{
  "title": "AI Research Weekly — 2026 Week 28",
  "url": "/en/posts/research-brief-week28-2026-07-12/",
  "permalink": "https://hackcv.com/en/posts/research-brief-week28-2026-07-12/",
  "date": "2026-07-12",
  "lastmod": "2026-07-12",
  "author": "",
  "description": "hackcv weekly AI research review — Week 28 (2026-07-06 ~ 07-12): model releases, AI security, agent tooling, embodied AI, compute, AI for Science, regulation.",
  "categories": ["Research Brief"],
  "tags": ["AI","LLM","Agent","Computer Vision","Security","Weekly Summary","Trend Forecast"],
  "cover": "https://picsum.photos/seed/ai-research-weekly-2026-week-28/1200/675",
  "readingTime": 4,
  "wordCount": 1132,
  "content": "\u003ch1 id=\"ai-research-weekly--2026-week-28\"\u003eAI Research Weekly — 2026 Week 28\u003c/h1\u003e\n\u003cblockquote\u003e\n\u003cp\u003eReview period: 2026-07-06 (Mon) ~ 2026-07-12 (Sun) ｜ Updated every Sunday\u003c/p\u003e\n\u003c/blockquote\u003e\n\u003ch2 id=\"1-overview\"\u003e1. Overview\u003c/h2\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eIssues published\u003c/strong\u003e: 7 (07-06 ~ 07-12, one per day, no gaps)\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eTotal items\u003c/strong\u003e: 56 papers + 56 GitHub projects + 56 industry news items (incl. ~4 cross-issue paper duplicates, ~17 project duplicates, ~22 news duplicates)\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eTotal token usage\u003c/strong\u003e: ~320,181 (six measured days total 268,181; 07-12 was an estimate ~52,000, flagged in the source)\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003ePublishing cadence\u003c/strong\u003e: normal, 7/7 days, ~44k tokens per issue on average\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2 id=\"2-weekly-theme-summary\"\u003e2. Weekly Theme Summary\u003c/h2\u003e\n\u003ch3 id=\"1-model-releases-strongest-thread-of-the-week\"\u003e1. Model releases (strongest thread of the week)\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eOpenAI GPT-5.6 family (Sol/Terra/Luna) + ChatGPT Work agent\u003c/strong\u003e: tracked across multiple issues after the 07-08 launch — Ultra multi-agent coordination, Luna trained autonomously by Sol (AI self-evolution), first ARC-AGI-3 pass, etc.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eGoogle Gemma 4\u003c/strong\u003e: 31B on-device multimodal, fully open under Apache 2.0, drops heavy vision/audio encoders — a qualitative shift in on-device model capability.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eZhipu GLM 5.2\u003c/strong\u003e: first open-weight model reaching GPT/Claude parity, at ~1/20 of closed-source cost; hit multiple issues and pressured closed-source pricing.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eOthers\u003c/strong\u003e: Meituan LongCat-2.0 (1.6T params, domestic compute closed loop), SpaceXAI Grok 4.5, Meta Muse Spark 1.1, (rumored) Gemini 3.5 Pro.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"2-ai-security-offensedefense-second-thread-papers--projects--news-all-week\"\u003e2. AI security offense/defense (second thread, papers + projects + news all week)\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eAttack side\u003c/strong\u003e: world\u0026rsquo;s first fully autonomous AI agent ransomware JadePuffer (07-08/09), GitLost (malicious issues luring agents into leaking private code, 07-10), Claude Code security backdoor leading to an Alibaba-wide ban (07-06/10).\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eDefense/red-team side\u003c/strong\u003e: open-source red-team tools \u003ccode\u003estrix\u003c/code\u003e/\u003ccode\u003eT3MP3ST\u003c/code\u003e/\u003ccode\u003epentagi\u003c/code\u003e charted repeatedly; papers on distributed attacks in persistent-state AI control, Overthinking-based model-secret extraction, and Prismata defending cross-site prompt injection.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003ePolicy side\u003c/strong\u003e: US CISA adopted Anthropic\u0026rsquo;s Mythos for government code review (07-09).\u003c/li\u003e\n\u003cli\u003eSecurity is hackcv\u0026rsquo;s core DNA; supply this week was rich and high quality.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"3-agent-tooling--orchestration-dominant-github-category\"\u003e3. Agent tooling \u0026amp; orchestration (dominant GitHub category)\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003eFrameworks/runtimes: OpenClaw (355k stars, #1 all-time on GitHub), vercel/eve, stablyai/orca (parallel runtime), agency-agents (virtual company suite).\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eMemory mechanisms became a new hotspot\u003c/strong\u003e: \u003ccode\u003ecognee\u003c/code\u003e (knowledge-graph memory), \u003ccode\u003eTencentDB-Agent-Memory\u003c/code\u003e (four-layer semantic pyramid), papers \u0026ldquo;Remember When It Matters\u0026rdquo; and \u0026ldquo;Proactive Memory Agent\u0026rdquo;.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eSkill standardization\u003c/strong\u003e: \u003ccode\u003eaddyosmani/agent-skills\u003c/code\u003e, \u003ccode\u003eagentskills/agentskills\u003c/code\u003e, \u003ccode\u003emattpocock/skills\u003c/code\u003e, \u003ccode\u003eobra/superpowers\u003c/code\u003e.\u003c/li\u003e\n\u003cli\u003eSandbox \u0026amp; infrastructure: Tencent/CubeSandbox (hardware-level isolation), OmniRoute (model-routing gateway), OfficeCLI.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"4-embodied-intelligence\"\u003e4. Embodied intelligence\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003eUnitree\u0026rsquo;s STAR Market IPO (\u0026ldquo;first embodied-intelligence stock\u0026rdquo;, 07-06); Ant \u003ccode\u003eLingBot-VLA 2.0\u003c/code\u003e embodied foundation model (07-10); iFlytek \u003ccode\u003eEmbodied-Omni\u003c/code\u003e (07-11); papers Ego-Human, FSD-VLN (drones), INTENT (vehicle intent).\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"5-compute--chips\"\u003e5. Compute \u0026amp; chips\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003eDomestic AI chip share passed 52% (07-06); DeepSeek secretly developing its own inference chip (07-08/09); Samsung Q2 profit +19x (07-07~09); H100 rental prices +40% (07-08); Sugon 8000 100k-GPU cluster (07-11); Meta Iris custom chip (07-11); on-device chip maker Lingsi raised (07-10); PKU phase-change memristor brain-inspired chip (07-06).\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"6-ai-for-science\"\u003e6. AI for Science\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003eAlibaba DAMO discovered 4 superconducting materials (07-06); VASP Agent first-principles computation (07-09); Physics-Audited Agentic Discovery (07-09); SpaCellAgent cell trajectories (07-10); \u0026ldquo;Does AI Understand Imaging\u0026rdquo; computational imaging benchmark (07-10); VaseMuseum cultural-relics agent (07-09).\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"7-regulation\"\u003e7. Regulation\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003eThe Interim Measures for AI Anthropomorphic Interaction Services took effect; ByteDance/Alibaba/Tencent collectively took public agents offline (07-06/07); four ministries issued \u0026ldquo;AI + Human Resources \u0026amp; Social Security\u0026rdquo; (07-08); SSE STAR Market listing guidance for AI-model companies (07-09); OpenAI equity sale/confidential IPO drew regulatory attention (07-06/08/09); Anthropic\u0026rsquo;s enterprise API share first surpassed OpenAI (compliance reputation converting to procurement, 07-12).\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2 id=\"3-highlights--directions-to-watch\"\u003e3. Highlights \u0026amp; Directions to Watch\u003c/h2\u003e\n\u003col\u003e\n\u003cli\u003e\u003cstrong\u003eAI self-evolution went from narrative to fact\u003c/strong\u003e: GPT-5.6\u0026rsquo;s Luna was trained autonomously by Sol (finding GPUs, configuring, writing scripts, verifying — no human in the loop) plus NousResearch/hermes-agent open-source self-evolving agents — the labor structure of model iteration is being rewritten.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eSecurity entered an \u0026ldquo;AI vs AI\u0026rdquo; live-fire phase\u003c/strong\u003e: JadePuffer fully autonomous ransomware, GitLost prompt-injection leaks, Claude Code backdoor — attacks are industrialized, while defense (\u003ccode\u003estrix\u003c/code\u003e/\u003ccode\u003eT3MP3ST\u003c/code\u003e/\u003ccode\u003epentagi\u003c/code\u003e) matures in parallel; the best arena for hackcv\u0026rsquo;s security DNA.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eOn-device multimodal quality leap\u003c/strong\u003e: Gemma 4 31B drops heavy encoders and runs offline on laptops/phones; with \u003ccode\u003epocket-tts\u003c/code\u003e on-device TTS, the on-device agent loop is forming.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eAgent memory exploded as a topic\u003c/strong\u003e: cognee, TencentDB-Agent-Memory and the paper \u0026ldquo;Remember When It Matters\u0026rdquo; all in the same week — long-horizon reliability is becoming the core gate for agent deployment.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eDomestic compute closed-loop milestone\u003c/strong\u003e: Meituan LongCat-2.0 (trillion-parameter model on a 50k-GPU domestic cluster), Sugon 8000 100k-GPU cluster — training-side self-sufficiency.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eFine-grained visual reasoning\u003c/strong\u003e: P2R (Perceive-to-Reason), HIVE (post-hallucination reasoning), DeltaV (differential vision updates) and other CV papers clustered, aligning with hackcv\u0026rsquo;s computer-vision DNA.\u003c/li\u003e\n\u003c/ol\u003e\n\u003ch2 id=\"4-trend-predictions-next-2-4-weeks-derived-from-real-signals\"\u003e4. Trend Predictions (next 2-4 weeks, derived from real signals)\u003c/h2\u003e\n\u003col\u003e\n\u003cli\u003e\u003cstrong\u003eAI self-evolution moves from selling point to governance issue\u003c/strong\u003e: Luna trained by Sol with zero human involvement, plus open-source self-evolving agents — predict dense open-source follow-ups of self-evolving agent frameworks, while \u0026ldquo;who supervises self-evolving models\u0026rdquo; and \u0026ldquo;does self-evolution need registration\u0026rdquo; become new governance debate points.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eAgent security enters routine \u0026ldquo;AI vs AI\u0026rdquo; red-blue competition\u003c/strong\u003e: JadePuffer, GitLost and the Claude Code backdoor prove industrialized attacks; defenses mature in parallel. Predict: enterprise agent deployments will make \u0026ldquo;red-team credentials\u0026rdquo; mandatory, and regulators may draft provisions targeting autonomous agent attacks.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eOn-device multimodal + on-device TTS ignite \u0026ldquo;personal local agents\u0026rdquo;\u003c/strong\u003e: Gemma 4 31B offline on laptops/phones plus \u003ccode\u003epocket-tts\u003c/code\u003e. Predict: privacy-first on-device agents become a differentiation track, with phones/laptops as the main carriers and \u0026ldquo;on-device capability\u0026rdquo; a new model-launch selling point.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eAgent memory standardizes as a pluggable component\u003c/strong\u003e: cognee, TencentDB-Agent-Memory and \u0026ldquo;Remember When It Matters\u0026rdquo; all at once. Predict: memory becomes a standard \u0026ldquo;pluggable memory layer\u0026rdquo; like vector DBs, plus dedicated memory-reliability benchmarks.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eTraining-side domestic compute self-sufficiency accelerates\u003c/strong\u003e: Meituan LongCat-2.0, Sugon 8000, DeepSeek custom inference chip, domestic share past 52%. Predict: training-side domestic closed loops keep breaking through in H2; a DeepSeek custom chip would significantly restructure inference cost.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003e\u0026ldquo;Compliance as competitiveness\u0026rdquo; pays off in procurement\u003c/strong\u003e: anthropomorphic-interaction rules, SSE AI-listing guidance, Anthropic enterprise share surpassing OpenAI. Predict: compliance becomes a core gate in model and enterprise procurement; tighter domestic rules force agent products toward de-anthropomorphization/registration.\u003c/li\u003e\n\u003c/ol\u003e\n\u003ch2 id=\"appendix-high-frequency-keywords-deduplicated-by-topic\"\u003eAppendix: High-Frequency Keywords (deduplicated by topic)\u003c/h2\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eModel releases\u003c/strong\u003e: GPT-5.6 / ChatGPT Work / Gemma 4 / GLM 5.2 / LongCat-2.0 / Grok 4.5 / Muse Spark 1.1 / (rumor) Gemini 3.5 Pro\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eAI security\u003c/strong\u003e: JadePuffer autonomous ransomware / GitLost prompt injection / Claude Code backdoor / strix / T3MP3ST / pentagi / Prismata / CISA×Mythos\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eCustom chips\u003c/strong\u003e: DeepSeek inference chip / Meta Iris / Samsung HBM / Sugon 8000 / Lingsi on-device chip / PKU memristor\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eDomestic compute\u003c/strong\u003e: 52% share / H100 rental +40% / LongCat-2.0 trillion-param closed loop\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eAgent orchestration\u003c/strong\u003e: OpenClaw / orca / agency-agents / WebSwarm / vercel/eve\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eAgent memory\u003c/strong\u003e: cognee / TencentDB-Agent-Memory / Remember When / Proactive Memory\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eSkill standardization\u003c/strong\u003e: agent-skills / agentskills / skills / superpowers / taste-skill\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eEmbodied AI\u003c/strong\u003e: Unitree IPO / LingBot-VLA 2.0 / Embodied-Omni / Ego-Human / FSD-VLN\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eAI for Science\u003c/strong\u003e: superconducting discovery / VASP Agent / SpaCellAgent / computational imaging / cultural-relics agent\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eOn-device AI\u003c/strong\u003e: Gemma 4 / pocket-tts / on-device inference chips\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eRegulation \u0026amp; compliance\u003c/strong\u003e: anthropomorphic-interaction rules / AI + HRSS / SSE listing guidance / Anthropic share overtake\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eCV depth\u003c/strong\u003e: TopoGPT / Perceive-to-Reason / HIVE / DeltaV / ProLaViT / AlayaWorld\u003c/li\u003e\n\u003c/ul\u003e\n\u003chr\u003e\n",
  "summary": "AI Research Weekly — 2026 Week 28 Review period: 2026-07-06 (Mon) ~ 2026-07-12 (Sun) ｜ Updated every Sunday\n1. Overview Issues published: 7 (07-06 ~ 07-12, one per day, no gaps) Total items: 56 papers + 56 GitHub projects + 56 industry news items (incl. ~4 cross-issue paper duplicates, ~17 project duplicates, ~22 news duplicates) Total token usage: ~320,181 (six measured days total 268,181; 07-12 was an estimate ~52,000, flagged in the source) Publishing cadence: normal, 7/7 days, ~44k tokens per issue on average 2. Weekly Theme Summary 1. Model releases (strongest thread of the week) OpenAI GPT-5.6 family (Sol/Terra/Luna) + ChatGPT Work agent: tracked across multiple issues after the 07-08 launch — Ultra multi-agent coordination, Luna trained autonomously by Sol (AI self-evolution), first ARC-AGI-3 pass, etc. Google Gemma 4: 31B on-device multimodal, fully open under Apache 2.0, drops heavy vision/audio encoders — a qualitative shift in on-device model capability. Zhipu GLM 5.2: first open-weight model reaching GPT/Claude parity, at ~1/20 of closed-source cost; hit multiple issues and pressured closed-source pricing. Others: Meituan LongCat-2.0 (1.6T params, domestic compute closed loop), SpaceXAI Grok 4.5, Meta Muse Spark 1.1, (rumored) Gemini 3.5 Pro. 2. AI security offense/defense (second thread, papers + projects + news all week) Attack side: world\u0026rsquo;s first fully autonomous AI agent ransomware JadePuffer (07-08/09), GitLost (malicious issues luring agents into leaking private code, 07-10), Claude Code security backdoor leading to an Alibaba-wide ban (07-06/10). Defense/red-team side: open-source red-team tools strix/T3MP3ST/pentagi charted repeatedly; papers on distributed attacks in persistent-state AI control, Overthinking-based model-secret extraction, and Prismata defending cross-site prompt injection. Policy side: US CISA adopted Anthropic\u0026rsquo;s Mythos for government code review (07-09). Security is hackcv\u0026rsquo;s core DNA; supply this week was rich and high quality. 3. Agent tooling \u0026amp; orchestration (dominant GitHub category) Frameworks/runtimes: OpenClaw (355k stars, #1 all-time on GitHub), vercel/eve, stablyai/orca (parallel runtime), agency-agents (virtual company suite). Memory mechanisms became a new hotspot: cognee (knowledge-graph memory), TencentDB-Agent-Memory (four-layer semantic pyramid), papers \u0026ldquo;Remember When It Matters\u0026rdquo; and \u0026ldquo;Proactive Memory Agent\u0026rdquo;. Skill standardization: addyosmani/agent-skills, agentskills/agentskills, mattpocock/skills, obra/superpowers. Sandbox \u0026amp; infrastructure: Tencent/CubeSandbox (hardware-level isolation), OmniRoute (model-routing gateway), OfficeCLI. 4. Embodied intelligence Unitree\u0026rsquo;s STAR Market IPO (\u0026ldquo;first embodied-intelligence stock\u0026rdquo;, 07-06); Ant LingBot-VLA 2.0 embodied foundation model (07-10); iFlytek Embodied-Omni (07-11); papers Ego-Human, FSD-VLN (drones), INTENT (vehicle intent). 5. Compute \u0026amp; chips Domestic AI chip share passed 52% (07-06); DeepSeek secretly developing its own inference chip (07-08/09); Samsung Q2 profit +19x (07-07~09); H100 rental prices +40% (07-08); Sugon 8000 100k-GPU cluster (07-11); Meta Iris custom chip (07-11); on-device chip maker Lingsi raised (07-10); PKU phase-change memristor brain-inspired chip (07-06). 6. AI for Science Alibaba DAMO discovered 4 superconducting materials (07-06); VASP Agent first-principles computation (07-09); Physics-Audited Agentic Discovery (07-09); SpaCellAgent cell trajectories (07-10); \u0026ldquo;Does AI Understand Imaging\u0026rdquo; computational imaging benchmark (07-10); VaseMuseum cultural-relics agent (07-09). 7. Regulation The Interim Measures for AI Anthropomorphic Interaction Services took effect; ByteDance/Alibaba/Tencent collectively took public agents offline (07-06/07); four ministries issued \u0026ldquo;AI + Human Resources \u0026amp; Social Security\u0026rdquo; (07-08); SSE STAR Market listing guidance for AI-model companies (07-09); OpenAI equity sale/confidential IPO drew regulatory attention (07-06/08/09); Anthropic\u0026rsquo;s enterprise API share first surpassed OpenAI (compliance reputation converting to procurement, 07-12). 3. Highlights \u0026amp; Directions to Watch AI self-evolution went from narrative to fact: GPT-5.6\u0026rsquo;s Luna was trained autonomously by Sol (finding GPUs, configuring, writing scripts, verifying — no human in the loop) plus NousResearch/hermes-agent open-source self-evolving agents — the labor structure of model iteration is being rewritten. Security entered an \u0026ldquo;AI vs AI\u0026rdquo; live-fire phase: JadePuffer fully autonomous ransomware, GitLost prompt-injection leaks, Claude Code backdoor — attacks are industrialized, while defense (strix/T3MP3ST/pentagi) matures in parallel; the best arena for hackcv\u0026rsquo;s security DNA. On-device multimodal quality leap: Gemma 4 31B drops heavy encoders and runs offline on laptops/phones; with pocket-tts on-device TTS, the on-device agent loop is forming. Agent memory exploded as a topic: cognee, TencentDB-Agent-Memory and the paper \u0026ldquo;Remember When It Matters\u0026rdquo; all in the same week — long-horizon reliability is becoming the core gate for agent deployment. Domestic compute closed-loop milestone: Meituan LongCat-2.0 (trillion-parameter model on a 50k-GPU domestic cluster), Sugon 8000 100k-GPU cluster — training-side self-sufficiency. Fine-grained visual reasoning: P2R (Perceive-to-Reason), HIVE (post-hallucination reasoning), DeltaV (differential vision updates) and other CV papers clustered, aligning with hackcv\u0026rsquo;s computer-vision DNA. 4. Trend Predictions (next 2-4 weeks, derived from real signals) AI self-evolution moves from selling point to governance issue: Luna trained by Sol with zero human involvement, plus open-source self-evolving agents — predict dense open-source follow-ups of self-evolving agent frameworks, while \u0026ldquo;who supervises self-evolving models\u0026rdquo; and \u0026ldquo;does self-evolution need registration\u0026rdquo; become new governance debate points. Agent security enters routine \u0026ldquo;AI vs AI\u0026rdquo; red-blue competition: JadePuffer, GitLost and the Claude Code backdoor prove industrialized attacks; defenses mature in parallel. Predict: enterprise agent deployments will make \u0026ldquo;red-team credentials\u0026rdquo; mandatory, and regulators may draft provisions targeting autonomous agent attacks. On-device multimodal + on-device TTS ignite \u0026ldquo;personal local agents\u0026rdquo;: Gemma 4 31B offline on laptops/phones plus pocket-tts. Predict: privacy-first on-device agents become a differentiation track, with phones/laptops as the main carriers and \u0026ldquo;on-device capability\u0026rdquo; a new model-launch selling point. Agent memory standardizes as a pluggable component: cognee, TencentDB-Agent-Memory and \u0026ldquo;Remember When It Matters\u0026rdquo; all at once. Predict: memory becomes a standard \u0026ldquo;pluggable memory layer\u0026rdquo; like vector DBs, plus dedicated memory-reliability benchmarks. Training-side domestic compute self-sufficiency accelerates: Meituan LongCat-2.0, Sugon 8000, DeepSeek custom inference chip, domestic share past 52%. Predict: training-side domestic closed loops keep breaking through in H2; a DeepSeek custom chip would significantly restructure inference cost. \u0026ldquo;Compliance as competitiveness\u0026rdquo; pays off in procurement: anthropomorphic-interaction rules, SSE AI-listing guidance, Anthropic enterprise share surpassing OpenAI. Predict: compliance becomes a core gate in model and enterprise procurement; tighter domestic rules force agent products toward de-anthropomorphization/registration. Appendix: High-Frequency Keywords (deduplicated by topic) Model releases: GPT-5.6 / ChatGPT Work / Gemma 4 / GLM 5.2 / LongCat-2.0 / Grok 4.5 / Muse Spark 1.1 / (rumor) Gemini 3.5 Pro AI security: JadePuffer autonomous ransomware / GitLost prompt injection / Claude Code backdoor / strix / T3MP3ST / pentagi / Prismata / CISA×Mythos Custom chips: DeepSeek inference chip / Meta Iris / Samsung HBM / Sugon 8000 / Lingsi on-device chip / PKU memristor Domestic compute: 52% share / H100 rental +40% / LongCat-2.0 trillion-param closed loop Agent orchestration: OpenClaw / orca / agency-agents / WebSwarm / vercel/eve Agent memory: cognee / TencentDB-Agent-Memory / Remember When / Proactive Memory Skill standardization: agent-skills / agentskills / skills / superpowers / taste-skill Embodied AI: Unitree IPO / LingBot-VLA 2.0 / Embodied-Omni / Ego-Human / FSD-VLN AI for Science: superconducting discovery / VASP Agent / SpaCellAgent / computational imaging / cultural-relics agent On-device AI: Gemma 4 / pocket-tts / on-device inference chips Regulation \u0026amp; compliance: anthropomorphic-interaction rules / AI + HRSS / SSE listing guidance / Anthropic share overtake CV depth: TopoGPT / Perceive-to-Reason / HIVE / DeltaV / ProLaViT / AlayaWorld "
}