GTL Technology Reading Kit Edition 1.0 · Global Tech Live https://globaltechlive.com/reading-kit Read technology claims with more context. A compact glossary, a verification checklist and practical templates for your next reporting idea. Free to read and download. No account, payment or invitation required. An educational reading aid, not a professional qualification, press credential or substitute for specialist advice. Terms can vary by context; check the original document. A glossary for careful readers 1. Training / 训练 Learning model parameters from examples. A training result is not proof that the model will perform equally well on new material. Further reading: https://developers.google.com/machine-learning/glossary 2. Inference / 推理 Using a trained model to produce an output. In AI reporting, this does not mean the output is a logically verified conclusion. Further reading: https://developers.google.com/machine-learning/glossary 3. Fine-tuning / 微调 Additional training of an existing model for a particular task or style. It changes parameters, unlike simply writing a different prompt. Further reading: https://developers.google.com/machine-learning/crash-course/llm/tuning 4. Distillation / 知识蒸馏 Training a student model using signals from a teacher model. A smaller student does not automatically retain every teacher capability. Further reading: https://developers.google.com/machine-learning/crash-course/llm/tuning 5. Token / 词元 A unit of text handled by a language model: sometimes a word, fragment or character. Token counts are not interchangeable with word counts. Further reading: https://developers.google.com/machine-learning/crash-course/llm 6. Context window / 上下文窗口 The token capacity available for a model interaction. A larger limit does not guarantee reliable recall of every supplied detail. Further reading: https://ai.google.dev/gemini-api/docs/long-context 7. Retrieval-augmented generation (RAG) / 检索增强生成 Retrieving external material to help a model answer. Check whether the retrieved sources actually support the answer; retrieval alone is not verification. Further reading: https://cloud.google.com/use-cases/retrieval-augmented-generation 8. Hallucination / 幻觉 A plausible-sounding model output that is unsupported or incorrect. Fluent wording, confidence and a citation-shaped link do not establish truth. Further reading: https://cloud.google.com/discover/what-are-ai-hallucinations 9. Open weights / 开放权重 Model weights made available for use. Inspect the license and restrictions: accessible weights alone do not establish that the whole system is open source. Further reading: https://opensource.org/ai/open-source-ai-definition 10. AI agent / AI 智能体 Software that can plan or take actions toward a goal, sometimes using tools. Ask what it may access, which actions need approval and how failures are handled. Further reading: https://cloud.google.com/discover/what-are-ai-agents 11. Benchmark / 基准测试 A defined task and measurement procedure for comparison. Read the workload, quality target and test conditions before calling one system better overall. Further reading: https://mlcommons.org/benchmarks/inference-datacenter/ 12. Latency / 延迟 Time taken for a specified request or operation. Ask where timing starts and ends, and whether queueing and network time are included. Further reading: https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/user_guide/metrics.html 13. Throughput / 吞吐量 Work completed per unit of time. Its unit and concurrency matter; high throughput does not necessarily mean short waits for each user. Further reading: https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/user_guide/metrics.html 14. Vulnerability / 安全漏洞 A weakness that can be exploited or triggered to cause harm. Distinguish a reported weakness, a demonstrated exploit and confirmed real-world abuse. Further reading: https://csrc.nist.gov/glossary/term/vulnerability 15. Encryption / 加密 Transforming readable data into ciphertext using cryptography. Ask which data is protected, who controls keys and where decryption happens. Further reading: https://csrc.nist.gov/glossary/term/encryption 16. Preprint / 预印本 A research manuscript shared before formal journal publication, often before peer review. Repository moderation is not a substitute for scientific peer review. Further reading: https://info.arxiv.org/help/moderation/index.html Before you share: verify the claim 1. Find the original document. Open the actual announcement, paper or filing, not just a screenshot or repost. Record its URL and relevant section. 2. Identify who is making the claim. Separate a company's statement from an independent finding; note funding, sponsorship and relevant interests. 3. Separate dates. Record when the event happened, when the document appeared and when it was updated. Retain the stated timezone. 4. Read numbers with their denominator. Check units, period, sample, baseline and geography. Do not turn an estimate, target or starting price into a confirmed result. 5. Inspect the method. For a test, record product/model version, workload, hardware and limitations. A chart without comparable conditions is not enough. 6. Corroborate independently. Two outlets repeating one press release are not two independent confirmations. Explain disagreements instead of silently choosing a convenient figure. 7. Check quotes and rights. Read the surrounding context; preserve meaning and attribution. Public access to an image or recording does not itself grant reuse permission. 8. Mark what remains unknown. Separate verified facts, analysis and unanswered questions. Keep a correction trail and never use an AI answer as the sole source. Prepare a pitch or interview A useful story pitch Fill in what you know; explicitly mark gaps. This is a preparation template, not an online submission form. 1. Working headline and one-sentence angle: [ ] 2. What changed, when, where and why now: [ ] 3. Who is affected, and what readers need to understand: [ ] 4. Original sources and independently corroborated facts: [URLs + notes] 5. Questions still open; people to interview; relevant expertise: [ ] 6. My actual role: interviewed / attended / tested / desk research: [ ] 7. AI assistance, relationships/conflicts and media reuse permissions: [ ] 8. Proposed format, delivery date and terms to agree before work: [ ] An evidence-led interview Agree recording, attribution and publication ground rules before starting. Ask open questions; do not invent answers or consent. 1. What precisely changed, and what was the previous situation? 2. Which original documents or reproducible evidence support that claim? 3. How was it measured: version, sample, comparison and limitations? 4. Who benefits, who may bear costs, and who could independently assess it? 5. What failed or remains uncertain, and what would change your conclusion? 6. What relevant commercial, funding or personal relationships should be disclosed? 7. What is the correct name/title/date, and how can we clarify factual ambiguities? Original references GTL wrote these short explanations and templates. The linked primary references offer further detail; they do not endorse GTL or certify this resource. Google — Machine Learning Glossary https://developers.google.com/machine-learning/glossary Google — Introduction to Large Language Models https://developers.google.com/machine-learning/crash-course/llm Google AI — Long context https://ai.google.dev/gemini-api/docs/long-context Google — Fine-tuning, distillation, and prompt engineering https://developers.google.com/machine-learning/crash-course/llm/tuning Google Cloud — Retrieval-Augmented Generation https://cloud.google.com/use-cases/retrieval-augmented-generation Google Cloud — AI hallucinations https://cloud.google.com/discover/what-are-ai-hallucinations Open Source Initiative — Open Source AI Definition https://opensource.org/ai/open-source-ai-definition Google Cloud — AI agents https://cloud.google.com/discover/what-are-ai-agents MLCommons — MLPerf Inference: Datacenter https://mlcommons.org/benchmarks/inference-datacenter/ NVIDIA — Triton Inference Server metrics https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/user_guide/metrics.html NIST — Vulnerability glossary https://csrc.nist.gov/glossary/term/vulnerability NIST — Encryption glossary https://csrc.nist.gov/glossary/term/encryption arXiv — Content moderation https://info.arxiv.org/help/moderation/index.html Choose a report to read, listen to or share: https://globaltechlive.com/#feed Discuss a pitch with the editorial desk: https://globaltechlive.com/join The current Join GTL route prepares an email enquiry. It does not create an online application, guarantee publication or issue credentials.