2023-06-06 at IAS
- Alondra Nelson
- Miranda Bogen
- Brian Christian
- Sorelle Friedler
- William Isaac
Nelson: WH OSTP; AI Bill of Rights
OSTP creation 1976 in 1976 by legislation mentioning "mitigate risk"
Nelson's intro includes lots of mention of "alignment"
-
"Thick Alignment" a la Geertz
-
panelists were at IAS for a working group on AI policy related to AI Bill of Rights
risks include:
- destabilization of worker experience
- bias/disc.
- security
- mis/dis info
- catastrophic outcomes (extinction risk)
Bogen -- Meta, partnership on AI
Christian -- "The Alignment Problem" book
Friedler - Haverford CS, OSTP, FAccT, Google X (and Swarthmore)
Isaac -- HR DAG, Deepmind
Bogen - expand notion of "user"
- 'future generations" "extremely long-term
Christian - RLHF, values from corpus, human raters (low-wage, but perhaps enfrachised in a way?)
Friedler- metrics, notions of difference / inclusion (demo. groups)
- non-LLM models, less attention but probably greater impact today
Isaac - stakeholders, public access - APIs for evaluation\
- independent eval.
-
Isaac - displacement, training data (compensation for artists, etc.)
-
augmentvs.displace
Friedler - how an input/correction from humans gets used
- e. g. "self-driving" cars that need drivers - can these jobs be worthwhile?\
- we have agency (writer's strike)
Christian - hiring will be changed
Bogen - system cards: whole system, not just isolated models
Bogen - fight for rights and liberties is a process, aiming for progress, never finished
How are these rights/fights manifest in the AI context?
- conversations about AI are a reminder of values we want to re-visit, protect
- leave room for improvement vs. "lock in"
Friedler - need new tools (not just technical) rather than new rights
- transparency mechanisms
shift from blockchain -> LLMs
public goods?
Friedler - convo crypto on energy use was encouraging, actually
- AI engergy use similar in scale? less transparency
Christian - climatet and AI risks are very similar
-
alignment → numerical objective == externalities
-
"climate is an alignment problem"
-
AI objective is capitalism, given incentives of the orgs that train them
- hard to think about values at societal level
Christian --
one heartenin gstory of AI progress: CV went from carefully hand-crafted features and training data to models that can learn everything from just images
-
RLHF is similar? learn values from positive/negative feedback without needing to articulate them explicitly. It seems to work!
-
Challenge: heterogeneous values. Current methods assume there's some average set of values that can be discovered and heterogeneity is just noise.
-
All the AI alignment folks he knows are now reading Poli Sci to try to understand collective decision-making.
Isaac:
- thinksofCVworkingwellinWesternhouseholds(to recongnize chair, table, etc), but not in non-Western contexts
at Deepmind:
-
rule-based
-
transparency about RLHF
encouraing progress? machine translation, working better and better, even for low-resource languages. works well now!
Friedler
- lots of good use cases for AI
- increasing public recognition that it's not magic
- e.g. help Dr.s, not replace
- human in loop, guidance
Bogen
- can make things faster, easier, less expensive
- does that make more room for important things?
- embed values in key systems that get widely used -> less need for many separate enforcement actions?
policy steps:
-
not one right thing
-
take actions today to raise the floor, move on to harder things
-
Christian - alignment will be messy, not elegant CS solutions
-
layers of law, agencies, etc. Bogen - lots of open Qs
-
don't let that stop us
-
act and learn more - over decades
Friedler - sector-specific approach - e.g. HHS, Dept. of Labor - domain experts, technical support
Isaac - many conversations mash together the tech and the application
- generic eval vs. application-specific eval
- encourage disclosure of evaluation data / metrics, transparency about performance
- invest in research! build capacity to understand not just CS, need social science etc.
Bogen: research suggests education not automatically effective
Friedler - ed. is not enough. need rules, not just personal responsibility/education
Christian- shows overton window has shifted
Isaac - large firms vs. community efforts: this needs to be taken into account in thinking about regs
Nelson - wonders if calls for facial rec. moratoria informed calls for AI pause?
e.g.
-
trolley prob. vs. impact of self-driving cars at scale on traffic patterns
-
Social media effects over time what infra do we need to deal w that?
Isaac - institution building
-
collective decision-making
-
governance /government
-
provide space for dynamic responce
Friedler - systems are monitored already\
- intervention point:include variety of value-based metrics
Bogen
-
metrics are value-laden
-
don't subsume that in technical decisions made by isolated engineers, need broader conversation
Isaac - they store, but can do more?
-
Still working out limits
-
not magic, but range is impressive
Bogen - AI is a variety of things
- remember there's more than one technology