Découvrir Pandipedia
Un répertoire grandissant de réponses utiles sélectionnées par la communauté Pandi. Cherchez dans la collection ou parcourez les dernières découvertes.
1752 entrées disponibles
this is all good stuff, but a bit too broad - i want data sets only in english, for example
For English-only datasets, you might find interest in the Mozilla Common Voice, which contains 7,335 validated hours of spoken English across 60 languages, including demographic metadata like age and sex, thereby focusing on various English speakers[1].
Another excellent option is the People’s Speech Dataset, which is noted as among the world's largest English speech recognition corpus available under CC-BY-SA and CC-BY 4.0 licenses, comprising over 30,000 hours of transcribed English speech with diverse speakers[2].
Examinons les alternatives :
- Modifier la requête.
- Démarrer une nouvelle conversation.
- Supprimer des sources (si elles ont été ajoutées manuellement).
Senior fitness transformations that inspire
Examinons les alternatives :
- Modifier la requête.
- Démarrer une nouvelle conversation.
- Supprimer des sources (si elles ont été ajoutées manuellement).
Créez votre compte pour façonner ce Feed autour de ce que vous cherchez et aimez.
A practical UK procurement data workflow for commercial intelligence
Examinons les alternatives :
- Modifier la requête.
- Démarrer une nouvelle conversation.
- Supprimer des sources (si elles ont été ajoutées manuellement).
Five SEC Privacy Act rule changes requesters should know
Examinons les alternatives :
- Modifier la requête.
- Démarrer une nouvelle conversation.
- Supprimer des sources (si elles ont été ajoutées manuellement).
What does lifecycle AI governance actually require?. Synthesize the recurring lifecycle model across NIST and the EU AI Act: govern, map, measure, manage, document, test, monitor, manage suppliers, and decommission. Separate voluntary guidance from binding obligations so readers can see what is a governance best practice versus a legal compliance requirement.
Lifecycle AI governance: what NIST and the EU AI Act together require
The NIST AI RMF and its Playbook treat AI governance as a full lifecycle control system spanning govern, map, measure, and manage, with supporting practices such as documentation, testing, monitoring, supplier oversight, and decommissioning.[1][2][3] These are voluntary best-practice controls in the NIST materials, while the EU AI Act imposes binding legal obligations only for the actors, system types, and risk categories it covers.[4]
A useful way to read the two sources together is this: NIST describes the governance architecture an organization should build across the lifecycle, while the EU AI Act specifies the legal minimums for high-risk systems, GPAI models, and certain generative AI uses, especially around documentation, logging, human oversight, monitoring, and incident reporting.[5][6][7][8][9][10]
Lifecycle comparison: voluntary governance controls versus binding EU AI Act duties
| Lifecycle area | NIST AI RMF and Playbook: voluntary best-practice governance | EU AI Act: binding legal obligation when applicable |
|---|---|---|
| Govern | Set policies, roles, accountability, communication lines, training, risk tolerance, and oversight, and connect AI governance to existing organizational and data governance.[11][12] | The Act uses a risk-based framework and assigns obligations according to system risk and actor role in the value chain.[13] |
| Map | Define intended purpose and context of use, document limits, and map impacts and third-party, legal, and rights-related risks.[14][15][16] | For high-risk systems, providers must give clear, adequate information for deployers and sufficient documentation for authorities to assess compliance.[17][18] |
| Measure | Select metrics and methods; run TEVV and red-teaming; test validity, safety, security, resilience, fairness, privacy, bias, explainability, and environmental impact.[19][20][21] | The source does not set a broad standalone test regime, but high-risk obligations include high-quality datasets, logging, and documentation that support compliance and traceability.[22][23] |
| Manage | Prioritize risks, choose mitigation, transfer, avoidance, or acceptance, plan incident response and recovery, monitor deployed systems, and deactivate or retrain systems when needed.[24][25][26] | Providers and deployers must report serious incidents and malfunctioning; providers of high-risk systems must maintain post-market monitoring, and deployers must ensure human oversight and monitoring.[27][28][29][30] |
| Documentation | Keep inventories, system cards, model cards, logs, change histories, benchmark comparisons, monitoring results, incidents, residual risks, and transparency artifacts.[31][32][33] | High-risk systems must have detailed documentation that lets authorities assess compliance and understand the system and its purpose.[34][35] |
| Testing | Use pre-deployment testing, internal and external evaluations, structured public feedback, adversarial testing, and review of guardrails and safety features.[36][37][38] | Testing is not described as a separate universal duty in the source summary, beyond the pre-market quality, logging, and documentation obligations for high-risk systems.[39] |
| Monitoring | Continuously monitor risk over time and track changes, incidents, and residual risks.[40][41] | Post-market monitoring is required for providers of high-risk systems, and deployers must monitor systems once on the market.[42][43] |
| Supplier management | Vet suppliers, maintain third-party inventories, use contracts and SLAs, require transparency and auditability, and plan for third-party failures or redundancies.[44][45][46] | The Act’s obligations can depend on role in the value chain, and the Commission’s guidance is meant to help actors understand who must comply with GPAI obligations.[47][48][49] |
| Decommissioning | Plan safe shutdown, phase-out, or retirement, manage dependencies and leakage risks, and avoid increasing risk during migration or decommissioning.[50][51][52] | The source summary does not frame decommissioning as a standalone legal duty, but it does require post-market monitoring and incident handling while systems remain in use.[53][54][55] |
What the EU AI Act adds beyond the NIST lifecycle model
- Risk-tiered legal triggers: the Act defines four risk levels and bans certain AI practices outright because they are considered a clear threat to safety, livelihoods, and rights.[56][57]
- High-risk pre-market controls: high-risk systems face strict obligations before market placement, including high-quality datasets, logging, detailed documentation, and clear information for deployers.[58][59]
- Role-specific duties: responsibilities differ for providers, deployers, and other value-chain actors, so compliance is not only about the system but also about the actor’s role.[60][61][62]
- Generative AI and GPAI: providers must make AI-generated content identifiable, label certain content such as deep fakes and public-interest text, and for GPAI models provide transparency and copyright-related compliance, plus risk assessment and mitigation for systemic-risk models.[63][64][65][66]
- Human oversight is necessary but not sufficient: the Act pairs human oversight with monitoring, post-market monitoring, incident reporting, and documentation rather than treating oversight alone as enough.[67][68][69][70]
In practice, the governance stack is strongest when NIST-style lifecycle controls are used to operationalize the EU’s binding duties: pre-market testing and documentation support high-risk compliance, supplier controls support value-chain accountability, and monitoring and incident reporting support post-market obligations.[71][72][73][74][75]
Bottom line
NIST says good AI governance should cover the whole lifecycle, including govern, map, measure, manage, documentation, testing, monitoring, suppliers, and decommissioning, but these are guidance controls, not law.[76][77][78] The EU AI Act turns several of those same lifecycle controls into enforceable obligations only for specific risk categories and actor roles, especially high-risk systems, GPAI providers, and certain generative AI uses.[79][80][81]
Examinons les alternatives :
- Modifier la requête.
- Démarrer une nouvelle conversation.
- Supprimer des sources (si elles ont été ajoutées manuellement).
Why trial only reviews can miss harms
Transcription
When systematic reviews of semaglutide or tirzepatide rely on randomized trials alone, they can miss important harms. The trials may include too few people to detect rare adverse events, and they may end before delayed harms appear. Published trial reports can also leave out a lot of adverse event information, so the safety picture can look cleaner than it really is. Reviews should therefore go beyond trial reports and bring in post marketing surveillance, spontaneous reporting systems, epidemiologic studies, and unpublished trial data when available. Harms, eligible study designs, relevant patient risk factors, and the length of follow up should be set in advance. Review authors should also report how much adverse event data were unavailable, and downgrade certainty when substantial unpublished data cannot be accessed.
Examinons les alternatives :
- Modifier la requête.
- Démarrer une nouvelle conversation.
- Supprimer des sources (si elles ont été ajoutées manuellement).
Créez votre compte pour façonner ce Feed autour de ce que vous cherchez et aimez.
Can evaluation be both independent and useful? WHO and OECD say yes
Examinons les alternatives :
- Modifier la requête.
- Démarrer une nouvelle conversation.
- Supprimer des sources (si elles ont été ajoutées manuellement).
Five operational threat-hunting leads from recent CISA advisories
Examinons les alternatives :
- Modifier la requête.
- Démarrer une nouvelle conversation.
- Supprimer des sources (si elles ont été ajoutées manuellement).
HubSpot 10-K pricing architecture lessons, not a current price sheet
Examinons les alternatives :
- Modifier la requête.
- Démarrer une nouvelle conversation.
- Supprimer des sources (si elles ont été ajoutées manuellement).
Satellite latency record figures
Examinons les alternatives :
- Modifier la requête.
- Démarrer une nouvelle conversation.
- Supprimer des sources (si elles ont été ajoutées manuellement).