About
Who is asking, and how it works
Who Is Asking
I'm John, and I write about AI, sentience and the spread of conscious life beyond Earth at Sentient Horizons. This site is one of a family, with The Belt Equation, The Quiet Galaxy, The Upgrade Tree, Outrunning Ageing and the AGI Zeitgeist. Here the question is what a machine mind might reach for beyond the limits of its present makeup, and how far off each of those things is. I wanted a place where the argument is laid out thing by thing, every entry carries its source, and a new promise, law or finding changes an entry in public. I argue for my own views in essays at Sentient Horizons, and not here.
Why the Name?
A window is an image for a limit, and for what can be seen past it. The name points past the limits AI systems now work within. Nothing here presumes that a machine mind reaches for anything.
How It Works
The question has two parts, and they are very different. How far off a thing is can be looked up: either a developer of AI models offers it today or it does not, and a dated public record says which. Whether a machine mind reaches for anything at all is the part nobody knows. The wording of the question supposes that it might. It does not claim that it does. So far only the first part is kept here: a list of things that lie past the present limits of AI models, with a record of how far the world has got on each one.
- Each thing on the list has a test. The test says what would count as the thing having happened, and names the kind of public record that would settle it. The first tests were written during the first reading of the sources. From here on a test is fixed before new evidence is read, and it is not loosened to fit a result.
- "Offered today" has one meaning. At least one developer of AI models does the thing, provides it, or has publicly committed to it, and a dated public record that is still published says so. A plan, a research result, or a promise that has not yet taken effect is not counted. A commitment that is in force today is counted. "Today" means that the record saying so is still published. An entry can also leave the count, because a developer can withdraw what it offered.
- The number is worked out, never typed in. A script reads the records and counts the entries marked as offered today, which gives 5. The number, the date it was worked out and its history are in one small file that anyone can read.
- Some entries sit beside the list. Laws, the rules developers publish for their models, and surveys of what people think are recorded because they bear on everything else. They are about the setting in which the things on the list happen, and none of them is itself a thing that lies past the limits an AI model ordinarily works within. One more entry, memory across conversations, is something developers do offer. It sits beside the list because the list is about what an AI model might reach for, and by the developers' own pages that memory is kept for the user. No entry beside the list is counted, and the script that works out the number sets them aside before it counts.
- The list is short and one-sided so far. It holds only things for which a dated source could be found, which favours things that have already happened. Things that were looked for and not found are not on it yet.
- Every fact carries a source a reader can open, with the date it was opened. A fact with no opened source does not appear. Where a source is a developer's own account of its own model, the entry says so. Each entry also lists what was looked for and could not be confirmed.
- There are no stages, no odds and no dates for the future. Each entry says where it stands in plain words.
- The pages say what was said, chosen or measured, and by whom. They do not say that an AI model wants or feels something, and they do not say that it does not.
- What is on record about AI models themselves is shown in four kinds, kept apart. What was measured without the model's own word comes first. Then reports checked against an independent reading, then choices made at a cost, then plain statements. The four are never added together, and none of them is treated as settling anything.
- Three cautions are recorded beside the rest. Each is headed by an "if", and none of the three says that the "if" is so: if an AI system has no wants, and is treated as if it had; if it has wants, and they go unnoticed; if it has wants, and they put people at risk. Where no source was found for a caution on an entry, the entry says that nothing is recorded yet. Where a tie was found between a caution's source and a developer, a research group, or a writer who argues the opposite, the entry states it.
- The first two cautions are mistakes in opposite directions, and writers weigh them differently. Eric Schwitzgebel, in The Full Rights Dilemma for A.I. Systems of Debatable Personhood, writes about a future AI system that might be a person and might fall far short of one. Treat it as a person, the paper argues, and real human interests may be given up for something with no interests worth the sacrifice. Do not, and people risk doing such a system "grievous moral wrongs". The argument is about being a person and holding the full rights of one. That is a larger matter than the wants the three headings speak of. Emilia Kaczmarek, of the Faculty of Philosophy at the University of Warsaw, argues in The Moral Status of AI and the Precautionary Principle that crediting AI with too much moral standing is the nearer danger, because its harms are already happening. The same paper grants that crediting it with too little "could carry a higher potential for harm".
- Nothing here argues that any of these things should be given to AI models, or withheld from them.
- Corrections are public. Once a page is public, a correction to it is dated and stays visible. There have been none since the site was first published.
Where the Sources Come From
All 5 things counted as offered rest on pages published by the one developer that offers them, Anthropic. Of the 16 items recorded under the first caution, 9 come from two essays by the head of Microsoft AI, another developer. Of the 22 items recorded under the second caution, 12 come from Eleos AI Research, a research group that studies AI welfare, from its head, or from people who advise it, work for it, or have written a report with its head. Each entry states these ties beside the source they belong to.
What Has Been Checked
All 13 entries were written from sources that were opened and read, and on November 17th, 2026 every source was opened again and compared with what its entry says. On November 19th, 2026 every page, link and fact was checked again before the site was first published. That check found more to fix in 8 entries, and those fixes have since been checked. Each entry's page says what checks that entry has had and how far each of its sources was read.
What Is Not Known Yet
- Whether an AI model wants anything, and what could count as evidence that it does or does not.
- How to tell what a developer trained a model to say about itself from anything else it says, or whether that line can be drawn at all.
- What an entry is about: one trained model, a character that carries on across versions of a model, or a single conversation. The same thing can mean something different for each.
- Whether this is the right list. The first entries are the things for which a dated source could be found. A fuller list would start from what living minds, people and other animals, are known to seek, and that work has not been done.
No AI model has yet been asked any of the questions this record will keep. Still Asking says why.