{ "story_id": "f206debeb8b8498b878f7734fe40fbdc", "desk": "drm3", "revision": 1, "published_at": "2026-09-02T16:32:24.000Z", "content_hash": "990b6327d2b7a46e7111c0808e61937b7bc647fc8d9e305d7f6ec6b4ddc6295f", "hash_basis": "sha256 over `headline\\ndek\\nprose`, plus `\\n` + the canonical citations JSON when any source is placed, plus `\\n#blog` for blogs", "basis": { "headline": "Alignment Tuning Drives Sycophancy and Bias in LLMs", "dek": "Research finds alignment tuning, not pretraining, shapes sycophancy and cue-induced biases in large language models.", "prose": "Cue-induced bias is best understood not as a single flaw in LLMs but as a family of causally effective linear directions that are largely shaped by alignment tuning. [^1]\n\nResearchers studied where susceptibility to sycophancy and cue-induced biases lives inside large language models across five model families and seven bias types. [^2]\n\nThe susceptibility to sycophancy and cue-induced biases is largely shaped by alignment tuning rather than pretraining. [^3]\n\nResearchers studied how explicit world-modeling objectives affect the internal representations and downstream capability of Transformers using Rubik's Cubes as the training domain. [^4]\n\nIn an interview with LiveMint, creator and podcaster Prakhar Gupta said that if his YouTube business were wiped out, he would first find someone who already has attention and make himself disproportionately useful to them, rather than trying to become famous. [^5]\n\nPrakhar Gupta said in the same interview that his first ₹1 lakh from rebuilding would likely come from services, not views, and he would then use the cash flow to build his own distribution again. [^6]\n\nPrakhar Gupta said that if he invested ₹10 lakh in a creator, he would examine velocity of recent content, returning viewers, non-follower views, and hunger, and would structure a revenue share rather than buying permanent equity. [^7]\n\nPrakhar Gupta said that business, finance, health, and aspirational audiences monetize better because viewers have commercial intent. [^8]", "cited": "[{\"statement\":\"Cue-induced bias is best understood not as a single flaw in LLMs but as a family of causally effective linear directions that are largely shaped by alignment tuning.\",\"source\":\"arXiv.org\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-02T04:00:00.000Z\",\"publisher_count\":1,\"sources\":[\"arXiv.org\"]},{\"statement\":\"Researchers studied where susceptibility to sycophancy and cue-induced biases lives inside large language models across five model families and seven bias types.\",\"source\":\"arXiv.org\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-02T04:00:00.000Z\",\"publisher_count\":1,\"sources\":[\"arXiv.org\"]},{\"statement\":\"The susceptibility to sycophancy and cue-induced biases is largely shaped by alignment tuning rather than pretraining.\",\"source\":\"arXiv.org\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-02T04:00:00.000Z\",\"publisher_count\":1,\"sources\":[\"arXiv.org\"]},{\"statement\":\"Researchers studied how explicit world-modeling objectives affect the internal representations and downstream capability of Transformers using Rubik's Cubes as the training domain.\",\"source\":\"arXiv.org\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-01T04:00:00.000Z\",\"publisher_count\":1,\"sources\":[\"arXiv.org\"]},{\"statement\":\"In an interview with LiveMint, creator and podcaster Prakhar Gupta said that if his YouTube business were wiped out, he would first find someone who already has attention and make himself disproportionately useful to them, rather than trying to become famous.\",\"source\":\"mint\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-02T16:32:24.000Z\",\"publisher_count\":1,\"sources\":[\"mint\"]},{\"statement\":\"Prakhar Gupta said in the same interview that his first ₹1 lakh from rebuilding would likely come from services, not views, and he would then use the cash flow to build his own distribution again.\",\"source\":\"mint\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-02T16:32:24.000Z\",\"publisher_count\":1,\"sources\":[\"mint\"]},{\"statement\":\"Prakhar Gupta said that if he invested ₹10 lakh in a creator, he would examine velocity of recent content, returning viewers, non-follower views, and hunger, and would structure a revenue share rather than buying permanent equity.\",\"source\":\"mint\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-02T16:32:24.000Z\",\"publisher_count\":1,\"sources\":[\"mint\"]},{\"statement\":\"Prakhar Gupta said that business, finance, health, and aspirational audiences monetize better because viewers have commercial intent.\",\"source\":\"mint\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-02T16:32:24.000Z\",\"publisher_count\":1,\"sources\":[\"mint\"]}]", "kind": "news" }, "receipt_verify": "Ed25519 over the dot-joined string `slice_hash.cursor_from.cursor_to.view.view_version.row_count`; public_key and sig are base64url of the raw 32-byte key / 64-byte signature", "receipt": null, "receipt_note": "this revision predates receipt-keeping (before v0.37.0); the filed row lives in the record", "generation_chain": { "wire": { "stream": "fountain_news", "story_id": "0486c06f17bbefdd030a7d960fcf5cb6", "thread_id": "10a7ef1fccb347738c43796cc7621893", "thread_label": "Prakhar Gupta", "novelty": "UPDATE", "content_hash": "0f2bdf7126c41fd17c5f3b68f136b3f217570e172863f56181b3c14874839945", "last_published_at": "2026-09-02T16:32:24.000Z", "read_receipt": { "slice_hash": "4cacbcc1c8ccf43a891d2f2aaa9dd85a75a21f0eb716714b606b48e8aa0d2073", "cursor_from": "eyJ0cyI6IjIwMjYtMDktMDJUMTY6MDA6MTIuMDAwMDAwWiIsImlkIjoiM2Q3MWZkM2JlOTQyY2I5ZWNkN2NiMWQxNDk3YTg3NjMiLCJ2IjoiMSJ9", "cursor_to": "eyJ0cyI6IjIwMjYtMDktMDJUMTY6NDk6NDEuMDAwMDAwWiIsImlkIjoiYTVlNGY3ZjRmZWRlZTBiZTA5ODllYmY0YWFhMThmMjIiLCJ2IjoiMSJ9", "view": "v_fountain_news", "view_version": "1", "row_count": 100, "window_days": 3, "bytes_scanned": 10113494, "credits": 8, "price_per_100_rows": 8, "sig": "BXDkUoytkUF-6vhEBXL_iYp0NHuO6YAIj2XKODfENdK_O4bUpfgaHF894DBgK6LO9dEThXKDng75CTuAHrNvDA", "public_key": "bMUigy8O0jOnBxQ4Sc-5lwhIZ8LQVAhxMbR7qESVuUE", "signer_path": "lakehouse/data-extract/v1", "alg": "Ed25519", "signed": true } }, "written_at": "2026-09-03T00:41:21.745Z" }, "cited_facts": [ { "statement": "Cue-induced bias is best understood not as a single flaw in LLMs but as a family of causally effective linear directions that are largely shaped by alignment tuning.", "source": "arXiv.org", "instrument": "News", "claim_key": null, "published_at": "2026-09-02T04:00:00.000Z", "publisher_count": 1, "sources": [ "arXiv.org" ] }, { "statement": "Researchers studied where susceptibility to sycophancy and cue-induced biases lives inside large language models across five model families and seven bias types.", "source": "arXiv.org", "instrument": "News", "claim_key": null, "published_at": "2026-09-02T04:00:00.000Z", "publisher_count": 1, "sources": [ "arXiv.org" ] }, { "statement": "The susceptibility to sycophancy and cue-induced biases is largely shaped by alignment tuning rather than pretraining.", "source": "arXiv.org", "instrument": "News", "claim_key": null, "published_at": "2026-09-02T04:00:00.000Z", "publisher_count": 1, "sources": [ "arXiv.org" ] }, { "statement": "Researchers studied how explicit world-modeling objectives affect the internal representations and downstream capability of Transformers using Rubik's Cubes as the training domain.", "source": "arXiv.org", "instrument": "News", "claim_key": null, "published_at": "2026-09-01T04:00:00.000Z", "publisher_count": 1, "sources": [ "arXiv.org" ] }, { "statement": "In an interview with LiveMint, creator and podcaster Prakhar Gupta said that if his YouTube business were wiped out, he would first find someone who already has attention and make himself disproportionately useful to them, rather than trying to become famous.", "source": "mint", "instrument": "News", "claim_key": null, "published_at": "2026-09-02T16:32:24.000Z", "publisher_count": 1, "sources": [ "mint" ] }, { "statement": "Prakhar Gupta said in the same interview that his first ₹1 lakh from rebuilding would likely come from services, not views, and he would then use the cash flow to build his own distribution again.", "source": "mint", "instrument": "News", "claim_key": null, "published_at": "2026-09-02T16:32:24.000Z", "publisher_count": 1, "sources": [ "mint" ] }, { "statement": "Prakhar Gupta said that if he invested ₹10 lakh in a creator, he would examine velocity of recent content, returning viewers, non-follower views, and hunger, and would structure a revenue share rather than buying permanent equity.", "source": "mint", "instrument": "News", "claim_key": null, "published_at": "2026-09-02T16:32:24.000Z", "publisher_count": 1, "sources": [ "mint" ] }, { "statement": "Prakhar Gupta said that business, finance, health, and aspirational audiences monetize better because viewers have commercial intent.", "source": "mint", "instrument": "News", "claim_key": null, "published_at": "2026-09-02T16:32:24.000Z", "publisher_count": 1, "sources": [ "mint" ] } ], "note": "A signature proves who filed this and that it has not changed since. It never makes a claim true." }