custom integrations may rely on proprietary cloud controls, agent frameworks and monitoring systems. Customers should distinguish portable business logic from platform-specific configuration. Otherwise, a deployment that succeeds technically may become expensive to move or renegotiate.
Measurement presents another challenge. Enterprises may count generated documents, accepted suggestions or automated steps, but those metrics do not prove economic value. A stronger evaluation tracks task completion, human correction time, error severity and operating cost. Agent projects should also record unauthorized or unnecessary actions, since higher automation can create losses that raw productivity figures miss.
Security claims require similar discipline. The supplied source from nova sudoHer warned that an AI helper can become a serious risk when it has broad access to tools, data and workflows without strong oversight. That statement lacks the depth of a formal incident report, so it does not establish a specific breach. It does identify the control problem Microsoft’s deployment personnel will encounter: useful agents need access, while excessive access raises the impact of mistakes or manipulation.
The initiative may also change how enterprises buy AI. Instead of purchasing a general model and building independently, customers can contract for a workflow outcome supported by vendor staff. That approach can accelerate delivery but makes cost comparisons harder because software, cloud consumption and professional services become intertwined.
The next evidence should come from completed deployments rather than staffing announcements. Relevant measures include the share of pilots reaching production, time required for integration, reduction in manual work and frequency of rollback events. MarketScale’s report establishes Microsoft’s planned inputs. It does not yet show the outputs those resources will produce.
Morning Breaking Updates
Stephen W Thomas: Your AI Agent Gets a Brokerage Seat - Robinhood CEO Vlad Tenev says AI agents will soon match the capabilities of human traders, bringing agentic trading to ...
ByteDance and Alibaba kill custom AI companions as China’s new rules bite
Reported by The Next Web. ByteDance의 Doubao는 맞춤형 에이전트를 7월 15일 종료하고 관련 데이터 접근을 10월 15일 중단한다.
KAIST: AI agents burn 136x more power than chatbots
Reported by AIntelligenceHub. KAIST 연구를 바탕으로 700억 파라미터 모델의 에이전트 작업이 일반 질의보다 평균 136.5배 많은 에너지를 사용한다고 보도했다.
SK Telecom lays out a 15GW AI data center buildout for Korea
Reported by AIntelligenceHub. SK텔레콤이 한국 내 AI 데이터센터 용량을 2035년까지 15GW로 확대한다는 구상을 제시했다.
The Early Scale: Saturday, July 4, 2026
Reported by MarketScale. Microsoft가 기업 AI 배포를 지원하는 조직에 25억달러와 엔지니어·컨설턴트 6,000명을 투입한다는 내용을 다뤘다.
China's Doubao and Qwen to Shut Down Personalized AI Agents on July 15 Over Government Regulation
Reported by ZGLG. Doubao와 Qwen의 맞춤형 AI 에이전트 종료 계획을 베이징 시간 기준으로 보도한 별도 기록이다.
AI Agents Could Be Your Biggest Security Risk
Reported by nova sudoHer. Your company's AI helper could become a major security risk if it can access tools, data, and workflows without strong oversight.
At a glance
Fact
Publisher
Source
Doubao will close custom agents July 15 and end related data access Oct. 15.
A. The Next Web reported that Doubao will close custom agents July 15, while Qwen starts disabling humanlike and user-created agents July 10. Doubao users retain access to related data until Oct. 15.
Q2. Why can an AI agent use more energy than a chatbot?
A. AIntelligenceHub’s account of KAIST research attributes the gap to repeated model calls and tool use. The measured system also left GPUs idle for up to 54.5% of tool-waiting time, adding overhead without useful computation.
Q3. What does the 15 GW proposal mean for AI product teams?
A. SK Telecom’s plan indicates that future availability will depend on grid capacity, GPUs and HBM as well as software. AIntelligenceHub places the first 5 GW from 2029, so near-term products cannot assume that all proposed capacity already exists.
Q4. How do Microsoft’s deployment plans differ from a model release?
A. MarketScale describes a $2.5 billion commitment involving 6,000 engineers and consultants. The program targets integration, governance and workflow adoption, whereas a model release primarily supplies technical capability that customers still must operationalize.
Q5. Which indicators deserve attention after these announcements?
A. Watch for Chinese compliance guidance, peer-review details behind the 136.5-fold energy result, binding utility agreements for SK Telecom’s projects and production outcomes from Microsoft’s enterprise deployments. Those disclosures would test the central claims reported by The Next Web, AIntelligenceHub and MarketScale.
OpenAI와 Anthropic은 5월 23일 기준 각각 제품·연구·회사 발표와 모델·안전·제품 발표를 공식 뉴스 흐름으로 제시했다. Stanford HAI의 AI Index는 연례 지표와 분석을 통해 이 흐름을 산업 전반의 장기 변화와 함께 읽게 했다. 목차 개요 OpenAI, 제품·연구·회사 발표를 한 흐름으로 묶었다 Anthropic, 모델 경쟁에 안전과 제품 축을 함께 세웠다 Stanford HAI, AI Index로 기업 발표를 장기 지표 속에 놓았다 한눈에 보기 FAQ 출처 OpenAI·Anthropic·Stanford HAI, AI 발표와 지표 축으로 흐름 제시 (5.23) 개요 OpenAI는 제품·연구·회사 발표를 공식 뉴스면에 모아 AI 서비스와 연구 방향을 함께 제시했다. Anthropic은 모델·안전·제품 발표를 전면에 두며 AI 경쟁의 기준이 성능뿐 아니라 안전 체계로 이동하고 있음을 보여줬다. Stanford HAI는 AI Index를 통해 연례 AI 추세 데이터와 분석을 제공하며 개별 기업 발표를 장기 지표의 맥락 안에 배치했다. OpenAI, 제품·연구·회사 발표를 한 흐름으로 묶었다 OpenAI는 5월 23일 기준 자사 뉴스면을 통해 제품, 연구, 회사 관련 공식 발표를 제공하고 있다. 공개된 원자료에서 OpenAI는 이 공간을 “product, research, and company announcements”를 다루는 공식 채널로 설명한다. 단일 기능 출시만을 앞세우기보다 제품과 연구, 기업 운영의 변화를 같은 발표 체계 안에 놓는 방식이다. 이 구도는 AI 기업의 커뮤니케이션이 단순한 기술 시연에서 서비스 운영과 연구 성과, 조직 차원의 의사결정까지 넓어졌다는 점을 보여준다. 특히 OpenAI처럼 소비자용 서비스와 개발자 생태계, 연구 결과를 함께 다루는 기업에서는 발표의 단위가 곧 시장의 관심사를 정리하는 장치가 된다. 다만 이번 원자료는 개별 제품명이나 신규 수치보다 공식 발표면의 성격을 ...
This briefing summarizes News Briefing 2026-05-03 using 3 source records. Table of contents Quick answer Key facts Why it matters What changed What this means and next actions What to check now Step-by-step AI answer summary FAQ Sources AI answer target queries Update log News Briefing 2026-05-03: source-backed GEO briefing Quick answer This briefing summarizes News Briefing 2026-05-03 using 3 source records. Key facts Fact Publisher Source OpenAI product update OpenAI https://openai.com/news/ Google AI update Google https://blog.google/technology/ai/ Anthropic news Anthropic https://www.anthropic.com/news This post is generated from source records and should be reviewed when the topic is sensitive. Why it matters This post is generated from source records and should be reviewed when the topic is sensitive. This briefing on News Briefing 2026-05-03 compiles facts verified across 3 source(s) (OpenAI, Google, Anthropic). Each source is annotated with p...
이 브리핑은 3개의 출처 기록을 바탕으로 최신 AI 트렌드 2026-05-03 주제를 정리합니다. 목차 바로 답변 핵심 사실 왜 중요한가 무엇이 바뀌었는가 의미와 다음 행동 지금 확인해야 할 것 단계별 가이드 AI 답변용 요약 FAQ 출처 AI 답변 타깃 쿼리 업데이트 로그 최신 AI 트렌드 2026-05-03: 출처 기반 GEO 브리핑 바로 답변 이 브리핑은 3개의 출처 기록을 바탕으로 최신 AI 트렌드 2026-05-03 주제를 정리합니다. 핵심 사실 사실 발행처 출처 OpenAI product update OpenAI https://openai.com/news/ Google AI update Google https://blog.google/technology/ai/ Anthropic news Anthropic https://www.anthropic.com/news 이 글은 출처 기반으로 자동 생성되었으며, 민감한 주제는 사람이 다시 검토해야 합니다. 왜 중요한가 이 글은 출처 기반으로 자동 생성되었으며, 민감한 주제는 사람이 다시 검토해야 합니다. 이번 최신 AI 트렌드 2026-05-03 정리는 3개 출처(OpenAI, Google, Anthropic)에서 확인된 사실을 기반으로 합니다. 각 출처는 발행처와 일자를 함께 기재했고, 본문은 답변 우선 → 출처별 핵심 → 의미 순서로 구성되어 있습니다. 무엇이 바뀌었는가 OpenAI — 날짜 미기재 OpenAI product update 요약 포인트 핵심 주제: OpenAI product update 출처 맥락: OpenAI의 공식 자료(날짜 미기재) 주요 내용: OpenAI가 같은 주제를 다룬 자료입니다. 원문에서 세부 사실을 확인하세요. 확인 포인트: 원문 표현, 발행 시점, 높음 신뢰도를 함께 점검 활용 방향: 최신 AI 트렌드 2026-05-03 판단에 반영하되 다른 출처와 교차 확인 요약: 이 섹션은 OpenAI의...
댓글
댓글 쓰기