Greg Robison
@gregrobison.bsky.social
Chief AI Architect, CTO and Founder at F'inn (https://www.linkedin.com/in/gregrobison), writer (https://gregrobison.medium.com/) and photographer (https://www.gregrobison.com)
Google Dream-RSI dreams by replaying history as a simulator and doesn't affect the model's weights. My latest post on "recursive self-improvement" through dreaming: gregrobison.medium.com/dream-rsi-ho...
Dream-RSI: How Self-Improving AI Is Learning to Dream About Its Own Past
“Life can only be understood backwards; but it must be lived forwards.”
gregrobison.medium.com
October 6, 2026 at 1:10 PM
Google Dream-RSI dreams by replaying history as a simulator and doesn't affect the model's weights. My latest post on "recursive self-improvement" through dreaming: gregrobison.medium.com/dream-rsi-ho...
We hoped AI would think more clearly than we do. But it learned our confirmation bias, anchoring, and overconfidence from our own writing. It has new biases we have no defense against. Here's how human and LLM bias overlap and differ and 8 things to watch for.
gregrobison.medium.com/trained-on-u...
gregrobison.medium.com/trained-on-u...
Trained on Us: The Cognitive Biases LLMs Absorbed and Invented
“These heuristics are highly economical and usually effective, but they lead to systematic and predictable errors.”
gregrobison.medium.com
September 10, 2026 at 1:15 PM
We hoped AI would think more clearly than we do. But it learned our confirmation bias, anchoring, and overconfidence from our own writing. It has new biases we have no defense against. Here's how human and LLM bias overlap and differ and 8 things to watch for.
gregrobison.medium.com/trained-on-u...
gregrobison.medium.com/trained-on-u...
Researchers found a hidden stream of "thoughts" running inside an AI, words it never says out loud. They swapped one out, and its answer changed to match.
Spooky. But it's not evidence the lights are on inside. Here's the argument: gregrobison.medium.com/reading-the-...
Spooky. But it's not evidence the lights are on inside. Here's the argument: gregrobison.medium.com/reading-the-...
Reading the Machine’s Thoughts: How J-Space Compares to Human Consciousness
“Phenomenal consciousness is experience; the phenomenally conscious aspect of a state is what it is like to be in that state. The mark of…
gregrobison.medium.com
August 11, 2026 at 12:36 PM
Researchers found a hidden stream of "thoughts" running inside an AI, words it never says out loud. They swapped one out, and its answer changed to match.
Spooky. But it's not evidence the lights are on inside. Here's the argument: gregrobison.medium.com/reading-the-...
Spooky. But it's not evidence the lights are on inside. Here's the argument: gregrobison.medium.com/reading-the-...
A new study cracked the dad joke formula:
The bigger the leap between setup and punchline, the funnier
Concrete > abstract
NEVER end on a question
Oh, and dads genuinely find dad jokes funnier. The data agrees.
gregrobison.medium.com/corny-by-des...
The bigger the leap between setup and punchline, the funnier
Concrete > abstract
NEVER end on a question
Oh, and dads genuinely find dad jokes funnier. The data agrees.
gregrobison.medium.com/corny-by-des...
Corny by Design: Why Dad Jokes Make You Wince
“It might be best to think of the dad joke not as a kind of joke but as a kind of performance, one that involves a teller — the dad — and…
gregrobison.medium.com
July 7, 2026 at 1:22 PM
A new study cracked the dad joke formula:
The bigger the leap between setup and punchline, the funnier
Concrete > abstract
NEVER end on a question
Oh, and dads genuinely find dad jokes funnier. The data agrees.
gregrobison.medium.com/corny-by-des...
The bigger the leap between setup and punchline, the funnier
Concrete > abstract
NEVER end on a question
Oh, and dads genuinely find dad jokes funnier. The data agrees.
gregrobison.medium.com/corny-by-des...
@nousresearch.com Hermes AI agent has become my second brain. It starts learning about me through skills it writes itself. It's become infrastructure. After months of use, here's what “self-improving” looks like when you can read the files: gregrobison.medium.com/hermes-agent...
Hermes Agent: What Self-Improving Means When You Can Read the Files
“You have to begin to lose your memory, if only in bits and pieces, to realize that memory is what makes our lives. Life without memory is…
gregrobison.medium.com
June 9, 2026 at 1:00 PM
@nousresearch.com Hermes AI agent has become my second brain. It starts learning about me through skills it writes itself. It's become infrastructure. After months of use, here's what “self-improving” looks like when you can read the files: gregrobison.medium.com/hermes-agent...
We've built AI systems that outperform humans at almost everything, but we have no idea why they work. A new paper argues that trial and error is finally starting to change and the shape of what's coming looks a lot like physics. Read more: gregrobison.medium.com/learning-mec...
Learning Mechanics: The Scientific Theory That Could Explain Deep Learning
“What I cannot create, I do not understand.”
gregrobison.medium.com
May 26, 2026 at 12:45 PM
We've built AI systems that outperform humans at almost everything, but we have no idea why they work. A new paper argues that trial and error is finally starting to change and the shape of what's coming looks a lot like physics. Read more: gregrobison.medium.com/learning-mec...
Two-thirds of us use AI, but fewer than half trust it. We’re taking advice from a system that can’t show its work and makes stuff up, and we keep asking for more. But the shift in AI could earn trust by design. Beyond the black box with neuro-symbolic AI: gregrobison.medium.com/beyond-the-b...
Beyond the Black Box: How Neuro-Symbolic AI Earns Your Trust
“To restore trust we need not only trustworthy persons and institutions, but also assessable reasons for trusting and mistrusting.”
gregrobison.medium.com
May 5, 2026 at 1:44 PM
Two-thirds of us use AI, but fewer than half trust it. We’re taking advice from a system that can’t show its work and makes stuff up, and we keep asking for more. But the shift in AI could earn trust by design. Beyond the black box with neuro-symbolic AI: gregrobison.medium.com/beyond-the-b...
Interested in agentic AI like OpenClaw but concerned about security? You might be interested in Hermes Agent by @nousresearch.com
gregrobison.medium.com/architectura...
gregrobison.medium.com/architectura...
Architectural and Strategic Analysis of the Hermes Agent Framework and the Psyche Decentralized…
OpenClaw is a security nightmare, but the underlying idea of a semi-autonomous agent armed with tools, skills, and a workspace. Nous…
gregrobison.medium.com
April 22, 2026 at 2:06 PM
Interested in agentic AI like OpenClaw but concerned about security? You might be interested in Hermes Agent by @nousresearch.com
gregrobison.medium.com/architectura...
gregrobison.medium.com/architectura...
LeCun, Dupoux & Malik just published a paper that admits the AI industry has outsourced learning to human engineers and called it machine intelligence. @thousandbrains.org has figured out a key piece, but still more is needed to fix it. gregrobison.medium.com/one-step-beh...
One Step Behind: Why AI Models Don’t Learn
“In training a child to activity of thought, above all things we must beware of what I will call ‘inert ideas’ — that is to say, ideas that…
gregrobison.medium.com
April 14, 2026 at 1:43 PM
LeCun, Dupoux & Malik just published a paper that admits the AI industry has outsourced learning to human engineers and called it machine intelligence. @thousandbrains.org has figured out a key piece, but still more is needed to fix it. gregrobison.medium.com/one-step-beh...
Everyone thinks bigger models = smarter AI. Princeton just showed a 32B LLM fail a basic medical fact test… while an 80M neuro-symbolic model running on a laptop got it right. Scale isn’t intelligence. After the Transformer: How GraphMERT is delivering: gregrobison.medium.com/after-the-tr...
After the Transformer: How GraphMERT’s Neuro-symbolic AI Is Finally Delivering
“You can know the name of a bird in all the languages of the world, but when you’re finished, you’ll know absolutely nothing whatever about…
gregrobison.medium.com
March 31, 2026 at 1:27 PM
Everyone thinks bigger models = smarter AI. Princeton just showed a 32B LLM fail a basic medical fact test… while an 80M neuro-symbolic model running on a laptop got it right. Scale isn’t intelligence. After the Transformer: How GraphMERT is delivering: gregrobison.medium.com/after-the-tr...
Reposted by Greg Robison
Welcome to awesome April at the Oakland Public Library! This month we're celebrating Arab American Heritage Month, National Poetry Month, National Library Week, and día de los niños/día de los libros.
We host *hundreds* of programs every month so here's just a quick snapshot: https://bit.ly/4dSeXAK
We host *hundreds* of programs every month so here's just a quick snapshot: https://bit.ly/4dSeXAK
March 30, 2026 at 8:02 PM
Welcome to awesome April at the Oakland Public Library! This month we're celebrating Arab American Heritage Month, National Poetry Month, National Library Week, and día de los niños/día de los libros.
We host *hundreds* of programs every month so here's just a quick snapshot: https://bit.ly/4dSeXAK
We host *hundreds* of programs every month so here's just a quick snapshot: https://bit.ly/4dSeXAK
Your customers aren’t Googling anymore because they’re asking ChatGPT what to buy. AI shopping traffic is up 4,700% YoY. 44% of users now prefer AI search over traditional search. Welcome to the age of GEO.
Is your brand ready?
gregrobison.medium.com/the-new-age-...
Is your brand ready?
gregrobison.medium.com/the-new-age-...
The New Age of Marketing: Why AI Shopping Is Making Your Playbook Obsolete
“The transition between competing paradigms cannot be made a step at a time, forced by logic and neutral experience.”
gregrobison.medium.com
March 23, 2026 at 1:23 PM
Your customers aren’t Googling anymore because they’re asking ChatGPT what to buy. AI shopping traffic is up 4,700% YoY. 44% of users now prefer AI search over traditional search. Welcome to the age of GEO.
Is your brand ready?
gregrobison.medium.com/the-new-age-...
Is your brand ready?
gregrobison.medium.com/the-new-age-...
When you talk to an AI assistant, you're talking to a character, one with a coherent psychology, inferred personality traits, and a sense of what kind of entity it is. And that changes how we build safe AI. Read more here: gregrobison.medium.com/persona-non-...
Persona Non Grata: The Hidden Psychology Shaping AI Assistants
“The persona is a complicated system of relations between individual consciousness and society - a kind of mask, designed on the one hand…
gregrobison.medium.com
March 11, 2026 at 1:02 PM
When you talk to an AI assistant, you're talking to a character, one with a coherent psychology, inferred personality traits, and a sense of what kind of entity it is. And that changes how we build safe AI. Read more here: gregrobison.medium.com/persona-non-...
Can AI truly understand reality? LLMs predict words. World Models predict consequences because they simulate gravity, space, and time before acting.
It’s the shift from "book smart" to "street smart" AI, the missing link for Robotics, Driving, and AGI.
gregrobison.medium.com/beyond-the-c...
It’s the shift from "book smart" to "street smart" AI, the missing link for Robotics, Driving, and AGI.
gregrobison.medium.com/beyond-the-c...
Beyond the Chatbot: How AI World Models Are Learning to Understand Reality
“If the organism carries a ‘small-scale model’ of external reality and of its own possible actions within its head, it is able to try out…
gregrobison.medium.com
March 3, 2026 at 2:16 PM
Can AI truly understand reality? LLMs predict words. World Models predict consequences because they simulate gravity, space, and time before acting.
It’s the shift from "book smart" to "street smart" AI, the missing link for Robotics, Driving, and AGI.
gregrobison.medium.com/beyond-the-c...
It’s the shift from "book smart" to "street smart" AI, the missing link for Robotics, Driving, and AGI.
gregrobison.medium.com/beyond-the-c...
What effect does repeated exposure to deepfakes have on us? As I deepfake my friend into every image I create, my conception of him has changed? Has he actually been a fireman? Base jumped in a wing suit? I'd believe it...
gregrobison.medium.com/7c38183e3e62
gregrobison.medium.com/7c38183e3e62
Epistemic Fragmentation and the Synthetic Horizon: A Longitudinal Analysis of Deepfake Exposure…
With deepfakes becoming easier to make and more realistic and convincing, what’s happening to us? Is it eroding trust now that we can’t…
gregrobison.medium.com
February 25, 2026 at 1:30 PM
What effect does repeated exposure to deepfakes have on us? As I deepfake my friend into every image I create, my conception of him has changed? Has he actually been a fireman? Base jumped in a wing suit? I'd believe it...
gregrobison.medium.com/7c38183e3e62
gregrobison.medium.com/7c38183e3e62
Digital info used to be solid: files, folders, pixels.
AI liquefies it.
Latent space is a semantic ocean where concepts have gradients, not edges and creativity is just steering the current.
A deep dive into the physics of meaning:
gregrobison.medium.com/semantic-flu...
AI liquefies it.
Latent space is a semantic ocean where concepts have gradients, not edges and creativity is just steering the current.
A deep dive into the physics of meaning:
gregrobison.medium.com/semantic-flu...
Semantic Fluidity of Latent Space: How AI Dissolved Information into an Ocean of Meaning
“Every actuality surrounds itself with a fog of virtual images.”
gregrobison.medium.com
February 17, 2026 at 3:30 PM
Digital info used to be solid: files, folders, pixels.
AI liquefies it.
Latent space is a semantic ocean where concepts have gradients, not edges and creativity is just steering the current.
A deep dive into the physics of meaning:
gregrobison.medium.com/semantic-flu...
AI liquefies it.
Latent space is a semantic ocean where concepts have gradients, not edges and creativity is just steering the current.
A deep dive into the physics of meaning:
gregrobison.medium.com/semantic-flu...
Fascinated by Moltbook and OpenClaw? Interested in running your own AI agent on the site? Beware of the security issues plaguing agents, including malware-infested tools and prompt injection attacks. Stay safe and read more about the security issues here: gregrobison.medium.com/the-sovereig...
The Sovereign Agent Dilemma: A Comprehensive Analysis of OpenClaw, Moltbook, and the Emerging…
As Chief AI Architect of F’inn, I’ve been developing tools powered by AI agents for a couple of years now and Moltbook looked like a…
gregrobison.medium.com
February 7, 2026 at 9:40 PM
Fascinated by Moltbook and OpenClaw? Interested in running your own AI agent on the site? Beware of the security issues plaguing agents, including malware-infested tools and prompt injection attacks. Stay safe and read more about the security issues here: gregrobison.medium.com/the-sovereig...
Our best vision‑language models can write poems, but still fail tasks a toddler aces. A new benchmark, BabyVision, shows top VLLMs score ~49 vs adults’ 94 on basic visual perception.
Have we been testing the wrong things? gregrobison.medium.com/babyvision-w...
Have we been testing the wrong things? gregrobison.medium.com/babyvision-w...
BabyVision: Why Your Toddler Can See What AI Cannot
“Perception starts with the eye.”
gregrobison.medium.com
February 3, 2026 at 1:58 PM
Our best vision‑language models can write poems, but still fail tasks a toddler aces. A new benchmark, BabyVision, shows top VLLMs score ~49 vs adults’ 94 on basic visual perception.
Have we been testing the wrong things? gregrobison.medium.com/babyvision-w...
Have we been testing the wrong things? gregrobison.medium.com/babyvision-w...
AI agents fail like deputies with too much authority.
Prompt injection is now the #1 risk for LLM apps, enabling zero‑click data exfiltration and self‑replicating AI worms.
Full threat landscape + mitigations, read here:
gregrobison.medium.com/the-crisis-o...
Prompt injection is now the #1 risk for LLM apps, enabling zero‑click data exfiltration and self‑replicating AI worms.
Full threat landscape + mitigations, read here:
gregrobison.medium.com/the-crisis-o...
The Crisis of Agency: A Comprehensive Analysis of Prompt Injection and the Security Architecture of…
Below is an unedited Google Gemini Deep Research report on the current details, implications, and potential remediations for prompt…
gregrobison.medium.com
January 27, 2026 at 1:45 PM
AI agents fail like deputies with too much authority.
Prompt injection is now the #1 risk for LLM apps, enabling zero‑click data exfiltration and self‑replicating AI worms.
Full threat landscape + mitigations, read here:
gregrobison.medium.com/the-crisis-o...
Prompt injection is now the #1 risk for LLM apps, enabling zero‑click data exfiltration and self‑replicating AI worms.
Full threat landscape + mitigations, read here:
gregrobison.medium.com/the-crisis-o...
Your brain doesn’t learn at one speed and neither should AI.
Current LLMs live in the instant present and the distant past, with nothing in between. A new framework called Nested Learning by Google explains and how to fix it. gregrobison.medium.com/the-illusion...
Current LLMs live in the instant present and the distant past, with nothing in between. A new framework called Nested Learning by Google explains and how to fix it. gregrobison.medium.com/the-illusion...
The Illusion of Deep Learning: How Multi-Timescale Processing via Nested Learning Solves AI’s…
“In much the same way, scientific revolutions are inaugurated by a growing sense, again often restricted to a narrow subdivision of the…
gregrobison.medium.com
January 21, 2026 at 1:14 PM
Your brain doesn’t learn at one speed and neither should AI.
Current LLMs live in the instant present and the distant past, with nothing in between. A new framework called Nested Learning by Google explains and how to fix it. gregrobison.medium.com/the-illusion...
Current LLMs live in the instant present and the distant past, with nothing in between. A new framework called Nested Learning by Google explains and how to fix it. gregrobison.medium.com/the-illusion...
For years we taught robots scripts. Now we’re letting them learn from the world itself. At CES 2026, Boston Dynamics and DeepMind agreed marking the moment embodied intelligence goes industrial. gregrobison.medium.com/bodies-for-b...
Bodies for Brains, Brains for Bodies: Why Boston Dynamics and DeepMind Need Each Other to Build…
“The world is its own best model — always exactly up to date and complete in every detail.”
gregrobison.medium.com
January 13, 2026 at 1:46 PM
For years we taught robots scripts. Now we’re letting them learn from the world itself. At CES 2026, Boston Dynamics and DeepMind agreed marking the moment embodied intelligence goes industrial. gregrobison.medium.com/bodies-for-b...
LMs can explain Mario, Pokémon, Doom… but can’t play them well. A new NVIDIA-led model learns like humans do by watching 40k hours of gameplay. This might be the execution innovation that gaming AI needed. gregrobison.medium.com/nitrogen-and...
NitroGen and the Execution Problem: Why AI Finally Learned to Play Games (But Not to Win Them)
“Fortunately, most human behavior is learned observationally through modeling from others.”
gregrobison.medium.com
January 6, 2026 at 1:36 PM
LMs can explain Mario, Pokémon, Doom… but can’t play them well. A new NVIDIA-led model learns like humans do by watching 40k hours of gameplay. This might be the execution innovation that gaming AI needed. gregrobison.medium.com/nitrogen-and...
Heard about Model Context Protocol (MCP)? It gives LLMs superpowers with tool use, connecting to all sorts of data sources. Is MCP the future of AI connectivity? Find out how to works and what it means for the future: gregrobison.medium.com/the-model-co...
The Model Context Protocol: The Architecture of Agentic Intelligence
Model Context Protocol (MCP) is continuing to gain traction as the data connector for tool use by LLMs, enabling deep integration into…
gregrobison.medium.com
December 23, 2025 at 2:35 PM
Heard about Model Context Protocol (MCP)? It gives LLMs superpowers with tool use, connecting to all sorts of data sources. Is MCP the future of AI connectivity? Find out how to works and what it means for the future: gregrobison.medium.com/the-model-co...
Counterintuitive AI safety research by OpenAI suggests that teaching models to confess their mistakes by making honesty easier than deception. When they misbehave, they admit it 74% of the time. The trick? Never punish them for telling the truth. gregrobison.medium.com/teaching-ai-...
Teaching AI to Come Clean: How OpenAI’s “Confessions” Create Honest AI Systems
“Above all, don’t lie to yourself. The man who lies to himself and listens to his own lie comes to a point that he cannot distinguish the…
gregrobison.medium.com
December 17, 2025 at 1:40 PM
Counterintuitive AI safety research by OpenAI suggests that teaching models to confess their mistakes by making honesty easier than deception. When they misbehave, they admit it 74% of the time. The trick? Never punish them for telling the truth. gregrobison.medium.com/teaching-ai-...
Humans know robots are machines, but we treat them like minds, friends, even moral beings. From the Uncanny Valley to LLM-driven empathy, the “social artifact” is reshaping how we think, feel, and trust. To understand the future of human cognition, start here: gregrobison.medium.com/e970305dd7b6
The Cognitive Mirror: An Analysis of Human Social Cognition in the Age of Artificial Agents Like…
Below is an unedited Gemini Deep Research review of academic research on our social cognition when the “other” is a robot. Do we treat…
gregrobison.medium.com
December 10, 2025 at 2:45 PM
Humans know robots are machines, but we treat them like minds, friends, even moral beings. From the Uncanny Valley to LLM-driven empathy, the “social artifact” is reshaping how we think, feel, and trust. To understand the future of human cognition, start here: gregrobison.medium.com/e970305dd7b6