Kerry Champion

Human abstraction ability applies not just to language but across all of the subjects we reason about.

AI won’t reach its potential till we learn to blend symbolic and causal capabilities with the statistical pattern matching that powers today’s LLMs.

#AI #NeurosymbolicAI #CausalAI

Instead of relying only on patterns in input, humans:

+ Form internal, rule-based models of language structure (e.g. grammar, syntax).

+ Infer underlying rules even when they’re not explicitly taught.

+ Use these abstractions to generalize beyond what they’ve directly heard.

2/n

Table comparing human and LLM language learning patterns

LLMs are missing “theoretical abstraction” capability we see in children.

Multiple folks pointed this out, for example @teppofelin.bsky.social and Holweg in “Theory Is All You Need: AI, Human Cognition, and Causal Reasoning”. papers.ssrn.com/sol3/papers....

#AI #CausalAI #SymbolicAI

1/n

Cleaning Up Policy Sludge: An AI Statutory Research System | Stanford HAI This brief introduces a novel AI tool that performs statutory surveys to help governments—such as the San Francisco City Attorney Office—identify policy sludge and accelerate legal reform.

Legal reform can get bogged down by outdated or cumbersome regulations. Our latest brief with Stanford RegLab scholars presents an AI tool that helps governments—such as the San Francisco City Attorney's Office—identify and eliminate such “policy sludge.” hai.stanford.edu/policy/clean...

Enjoying the graphic.

On your list of ways to address these concerns, where would you put implementation neurosymbolic AI?

Seems to me that combining deep learning (LLMs) with symbolic/causal models could go a long way to creating more reliable, auditable, and aligned AI.

#AI

@jwmason.bsky.social The other thing worth knowing is that bigger LLMs are not the only path forward for AI. Combining LLMs with symbolic/causal models has the promise of creating hybrid AI systems that are much more reliable in reflecting the world as it is.

#AI #SymbolicAI #CausalAI

LLMs form a semi-accurate representation of the world as it is reflected in the writing they train on. A next step would be to create hybrid AIs that combine LLMs with symbolic and causal models that have explicit (and more accurate/auditable) representations of the world #AI #SymbolicAI #CausalAI

Agentic AI: Be Careful What You Wish For 🧞 Wishes have consequences. Especially when they run in production.

For more from Cassie ‪@decisionleader.bsky.social‬ :
decision.substack.com/p/agentic-ai...

The opportunity here is for us to perfect hybrid systems that integrate deep learning with symbolic reasoning and causal understanding. This will reduce our dependence on filtering bad consequences, by having models that are inherently more reliable.

#AI #CausalAI #SymbolicAI

Progress on the "control layer" feels far behind our breakthroughs with the "genie".

Having the control layer be a smart filter on the input and output is helpful but in the end seems fundamentally wrong headed.

2/n

“Wishes have consequences. Especially when they run in production.” - ain't that a fact!

Cassie Kozyrkov's genie metaphor rings true:

1/n

First, through a think-aloud study (N=16) in which participants use ChatGPT to answer objective questions, we identify 3 features of LLM responses that shape users' reliance: #explanations (supporting details for answers), #inconsistencies in explanations, and #sources.

2/7

‪@rohanpaul.bsky.social‬ this post is feeling lonely ;-)

Why not cross post on both X and Bluesky?

Table contrasting symbolic reasoning and causal models as next steps in AI evolution.

#SymbolicAI and #CausalAI companions in search for next #AI breakthrough

@jessicahullman.bsky.social‬ persuasively argues that current AI is poor tool for decisions that fit FIRE (forward-looking, individual/idiosyncratic, require reasoning or experimentation/intervention) profile

Hmm … does this call for #CausalAI

statmodeling.stat.columbia.edu/2025/06/05/w...

For example, Apple’s approach to callbacks to app code from the model as it reasons and their support for multi-layered guardrails illustrates what they have learned about needing components with use case specific checks and balances.

#PervasiveAI #AgenticAI #AppleAI #AISafety
2/n

Apple opening on device LLM take-aways:

1) Medium term “Pervasive AI” will have more reach than “Agentic AI”

2) AI best implemented through systems of components and not a single blackbox neural net

3) Use case specific adjustments is needed to balance latency, cost, reliability and safety
1/n