Floating-point arithmetic non-determinism can cause LLM inference to produce different outputs across runs, hardware, or provider routing, according to an analysis of numerical execution in large lang…
Provider routing can change an LLM's output when a request reaches a different model version, fallback model, parameter configuration, precision level, inference engine, region, or runtime environment…
A hosted LLM's output can change when providers update models, safety controls, routing, or serving infrastructure without code changes, breaking reproducibility. Google's Gemini API documentation exp…
OpenAI's API documentation explains that the `system_fingerprint` field in LLM responses is a provider-generated marker for the backend configuration serving a request, not a reflection of the user's …
Changing a tokenizer can alter an LLM's output, ranging from no visible difference to lower quality, different formatting, shorter usable context, or complete inference failure, according to Hugging F…
The same visible prompt is not always the same input received by a large language model (LLM), according to an analysis of provider routing and request assembly. Identical visible prompts do not estab…
Testing a nondeterministic LLM application requires treating it as a workflow with behavioral requirements across repeated runs, not as a function returning one exact string, according to a guide that…
VLLM's batch-invariance documentation defines a feature ensuring a request produces the same inference result regardless of batch size, composition, request order, or scheduling under a fixed hardware…
OpenAI reported that its strict Structured Outputs feature achieved 100% schema-matching reliability in an internal evaluation for gpt-4o-2024-08-06, but the company and other providers such as Amazon…
A new analysis from the Journal of Machine Learning Research explains that Mixture-of-Experts (MoE) routing can cause large language models to produce different outputs across runs due to discrete exp…
Setting a seed does not guarantee identical large language model (LLM) output, according to an analysis of LLM inference. A seed initializes the pseudorandom number generator used in token sampling bu…
A model alias is a named pointer that identifies a model's role and can be reassigned, while a versioned model ID identifies a specific release, according to guidance from OpenAI, Google Cloud, and Mi…
A new large language model (LLM) version can alter production behavior in unpredictable ways, so organizations should use a gated, progressive, reversible rollout that includes defining a production c…
A developer reports that switching from OpenAI's GPT-5.5 to GPT-5.6 in Codex, the company's AI coding agent, has made him less productive, with tasks spinning indefinitely and draining his three $200-…
AI has become a commodity, with DeepSeek V4 Flash costing $0.13 per million input tokens and $0.26 per million output tokens via OpenRouter, and TIME is serving AI crawlers a stripped-down markdown ve…
TIME is serving AI crawlers a stripped-down markdown version of its website with ads embedded, while human visitors receive the full HTML page, according to a test by an unnamed researcher. The markdo…
OpenAI's GPT-5.6 Sol xhigh uses 16.45M tokens per session, 2.25 times the 7.30M tokens per session of GPT-5.5 xhigh, according to a developer's analysis of 1,715 sessions over 14 days. Total token usa…
OpenAI's Codex CLI now renders responsive Markdown tables in the terminal, adapting to window width by switching from a grid to a record layout when columns are too narrow. The first implementation la…
Google Lighthouse 13.3 adds an experimental Agentic Browsing category that audits pages for automated client compatibility, while Cloudflare introduces AI traffic controls letting site operators block…
Google's Gemini 3.6 Flash and Gemini 3.5 Flash-Lite models ignore the temperature, top_p, and top_k sampling parameters, according to Google's current migration guide. The three parameters are depreca…