Kanika leaks 50 LLM questions
- An X post by Kanika repackaged a public 50-question large language model study guide, and the source trail points back to a May 2025 PDF by Hao Hoang rather than a verified Big Tech leak. - The 50-question list spans tokenization, attention, context windows, LoRA and QLoRA fine-tuning, beam search, retrieval-augmented generation, safety, and deployment, then spread into GitHub study guides and interview-prep sites. - By April 2026, the list had become a reusable prep template across blogs, GitHub, and Substack, not confirmed company interview material. (github.com)
A viral X post framed 50 large language model interview questions as a Big Tech “leak,” but the material appears to trace back to a public study guide by Hao Hoang. (github.com) (altervista.org) The clearest source now online is a PDF titled “Top 50 Large Language Model (LLM) Interview Questions,” dated May 2025 and credited to Hao Hoang. Its first questions cover tokenization, attention, context windows, LoRA, QLoRA, and beam search. (altervista.org) A separate GitHub repo built after the X post says the thread described “Big Tech” and “50 questions” with a “leaked” framing, but warns that any social-media list should be treated as unverified. The repo says it is “not a dump of proprietary ‘leaked’ questions.” (github.com) Large language model interviews usually test whether candidates can explain how text gets split into tokens, how attention links words across a sentence, and how a context window limits what the model can “remember” at one time. Those are the opening topics in Hoang’s guide. (altervista.org) The list then moves from model basics to practical engineering: low-rank adaptation, or LoRA, for cheaper fine-tuning; QLoRA for lower-memory tuning with quantization; and decoding methods such as beam search for text generation. The PDF presents them as interview prompts with worked answers, not screenshots of employer documents. (altervista.org) By June 8, 2025, AI Engineer Guide had already linked readers to Hoang’s document as a useful public resource. By the time the Kanika post circulated, the same 50-question structure had already escaped into the wider interview-prep ecosystem. (aiengineerguide.com) Scott Hanselman later turned the material into a browsable GitHub study guide with progress tracking and markdown pages for the 50 questions and answers. His repository credits Hoang for the original PDF content. (github.com) Hao Hoang also runs an “AI Interview Prep” Substack that publishes recurring posts on large language model system-design traps, including tokenizer behavior, precision choices, data loading, and decoding bottlenecks. That makes the viral list look more like the front door to a broader interview-prep brand than a one-off leak. (substack.com) The result is a familiar internet pattern: a public study guide gets relabeled as insider material once it hits X. In this case, the evidence available on April 26, 2026 points to a widely copied public question bank, not confirmed confidential interview prompts from Google, Meta, or OpenAI. (github.com 1) (github.com 2)
