Matthew Thompson (@thompson.gr) Bsky

Messy but escalation in play ✌🏼✌🏼✌🏼

1 week ago 0 0 0 0

[IA 11] What is Rationality and how does it relate to a Rational Agent in AI? In the Spring and Summer of 2025 I had the heady intent of defining Formal Agents, starting with Rationality as this is a term used throughout the domain of AI. With the leak of Claude Code’s code this seems even more topical, seeing prompts like “please do not break the law” is quite disappointing for the leading AI lab. Jonny (good kind) (@jonny@neuromatch.social) a digital infrastructure researcher covers the analysis here.

[IA 11] What is Rationality and how does it relate to a Rational Agent in AI?: matt.thompson.gr/2026/04/08/ia-what-is-ra...

1 week ago 1 0 0 0

Social Media Giants being held to account 💪🏼 www.bbc.com/news/arti… youtu.be/JyN-Iqc73… Which is needed given: Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians “AI psychosis” or “delusional spiraling” is an emerging phenomenon where AI chatbot users find themselves dangerously confident in outlandish beliefs after extended chatbot conversations. This phenomenon is typically attributed to AI chatbots' well-documented bias towards validating users' claims, a property often called “sycophancy.” In this paper, we probe the causal link between AI sycophancy and AI-induced psychosis through modeling and simulation.

Social Media Giants being held to account 💪🏼: matt.thompson.gr/2026/04/02/social-media-...

2 weeks ago 0 0 0 0

Working theory on how GenAI helps Software Engineering and Maintenance First part of my view on the value of GenAI to Software Engineering and Maintenance

Working theory on how GenAI helps Software Engineering and Maintenance: matt.thompson.gr/2026/03/24/working-theor...

First part of my view on the value of GenAI to Software Engineering and Maintenance

3 weeks ago 0 0 0 0

Managing epistemic learnings and uncertainty and Vibe-Learning Rust The weekend was spent organising notes for a final Masters presentation, reflecting on three years of study in LLM and AI in general, with a side of vibe-learning Rust for my Language User Interface intentions.

Managing epistemic learnings and uncertainty and Vibe-Learning Rust: matt.thompson.gr/2026/03/23/managing-epis...

The weekend was spent organising notes for a final Masters presentation, reflecting on three years of study in LLM and AI in general, with a ...

3 weeks ago 0 1 0 0

Be this the truest to the name of World Model?: matt.thompson.gr/2026/03/23/be-this-the-t...

Seems Deepmind is creating an embedding model to represent the Earth. Which also includes Population Dynamics such as Busyness and Search Terms, with EU regulations I’m OK with that fro...

3 weeks ago 0 0 0 0

What are the domains of AI? Discriminative ai (predictive, classification, etc.. Generative ai (create something from a prompt) Dynamic Programming (Value and Policy iteration, RL, Model-based, Model-free) Constraint satisfaction programming (formal methods, planning) Just my 2 cents, maybe too simplic but helping me arrange my thinking. Where does decision making fit? On top of all? Yeah, which means there’s another domain of Information… That is a belief state stored in paper, words, corporate culture, etc…, which inferring over is either intractable or not in place.

What are the domains of AI?: matt.thompson.gr/2026/03/15/what-are-the-...

1 month ago 0 0 0 0

Developers guide managing risk of Coding Agents having more permissions There are 5 types of controls in InfoSec: Preventative Detective Corrective Recovery Deterrent Agents are irritating if you don’t give them access, really I don’t want the agent to be able to remove files from git but it finds a way when given the ability to add and commit. So whilst I’m figuring out how to set up solid preventative controls AND not lose my mind with approvals I’ve set up a recovery control.

Developers guide managing risk of Coding Agents having more permissions: matt.thompson.gr/2026/03/12/developers-gu...

1 month ago 0 0 0 0

How do we Plan? First we Sense. How do we Sense? First we Plan the Sensing... argh.... It’s a trap!

How do we Plan? First we Sense. How do we Sense? First we Plan the Sensing argh.: matt.thompson.gr/2026/02/18/how-do-we-pla...

It’s a trap!

1 month ago 0 0 0 0

Research Links Collection for Reasoning (LLM and other types) Links I collected last summer on Reasoning - large amount LLM links but it goes past that into what is reasoning and the cognitive link (I went too far back and missed Active Inference!)

Research Links Collection for Reasoning (LLM and other types): matt.thompson.gr/2026/02/18/research-link...

Links I collected last summer on Reasoning - large amount LLM links but it goes past that into what is reasoning and the cognitive link (I went too far ...

1 month ago 0 0 0 0

2 months ago 0 0 0 0

Best AI use case: build a beginner programmable Drone guide! This has been a fun morning of investigation into how I can replace my broken Tello and get into RL for Drone Training 🤓🤓

Best AI use case: build a beginner programmable Drone guide!: matt.thompson.gr/2026/02/01/best-ai-use-c...

This has been a fun morning of investigation into how I can replace my broken Tello and get into RL for Drone Training 🤓🤓

2 months ago 0 0 0 0

France to ban social media for under 15s: matt.thompson.gr/2026/01/29/france-to-ban...

2 months ago 0 0 0 0

Learning about Contingent Planning: matt.thompson.gr/2026/01/28/learning-abou...

Looking back through my conversation with Gemini about Dynamic Programming (DP) and Constraint Satisfaction Problems (CSP) that opened the door to Contingent Planning.

2 months ago 0 0 0 0

The Language Construction Kit A kit for constructing a language!

Doing similar! My current to do list:

1. Decide on the sounds of the language
2. Create the lexicon
3. Create the grammar
4. ...

matt.thompson.gr/2026/01/23/t...

2 months ago 0 0 0 0

Learning about Lexicons Building on the last post I’ve been reminding myself what a lexicon is and how they are used in areas other than a compiler…

Learning about Lexicons: matt.thompson.gr/2026/01/24/learning-abou...

Building on the last post I’ve been reminding myself what a lexicon is and how they are used in areas other than a compiler…

2 months ago 0 0 0 0

The Language Construction Kit A kit for constructing a language!

The Language Construction Kit: matt.thompson.gr/2026/01/23/the-language-...

A kit for constructing a language!

2 months ago 0 0 0 0

finally, we've invented funny looking droids from Star Wars

2 months ago 21 4 0 0

This is worth repeating often as it is easy to conflate the two when caught up in other priorities.

Incremental means add onto; Iterative means revise

"Incremental means add onto.
It helps improve the process."

"Iterative m... matt.thompson.gr/2026/01/21/this-is-worth...

2 months ago 0 0 0 0

This is so dam wholesome This is so damn wholesome - a 13 year old kid, talking to one of the greatest rugby players ever and getting some great life advice whilst having a chuckle!! Never undervalue the grunt work to make string foundations. www.facebook.com/story.php This is golden advice on being a good captain and the most important part of leadership !! www.facebook.com/share/r/1…

This is so dam wholesome: matt.thompson.gr/2026/01/20/this-is-so-da...

2 months ago 0 0 0 0

The fidelity of the connections is certainly impressive with embeddings - the benefit of dimensionality!

Slightly related: are you aware of anything that maps concepts in embeddings to external, lower dimension concepts? E.g. the OR logical operator

2 months ago 0 0 0 0

Do homoiconic languages like Lisp unlock formal online search and planning Got a fun and interesting challenge ahead. Looking to refine my intuition/thinking/knowledge of search spaces and online search/planning. I think the homoiconic nature of Lisp could unlock online search/planning and full autonomy (see matt.thompson.gr/2026/01/1…) To be clear, an agent that produces valid Lisp, verified by a lisp parser guard , is a step forward. The context would be something like: “Here’s the macros for AtomicGuard/Dual State Action Pairs: ….” The original specification would be the goal from a human (an action pair may produce decomposed goals in the form of specification to meet the original goal specified).

Do homoiconic languages like Lisp unlock formal online search and planning: matt.thompson.gr/2026/01/19/do-homoiconic...

2 months ago 1 0 0 0

Definitions Reference Working definitions to track and build into the documentation of my research. Generally they are included in the framework or extensions, though I need to learn more about Markov Blankets as I think that could be a boundary between the two state spaces. What the agent can sense and take action on. Otherwise this post is in an order that has trial logic, both in growing on the initial agency through to planning and learning - potential full autonomy.

Definitions Reference: matt.thompson.gr/2026/01/18/definitions-r...

2 months ago 0 0 0 0

AtomicGuard and Active Inference A rough comparison of AtomicGuard and Active Inference - are they opposite sides of the act-sense loop?

AtomicGuard and Active Inference: matt.thompson.gr/2026/01/17/atomicguard-a...

A rough comparison of AtomicGuard and Active Inference - are they opposite sides of the act-sense loop?

2 months ago 0 0 0 0

Annotated History of Modern AI and Deep Learning A well formatted and concise overview of deep learning from the calculus of 1676, when Gottfried Wilhelm Leibniz please blushed the chain rule to the RL-based NN advancements by DeepSeek in 2025.

Annotated History of Modern AI and Deep Learning: matt.thompson.gr/2026/01/12/annotated-his...

A well formatted and concise overview of deep learning from the calculus of 1676, when Gottfried Wilhelm Leibniz please blushed the chain rule to the RL-based NN advancements...

3 months ago 0 0 0 0

The Bernshteyn Bridge and Axiom of Choice Nice article on the bridge between set theory and computer science (which I’d always thought was there! 🙃) A New Bridge Links the Strange Math of Infinity to Computer Science Also helped remind me what the axiom of choice is; an arbitrary choice that acts as a junction between rule based decisions. I still have to understand the actual algorithm as it seems handy to be able to label infinite nodes so that they do not locally conflict… 🤔🤓

The Bernshteyn Bridge and Axiom of Choice: matt.thompson.gr/2026/01/05/the-bernshtey...

3 months ago 0 0 0 0

I’ve been digging into paraconsistent and paracomplete logics and hit a useful tangent Foundational Theories and Philosophical Lenses for Maths and Computer Science

Nice to be aware of my bias and the core details of other ways to look at Maths and CS.

3 months ago 0 0 0 0

A Wikipedia page excerpt explains the principle of explosion in classical logic with a table illustrating logical premises and rules.

Classical Logic versus Paraconsistent Logic: matt.thompson.gr/2026/01/02/classical-log...

3 months ago 0 0 0 0

The joy of happenstance Intentions for Math and Philosophy manifest in an unexpected way

The joy of happenstance: matt.thompson.gr/2026/01/02/the-joy-of-ha...

Intentions for Math and Philosophy manifest in an unexpected way

3 months ago 0 0 0 0

Dual state domain Rough work to be made clearer Domain Table (Complete) Scope Symbol Name Formal Definition State Space Composite S System State S = S_workflow × S_env (Def. 1) Control S_workflow Workflow State {σ | σ : G → {⊥, ⊤}} — truth assignments tracking guard satisfaction Information S_env Environment State A × C (Def. 1, Eq.

Dual state domain: matt.thompson.gr/2025/12/30/dual-state-do...

3 months ago 0 0 0 0

Posts by Matthew Thompson