AI skeptic.

AI cannot substitute for institutions. The development problems that institutions address remain.

“The Analytical Engine has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform.”

The extinction narrative has done real damage to the field by distracting from present harms and deployment reality.

The best way to predict the future is to invent it. We have not yet invented an AI worth predicting.

Hype is not a neutral description. It is a mode of governance.

When I think about existential risks to humans of AI, I don't know how AI could cause us to go extinct. I don't see it.

Don't confuse a benchmark score with a deployed product. The gap is bigger than you think.

The mind extends into the world, into tools, into others, and now into AI. The boundary of cognition is not the skull.

Context: Core claim of his 1998 paper with David Chalmers, applied to AI in subsequent work.

Most AI systems are far less capable than they are marketed to be. The conversation should be about specific deployed systems, not general 'AI'.

If we cannot reproduce the result, we cannot build policy on it.

Tech-savvy people are saying things about AI that, in any other context, would be classed as religious utterances.

I lost my bet. GPT-4 got an A on my labor economics midterm. I am publicly updating.

The reasonable response to AI in knowledge work is not to chase the latest hype cycle but to ask what kind of work makes sense in a world where these tools exist, and structure your day around that.

Most AI projects fail at the decision, not the model.

The AI boom feels less like a technological revolution and more like a cultural and political one, where the technology is the vehicle, not the driver.

“Moralizing AGI governance and innovation, calling some 'bad' and others 'good', is disingenuous. All players are selfish.”

The questions I sent to ChatGPT brought back results that ranged from outstanding to almost-correct to deeply wrong, all delivered with the same confidence.

LLMs do not understand. They do something else, which is impressive, but it is not understanding.

“By 'augmenting human intellect' we mean increasing the capability of a man to approach a complex problem situation, to gain comprehension to suit his particular needs, and to derive solutions to problems.”

OpenAI is a money pit propped up by VC delusion. The product doesn't pay for the compute, the compute doesn't produce a product worth its cost, and the entire thing is held together by hype.

Large language models present dangers such as environmental and financial costs, inscrutability leading to unknown dangerous biases, and potential for deception. They cannot understand the concepts underlying what they learn.

What we have is a real technological achievement combined with a layer of institutional capture that should not be confused with the technology itself. The two have to be analyzed separately.

We have fallen into the Turing Trap. Building AI to imitate humans concentrates power and displaces workers; building AI to augment humans does the opposite.

There is no such thing as 'AI'. There is only a set of political-economic choices about how data, labour, and capital are organised.

LLMs are not the path to AGI. They are impressive pattern-matchers, but they do not generalise to novel problems.

Every AI demo is the best version of the product. Every AI deployment is the worst.

Current large language models are not intelligent; they are stochastic compression of text at best.

AI cannot solve climate. AI plus politics might.

We do not start out with explicit rules and then learn how to apply them. We learn by example, by skill, by being in the world.

AIs are weird. They generalize from data in ways that humans don't, and the ways they fail tell us as much about them as the ways they succeed.

There is no AI. There is only a new form of social collaboration.

Intelligent machines will not have survival drives unless we give them. The alignment-extinction framing projects evolution onto systems that didn't evolve.

Imagine a person who knows no Chinese sitting in a room with a rule book. Slips of paper with Chinese characters come in, the person uses the rule book to send appropriate slips back. The room passes the Turing test, but the person inside understands no Chinese.

“No children should be having a relationship with AI. If we give our kids AI companions that they can order around and will always flatter them, we are creating people who no one will want to employ or marry.”

AI systems do not understand the world the way we do. Deployments that assume they do will fail in specific, predictable ways.

“There are certain tasks which computers ought not be made to do, independent of whether computers can be made to do them.”

Deep learning at present remains stuck at the bottom rung of the ladder of causation. It does observation, not intervention, and certainly not counterfactuals.

“AI is made from vast amounts of natural resources, fuel, and human labor. It is a technology of extraction.”

Context: Opening framing of Atlas of AI.

We argue that LLMs are good at formal linguistic competence but inconsistent at functional linguistic competence: the latter requires more than next-token prediction.

I'm tired of the hype. We don't have AGI. We have very useful pattern matchers, and the social effect of pretending otherwise is corrosive.

“Artificial intelligence does not exist.”

Context: Title and core thesis of his French-language book L'Intelligence Artificielle n'existe pas.

Robotics is hard. The gap between getting something to work in simulation and getting it to work in the real world has not narrowed nearly as much as the language-AI hype suggests.

Behaviour can imitate understanding without instantiating it. Cognitive science exists precisely to keep that distinction in view.

“We shape our tools and thereafter our tools shape us.”

Context: Often attributed to McLuhan via Understanding Media; the precise wording was John M. Culkin in 1967, but the framing is McLuhan's.

The most important thing AI does is reshape how we think about thinking. The economic effects come later.

Most of the questions we ask about AI, what it knows, whether it has a soul, what we owe it, were first asked by theologians. We have not stopped being theological; we have only forgotten that we are.

The real intelligence we want our machines to have, flexible, abstract, analogical reasoning, is far beyond current systems.

“My concerns are more about the people, institutions and incentives that are shaping AI than they are about the technology itself, or the idea that it could somehow become sentient or God-like.”

Arguments about existential AI risk are implicitly arguing that we need to wait until the people who are most privileged now, who are not threatened currently, are in fact threatened before we consider a risk big enough to care about.

The AI we have is not the AI we imagined. And the rhetorical conflation of statistical pattern recognition with intelligence is harmful.

“AI researchers have spent huge amounts of effort and money and repeatedly claimed to have made breakthroughs that bring the dream of intelligent machines within reach, only to have their claims exposed as hopelessly overoptimistic.”

Existing frontier models score under 50% on ARC-AGI puzzles that are easy for humans. The gap reveals what 'general intelligence' really demands beyond scale.

If AI writes the code, who audits the AI?

In a reality of hyper-concentrated power and wealth, AI is much more likely to become a tool of further dispossession and despoilation.

A world of deep fakes, mimicry loops and worsening inequality is not an inevitability but a set of policy choices.

Modern AI is a shortcut. We did not solve intelligence; we found ways to produce useful behaviour without it. The risks of this shortcut differ from the risks of human-like minds.

Disembodied LLMs can mimic many features of intelligence without acquiring the structural understanding that physical interaction with the world produces. We need both, but the latter is what we are skipping.

ChatGPT is basically high-tech plagiarism and a way of avoiding learning.

AI is a long way from being able to spontaneously form its own goals or acquire resources to pursue them.

Our minds are designed to attribute agency and feeling to things that move with apparent purpose. AI systems exploit those inclinations far better than we appreciate.

AI's greatest risk is not having enough of it.

The default narrative that transformer scale-up leads to AGI is probably wrong. Architectural diversity matters.

It is tempting to declare LLMs either trivially intelligent or trivially mindless. Neither verdict survives careful philosophical analysis of what these systems actually do.

Foom scenarios are extremely unlikely. AI will progress by ordinary competitive dynamics.

“LLMs do not reason, by any reasonable definition of reason.”

TinyStories shows that small models can produce coherent, grammatical, and creative text when trained on a constrained synthetic corpus. The dependency on scale is more about diversity of training distribution than fundamental capability.

The question is not whether AI is dangerous; it's whether we know what we want from it.

Ideas in AI often succeed or fail based on whether they happen to fit existing hardware, rather than their inherent merit.

Compute thresholds as a governance strategy have serious limitations and may miss the risks they were meant to catch.

I'm willing to take seriously that AI is a really big deal. I'm not willing to grant that the specific paths to doom you imagine have anything like the probability you assign to them.

“AI is the greatest assault on empathy I have ever seen.”

Intelligence is not the same as power. The doomsday scenarios conflate the two.

I take AI risk seriously, but I'm not sure the quantitative arguments for high p(doom) are as rigorous as they present.

Auto-regressive LLMs cannot, by themselves, do planning or self-verification. They are approximate knowledge sources, not reasoners.

When you obfuscate the names of actions and objects in planning problems, GPT-4's performance plummets.

AI ethics has been written in one philosophical tradition. Other traditions have things to say.

“ChatGPT is a blurry JPEG of the web.”

Will A.I. Become the New McKinsey? Applying A.I. to the real world is a form of economic outsourcing.

“The biggest risk is that those algorithms may not always work. We need to be conscious of this risk and create systems that can still function safely even when the AI components commit errors.”

We urge the signatories of the FLI letter to be mindful of the hype surrounding the power of AI, and to focus on the actual harms that are being done.

Context: DAIR statement in response to the FLI Pause letter.

“Right now we're all adopting this thing and we don't know what problems it causes.”

Most of the lessons of statistical learning still apply to neural networks: bias-variance trade-offs, regularization, distribution shift. Pretending these have been transcended is how we get unreliable systems in production.

Most AI breakthrough headlines wouldn't survive a rigorous reproduction.

“You're going to have to pardon my French, but that's complete B.S.”

Context: Response when asked by the Wall Street Journal whether AI could become smart enough to pose a threat to humanity.

Before we worry about controlling super-intelligent AI, we need to have the beginning of a hint of a design for a system smarter than a house cat.

When I read a new paper, the first question is always: is this real, or is the impressive number coming from something obvious about the evaluation? You'd be surprised how often it's the latter.