Bruce Pullman | Melbury

69. Oxford chemist → quant investor → AI founder. Abundant intelligence and a shortage of human agency. That is the problem I’m working on. For billions.

[Bruce Pullman | Melbury @bruce_pullman]

We help people solve real-life problems. As they use Melbury, we build a compact, evolving model of the individual — so AI understands them better over time. At scale, this creates a longitudinal picture of humanity largely missing from today’s AI.

[Bruce Pullman | Melbury @bruce_pullman]

In just 66 days, Anthropic's AI systems achieved an estimated sixfold increase in the speed of formalising Fermat's Last Theorem. Annualised, that is a staggering 20,000-fold increase in speed. We may already be on the steeply accelerating part of the AI development curve.

Meanwhile in the real world of my laptop....

[Bruce Pullman | Melbury @bruce_pullman]

@joelcomm 69 founder working on multiple new ideas. AI a great enabler. Hoe about a dynamic aye-book

[Bruce Pullman | Melbury @bruce_pullman]

@dioscuri This is the Waymo vs Tesla approach again. Will data and training cook everything?

[Bruce Pullman | Melbury @bruce_pullman]

@joelcomm Yup

[Bruce Pullman | Melbury @bruce_pullman]

@dioscuri paren......close paren

[Bruce Pullman | Melbury @bruce_pullman]

@joelcomm Chagpt and it produces a comprehensive report on 10 topics to my specifications

[Bruce Pullman | Melbury @bruce_pullman]

@joelcomm 69. All the research ideas I never had time, budget or patience to pursue

[Bruce Pullman | Melbury @bruce_pullman]

@joelcomm HP9845 in a room shared with a photocopier and fax machine aged 24 in 1981....now at 69

[Bruce Pullman | Melbury @bruce_pullman]

On the train use Wispr to drive chatgpt to draft the email you want to send to a group of people, then get it to research each of them on the web individually, including your WhatsApp with them and any email correspondence, and edit each email appropriately in a style tailored to them personally, package it as a .zip file.. When you get home send the zip to yourelf in WhatsApp. Download it to your laptop desktop. Then tell Codex to process it through your Outlook to send all the emails.

[Bruce Pullman | Melbury @bruce_pullman]

BBC science journalism has badly failed here.

@Vic_Gill writes that Kagan and Soai “answered the question of why life favours one of these mirror image molecules over the other.”

They did not.

Their work shows how chiral imbalance can arise and be amplified. It does not explain why terrestrial biology ended up with its particular handedness.

Gill is a BBC Science Correspondent and was previously a Chemistry World science correspondent. She should know the distinction.

In an AI age, checking this takes minutes.

The article is already being edited. I have the screenshots.

@BBCScienceNews @BBCNews @ChemistryWorld

https://t.co/ow9Vx4qAKf

[Bruce Pullman | Melbury @bruce_pullman]

I agree these systems are intelligent. But the question that matters to me is how far that intelligence extends to originating an inquiry and finding a new way to explain something.

They can follow a subject I introduce with extraordinary fluency. What I still want to see is a system independently noticing an unexplained pattern, deciding it deserves investigation, and developing the mathematical representation that makes it intelligible.

We explored part of this with a blind Kepler-style experiment. We generated artificial worlds governed by compact mathematical laws, then gave fresh sessions only angular observations from a moving observer. No distances, no hidden equations, and no instruction to assume familiar planetary motion. The task was to recover the geometry and motion rules and predict measurements we had withheld.

The results were mixed. One case appeared to recover almost the whole hidden system. Several others reconstructed parts of it, searched familiar mathematical families, then retreated into increasingly flexible fits or the observation that many explanations were possible.

That was the interesting boundary. Could the model recognise that its current representation was the obstacle, and find one in which the observations became simple?

So I wouldn’t say AI cannot discover laws. Our experiment suggested it sometimes can. But reliable conceptual invention remains a different claim from intelligent performance.

And the experiment still supplied the question. It didn’t test whether the system would originate that investigation itself.

My next question was whether a thousand attempts might produce one successful discoverer. Perhaps. But then another system must recognise the good theory without an answer key, using simplicity, new predictions and discriminating experiments.

That is the capability I want to understand before declaring the discussion over.

[Bruce Pullman | Melbury @bruce_pullman]

Henry, what specifically changed your mind about genuine understanding? Was there a particular interaction or result, or did the old explanation gradually become harder to sustain?

I ask because my own experience of using these systems increasingly makes “merely predicting the next word” feel inadequate as an account of what they can do. Though that feeling doesn’t itself settle what is happening inside them.

As a chemist by training, I’m comfortable with a system acquiring properties that aren’t obvious from a description of its components. But I still want to know what observation would distinguish understanding from an extraordinarily successful imitation of it.

And I wonder whether we’re setting AI a test we don’t consistently apply to each other. I infer your understanding from what you say, what you do, and how you cope when I change the question. I can’t inspect it directly.

Where would you now draw that line? And how much does accepting genuine understanding tell us about consciousness, if anything?

[Bruce Pullman | Melbury @bruce_pullman]

I recognise this very strongly. At 69 I’m building an AI company, and the list of things I could now attempt grows faster than the list of things I finish. It is exhilarating. It also makes it surprisingly difficult to leave anything alone.

But your phrase “for agentic people” is doing quite a lot of work here. Some people already know how to turn a possibility into a project. Many others have intelligence and curiosity but little confidence that they can act on either. What does this abundance do for them?

That’s the question behind what we’re building at Melbury. Can AI help someone see a possibility, judge whether it matters to them, and take a first step? And can it do that while leaving the judgement genuinely theirs?

Your dissatisfaction points to a further difficulty. Once the machine can give us another ten interesting things to do whenever we ask, we need some way of saying: this one matters enough to stay with. Even while the other nine remain possible.

I suspect an AI that helps us make and keep that choice would be more useful than one that endlessly expands the menu. Though deciding when helpful encouragement becomes steering is going to be quite a problem.

[Bruce Pullman | Melbury @bruce_pullman]

@dioscuri @LeverhulmeCFI Excellent. Whats going to happen to society when you wake up, and your ai then, of its "own volition", starts the conversation on a subject you dont know, in a way youve nrver considered, but which {by definition} your find intoxicatingly interesting. Auto-rabbit-holeing!

[Bruce Pullman | Melbury @bruce_pullman]

I think the really sharp issue is that, whatever the academics think, whatever the scientists think, whatever the professors think, ordinary people (and by ordinary people I mean those who don't necessarily have even a university education but who may be far from stupid) are going to very quickly treat it as reality. The thing has something inside it. In the same way as they kind of treat, in a sense, Coronation Street as real. Something I don't fully understand, but clearly it's a phenomenon.

[Bruce Pullman | Melbury @bruce_pullman]

Something rather revealing has happened this week.
Trump wants “Artificial Intelligence” renamed “Super Intelligence”. Pete Killingley’s new A Christian’s Guide to AI goes almost exactly the other way, preferring “apparent intelligence”.
I’m uncomfortable with both.
“Super intelligence” upgrades the technology before we have settled what sort of intelligence it possesses. “Apparent intelligence” seems to downgrade it for much the same reason: the terminology smuggles the conclusion into the definition.
And in the Christian case I’m not sure why we need to do that.
We should surely separate intelligence, understanding, agency, consciousness, personhood and the imago Dei. They are not interchangeable.
A machine could conceivably become genuinely intelligent — reasoning, abstracting, planning, discovering, solving problems beyond human ability — while still lacking phenomenal consciousness, moral personhood, a soul, or whatever Christians mean by humanity bearing the image of God.
If so, Christianity loses nothing.
Human dignity never depended upon our being the cleverest entities available. The theological claims about human beings concern creaturehood, embodiment, moral responsibility, relationship with God, sin, redemption, love and the imago Dei. None requires us to win an intelligence competition against our own machines.
That is why “apparent intelligence” worries me. It risks making Christian anthropology hostage to an empirical claim about what machines can never genuinely do. If capabilities keep advancing, Christians then have to keep moving the boundary: yes, it can do that, but it still isn’t really intelligence.
The reverse problem exists with Trump’s “Super Intelligence”. Capability gets promoted into ontology. Calling something super-intelligent tells us remarkably little about whether it understands as we do, experiences anything, possesses interests, deserves moral consideration or can bear responsibility.
There is also an important practical distinction. A system need not be conscious or a person before its agency becomes consequential. It can plan, select routes, use tools and pursue objectives without possessing anything resembling human interior life.
So perhaps the terminology at both ends is getting ahead of us.
Trump: “Super Intelligence.”
Killingley: “Apparent Intelligence.”
Why not simply allow the possibility of genuine non-human intelligence, and then do the much harder work of asking what kind of thing it is?
Pete, I’d be genuinely interested in your answer: what Christian theological claim would actually be threatened if machine intelligence turned out to be real?

[Pete Killingley 🇺🇦 @petekillingley]

IT IS LAUNCH DAY! My book, 'A Christian's Guide to AI', is now available at The Good Book Company website and all good retailers.

https://t.co/vPblFEyOUi

I can't quite believe this day has finally come. Thank you to everyone who has encouraged and supported me along the way!!

[Pete Killingley 🇺🇦 @petekillingley]

IT IS LAUNCH DAY! My book, 'A Christian's Guide to AI', is now available at The Good Book Company website and all good retailers.

https://t.co/vPblFEyOUi

I can't quite believe this day has finally come. Thank you to everyone who has encouraged and supported me along the way!!

[Bruce Pullman | Melbury @bruce_pullman]

What Shapiro is pointing to is a distinction we need to keep very clear in the AI consciousness debate.

An AI can become extraordinarily good at the functions associated with consciousness. It can perceive, remember, maintain a model of itself, reflect on its own behaviour, plan, revise its plans, respond emotionally, act consistently over time and tell you in convincing detail what it is supposedly experiencing.

We can add persistent memory, embodiment, multiple interacting agents, internal feedback loops, asynchronous processes, different internal clocks and vastly more compute. Eventually the resulting system may look, from the outside, increasingly like a conscious person.

But the central question remains.

Why should any of those functions produce an inner experience?

Why should processing information about red create the experience of redness? Why should a system describing pain actually feel pain? Why should a sophisticated self-model create a subject that experiences anything at all?

We are getting increasingly good at explaining how an AI could behave as though it were conscious.

The unanswered question is whether there is anyone home: whether anything is actually being experienced from the inside, or whether the system is simply producing all the outward signs of consciousness.

That is the unresolved problem.

[Bruce Pullman | Melbury @bruce_pullman]

What Shapiro is pointing to is a distinction we need to keep very clear in the AI consciousness debate.
An AI can become extraordinarily good at the functions associated with consciousness. It can perceive, remember, maintain a model of itself, reflect on its own behaviour, plan, revise its plans, respond emotionally, act consistently over time and tell you in convincing detail what it is supposedly experiencing.
We can add persistent memory, embodiment, multiple interacting agents, internal feedback loops, asynchronous processes, different internal clocks and vastly more compute. Eventually the resulting system may look, from the outside, increasingly like a conscious person.
But the central question remains.
Why should any of those functions produce an inner experience?
Why should processing information about red create the experience of redness? Why should a system describing pain actually feel pain? Why should a sophisticated self-model create a subject that experiences anything at all?
We are getting increasingly good at explaining how an AI could behave as though it were conscious.
The unanswered question is whether there is anyone home: whether anything is actually being experienced from the inside, or whether the system is simply producing all the outward signs of consciousness.
That is the unresolved problem.