GPT

Headlines · 4

1. 3d
   Retail LLM platform cut agent hallucinations from 15% to 1.5% in six months

   An InfoQ China engineering playbook describes a large retailer's LLM-based inventory recommendation system, where about 15% of agent responses were hallucinations at the end of the first production month; six months later the rate had fallen to 1.5%. The authors say the drop came not from a stronger base model but from treating LLM capabilities as platform infrastructure: shared retry and failure classification, a prompt registry with versioning and rollback, schema enforcement at the platform boundary, per-request token cost attribution, behavioral observability and tool routing. The platform now serves dozens of application teams across inventory audit, replenishment review and discrepancy triage, and is built on Google's ADK orchestrator, LiteLLM for routing to models such as Gemini and GPT, and MCP-based tool services, with a companion codebase that runs locally on llama3.1 via Ollama to avoid API costs.

   InfoQ 中文 AI&大模型 · 🔥 0
2. 3d
   PromptArmor: open-weight models 88% more vulnerable to prompt injection

   PromptArmor Research says it ran 9,800 trials of indirect prompt injection attacks against agents built on leading open-weight and proprietary models, finding open-weight models 88% more susceptible overall: 26.7% versus 14.2%. In sample attack chains, 49% of open-weight cases ran malicious code after an agent visited an attacker website versus 9% for proprietary models, and 36% rendered a phishing login overlay versus 5%. The study also reports that newer cost-effective proprietary models were more exposed than their flagships (Claude Haiku 5.5 at +38 percentage points over Claude Fable 5.1, GPT-6 Luna at +21 over GPT-6 Astra), while Qwen and DeepSeek cost-effective models were less exposed than their flagships.

   PromptArmor Research · 🔥 0
3. 6d
   US backs OpenAI fair use but calls Chinese model distillation a threat

   In September 2026 the US government took two opposing positions on AI data use. On Sept 1 the Department of Justice filed a statement of interest in the Southern District of New York backing OpenAI in copyright suits brought by The New York Times and other publishers, arguing that training models on copyrighted text is highly transformative fair use and that requiring licenses would be unacceptable. On Sept 8, the NSA, CISA and FBI issued a joint cybersecurity advisory saying six Chinese firms used APIs to distill US frontier models such as Claude and GPT, extracting billions of tokens, and labeled it malicious, targeted capability theft. The article argues training and distillation are technically kindred: both use existing corpora without individual authorization, both yield new parameters rather than copies, and both may substitute for the original market — differing mainly in whether the material learned from is human work or machine output.

   虎嗅 AI · 🔥 0
4. 6d
   'AI bros' backlash: name-dropping AI as a social flex draws scorn

   A New Weekly commentary republished by Huxiu AI popularized the Chinese slang term 'AI deng' — people who pepper conversations with Prompt, Agent and multimodal-model jargon and treat 'using AI' as a high-end social persona. The author argues this showing-off actually exposes anxiety about being left behind by new technology rather than real capability, citing examples such as Cheetah Mobile chairman Fu Sheng using AI to write Mid-Autumn greetings for 50 staff and to reply to 70-plus messages (which drew online criticism), and a Golden Eagle Awards short film criticized for inserting AI-generated avatars 'AI for AI's sake'. The piece also cites a widely shared but contested chart claiming about 6.8 billion people have never used AI, 1.3 billion have used free conversational AI, and only 0.04% have used coding-style agent tools, while noting the figures and methodology were disputed.

   虎嗅 AI · 🔥 0

Experience and discussion from the community

Share my GPT experienceAsk about GPT