Robin Li Says Gap Exists Between Compute Investment and Application Returns; Anthropic Revenue Growth Plummets to 47% · Issue 1117-836
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
The underwhelming results of AI application-layer deployment are prompting industry vigilance over bubble risks. At the APEC summit, Robin Li stated bluntly that a gap exists between compute investment and application returns [9]. Meanwhile, Anthropic's disclosed S-1 data shows quarter-over-quarter revenue growth plunging from 144% to 47%, while carrying nearly $900 billion in procurement and leasing commitments—seen as a typical signal of an "AI valuation reset period" [16]. On the technical side, approximately 90% of the Claude Code team's code is written by AI, with verification loops becoming key infrastructure for running Agents at scale [1].
🚀 Key Updates
- Microsoft global outage raises questions about AI coding reliability [0]: A large-scale service failure has prompted reflection on the reliability of AI coding systems and the dilution of team experience.
- Claude Code team publishes Agent verification methodology [1]: About 90% of code is written and reviewed by Claude; an independent grader can boost task success rates by up to 10 points.
- Jailbroken PS4 runs Qwen3.5-9B local model [2]: The open-source project Garlic LLM achieves up to 43 tokens per second, enabling voice conversations without internet connectivity.
- Robin Li warns of AI bubble risk [9]: Says compute and infrastructure investment is enormous, but application-layer deployment is falling short; remains optimistic long-term.
- OpenAI spends $5 billion on internal model analysis reports [12]: Purchases model decomposition services from Goodfire; three other safety assessment firms have signed similar agreements.
- Anthropic publishes Opus 5.5 prompt recommendations [15]: Recommends removing prompts like "think hard" in favor of adaptive reasoning mechanisms.
- fal launches ByteDance Seedance 2.5 video model [17]: Offers three entry points—text-to-video, image-to-video, and reference-to-video—plus a US-hosted 4K version [18].
- BAAI FlagOpen open-sources HumanCOIN dataset [6]: A LeRobot-based multi-sensor, multi-frequency extension with dynamic alignment by time window, requiring no pre-resampling.
🔗 Sources
[0] Microsoft global outage: how AI became the scapegoat again — https://www.bestblogs.dev/article/ceb7562cd8?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [1] How the Claude Code team verifies Agent outputs: three types of testing and verification loop practices — https://aihot.news/items/qhchrbfl6n5vq38zmpjdx84rc [2] Garlic LLM enables jailbroken PS4 to run Qwen3.5-9B local model at up to 43 tokens per second — https://aihot.news/items/kmro26ifgirc7gsea3wu9rzy8 [6] FlagOpen open-sources HumanCOIN: a LeRobot-based multi-sensor, multi-frequency dataset extension — https://aihot.news/items/cuik6d6r5zzzur4pkwk2hve2j [9] Baidu's Robin Li warns of AI bubble risk at APEC summit, says application deployment results fall short of expectations — https://aihot.news/items/ljzzfpr6dmqbqn519op491she [12] OpenAI pays Goodfire $5 billion for internal model analysis reports — https://www.bestblogs.dev/status/2122550043962163652?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item