Lifeboat Foundation: Safeguarding Humanity
Toggle light / dark theme

Blog

Category: robotics/AI

View a PDF of the paper titled When Models Manipulate Manifolds: The Geometry of a Counting Task, by Wes Gurnee and 6 other authors

When you look at text, you subconsciously track how much space remains on each line. If you’re writing “Happy Birthday” and “Birthday” won’t fit, your brain automatically moves it to the next line. You don’t calculate this—you *see* it. But AI models don’t have eyes. They receive only sequences of numbers (tokens) and must somehow develop a sense of visual space from scratch.

Inside your brain, “place cells” help you navigate physical space by firing when you’re in specific locations. Remarkably, Claude develops something strikingly similar. The researchers found that the model represents character counts using low-dimensional curved manifolds—mathematical shapes that are discretized by sparse feature families, much like how biological place cells divide space into discrete firing zones.

The researchers validated their findings through causal interventions—essentially “knocking out” specific neurons to see if the model’s counting ability broke in predictable ways. They even discovered visual illusions—carefully crafted character sequences that trick the model’s counting mechanism, much like optical illusions fool human vision.

2. Attention mechanisms are geometric engines: The “attention heads” that power modern AI don’t just connect related words—they perform sophisticated geometric transformations on internal representations.

1. What other “sensory” capabilities have models developed implicitly? Can AI develop senses we don’t have names for?

Language models can perceive visual properties of text despite receiving only sequences of tokens-we mechanistically investigate how Claude 3.5 Haiku accomplishes one such task: linebreaking in fixed-width text. We find that character counts are represented on low-dimensional curved manifolds discretized by sparse feature families, analogous to biological place cells. Accurate predictions emerge from a sequence of geometric transformations: token lengths are accumulated into character count manifolds, attention heads twist these manifolds to estimate distance to the line boundary, and the decision to break the line is enabled by arranging estimates orthogonally to create a linear decision boundary. We validate our findings through causal interventions and discover visual illusions—character sequences that hijack the counting mechanism.

How scientists are trying to use AI to unlock the human mind

Compared with conventional psychological models, which use simple math equations, Centaur did a far better job of predicting behavior. Accurate predictions of how humans respond in psychology experiments are valuable in and of themselves: For example, scientists could use Centaur to pilot their experiments on a computer before recruiting, and paying, human participants. In their paper, however, the researchers propose that Centaur could be more than just a prediction machine. By interrogating the mechanisms that allow Centaur to effectively replicate human behavior, they argue, scientists could develop new theories about the inner workings of the mind.

But some psychologists doubt whether Centaur can tell us much about the mind at all. Sure, it’s better than conventional psychological models at predicting how humans behave—but it also has a billion times more parameters. And just because a model behaves like a human on the outside doesn’t mean that it functions like one on the inside. Olivia Guest, an assistant professor of computational cognitive science at Radboud University in the Netherlands, compares Centaur to a calculator, which can effectively predict the response a math whiz will give when asked to add two numbers. “I don’t know what you would learn about human addition by studying a calculator,” she says.

Even if Centaur does capture something important about human psychology, scientists may struggle to extract any insight from the model’s millions of neurons. Though AI researchers are working hard to figure out how large language models work, they’ve barely managed to crack open the black box. Understanding an enormous neural-network model of the human mind may not prove much easier than understanding the thing itself.

Dyson Strawberry Farming: 5,127 Prototypes to 250% Yields

When James Dyson built his 5,127th prototype of a bagless vacuum cleaner, he had no idea that the same relentless engineering philosophy would one day transform him into Britain’s largest farmer. Today, Dyson strawberry farming represents one of the most ambitious applications of high-tech innovation to agriculture ever attempted in the United Kingdom.

The numbers tell an extraordinary story. After spending five years and creating over five thousand prototypes to perfect a single vacuum cleaner design, Dyson has now invested £140 million into a farming operation spanning 36,000 acres across five English counties. At the heart of this agricultural empire sits a 26-acre glasshouse in Lincolnshire, home to 1.25 million strawberry plants and technology that has increased yields by 250% compared to traditional farming methods.

This isn’t farming as your grandparents would recognize it. Inside Dyson’s facility, massive 5.5-meter “ferris wheel” structures rotate strawberry plants through optimal sunlight positions. Sixteen robotic arms delicately harvest ripe fruit using computer vision. UV-emitting robots patrol the aisles at night, destroying mould without chemicals. And all of it runs on renewable energy generated from an adjacent anaerobic digester.

Planes grounded as extreme solar radiation hits essential flight controls. 6,000 jets affected

The A320 involved suffered a flight-control issue that caused a sudden drop in altitude, leaving some passengers with non-life-threatening injuries. During the investigation, a vulnerability to solar flares emerged.

As the aviation industry grows more automated and electronics-dependent, understanding space-weather threats is increasingly vital.

Recent NASA studies suggest that space weather is becoming more intense and frequent, with the Sun currently in a stronger-than-expected activity cycle (solar cycle 25) and potentially entering a period of elevated activity that could last decades.

AGI Is Here: AI Legend Peter Norvig on Why it Doesn’t Matter Anymore

Are we chasing the wrong goal with Artificial General Intelligence, and missing the breakthroughs that matter now?

On this episode of Digital Disruption, we’re joined by former research director at Google and AI legend, Peter Norvig.

Peter is an American computer scientist and a Distinguished Education Fellow at the Stanford Institute for Human-Centered Artificial Intelligence (HAI). He is also a researcher at Google, where he previously served as Director of Research and led the company’s core search algorithms group. Before joining Google, Norvig headed NASA Ames Research Center’s Computational Sciences Division, where he served as NASA’s senior computer scientist and received the NASA Exceptional Achievement Award in 2001.He is best known as the co-author, alongside Stuart J. Russell, of Artificial Intelligence: A Modern Approach — the world’s most widely used textbook in the field of artificial intelligence.

Peter sits down with Geoff to separate facts from fiction about where AI is really headed. He explains why the hype around Artificial General Intelligence (AGI) misses the point, how today’s models are already “general,” and what truly matters most: making AI safer, more reliable, and human-centered. He discusses the rapid evolution of generative models, the risks of misinformation, AI safety, open-source regulation, and the balance between democratizing AI and containing powerful systems. This conversation explores the impact of AI on jobs, education, cybersecurity, and global inequality, and how organizations can adapt, not by chasing hype, but by aligning AI to business and societal goals. If you want to understand where AI actually stands, beyond the headlines, this is the conversation you need to hear.

In this episode:
00:00 Intro.
01:00 How AI evolved since Artificial Intelligence: A Modern Approach.
03:00 Is AGI already here? Norvig’s take on general intelligence.
06:00 The surprising progress in large language models.
08:00 Evolution vs. revolution.
10:00 Making AI safer and more reliable.
12:00 Lessons from social media and unintended consequences.
15:00 The real AI risks: misinformation and misuse.
18:00 Inside Stanford’s Human-Centered AI Institute.
20:00 Regulation, policy, and the role of government.
22:00 Why AI may need an Underwriters Laboratory moment.
24:00 Will there be one “winner” in the AI race?
26:00 The open-source dilemma: freedom vs. safety.
28:00 Can AI improve cybersecurity more than it harms it?
30:00 “Teach Yourself Programming in 10 Years” in the AI age.
33:00 The speed paradox: learning vs. automation.
36:00 How AI might (finally) change productivity.
38:00 Global economics, China, and leapfrog technologies.
42:00 The job market: faster disruption and inequality.
45:00 The social safety net and future of full-time work.
48:00 Winners, losers, and redistributing value in the AI era.
50:00 How CEOs should really approach AI strategy.
52:00 Why hiring a “PhD in AI” isn’t the answer.
54:00 The democratization of AI for small businesses.
56:00 The future of IT and enterprise functions.
57:00 Advice for staying relevant as a technologist.
59:00 A realistic optimism for AI’s future.

#ai #agi #humancenteredai #futureofwork #aiethics #innovation.

Continue reading “AGI Is Here: AI Legend Peter Norvig on Why it Doesn’t Matter Anymore” | >

AI expert warns we’re close to extinction

Connor Leahy discusses the motivations of AGI corporations, how modern AI is “grown”, the need for a science of intelligence, the effects of AI on work, the radical implications of superintelligence, and what you might be able to do about all of this. https://www.thecompendium.ai 00:00 The AI Race 02:14 CEOs Lying 04:02 The Entente Strategy 06:12 AI is grown, not built 07:39 Jobs 10:47 Alignment 14:25 What should you do? Original Podcast: • Connor Leahy on Why Humanity Risks Extinct… Editing: https://zeino.tv/

World Modeling Workshop — Day 1

A fundamental desideratum of AI is the ability to model environment dynamics and transitions in response to both their own actions and external control signals. This capability, commonly referred to as world modeling (WM), is essential for prediction, planning, and generalization. Learning world models using deep learning has been an active area of research for nearly a decade. In recent years, the field has witnessed significant breakthroughs driven by advances in deep neural architectures and scalable learning paradigms. Multiple subfields, including self-supervised learning (SSL), generative modeling, reinforcement learning (RL), robotics, and large language models (LLMs), have tackled aspects of world modeling, often with different tools and methodologies. While these communities address overlapping challenges, they frequently operate in isolation. As a result, insights and progress in one area may go unnoticed in another, limiting opportunities for synthesis and collaboration. This workshop aims to bridge this gap between subfields of world modeling by fostering open dialogue, critical discussion, and cross-disciplinary exchange. By bringing together researchers from diverse backgrounds, from early-career researchers to established experts, we hope to establish a shared vocabulary, identify common challenges, and surface synergies that can move the field of world modeling forward.

/* */