Show HN: Are You in the Weights?
A retro‑styled web toy that queries multiple large language models for “Who is \<name\>?” is amusing users by surfacing everything from accurate bios to wildly hallucinated identities, often turning ordinary people into fictional athletes, politicians, or even criminals. Commenters see it as a vivid illustration of how LLMs blend truth, guesswork, and bias—especially with rare or foreign‑sounding names—while also questioning what the site’s “strength” scores really indicate about being “in the weights.” Alongside the fun, many raise privacy and safety concerns about entering real names into a public, logged service and what misattributed or indelible AI-generated profiles could mean for reputation and any future “right to be forgotten.”
Overall reception
- Many find the site fun, ego-boosting, and visually appealing (retro/8‑bit aesthetic, clever concept).
- Others view it as a privacy trap or an instructive but worrying demo of LLM behavior.
Accuracy, hallucinations, and identity collisions
- Experiences range from eerily accurate summaries (especially for unique surnames, open‑source contributors, academics, or long‑used handles) to complete fabrications.
- Even people with globally unique names often get described as politicians, security researchers, or especially professional athletes and entertainers.
- Pseudonyms and long-lived online handles are sometimes recognized more accurately than legal names.
- The “hallucinations” section is imperfect: some spot‑on descriptions are labeled hallucinations, while many wrong ones appear in the “main” section.
Dangerous and defamatory misidentifications
- Several users are incorrectly labeled as terrorists, murderers, extremists, or crime victims.
- Some note that for Arabic or uncommon names, the tool often confuses them with sanctioned individuals or bombers.
- Commenters find these false positives “scary,” especially given reports that LLMs may be used in military or security decision-making.
Scoring, models, and clustering
- “Strength” is explained as a linear combination of model self‑reported confidence plus bonuses for cross‑model agreement; commenters note LLM confidence is poorly calibrated.
- The percentile (“Top N%”) is relative to all queries so far, not the broader population.
- A clusterer merges model outputs into entities and decides what is a hallucination, optimized for recall over precision, leading to many misclassifications.
- Prompting details are shared; all models use the same JSON‑only “Who is
?” prompt. Clustering runs on a cheaper model.
Privacy and data handling
- Strong criticism that all queries (including real names) were publicly listed via a “latest” leaderboard and accessible via API; later partially mitigated but data remain broadly accessible.
- Lack of a clear privacy policy and presence of tracking/Cloudflare checks raise suspicion about IP/name harvesting.
- Several argue one should assume any text submitted to random sites will be stored and reused, possibly in future training sets.
UX, cost, and design
- Praised for design and portraits (generated via an image model), but some report input bugs, intrusive sounds, and rate‑limit errors under load.
- Running many frontier and smaller models per query is acknowledged as costly; described by the author as a non‑commercial “fun hack and science experiment.”
Broader reflections
- Thread touches on “being in the weights” as a form of weird digital immortality and raises questions about the right to be forgotten.
- Some are relieved not to appear at all; others note how hard it is to keep real‑life identity separate from online traces once models ingest public data.