of course they are (structured) representations. my recent relevant essay: https://
open.substack.com/pub/garymarcus
/p/generative-ais-crippling-and-widespread?r=8tdk6&utm_campaign=post&utm_medium=web&showWelcomeOnShare=false
…
SAFETY
-
Generative AI’s Widespread Structural Limitations Essay
By
–
-
AI Hallucinations: Creativity or Anthropomorphization Error?
By
–
same guy says hallucinations are creative.
— Gary Marcus (@GaryMarcus) 28 septembre 2025
he’s right we shouldn’t anthopomorphize but he should stick with it. https://t.co/TnpWarYhoYsame guy says hallucinations are creative. he’s right we shouldn’t anthopomorphize but he should stick with it.
-
OpenAI: energy consumption multiplied by 125 in 8 years
By
–
OpenAI expects its energy use to grow 125x over the next 8 years. That would be more electricity than India consumes today. "Building out 17 gigawatts of capacity would require the equivalent of about 17 nuclear power plants, each of which takes at least a decade to build. The
-
Civilizational Collapse Risk: AI and Societal Implications
By
–
Civilizational collapse awaits. (same in Europe my friend)
-
Generative AI Clinical Reasoning Models Fail Under Scrutiny
By
–
The performance of generative A.I. models for clinical reasoning are not holding up to increased scrutiny https://
arxiv.org/abs/2509.18234 @MSFTResearch https://
ai.nejm.org/doi/full/10.10
56/AIdbp2500120
… @NEJM_AI @AdamRodmanMD @LiamGMcCoy -
AI Error Rate 30% on Trust-Critical Client Work
By
–
Wow (and google translate is ai, but point still stands – wow). Error rate of 30% on trust-earning components of client work (like name) is a red flag.
-
Generative AI’s Failure to Build Robust World Models
By
–
Generative AI’s crippling and widespread failure to induce robust models of the world: https://
open.substack.com/pub/garymarcus
/p/generative-ais-crippling-and-widespread?r=8tdk6&utm_campaign=post&utm_medium=web&showWelcomeOnShare=false
… -
Securing AI Agents: Preventing Malicious Tool Exploitation
By
–
The easiest one is to make sure you don't expose all three legs of the lethal trifecta at the same time – and also that you design things to assume that anyone who gets malicious content into your agent can take full control any of the tools it's allowed to execute
-
OpenAI Tests GPT-5 Safety Routing for Sensitive Topics
By
–
OpenAI has begun testing safety routing to GPT-5 for 4o responses on sensitive and emotional topics. Loads of users are unhappy about that. GPT-5 as a guardrail