Giving AI Some Common Sense
Schmidt Sciences
Ron Brachman
Northeastern University
April 2024
(with Hector Levesque)You’re running an errand
● You go the same way almost every time …and you arrive at a T intersection
● Taking a route you’ve taken 100’s of times
Normally you’d turn left, but…
● It’s 2:30PM and there’s an elementary school to the left
● You notice a beautiful tree on the street to your right
● You look left and see a moving van and a guy waving his arms
● The ground rumbles and you see to the left the asphalt buckling up
…so you turn right
You had a reason to turn right
● You thought of a relevant reason quickly, without much effort
○ That reason was a direct cause of your action
○ The pure frequency of prior turns didn’t matter
○ Wasn’t a skill, like smooth braking, lane following, avoiding curbs
● Your reason was commonsensical
○ Based on obvious, simple things, not deep thought or calculation
○ Involved a quick mental forward projection
Today’s AI doesn’t have articulable reasons
● Trained on vast volumes of text (sometimes images), not principles, explanations, or richly structured situations
○ Statistics drives deep learning
○ Good on unconscious skills, like braking, lane-following
● Even the systems’ designers don’t know why they do what they do
“…why is it doing that? And I would say not only do I not know, but no one knows because the people who designed it don’t know, the people who trained it don’t know.”
- Holden Karnofsky…and it doesn’t have common sense
Yejin Choi 2023 TED Talk:
“Why AI is Incredibly Smart and Shockingly Stupid”
● Fraught with inexplicable, sometimes bizarre decisions
AI without common sense
Self-driving cars
Source: https://jalopnik.com/this-billboard-that-confuses-tesla-autopilot-is-a-good-1846698527
https://www.kron4.com/news/bay-area/waymo-cruise-vehicles-have-impeded-emergency-vehicle-response-66-times-this-year-sffd/
AI without common sense
Image recognition
Source: arXiv: 1312.67199
https://www.independent.co.uk/tech/artificial-intelligence-adversarial-examplesoptical-illusion-ai-tricked-mit-labsix-turtle-rifle-guacamole-cat-a8034241.html
“ostrich” “ostrich” “ostrich” “ostrich” “ostrich” “ostrich”AI without common sense
Language-based systems
Ten-year-old: I’m bored. Please give me a challenge to do.
Alexa: Plug in a phone charger about halfway into a wall outlet, then touch a penny to the exposed prongs.
Source: https://www.cnbc.com/2021/12/29/amazons-alexa-told-a-child-to-do-a-potentially-lethal-challenge.html
https://gizmodo.com/chatgpt-gone-berserk-giving-nonsensical-responses-1851273889
Why does it matter?
● If technology is just for entertainment or nothing rides on the results, occasional failures are not an issue
● But when system is autonomous, failures can be fatal
● Unpredictability of failure undercuts trust
Trust is built on reliable behavior
“It’s right some of the time, but you don’t know which part of the time…”
- Gary Marcus (NYT, 4/11/2024)● Human decisions generally make sense, even when they’re wrong
○ There is generally a reason for a breakdown
○ Our behavior is sensible, even while not perfect
We have a critical gap to fill in AI
So, what is common sense?
● There are lot of thoughts over centuries
○ But usually not constructive – don’t help us build better AI systems
“Common sense is genius dressed in its working clothes”
– Ralph Waldo Emerson
“Common-sense appears to be only another name for the thoughtlessness of the unthinking.”
– W. Somerset Maugham
“…the dark matter of intelligence… the unspoken, implicit knowledge that you and I have.”
– Yejin ChoiThe term in common use
The term in common use
Common insight: things to know about <x>, not to become an expert, but to guide your actions in a sensible way in ordinary situations.
Characteristics of common sense
● Broadly known (“everybody knows that…”)
● About mundane things
● “Obvious”
● Simplicity/Occam’s Razor
● Based on what you know, experience
“common”
“…judgement to discover plain truths and palpable contradictions.”
– Baron d’Holbach
“…the collection of prejudices acquired by age eighteen.”
– Albert Einstein
● Quick, shallow, qualitative — but involves thinking; explicable
● Reasonable and appropriate conclusions
● Practical
“sense”
“Me is a common sense man. That mean when me explain things, me explain them in a very simple way; that mean if I explain it to a baby, the baby will understand too, you know.”
– Bob MarleyWhat type of knowledge is commonsensical?
● Commonsense knowledge usually contrasted with expert knowledge
“When you’re dead, you stay dead”
“Don’t put your hand over a burner to test whether it’s on”
“Leave enough travel time to get to your destination before your deadline”
“Two particles at a long distance can be quantumly entangled”
“Laminar flow has much lower skin friction drag compared to turbulent flow.
“CYP450 enzymes polymorphisms play a crucial role in the metabolism of drugs by catalyzing the oxidation of organic substances”
AI’s approach to common sense
● Prevalent view: common sense is mainly lots of accumulated facts
● Very large numbers of obvious rules; missing from expert systems
○ Plus: “commonsense reasoning” for inferences that are more qualitative, plausible than classical purely logical reasoning
● Cyc project (started in 1984)
○ Manually built, formally structured knowledge
○ 40,000+ predicates, 1.5M+ concepts, 25M+ axioms
● ChatGPT and other LLM’s
○ Trained from data, high-dimensional statical knowledge base
○ GPT-4: over 500B parameters – trained on trillions of documents
These efforts share a goal of breadth
Large KB’s and LLM’s are transactional
● Focus on large knowledge repositories
○ Also: inference — go beyond the mere storage of facts
○ Cyc, e.g., has many, many inference methods
● But systems by and large work as “fact calculators”
○ Enter a couple of facts and push the “INFER” button
Fact calculator à la Cyc
Impressive inferences…
…but local/transactional
Source: https://www.forbes.com/sites/cognitiveworld/2019/07/03/what-ai-can-learn-from-romeo--juliet/?sh=671a42a21bd0 © CYCORP 2019
Fact calculator à la ChatGPT
Impressive text prediction…
…but ChatGPT fabricates …and neither system does anything*
* but see below
Knowing vs. using
Making effective use
● An agent said to have common sense must go far beyond simply knowing lots of commonsense facts and rules
○ It’s about appropriate and timely use
● Having common sense =
the ability to make effective use of ordinary, everyday experiential knowledge in achieving ordinary, practical goals
○ Sternberg, et al.: Common sense = practical intelligence
“Genius dressed in its working clothes”
– Ralph Waldo Emerson● Knowledge needs to be put to work
○ Knowing that drunk driving is dangerous and illegal
○ Contradictory common sense: “Haste makes waste” vs. “He who hesitates is lost”
○ Overriding reptile brain urges: dropped cell phone on subway tracks
● Adjudication based on situations and goals vs. knowing lots of commonsense facts and principles
○ ChatGPT: “The key is to find the appropriate timing and approach for each situation, recognizing when to act swiftly and when to proceed with caution.”
Working clothes, cont’d.
A funny thing happened on the way to the grocery store…
What do I do now?
us stuck red light
Assume this is the first time this has happened to us -> no contingency plan
Music from afar?
stuck red light us
other grocery stores home us Alternate destinations?
Can I safely cross the intersection anyway?
Are there other stores?
Can I make a U-turn?
Legally? Safely?
How about a right/U/left maneuver?
How badly do I need the item? Can I just do this later?
other grocery stores home us
Diving deeper
What is common sense for and where does it come into play? analysis
● Rapid decision-making in everyday situations, short-circuiting the need for resource-intensive analysis/reflection
common sense
○ Make use of common experiences for quick, qualitative decisions
● Fast, but cognitively penetrable*
○ Not just a reflex (but not deep analysis)
routine behavior
● The importance of expectations and goals/desires
○ What’s normal/anomalous, what I’m trying to do
*The ability to influence beliefs (and therefore decisions and actions) via introduced facts/beliefs/observations
Common sense from the bottom up
Constant situation awareness
● Nonstop situation recognition: automatic, always on
○ Constant recognition of objects and relationships independent of current goals
● Expectations driven from remembered prototypical patterns
○ From the 70s: frames and scripts
● Rapid detection of unexpected/anomalous situations
○ Violation of expectation can jolt overall system out of routine behavior
○ Noticing when you’ve just done something silly/risky
Common sense from the top down
Mundane plans are based on common sense
● Everyday plans are built from commonsensical steps
○ Based on prior remembered sequences, intuitive sense of causality
○ Rapid forward mental simulation, quick cost/benefit/risk analysis
● Pre-structured everyday plans can still be richly textured
○ Not just linear sequences but branches and loops
○ Able to adjust to small variations discovered at runtime
● Invoke more thoughtful planning for major anomalies
General situation awareness and plan-specific expectations combine as bottom-up supports execution of top-down plans
Putting it all together
The bigger picture demands consideration of overall architecture
anomalies
Prototype plan library
Knowledge Manager
World Model plans/expectations
frames/expectations
anomalies
Fundamental needs Desires achieved goals top-down bottom-up
Plan Executor
Watcher
Prototype situation library
Baseline commonsense knowledge
Common sense as System 2a
Common sense is a System 1 -like version of System 2
● Rapidity and simplicity as in System 1
○ Not deeply analytical
● But cognitively penetrable as in System 2
○ Alternative actions are considered, relative costs, explainable decisions
○ Leans on common sense knowledge
● Sys 1/2 distinction not detailed enough
System 1
● Kahneman is silent about common sense
System 2
Intelligible AI and explainability
Common sense as the foundation
● Autonomous systems will need to be responsible for their actions
● Like people, AI systems should have good reasons for what they do
○ They should be understandable, even if we disagree with them
● Philosopher Daniel Dennett’s intentional stance will be important:
○ Agent x did A because it believed that P and wanted G
If the behavior of a sufficiently complex AI system is going to be understandable at all, it will be in intentional terms
Common sense is likely the best foundation for intelligible intentions
LLM’s have ”no human interpretability”
– Ben Levinstein
Being open to advice
Cognitive penetrability of commonsensical behavior
● AI must be capable of change when we disagree with its reasons
○ If we see P is false and alternative Q is true, we should be able to explain, and the system should adopt and adjust to Q
● Architecture should be such that it is possible to isolate a mistaken belief or goal, and change it
○ Generally favors symbolic knowledge* within the system
● Beliefs and goals should actually govern the behavior of the system
○ So a bolted-on, after-the-fact explanation module is not enough
* nuances and room for discussion here…
Being open to advice
Cognitive penetrability of commonsensical behavior
● AI must be capable of change when we disagree with its reasons
○ If we see P is false, we should be able to explain, and that alternative relevant Q is true
● Architecture should be such that it is possible to isolate a mistaken belief or goal, and change it
○ Generally favors symbolic knowledge within the system
● Beliefs and goals should actually govern the behavior of the system
○ So a bolted-on, after-the-fact explanation module is not enough
In conclusion
AI can’t ignore common sense
● Acting commonsensically overall is more than knowing (even millions of) commonsense things
● Commonsense beliefs and goals support explainability and instructability
○ Should directly influence behavior
○ Autonomous systems should not be unleashed without common sense
● Augment our research on KB’s and LLM’s to deal with the use of common sense in everyday situations
○ Cognitive system architecture is important
Let’s have Cyc, GPT, Gemini, at al., put on their working clothes
Urgency
We seem to be crossing the Rubicon now…
Further reading
TowardaNewScienceofCommonSense*
RonaldJ.Brachman,1 HectorJ.Levesque21 JacobsTechnion-CornellInstituteandCornellUniversity 2 Dept.ofComputerScience,UniversityofToronto ron.brachman@cornell.edu,hector@cs.toronto.edu
Abstract
CommonsensehasalwaysbeenofinterestinAI,buthas rarelytakencenterstage.DespiteitsmentioninoneofJohn McCarthy’searliestpapersandyearsofworkbydedicated researchers,arguablynoAIsystemwithaseriousamount ofgeneralcommonsensehaseveremerged.Whyisthat? What’smissing?ExamplesofAIsystems’failuresofcommonsenseabound,andtheypointtoAI’sfrequentfocuson expertiseasthecause.Thoseattemptingtobreaktheresultingbrittlenessbarrier,eveninthecontextofmoderndeep learning,havetendedtoinvesttheirenergyinlargenumbers ofsmallbitsofcommonsenseknowledge.Whileimportant, allthecommonsenseknowledgefragmentsintheworlddon’t adduptoasystemthatactuallydemonstratescommonsense inahuman-likeway.Weadvocateexaminingcommonsense fromabroaderperspectivethaninthepast.Commonsense shouldbeconsideredinthecontextofafullcognitivesystem
wouldnormallycall commonsense.Butrecentcallsfora newgenerationofpost-modernAIsystemswithcommon sensegivenorealprescriptionforgettingthereoraclear ideaofwhatitwouldreallymeanforanAIsystemtohave it.Ourintentioninthispaperistostimulatethefieldinto closingthiscriticalgap.
ExpertiseandtheBrittlenessChallenge
Despitethenameofthefield,ArtificialIntelligence’sbiggest accomplishmentshavegenerallycomefromexpertiserather thananymoregeneralkindofintelligence.Ourgreatestsuccesseshavebeenontasksinnarrowdomainsorcircumscribedchallengeproblems,suchasGo,facialrecognition, infectiousdiseasediagnosis,andthelike.Themostobviouslimitofthisiswhatsomehavecalled“brittleness”—
thefailuretoproducereasonableoutcomes(or,inmany