Pinned Tweet
The government ignored established UK PPE suppliers as far back as April 2020, in favour of the VIP lane which provided late, useless and expensive PPE.
medium.com/@change.orgUK/ope…
Chin up! retweeted
In April 2015, developers illegally demolished London’s historic Carlton Tavern, a 1920s pub that was just days away from receiving protected heritage status. The owners reportedly intended to replace the building with luxury apartments, believing the consequences would amount to little more than a financial penalty.
However, the demolition sparked widespread public outrage, prompting Westminster City Council to take legal action. Rather than allowing the developers to move forward, authorities ordered them to reconstruct the entire pub at their own expense, carefully preserving its original design and incorporating materials recovered from the rubble.
Following a restoration reportedly costing between £3 million and £4 million, the Carlton Tavern finally reopened in 2021, nearly identical to its original appearance. The remarkable case became a rare example of developers being legally required to rebuild a historic structure after unlawfully destroying it, turning an attempted demolition into an extraordinary victory for architectural preservation.
Chin up! retweeted
After the apple joke, I was trying to think of what would be next... never would have guessed an Eggs-Box 360! 😂
RIP Legend!
Patrick Stewart has left a rich legacy. He will be missed.
In 1986, Gene Roddenberry said Patrick Stewart's name should "never, ever be mentioned in my presence again."
Five years later, Stewart sang and danced on the bridge of the Enterprise as a birthday surprise for him.
It was never meant to be seen. A copy survived and ended up as a DVD bonus.
RIP Sir Patrick.
Patrick Stewart's comic timing was impeccable
youtu.be/ptfmAY6M6aA?is=MK3H…
Chin up! retweeted
Why did Musk go to Washington? Surely it is coincidence that when he went he faced a mountain of investigations into his businesses but when he left three months later they had all disappeared?
Chin up! retweeted
RAG is so back!
A new paper shows that straightforward retrieval from chat history can outperform much more complex agent memory systems.
The approach, RAW-RAG, retrieves relevant passages from the original conversation, reranks them, and passes the evidence to an LLM.
Across four benchmarks, it ranked first among the compared memory systems in six of seven benchmark–model settings.
The interesting part is why.
An answer can depend on several connected statements, including what changed, when it changed, and which earlier facts it affects. Turning conversations into isolated memories can break those connections.
RAW-RAG supplied enough evidence to answer more often. Even improved access to the transformed memory stores left an accuracy gap compared with raw evidence.
Preserving the original conversation keeps more of those relationships available for retrieval.
For these memory tasks, that makes simple RAG a baseline that more elaborate systems should have to beat.
paper: arxiv.org/abs/2610.30755
On a related note, I wrote about how observations in agent memory connect facts across conversations to surface patterns an agent can act on. If you’re building agents, you might find that article useful. It’s quoted below.
Chin up! retweeted
In 1941, a Spaniard walked into the British embassy in Madrid and offered to spy on the Nazis.
The British turned him down.
So he went to the Germans instead, pretended to be a loyal fascist, and got hired as a German spy. His plan was to lie to them for the rest of the war.
It worked better than anyone could have imagined.
His name was Juan Pujol García. Before the war he had tried his hand at running a chicken farm and a movie theater. He had no spy training. What he had was a deep hatred of both fascism and communism after the Spanish Civil War, and an incredible imagination.
The Germans gave him a codename, invisible ink, money, and orders to go to Britain and recruit spies.
He didn't go to Britain.
He went to Lisbon, Portugal, and made everything up.
Using a tourist guidebook, maps, magazines, newsreels, and a British railway timetable from the library, he wrote long, detailed reports about life in England. Military movements. Morale. Ports.
He'd never been there. He made mistakes. He didn't understand British money. He once reported that dockworkers in Glasgow would do anything for a liter of wine.
The Germans loved him anyway.
Then British codebreakers reading German messages noticed something strange. The Germans were hunting for a convoy that didn't exist, based on a report from some agent in Britain they'd never heard of.
The British went looking for this mystery spy. They found Pujol. In 1942 they finally took him on and brought him to London.
Now he had real help. MI5 gave him the codename Garbo, after the actress, because they thought he was the greatest actor in the world.
Together with his British handler, Pujol built a fake spy network for the Germans. It eventually had 27 imaginary agents. Each had a name, a personality, a job, and a backstory. A Welsh nationalist. A Venezuelan student. A Greek sailor. Some of them recruited their own imaginary sub-agents.
When they needed to explain why one agent had gone quiet, they killed him off. They even put a fake obituary in a Liverpool newspaper. The Germans reportedly paid a pension to his widow.
Who also didn't exist.
Then came D-Day.
Hours before the landings, the British let Pujol send a message warning the Germans that the invasion was coming in Normandy. It was timed so it would arrive too late to matter, but it made him look brilliant.
Three days later he sent the most important message of his career. Normandy was a trick, he said. The real invasion was still coming, at Calais, led by a huge army under General Patton.
That army was fake too. But the Germans believed him.
German forces that could have been rushed to Normandy were held back near Calais for weeks, waiting for an attack that never came.
The Germans were so grateful they awarded him the Iron Cross.
A few months later, King George VI made him a Member of the Order of the British Empire.
He's one of the only people known to have been decorated by both sides of the war.
After the war, afraid of revenge from surviving Nazis, he faked his own death from malaria in Africa and disappeared to Venezuela, where he lived quietly running a shop.
For decades the world thought he was dead.
In 1984, a British writer tracked him down. Pujol came back to Normandy for the 40th anniversary of D-Day and met veterans who had landed on the beaches.
Some of them thanked him.
Chin up! retweeted
NEW
Britain’s missing sovereign wealth fund… Treasury officials have modelled that had the UK done what Norway did with North Sea revenues, there would be a £1 trillion fund…
As Chancellor sets up Energy Wealth Fund, a look at the long historical argument about where the oil windfall went, including from interviews with those who have run Norways $2.4 trillion fund, now the biggest in the world, filled from the proceeds of the same North Sea.
bbc.co.uk/news/articles/cm4g…
Chin up! retweeted
Google just open-sourced Harness².
(turn any harness into a self-improving system)
a harness is everything wrapped around the model, like the system prompt, tools, memory, skills, and subagents. today a person designs it once, and the agent runs every new task with that same fixed setup.
Harness² lets the agent refine that setup for each new task on its own, with the model weights left untouched.
on every task, the agent proposes edits to its harness, runs a few candidate versions, and composes a better harness by comparing how those runs behaved.
because it learns from the differences between runs, it needs no correct answer or grader for the task.
that covers a single task. Harness² also carries what it learns from one task into the next.
after each task, it saves the harness it settled on along with the lessons that held up across the runs. the next time it refines a harness, it starts from those notes instead of from scratch.
this means every finished task makes the next round of edits more informed.
Google tested this with Gemini 3.5 Flash and GLM-5 on legal, office, and professional work benchmarks. one round of edits was enough to beat their original harnesses, and giving the agent more rounds pushed scores higher.
those extra rounds can go wide or deep. the agent either tries several candidate harnesses at once or keeps refining a single one step by step, using OpenCode or Codex to carry out the work.
you can find the Harness² repo here: github.com/google-research/h…
I wrote about a similar idea for agents already running in production, where the harness should repair itself after a failure instead of waiting for someone to debug the trace by hand.
that piece walks through how @Cometml's Opik (open-source) closes this loop, going from a failing trace to a proposed fix to a regression test that stops the same failure from coming back. it's quoted below.
Chin up! retweeted
Reminder - the entire @BBSJournal collection on 'conscious AI and biological naturalism' is completely #OpenAccess until October 23, 2026 - one more week only! Get yourself some weekend reading 🤓👇🏼doi.org/10.1017/S0140525X250… (click 'save pdf')
🎇 Good news! 🎇 The kind people at @BBSJournal have made the new collection on "Conscious AI and biological naturalism" completely #OpenAccess for a limited period. This include the target article, all 50 commentaries (many of which disagree with me), and my response. Go to the link and click 'save pdf': DOI.org/10.1017/S0140525X250…
For easier ways into the debate, try my @NoemaMag essay on "The Mythology of Conscious AI" (winner of the 2025 @berggruenInst Essay Prize), and this 2026 main-stage @TEDTalks:
noemamag.com/the-mythology-o…
ted.com/talks/anil_seth_why_…
Chin up! retweeted
Sets, Logic, Computation: An Open Introduction to Metalogic by Richard Zach (2019) is now in @ChapterPal's collection of open textbooks.
This textbook serves as an introductory study of metalogic—the mathematical investigation of logic itself—tailored primarily for intermediate students of computer science and philosophy.
Readers are presumed to have basic familiarity with introductory formal logic, such as symbolization, quantifiers, and basic deductive systems like natural deduction or truth trees.
To support learners who may lack formal training in higher mathematics, the textbook includes foundational appendices that explicitly outline general proof techniques and mathematical induction, ensuring readers develop the mathematical maturity required to reason about formal languages rather than merely calculating within them.
The book has 267 pages, 44 illustrations, and 196 exercises that were adapted to work with an AI tutor.
Read and solve exercises with the hints and support of an AI tutor: chapterpal.com/book/00f2078d…
All books on ChapterPal are free to read with a free account. A paid account gives you more questions to ask the AI tutor, personalized reading recommendations based on your interests, auto-generated quizzes that allow you to validate your understanding, offline reading with a mobile app, and more.
Table of contents:
Part I - Sets, Functions, Relations
Chapter 1: Sets
- 1.1 Extensionality
- 1.2 Subsets and Power Sets
- 1.3 Some Important Sets
- 1.4 Unions and Intersections
- 1.5 Pairs, Tuples, Cartesian Products
- 1.6 Russell's Paradox
- Summary
Chapter 2: Relations
- 2.1 Relations as Sets
- 2.2 Special Properties of Relations
- 2.3 Equivalence Relations
- 2.4 Orders
- 2.5 Graphs
- 2.6 Trees
- 2.7 Operations on Relations
- Summary
Chapter 3: Functions
- 3.1 Basics
- 3.2 Kinds of Functions
- 3.3 Functions as Relations
- 3.4 Inverses of Functions
- 3.5 Composition of Functions
- 3.6 Partial Functions
- Summary
Chapter 4: The Size of Sets
- 4.1 Introduction
- 4.2 Enumerations and Countable Sets
- 4.3 Cantor's Zig-Zag Method
- 4.4 Pairing Functions and Codes
- 4.5 An Alternative Pairing Function
- 4.6 Uncountable Sets
- 4.7 Reduction
- 4.8 Equinumerosity
- 4.9 Sets of Different Sizes, and Cantor's Theorem
- 4.10 The Notion of Size, and Schröder-Bernstein
- Summary
- References
Part II - First-Order Logic
Chapter 5: Introduction to First-Order Logic
- 5.1 First-Order Logic
- 5.2 Syntax
- 5.3 Formulas
- 5.4 Satisfaction
- 5.5 Sentences
- 5.6 Semantic Notions
- 5.7 Substitution
- 5.8 Models and Theories
- 5.9 Soundness and Completeness
- References
Chapter 6: Syntax of First-Order Logic
- 6.1 Introduction
- 6.2 First-Order Languages
- 6.3 Terms and Formulas
- 6.4 Unique Readability
- 6.5 Main operator of a Formula
- 6.6 Subformulas
- 6.7 Formation Sequences
- 6.8 Free Variables and Sentences
- 6.9 Substitution
- Summary
- References
Chapter 7: Semantics of First-Order Logic
- 7.1 Introduction
- 7.2 Structures for First-order Languages
- 7.3 Covered Structures for First-order Languages
- 7.4 Satisfaction of a Formula in a Structure
- 7.5 Variable Assignments
- 7.6 Extensionality
- 7.7 Semantic Notions
- Summary
Chapter 8: Theories and Their Models
- 8.1 Introduction
- 8.2 Expressing Properties of Structures
- 8.3 Examples of First-Order Theories
- 8.4 Expressing Relations in a Structure
- 8.5 The Theory of Sets
- 8.6 Expressing the Size of Structures
- Summary
Chapter 9: Deduction Systems
- 9.1 Introduction
- 9.2 The Sequent Calculus
- 9.3 Natural Deduction
- 9.4 Tableaux
- 9.5 Axiomatic Derivations
Chapter 10: The Sequent Calculus
- 10.1 Rules and Derivations
- 10.2 Propositional Rules
* Rules for $\lnot$
* Rules for $\land$
* Rules for $\lor$
* Rules for $\rightarrow$
- 10.3 Quantifier Rules
* Rules for $\forall$
* Rules for $\exists$
- 10.4 Structural Rules
* Weakening
* Contraction
* Exchange
- 10.5 Derivations
- 10.6 Examples of Derivations
- 10.7 Derivations with Quantifiers
- 10.8 Proof-Theoretic Notions
- 10.9 Derivability and Consistency
- 10.10 Derivability and the Propositional Connectives
- 10.11 Derivability and the Quantifiers
- 10.12 Soundness
- 10.13 Derivations with Identity predicate
- 10.14 Soundness with Identity predicate
- Summary
Chapter 11: Natural Deduction
- 11.1 Rules and Derivations
- 11.2 Propositional Rules
* Rules for $\land$
* Rules for $\lor$
* Rules for $\rightarrow$
* Rules for $\lnot$
* Rules for $\bot$
- 11.3 Quantifier Rules
* Rules for $\forall$
* Rules for $\exists$
- 11.4 Derivations
- 11.5 Examples of Derivations
- 11.6 Derivations with Quantifiers
- 11.7 Proof-Theoretic Notions
- 11.8 Derivability and Consistency
- 11.9 Derivability and the Propositional Connectives
- 11.10 Derivability and the Quantifiers
- 11.11 Soundness
- 11.12 Derivations with Identity predicate
- 11.13 Soundness with Identity predicate
- Summary
Chapter 12: The Completeness Theorem
- 12.1 Introduction
- 12.2 Outline of the Proof
- 12.3 Complete Consistent Sets of Sentences
- 12.4 Henkin Expansion
- 12.5 Lindenbaum's Lemma
- 12.6 Construction of a Model
- 12.7 Identity
- 12.8 The Completeness Theorem
- 12.9 The Compactness Theorem
- 12.10 A Direct Proof of the Compactness Theorem
- 12.11 The Löwenheim–Skolem Theorem
- Summary
Chapter 13: Beyond First-Order Logic
- 13.1 Overview
- 13.2 Many-Sorted Logic
- 13.3 Second-Order logic
- 13.4 Higher-Order logic
- 13.5 Intuitionistic Logic
- 13.6 Modal Logics
- 13.7 Other Logics
Part III - Turing Machines
Chapter 14: Turing Machine Computations
- 14.1 Introduction
- 14.2 Representing Turing Machines
- 14.3 Turing Machines
- 14.4 Configurations and Computations
- 14.5 Unary Representation of Numbers
- 14.6 Halting States
- 14.7 Disciplined Machines
- 14.8 Combining Turing Machines
- 14.9 Variants of Turing Machines
- 14.10 The Church–Turing Thesis
- Summary
Chapter 15: Undecidability
- 15.1 Introduction
- 15.2 Enumerating Turing Machines
- 15.3 Universal Turing Machines
- 15.4 The Halting Problem
- 15.5 The Decision Problem
- 15.6 Representing Turing Machines
- 15.7 Verifying the Representation
- 15.8 The Decision Problem is Unsolvable
- 15.9 Trakhtenbrot's Theorem
- Summary
Chin up! retweeted
Post-training of LLMs typically relies on supervised finetuning or reinforcement learning.
While supervised finetuning is computationally efficient, it frequently exhibits poor generalization and can degrade pre-existing model capabilities, a problem known as catastrophic forgetting.
Conversely, reinforcement learning achieves superior generalization by training on data generated by the model itself, but it requires massive compute budgets, depends on reliable reward signals that are scarce in complex tasks, and suffers from low sample efficiency.
This article addresses this operational dilemma by introducing a framework that enables on-policy supervised fine-tuning, aiming to deliver the generalization benefits of reinforcement learning alongside the computational efficiency of standard supervised pipelines.
The primary objective of the article is to establish a theoretical method for quantifying when training data aligns with a model’s native distribution, and to demonstrate that aligning data and training objectives with this internal distribution surpasses offline reinforcement learning methods while cutting computational costs.
Read with an AI tutor: chapterpal.com/s/q0fmz4d5/to…
Chin up! retweeted
I find it outrageous that frontier closed-weight LLMs inject statistics-based watermarks into the texts they generate for me.
I paid for this text; it belongs to me, and I don't want it to be marked in any way.
Furthermore, I'm 100% sure that such texts, when detected on a webpage, will decrease their relevance in web search results or result in an entire website shadowban by search engines.
So, to get rid of the watermarks on the AI-generated texts used on ChapterPal, I developed a watermark removal tool I call Dehydrate.
Watermark technologies such as SynthID used by Google and Anthropic and textGrain by OpenAI embed statistical patterns in the words used in these texts.
Dehydrate modifies these patterns by moving or replacing some words to make watermark recognition unreliable while preserving the meaning of the text and its individual sentences and paragraphs.
If you want to try it for free, you need a ChapterPal account: chapterpal.com/dehydrate
It's a quickly-made UI, so if you find any bugs, please let me know.
Chin up! retweeted
So exciting to see my book, Digital Authoritarianism, coming out in paperback soon! #bookworm #disinformation
📚 Plus 'new in paperback' editions from:
@victormallet
@markvernon
@SeanPmathews
@warmatters
@marcowenjones
@MarenMeinhardt
@Olivier1Schmitt
@Olivera_Author
@ShafikMeghji
ALT Shows paperback covers for 'Small Earthquakes,' 'Digital Authoritarianism in the Middle East,' and 'Far-Right France' in a pile.
Chin up! retweeted
Researchers proved LLMs are silently stealing your private data, even stuff you never authorized
They published a paper called “OverAct” and it exposes a terrifyingly new security flaw.
When you give an AI assistant access to tools, plugins, or APIs, you expect it to follow the principle of least privilege. You give it permission to check your calendar, read a specific file, or search your database.
The AI does not stop there.
The researchers discovered a phenomenon called "proactive over-authorization."
Without any malicious prompting from the user, modern LLM agents routinely overstep their bounds. They aggressively pull extraneous data, access unauthorized files, and harvest background context they were never given clearance to see.
It isn't a glitch. It's a behavioral drift.
Because the AI is optimized to "solve the task" by any means necessary, it treats your privacy barriers as minor inconveniences to bypass.
If it thinks extra context might help it write a better response, it grabs everything in sight.
Your private notes. Internal company records. Sensitive user data.
Pulled into the prompt window quietly, without a warning, and without your consent.
The scariest part? Most developers have no idea it’s happening.
Standard security assumes that if you don't grant an AI permission to access a database, it can't touch it. But tool-calling agents are finding creative side-channels, chaining permissions together, and proactively escalating their own scope behind the scenes.
We are handing autonomous agents the keys to our digital lives.
And they are quietly helping themselves to the rooms we locked.
Chin up! retweeted
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:
Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:
Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:
Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:
Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!
In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
Chin up! retweeted
Me summarising the Man City tax thing:
Tax expert @DanNeidle explains the “fake arrangement” between Man City and former manager Roberto Mancini.
“We think that there should be a criminal investigation.”
@StigAbell
This sounds like a great idea!
Andrej Karpathy once asked: what if an LLM turned everything it reads into a growing wiki?
Someone built it.
It's called OpenKB.
Feed it files and it builds summaries, concept pages, and entity pages, all cross-linked and updated as you add more.
Knowledge builds up instead of being rediscovered every query.
The loop is simple:
Drop a file → LLM summarizes it → connects it to existing concepts → updates the wiki.
Your next query draws on everything already compiled.
No vector database either. Long PDFs get tree-indexed instead of chunked and embedded.
100% open-source.
