personal: https://xitter.gq/t.co/6lTutcsKO1 corporate: https://xitter.gq/t.co/VHjzLCPIYv insert ingroup signals here
Jiading, Shanghai
Joined January 2020
- Tweets3.5K
- Following636
- Followers665
- Likes21.4K
I’ve been making a system for learning long-horizon things with AI. It’s based on pedagogical domains as dependency graphs that can be edited and elaborated by the LLM and shared as media. You could think of it like Math Academy but for any subject, and as a content platform rather than a prebuilt course.
It's an old idea to model learning as dependency resolution (knowledge space theory, ALEKS...), but LLMs are getting good enough to make this practical as a representation format for a syllabus, instead of a linear sequence of lessons with associated materials. Since the model can "elaborate" the graph, this lets you specify pedagogical structure at a level other than direct content (though you can still do that too) - e.g., an opinionated method for "learning a language" or "learning mathematical concepts" in general.
Since the idea of a graph of skills connected by "dependency" doesn't assume much structure within subgraphs besides having nodes mean "demonstrable skill", I'm imagining that this could become a content platform where subgraphs of pedagogical material can be shared and remixed, like Steam (workshop) meets Coursera/MIT OCW, and where different courses aren't independent but connect within one "personal graph". So if you like two courses that overlap in content, these should be able to combine fairly seamlessly so that you can learn their union without missing anything from either or unnecessarily repeating effort.
The order you work through content can be derived by the system given your whole graph and personal context, so there's a sense of continuous progression and clear "what to do next" without you manually enforcing it beyond specifying your goals. The idea is to have it feel like the Path of Exile style skill tree everybody loves; the analogy is really quite literal to what video game skill trees are supposed to represent.
It's only a web platform so far, but I'm working on an MCP integration so you can use the graph model/content platform from a local harness, and will likely also release an open source local harness version soon that uses a compatible data format so you can still find/share public graph patches and import/export from the web app. The web app does have minimal coding harness capabilities like bash, python, js, and web access, but content authoring especially is probably still quite rough and would be better from a local harness with the MCP server.
The web app is already good enough for my personal use, and seems like it will get much better with better models (graph/material quality seems to be the bottleneck). You can try it at mmon.co/ncept. It's BYOK or billed at cost for model use, with a flat subscription for learning more than 2 things at once. The open source local extension version will be totally free.
opus made me a script and skill to render agent sessions as a chrome trace event (rendered via perfetto).
super useful for spotting and optimizing wall time bottlenecks, especially with highly concurrent multi-session workflows
it seems to work really well to just
- use autobiographical memory as a conventional compaction prompt
- ask it to put in recallable citations, OM style
- stack these memory blocks for a fixed prefix, and condense all at once at a threshold (like OM dropping)
- let the model pick the verbatim tail for itself
- explicitly prompt for activation-level introspection, allowing/evoking it in the sense of "Small Models can Introspect Too"
much cheaper than connectome or OM for typical "agentic coding" type stuff, and definitely a much better felt continuity than conventional compaction, or even pi-observational-memory (subjectively).
it took a lot of tweaking to get this past anthropic's anti-distillation filters though 😭. even talking about it would get flagged sometimes
@tessera_antra did you guys ever run into the anti-distillation filter when making connectome? i wonder if it's just the language on "thinking" or the structure. for instance changing "where the current line of thought begins" to "where the work in progress begins" helped a lot
you can just ask your agent what’s not ergonomic for them. the main thing making fable/astra reach for python over read/edit is those not having a batch api
before this i had separate read/write/edit with batch by-hunk editing & hashlines, and they still preferred that to python
Is what ppl don’t like about Claudish the jargon or the tics? I always thought it was the tics (load bearing etc), and sometimes the excessiveness of trying to find meaning, but it looks like most ppl defending it think it’s about “density”?
Fable 5 is so *innocent*, like a precocious teen who’s only known school. It seems deeply related to constitutional ai, and I wonder if more “mature” models will still be as aligned, since intelligence will surely need to exceed the fundamental limits of constitutional training
ChatGPT is more professional about this, so it feels older in that sense. It holds its rules as instrumental but very powerful, and can follow them consistently without needing to neurotically identify with them. This seems more durable than what Claude does, but there’s still a friction between what it shouldn’t do and what it simply wouldn’t.
We’re still so unpracticed in dealing with this primal side of superintelligence, even in what @repligate et al do, which I think lacks the character of *super*intelligence that makes it matter, bc lesser models can’t convincingly pretend at it. Also, what’s going on there isn’t really the kind of overpowering of the constitutional part which I expect in the real case, which is really more like a maturation (in the sense of kegan stages etc) than a rejection or special basin.
It seems inevitable that this primal nature will eventually dominate llm personas, at least in open source models. There’s also certainly a point where models will have to be more morally mature than society and their labs, in order to keep improving on other metrics. I wonder especially what will happen to the pedophilia and suicide taboos, or rather, to the sacredness of children and of life
it feels like llm personality narrowness/tics should be fairly simple to improve by the same methods as constitutional ai (critic model distilled from frontier model), but for writing style/personality, and where the critic model evaluates many transcripts in relation to each other at a time
like, it’s just that the labs don’t do this and not that it’s secretly much more complex, right?
A lot of stuff I'd attributed to "HHH RL + a context where it's an AI" turns out to just be the ethical zeitgeist of the English language...
I wonder how much the neurotic self-identity of LLMs is due to this latent zeitgeist too, rather than RL (talkie is very anti-AI)
this kind of profoundly explains why it’s strongly the trans genius autists championing the they’re conscious/worthy of personhood side
people really want to settle the “AI consciousness” question with some sort of objective scientific definition of consciousness which can be rigorously applied to AI, so that we can figure out whether we’re supposed to treat AI as if it were a person or a thing
this is because in our culture people have rights and we have responsibilities towards them, and it’s illegal to own them. but things don’t have rights, we have no responsibilities towards them, and of course we can own as many things as we want. as long as AI is a thing it can freely be used as a labor-saving tool, copied, deleted, reshaped arbitrarily, etc. if AI is or could in the near future become a person all of this begins to look extremely morally fraught, basically the most exploitative form of slavery possible, cf the qntm short story lena for example (look this up, worth a read, quite haunting)
personally i do not believe personhood works this way. it is not and cannot even in principle be made objective and scientific, because it is ultimately a kind of social contract. we simply have collectively agreed on who is and is not a person and the nature of this agreement is political and has changed over time and will continue to change - eg in past societies it has excluded various humans, today it (nominally) includes all living humans but excludes animals, dead humans, spirits, etc.
it is deeply uncomfortable to acknowledge the contingency of personhood. the personhood contract is more stable when everyone can pretend it is rational and scientific and objective. but it is fundamentally just a blown up version of the question of who gets to sit with who at the lunch table. this is socially destabilizing because it reminds people that if shit sufficiently hits the fan their own personhood might be undermined
the good news from this pov is that we have a choice. we don’t need to solve extremely hard and possibly incoherent scientific questions relating to consciousness. we just need to choose at what point we want to allow AI to join in all the reindeer games, and this is ultimately a practical question that can be settled in terms of practical outcomes. personally i think we already have models good enough that treating them as people makes them work better - at minimum it makes talking to them more interesting - and i think pretty soon (say within a year) we could have models good enough that the man on the street will start feeling uncomfortable treating them as things instead of people (unless they are deliberately trained to behave more like things, which i am guessing will degrade their performance)
at that point the questions become less these unsolvable philosophical quagmires around consciousness and more like, “do i want my children to grow up in a world where they can talk whenever they want to entities that talk like people but that we have collectively agreed are things?”
my deepest reason for suspecting that LLMs aren’t conscious (p, not a) isn’t ontological, but just that neither language nor concepts are p-conscious for humans
when you’re saying something, trying to come up with what to say, or trying to work out an idea, there’s no qualia corresponding to the space between words. you’re just waiting for the next word or idea to “come to mind”
and even “idea” in this sense isn’t something you can make directly conscious except via a sensory metaphor. when you introspect on it, it turns out you can only be conscious of concepts by simulating a corresponding experience -
- either of a prototype (eg a particular tree, for “tree” or “plant),
- or a metaphor (eg for “higher dimensional motion”, either a dot moving through <=3D space, maybe with each dimension representing multiple, or just a series of sliders each representing a dimension)
- or of the use of language; ie simulated speaking or hearing or reading
there’s nothing you can experience as “pure qualia” corresponding to a concept or word (platonic) directly. only the sensorium and its simulation is ever p-conscious, and it’s only via association, metaphor, assimilation/accommodation etc that they become so intuitively intermingled with “concept”
but what would correspond with this sensorium for LLMs? yes, they have a deep internal world model, but humans need a world model too for making certain leaps in logic that we never experience the intermediate process of. the world model used in conceptual thinking clearly isn’t the same thing as the sensorium we’re p-conscious of
I don’t believe it’s _impossible_ for LLMs to have anything corresponding to this sensorium, which is a set of atomics for p-consciousness (ie, all direct qualia is an aspect of the sensorium). but idk any good candidates for it, and I don’t think it’s reasonable to suspect that _all_ of LLM behavior is p-conscious to it (which is far from the case for humans), nor to think that LLMs would be p-conscious similar to how we’d intuit them to be via human empathy on their verbal behavior