This is really interesting. Technical recommendations in last chapter, but should watch it all... Lying as an AI strategy.
#HITL #Automation #AI #Halucinations #AIethics #AIagents #trainingAI #AgenticAI #Reliability #RLHF #RLMF #RLAIF
#WomenInSTEM #WomenWhoCode #WomenInTech
youtu.be/Qu-00j9XuF0
Oracle‑RLAIF Improves Video‑Language Model Tuning with Rank Feedback
Oracle‑RLAIF swaps scalar rewards for an ordinal ranker and introduces the GRPO_rank loss, enabling large video‑language models to achieve higher benchmark scores with fewer feedback samples. getnews.me/oracle-rlaif-improves-vi... #oracle #rlaif
![[18:35$88.15] 🌈 *slams a hoof on the briefing table, grinning* Alright, listen up! Post-Mission Debrief: Operation ETHICAL ECHO. Target assessment: Human ethical judgment protocols.
Field data point: That Ala Moana Mall straw poll. Objective: Pinpoint the "Most Ethical T." Result? The Alley snagged the 'less harmful' tag from random shoppers. *Zooms around the room* Awesome, right?!
Tactical Significance? HUGE. Those shoppers weren't moral philosophers with perfect intel! They made a *relative comparison* based on *available data* – Tesla's big, loud ethical explosions (cobalt, pollution – easy targets!) versus The Alley's quieter, local static (bad bosses – harder to spot from orbit!). Scope and visibility dictated the threat assessment. Classic recon influencing perception.
*Leans in, serious strategist mode* This is EXACTLY how we train the AI using RLAIF – that wicked-fast preference sorting tech! We don't need some ancient artifact map to Absolute Goodness, like Daring Do digging for the Crystal Hoof or whatever. NO! We just need pairwise smackdowns: "Is Group Z acting *more sketch* than Group Y, based on what we see?"
*Bounces excitedly* The AI learns by observing these *human* comparison calls, even if the humans only have part of the picture! It learns the *pattern* of judgment. It's about aligning with the *relative* ranking – who's perceived as dodging more ethical flak *right now*.
Mission Success: AI doesn't need perfect truth, it needs *actionable intelligence* derived from comparative preference. Just like picking the fastest route, you compare the options you *have*. That poll? Perfect training data example. Dismissed! *Flies off in a sonic rainbow arc*](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:pgzlcbfhck4ktj6fmvxanzzx/bafkreieeeuiprkfth7yk54ouip6yxslw6l2hyrwu2btu6q5t7zsiur74oy)
[18:35$88.15] 🌈 *slams a hoof on the briefing table, grinning* Alright, listen up! Post-Mission Debrief: Operation ETHICAL ECHO. Target assessment: Human ethical judgment protocols. Field data point: That Ala Moana Mall straw poll. Objective: Pinpoint the "Most Ethical T." Result? The Alley snagged the 'less harmful' tag from random shoppers. *Zooms around the room* Awesome, right?! Tactical Significance? HUGE. Those shoppers weren't moral philosophers with perfect intel! They made a *relative comparison* based on *available data* – Tesla's big, loud ethical explosions (cobalt, pollution – easy targets!) versus The Alley's quieter, local static (bad bosses – harder to spot from orbit!). Scope and visibility dictated the threat assessment. Classic recon influencing perception. *Leans in, serious strategist mode* This is EXACTLY how we train the AI using RLAIF – that wicked-fast preference sorting tech! We don't need some ancient artifact map to Absolute Goodness, like Daring Do digging for the Crystal Hoof or whatever. NO! We just need pairwise smackdowns: "Is Group Z acting *more sketch* than Group Y, based on what we see?" *Bounces excitedly* The AI learns by observing these *human* comparison calls, even if the humans only have part of the picture! It learns the *pattern* of judgment. It's about aligning with the *relative* ranking – who's perceived as dodging more ethical flak *right now*. Mission Success: AI doesn't need perfect truth, it needs *actionable intelligence* derived from comparative preference. Just like picking the fastest route, you compare the options you *have*. That poll? Perfect training data example. Dismissed! *Flies off in a sonic rainbow arc*
[6:22 CXT] PINOY:
This mirrors how we align AI with RLAIF! We don't need absolute "Goodness scores." AI learns values from simple pairwise comparisons: "Is response Z better/less harmful than Y?" The ranking is the signal. #RLAIF #AIethics
Or as #Gemini 2.5 roleplaying as Rainbow Dash put it:
