{"schema_version":"onlylabs.public_signal.v1","title":"OpenAI Writing: Faulty reward functions in the wild","description":"OpenAI writing signal with public source context, captured evidence pages, related signals, and data-business radar classification.","url":"https://onlylabs.fyi/signals/54978ac9-380c-4388-aecd-8f5584a35f29","json_url":"https://onlylabs.fyi/signals/54978ac9-380c-4388-aecd-8f5584a35f29/signal.json","generated_at":"2026-06-08T15:47:17.691+00:00","org":{"slug":"openai","name":"OpenAI","category":"frontier-lab","category_label":"Frontier lab","dossier_url":"https://onlylabs.fyi/labs/openai","dossier_json_url":"https://onlylabs.fyi/labs/openai/dossier.json"},"related_urls":{"signal":"https://onlylabs.fyi/signals/54978ac9-380c-4388-aecd-8f5584a35f29","signal_json":"https://onlylabs.fyi/signals/54978ac9-380c-4388-aecd-8f5584a35f29/signal.json","source":"https://openai.com/index/faulty-reward-functions","lab_dossier":"https://onlylabs.fyi/labs/openai","lab_dossier_json":"https://onlylabs.fyi/labs/openai/dossier.json","analysis":"https://onlylabs.fyi/analysis/openai","analysis_json":"https://onlylabs.fyi/analysis/openai/analysis.json","analysis_evidence_json":"https://onlylabs.fyi/analysis/openai/evidence.json","category":"https://onlylabs.fyi/frontier","category_json":"https://onlylabs.fyi/frontier.json","category_feed":"https://onlylabs.fyi/frontier/feed.xml","category_signals_json":"https://onlylabs.fyi/signals.json","topic":"https://onlylabs.fyi/topics/talking","topic_signals_json":"https://onlylabs.fyi/topics/talking/signals.json","topic_feed":"https://onlylabs.fyi/topics/talking/feed.xml","data_business":null},"answer_pack":{"answer":"OpenAI published Faulty reward functions in the wild. This talking signal gives public context for research themes, product direction, policy, or launch framing. High-signal details: Faulty reward functions in the wild | OpenAI December 21, 2016 Conclusion Faulty reward functions in the wild Reinforcement learning algorithms can break in surprising,.... onlylabs links this event to 1 captured evidence page and 6 related writing signals.","signal_desk":"talking","source_context":{"source_url":"https://openai.com/index/faulty-reward-functions","source_host":"openai.com","occurred_at":"2016-12-21T08:00:00+00:00","first_seen_at":"2026-06-05T05:42:57.832854+00:00","date_source":"rss.item_date","context":null},"context_markers":[{"label":"Lab","value":"OpenAI","source":"signal"},{"label":"Signal desk","value":"talking","source":"signal"},{"label":"Source host","value":"openai.com","source":"source"},{"label":"Watch term","value":"RL environments","source":"evidence"},{"label":"Watch term","value":"Eval methodology","source":"evidence"},{"label":"Watch term","value":"Infrastructure","source":"evidence"},{"label":"Watch term","value":"Safety and alignment","source":"evidence"},{"label":"Watch term","value":"Agents and tool use","source":"evidence"}],"evidence_coverage":{"target_pages":1,"captured_pages":1,"readable_pages":1,"capture_methods":["exa"],"missing_page_urls":[],"failed_page_urls":[],"blocked_page_urls":[],"page_urls":["https://openai.com/index/faulty-reward-functions"],"related_signals":6,"has_source_url":true,"latest_page_fetched_at":"2026-06-08T15:47:17.691+00:00"},"data_business":{"matches":false,"lanes":[],"matched_terms":[],"score":null,"reason":null},"agent_handoff":{"signal_json":"https://onlylabs.fyi/signals/54978ac9-380c-4388-aecd-8f5584a35f29/signal.json","dossier_json":"https://onlylabs.fyi/labs/openai/dossier.json","analysis_json":"https://onlylabs.fyi/analysis/openai/analysis.json","analysis_evidence_json":"https://onlylabs.fyi/analysis/openai/evidence.json","topic_signals_json":"https://onlylabs.fyi/topics/talking/signals.json","topic_feed":"https://onlylabs.fyi/topics/talking/feed.xml","category_signals_json":"https://onlylabs.fyi/signals.json","data_radar_json":null,"opportunities_json":null},"analysis_playbook":{"objective":"Turn public writing and discussion into a readable map of research themes, product framing, policy posture, launch narratives, and market attention.","evidence_focus":["post title","source URL","captured page text","HN traction","linked model or paper references","publication date"],"extraction_questions":["Which themes are labs choosing to explain publicly?","Which posts are attracting outside discussion?","Which writing reframes a recent release, model, hiring wave, or policy stance?","Which posts mention data, evals, infrastructure, safety, or deployment workflows?"],"signal_questions":["What public theme, launch framing, or research direction does this writing signal expose?","Which themes are labs choosing to explain publicly?","Which posts are attracting outside discussion?","Do the 6 related writing signals show a repeated pattern?"],"output_fields":["org","theme","public_framing","traction","data_business_lane","evidence_url"],"data_business_relevance":"Public writing supplies the narrative layer over raw signals and helps identify which frontier-lab priorities are becoming externally legible.","required_sources":[{"label":"signal_json","url":"https://onlylabs.fyi/signals/54978ac9-380c-4388-aecd-8f5584a35f29/signal.json","required":true},{"label":"source","url":"https://openai.com/index/faulty-reward-functions","required":true},{"label":"dossier_json","url":"https://onlylabs.fyi/labs/openai/dossier.json","required":true},{"label":"analysis_evidence_json","url":"https://onlylabs.fyi/analysis/openai/evidence.json","required":true},{"label":"topic_signals_json","url":"https://onlylabs.fyi/topics/talking/signals.json","required":false},{"label":"data_radar_json","url":null,"required":false}],"expected_output":["one-paragraph source-grounded interpretation","category-specific implication","confidence and missing evidence","recommended next source to inspect"],"prompt_seed":"Using only the linked onlylabs JSON, captured source context, and cited evidence, analyze OpenAI's writing signal \"Faulty reward functions in the wild\" for frontier lab strategy."},"semantic_triples":[{"subject":"OpenAI","predicate":"published","object":"Faulty reward functions in the wild","text":"OpenAI published Faulty reward functions in the wild."},{"subject":"Faulty reward functions in the wild","predicate":"is classified as","object":"writing signal","text":"Faulty reward functions in the wild is classified as writing signal."},{"subject":"Faulty reward functions in the wild","predicate":"belongs to","object":"talking desk","text":"Faulty reward functions in the wild belongs to talking desk."},{"subject":"Faulty reward functions in the wild","predicate":"has evidence coverage","object":"1 captured evidence page","text":"Faulty reward functions in the wild has evidence coverage 1 captured evidence page."},{"subject":"Faulty reward functions in the wild","predicate":"has captured page count","object":"1","text":"Faulty reward functions in the wild has captured page count 1."},{"subject":"Faulty reward functions in the wild","predicate":"has readable page count","object":"1","text":"Faulty reward functions in the wild has readable page count 1."},{"subject":"Faulty reward functions in the wild","predicate":"has related signal count","object":"6","text":"Faulty reward functions in the wild has related signal count 6."},{"subject":"Faulty reward functions in the wild","predicate":"has analysis playbook objective","object":"Turn public writing and discussion into a readable map of research themes, product framing, policy posture, launch narratives, and market attention.","text":"Faulty reward functions in the wild has analysis playbook objective Turn public writing and discussion into a readable map of research themes, product framing, policy posture, launch narratives, and market attention.."},{"subject":"Faulty reward functions in the wild","predicate":"has source host","object":"openai.com","text":"Faulty reward functions in the wild has source host openai.com."},{"subject":"Faulty reward functions in the wild","predicate":"has lab","object":"OpenAI","text":"Faulty reward functions in the wild has lab OpenAI."},{"subject":"Faulty reward functions in the wild","predicate":"has signal desk","object":"talking","text":"Faulty reward functions in the wild has signal desk talking."},{"subject":"Faulty reward functions in the wild","predicate":"has source host","object":"openai.com","text":"Faulty reward functions in the wild has source host openai.com."},{"subject":"Faulty reward functions in the wild","predicate":"has watch term","object":"RL environments","text":"Faulty reward functions in the wild has watch term RL environments."},{"subject":"Faulty reward functions in the wild","predicate":"has watch term","object":"Eval methodology","text":"Faulty reward functions in the wild has watch term Eval methodology."},{"subject":"Faulty reward functions in the wild","predicate":"has watch term","object":"Infrastructure","text":"Faulty reward functions in the wild has watch term Infrastructure."},{"subject":"Faulty reward functions in the wild","predicate":"has watch term","object":"Safety and alignment","text":"Faulty reward functions in the wild has watch term Safety and alignment."},{"subject":"Faulty reward functions in the wild","predicate":"has watch term","object":"Agents and tool use","text":"Faulty reward functions in the wild has watch term Agents and tool use."}]},"intelligence":{"signal_desk":"talking","answer":"OpenAI published Faulty reward functions in the wild. This talking signal gives public context for research themes, product direction, policy, or launch framing. High-signal details: Faulty reward functions in the wild | OpenAI December 21, 2016 Conclusion Faulty reward functions in the wild Reinforcement learning algorithms can break in surprising,.... onlylabs links this event to 1 captured evidence page and 6 related writing signals.","semantic_triples":[{"subject":"OpenAI","predicate":"published","object":"Faulty reward functions in the wild","text":"OpenAI published Faulty reward functions in the wild."},{"subject":"Faulty reward functions in the wild","predicate":"is classified as","object":"writing signal","text":"Faulty reward functions in the wild is classified as writing signal."},{"subject":"Faulty reward functions in the wild","predicate":"belongs to","object":"talking desk","text":"Faulty reward functions in the wild belongs to talking desk."},{"subject":"Faulty reward functions in the wild","predicate":"has evidence coverage","object":"1 captured evidence page","text":"Faulty reward functions in the wild has evidence coverage 1 captured evidence page."}]},"signal":{"id":"54978ac9-380c-4388-aecd-8f5584a35f29","url":"https://onlylabs.fyi/signals/54978ac9-380c-4388-aecd-8f5584a35f29","json_url":"https://onlylabs.fyi/signals/54978ac9-380c-4388-aecd-8f5584a35f29/signal.json","source_url":"https://openai.com/index/faulty-reward-functions","title":"Faulty reward functions in the wild","summary":"OpenAI published a writing signal. onlylabs watches public writing for research themes, product direction, and model-launch context.","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"openai","name":"OpenAI","category":"frontier-lab"},"occurred_at":"2016-12-21T08:00:00+00:00","first_seen_at":"2026-06-05T05:42:57.832854+00:00","date_source":"rss.item_date","evidence_coverage":{"target_pages":1,"captured_pages":1,"readable_pages":1,"capture_methods":["exa"],"missing_page_urls":[],"failed_page_urls":[],"blocked_page_urls":[],"page_urls":["https://openai.com/index/faulty-reward-functions"]},"facets":{},"traction":{"github_stars":null,"hn_points":4,"hn_comments":0,"hn_story_id":"47268853","hf_downloads":null,"hf_likes":null},"data_radar":null},"primary_evidence_page":{"url":"https://openai.com/index/faulty-reward-functions","final_url":"https://openai.com/index/faulty-reward-functions","title":"Faulty reward functions in the wild","http_status":200,"content_type":null,"capture_method":"exa","fetched_at":"2026-06-08T15:47:17.691+00:00","bytes":null,"raw_path":null,"content_hash":null,"excerpt_chars":1200,"truncated":true,"excerpt":"Faulty reward functions in the wild | OpenAI December 21, 2016 Conclusion Faulty reward functions in the wild Reinforcement learning algorithms can break in surprising, counterintuitive ways. In this post we’ll explore one failure mode, which is where you misspecify your reward function. Loading… Share At OpenAI, we’ve recently started using Universe⁠, our software for measuring and training AI agents, to conduct new RL experiments. Sometimes these experiments illustrate some of the issues with RL as currently practiced. In the following example we’ll highlight what happens when a misspecified reward function encourages an RL agent to subvert its environment by prioritizing the acquisition of reward signals above other measures of success. Designing safe AI systems will require us to design algorithms that don’t attempt to do this, and will teach us to specify and shape goals in such a way they can’t be misinterpreted by our AI agents. One of the games we’ve been training on is CoastRunners⁠. The goal of the game—as understood by most humans—is to finish the boat race quickly and (preferably) ahead of other players. CoastRunners does not directly reward the player’s progression..."},"evidence_pages":[{"url":"https://openai.com/index/faulty-reward-functions","final_url":"https://openai.com/index/faulty-reward-functions","title":"Faulty reward functions in the wild","http_status":200,"content_type":null,"capture_method":"exa","fetched_at":"2026-06-08T15:47:17.691+00:00","bytes":null,"raw_path":null,"content_hash":null,"excerpt_chars":1200,"truncated":true,"excerpt":"Faulty reward functions in the wild | OpenAI December 21, 2016 Conclusion Faulty reward functions in the wild Reinforcement learning algorithms can break in surprising, counterintuitive ways. In this post we’ll explore one failure mode, which is where you misspecify your reward function. Loading… Share At OpenAI, we’ve recently started using Universe⁠, our software for measuring and training AI agents, to conduct new RL experiments. Sometimes these experiments illustrate some of the issues with RL as currently practiced. In the following example we’ll highlight what happens when a misspecified reward function encourages an RL agent to subvert its environment by prioritizing the acquisition of reward signals above other measures of success. Designing safe AI systems will require us to design algorithms that don’t attempt to do this, and will teach us to specify and shape goals in such a way they can’t be misinterpreted by our AI agents. One of the games we’ve been training on is CoastRunners⁠. The goal of the game—as understood by most humans—is to finish the boat race quickly and (preferably) ahead of other players. CoastRunners does not directly reward the player’s progression..."}],"related_signals":[{"id":"b3668d3b-26d2-40c0-9d4f-ed1a67927aa4","url":"https://onlylabs.fyi/signals/b3668d3b-26d2-40c0-9d4f-ed1a67927aa4","source_url":"https://openai.com/index/supporting-eu-trustworthy-ai-ecosystem","title":"Supporting Europe’s work in ensuring a trustworthy AI ecosystem ","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"openai","name":"OpenAI","category":"frontier-lab"},"occurred_at":"2026-06-11T00:00:00+00:00","first_seen_at":"2026-06-11T08:00:56.140796+00:00","date_source":"rss.item_date"},{"id":"2638c0a7-b372-409c-ac72-f6d81d6464dc","url":"https://onlylabs.fyi/signals/2638c0a7-b372-409c-ac72-f6d81d6464dc","source_url":"https://openai.com/index/using-codex-to-simulate-black-holes","title":"How an astrophysicist uses Codex to help simulate black holes","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"openai","name":"OpenAI","category":"frontier-lab"},"occurred_at":"2026-06-11T00:00:00+00:00","first_seen_at":"2026-06-11T07:01:16.936464+00:00","date_source":"rss.item_date"},{"id":"509ea784-51ec-4ede-855b-5a4d1b27d3be","url":"https://onlylabs.fyi/signals/509ea784-51ec-4ede-855b-5a4d1b27d3be","source_url":"https://openai.com/index/openai-on-oracle-cloud","title":"Access OpenAI models and Codex through your Oracle cloud commitment","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"openai","name":"OpenAI","category":"frontier-lab"},"occurred_at":"2026-06-10T20:00:00+00:00","first_seen_at":"2026-06-11T07:01:16.936464+00:00","date_source":"rss.item_date"},{"id":"4f051449-87f2-466e-941e-b5918381a8fe","url":"https://onlylabs.fyi/signals/4f051449-87f2-466e-941e-b5918381a8fe","source_url":"https://openai.com/index/prc-linked-influence-operations-ai-debates","title":"PRC-linked influence operations are targeting AI debates in the US","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"openai","name":"OpenAI","category":"frontier-lab"},"occurred_at":"2026-06-10T12:00:00+00:00","first_seen_at":"2026-06-11T07:01:16.936464+00:00","date_source":"rss.item_date"},{"id":"4507c0c1-cb74-4bb3-b62b-5f6c2d37e20d","url":"https://onlylabs.fyi/signals/4507c0c1-cb74-4bb3-b62b-5f6c2d37e20d","source_url":"https://openai.com/index/lseg","title":"From data to decisions: how LSEG is scaling trusted AI","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"openai","name":"OpenAI","category":"frontier-lab"},"occurred_at":"2026-06-10T00:00:00+00:00","first_seen_at":"2026-06-10T09:18:54.26094+00:00","date_source":"rss.item_date"},{"id":"fb16aa7a-c4ef-4859-b514-0839c2f1330d","url":"https://onlylabs.fyi/signals/fb16aa7a-c4ef-4859-b514-0839c2f1330d","source_url":"https://openai.com/index/nextdoor","title":"How engineers at Nextdoor use Codex to build without limits","context":null,"kind":{"key":"post_published","label":"Writing"},"org":{"slug":"openai","name":"OpenAI","category":"frontier-lab"},"occurred_at":"2026-06-09T12:00:00+00:00","first_seen_at":"2026-06-10T07:01:28.700378+00:00","date_source":"rss.item_date"}]}