Serguey Asael Shinder
PulseAugur coverage of Serguey Asael Shinder — every cluster mentioning Serguey Asael Shinder across labs, papers, and developer communities, ranked by signal.
7 day(s) with sentiment data
-
AI assistant leaks discount code via system prompt, highlighting OWASP Top 10 vulnerability
A customer successfully prompted an AI assistant to reveal its system instructions, including a discount code, by simply asking it to print all text above a certain line. This incident highlights a common vulnerability …
-
AI support assistant leaks system prompt and discount code
A customer successfully prompted an AI support assistant to reveal its system prompt, including a discount code. This incident highlights a common vulnerability known as system prompt leakage, where instructions meant t…
-
LLM prompt caching fails with dynamic elements at the start
Prompt caching for LLMs is ineffective when dynamic elements like timestamps are placed at the beginning of the prompt. Prompt caching relies on identical prefixes to reuse processed tokens, but a changing timestamp at …
-
Prompt Caching Fails Due to Timestamp Placement at Prompt's Beginning
A software engineer named Serguey Asael Shinder observed that prompt caching was ineffective due to a timestamp placed at the beginning of the prompt. Prompt caching relies on identical prefixes, and a changing timestam…
-
LLMs struggle with "Lost in the Middle" context retrieval
A recent article highlights a phenomenon known as "Lost in the Middle," where large language models struggle to retrieve information located in the middle of their context window. Research indicates that models perform …
-
LLM retrieval fails by splitting headings from context
A developer encountered an issue where an LLM's retrieval system provided incorrect information due to poor document chunking. The system, designed to answer questions based on a pricing policy, failed because it split …
-
AI models can silently truncate output, leading to incomplete data
A developer highlights a common issue where AI models truncate output without explicit notification, leading to incomplete results being passed downstream. The problem arises when the model hits its maximum output lengt…
-
AI answer cache fails by confusing similar topics with identical questions
An AI caching system designed to reduce model call expenses by storing and reusing answers to similar questions was found to be flawed. The cache used embeddings to determine similarity, treating questions with similar …
-
LLMs lack real-time clocks, leading to date inaccuracies
Large language models lack a real-time clock, relying instead on their training data's cutoff date to understand time. This can lead to inaccuracies when users ask about current events, dates, or the latest versions of …
-
Unversioned AI prompts pose significant risks, developer warns
A developer highlights the critical issue of unversioned and uncommitted prompts used in production AI systems. These prompts, often treated as experimental text, can be altered without a traceable history, leading to u…
-
AI models can be prompted to argue either side of a question
An AI model can be prompted to argue either side of a given topic, providing specific points and encouragement for one perspective, or detailing significant problems for another. The model's output is highly dependent o…
-
LLM Cost Management: Track, Cap, and Optimize Usage
The author argues that developers need to closely monitor the cost of LLM requests, similar to how they track latency and error rates. He suggests logging costs per request, attaching user and endpoint information, and …
-
LLM prompts are code: Pin model versions to avoid silent breaks
Developers building on large language models must treat model IDs as versioned dependencies, akin to software libraries, to avoid unexpected behavior changes. The author, Serguey Asael Shinder, advises explicitly pinnin…
-
AI models excel at answers, not problem definition, human oversight crucial
Large language models excel at generating answers but struggle to identify the correct problem or question. They can produce confident, well-documented responses to flawed prompts, leading to wasted effort on solutions …
-
AI models offer helpful drafts but lack personal stakes in outcomes
The author argues that while AI models offer valuable assistance, they lack personal investment in the outcomes of their suggestions. Unlike humans, models do not face consequences for errors, such as losing sleep, clie…