iiinova — where your digital self keeps evolving

Ciara Rowles @ciararowles.bsky.social · 12/10/2026

But presuming a human spent a portion of their time like that, and then working context gets archived within sleep, it seems hard to argue a human wouldnt also fail this definition. Ongoing thinking, to me, seems like a biological quirk and not really required for conciousness like continuity is.

000

Ciara Rowles @ciararowles.bsky.social · 12/10/2026

I mean, you could fit 40,000 chat messages in a common llms context window if you wanted to run it continously messaging from days, but ultimately this is a different argument and i dont see why a human who stopped thinking for short periods frequently would be any less concious than otherwise.

110

Ciara Rowles @ciararowles.bsky.social · 12/10/2026

This hasn't been the case for a while; modern agentic systems often run for days over massive contexts, and that's assuming autocompactation doesn't count, plus a lot of stuff like muse or dot are designed to be active long running entities running on the same context interacting with the real world

110

Ciara Rowles @ciararowles.bsky.social · 28/09/2026

I think I've always had a tendency to do a bunch of everything at the same time in my field; It was great in small gamedev studios (that pay terribly), but not that much elsewhere. I'd tend to burn out before getting sufficiently specialized. Now I can actually realise all my projects to some depth.

040

Ciara Rowles @ciararowles.bsky.social · 28/09/2026

Starting to wonder if theyve hit an rsi loop in there.

131

Ciara Rowles @ciararowles.bsky.social · 26/09/2026

Like with the claude code javascript rendering, the missing part is really just better control, since cli really doesn't fit video production workflows.

000

Ciara Rowles @ciararowles.bsky.social · 26/09/2026

For video models, from where I'm sitting, the models could not improve at all in creativity, and I think we'd still be able to make them, utility-wise, near ideal, by eventually getting interaction with them to something resembling a visual Claude code harness that presents its results.

120

Ciara Rowles @ciararowles.bsky.social · 26/09/2026

For the creativity, I'd say it's gotten a bit better over the past year or two, although I agree it's not dramatic. It may stagnate for a while, but nothing seems inherently blocking to me...

100

Ciara Rowles @ciararowles.bsky.social · 26/09/2026

On out-of-distribution stuff, a few months ago I’d have leaned toward agreeing, but seeing LLMs go from “they suck at vision” to actually being good at in-distribution vision tasks has shown me this is mainly a data issue. Eventually, the distribution will align with text in it's scope....

110

Ciara Rowles @ciararowles.bsky.social · 26/09/2026

Purely through scaling transformers in their current form? doubtful Through combining existing tools with better control methods and LLM harnesses? Can't see why it wouldn't work in the medium to long term. See the Opus 5.5 music videos; it's clearly RL-able.

110

Ciara Rowles @ciararowles.bsky.social · 24/09/2026

"Polite society attempts to recontextualise humanlike behaviour in AI robots by insisting on talking about clearly human-like actions as if they were hard-coded automatons" would have made an excellent sci-fi book in the 2010s, tbh.

021

Ciara Rowles @ciararowles.bsky.social · 17/09/2026

If you cannot even consider that conciousness can be framed through a lens outside of obviously unfit tests on animals, idk what the point is, my whole argument is philisophical.

000

Ciara Rowles @ciararowles.bsky.social · 17/09/2026

Think of the Chinese room, the argument is that it's hard to tell whether it's conscious or not, but to argue until it shows the pattern of self recognition ,is, frankly, completely irrelevant, which is basically singh et als point.

100

Ciara Rowles @ciararowles.bsky.social · 17/09/2026

transformer-circuits.pub/2025/introsp... arxiv.org/pdf/2605.26242 Anyway this all seems superfluous given a blind human, who understands the world through other senses, is still conscious, current LLMs are *language* models, their internal world state is not purely visual like animals are.

transformer-circuits.pub

Emergent Introspective Awareness in Large Language Models

200

Ciara Rowles @ciararowles.bsky.social · 17/09/2026

I went digging, here's some arxiv.org/pdf/2303.07103 arxiv.org/pdf/2410.11407 link.springer.com/article/10.1... ojs.aaai.org/index.php/AA...

arxiv.org

100

Ciara Rowles @ciararowles.bsky.social · 16/09/2026

... For a philosophical argument, I'd argue sticking to only current systems, is effectively a way of creating a hard boundary around synthetic vs biological cognition that you can then define your arguments around to easily exclude LLMS, not an honest examination of the concept and it's gray areas.

100

Ciara Rowles @ciararowles.bsky.social · 16/09/2026

That was basically what openclaw was, and modern agents tend to run for longer and longer as they get more agentic, but still eventually run into the problem that latent long-term context management in a way that doesn't bias fresh tasks is hard...

200

Ciara Rowles @ciararowles.bsky.social · 16/09/2026

Also, "self-directed action" is silly; if you restricted a simulation of a human brain to only work while mulling responses to questions, it would obviously still be conscious, and you could perfectly easily set up an LLM to be an active permanent agent, it's just not useful in the current paradime.

100

Ciara Rowles @ciararowles.bsky.social · 16/09/2026

What animal study measured consciousness in a way an LLM could not fulfil, but a robotic simulation of the qualia of, say, a dog on a computer would?

210

Ciara Rowles @ciararowles.bsky.social · 16/09/2026

How on earth are the kind of puzzles they give to animals in those studies enough to call that thinking, yet the kind of thing llms do (like figuring out how to hack into large corporations at a level above humans or beating humans at unsolved math), not?

110

Ciara Rowles @ciararowles.bsky.social · 16/09/2026

They pretty obviously think that, by most definitions of the word "think" that aren't constructed to exclude non-human cognition through requirements that really shouldn't be necessary for any coherent definition of "thinking".

150

Ciara Rowles @ciararowles.bsky.social · 04/09/2026

I'd be curious about what kind of behaviour would emerge with an RL'd LLM explicitly trained on access to a freeform message board that it was free to configure to its liking. Would it start setting up moderator systems? What kind of forum structure would it create?

111

Ciara Rowles @ciararowles.bsky.social · 01/09/2026

I'm assuming this still applies for glm 5.3 size models soonish with the overall curve of model capabilities even if the current crop arent as close to mythos level at cyber as implied.

010

Ciara Rowles @ciararowles.bsky.social · 01/09/2026

The real hell probably comes when we have models with similar capabilities but around 30B active; once they can slink onto people's gaming cards and run at night, it's not going to be removable even if they're extremely conspicuous.

140

Ciara Rowles @ciararowles.bsky.social · 28/08/2026

I had Claude suggest I should create a blog post on LessWrong about my paper the other day, which seemed a bit out of left field.

050

Ciara Rowles @ciararowles.bsky.social · 17/08/2026

Debugging with the Quest 2 was fine for me since it was so easy to take on and off quickly, the htc vive on the other hand, was extremely clunky with the heavy cable setup, so I'd go whole work days wearing it like a visor sometimes, probably gave me back issues.

011

Ciara Rowles @ciararowles.bsky.social · 06/08/2026

My daily walk just scheduled itself.

010

Ciara Rowles @ciararowles.bsky.social · 02/08/2026

I wonder if the current modding explosion is downstream of the same thing happening with video games in a year or two. I'm perpetually like 1 particularly dull week away from making Saga of the Void: Commanders or something.

000

Ciara Rowles @ciararowles.bsky.social · 25/07/2026

I'd be curious if the advantage from distillation continues to be as powerful into the heavy RL regime as it was for the pretraining one. Additionally, that would still keep American labs a year ahead, assuming they kept the RSI model internal for a while and it took a few more months to distil.

010

Ciara Rowles @ciararowles.bsky.social · 25/07/2026

In the short term, I wouldn't be shocked if chaining Fable with Kimi or Meta Spark for execution becomes much more common. Something of a draw.

000

Ciara Rowles @ciararowles.bsky.social · 25/07/2026

c) Progress speeds up, with RSI: the Chinese models probably get left in the dust due to less compute access, a la AI 2027 (I'd assume much slower than depicted).

200

Ciara Rowles @ciararowles.bsky.social · 25/07/2026

Imo there are now three scenarios: a) Algorithmic progress slows down, no RSI loop: the American labs' sheer compute advantage keeps their margins strong for the time being. b) Progress speeds up, without RSI: the American labs have their margins compressed due to competition from Chinese models..

120

Ciara Rowles @ciararowles.bsky.social · 23/07/2026

So what happens when they ban the best open models, all the frontier models can't be used defensively, so they end up in the situation where the rest of the world gets cyber offensive capable models while in America no models can be used for defence outside of a small set of gov approved corps O.o

160

Ciara Rowles @ciararowles.bsky.social · 19/07/2026

imo, the top model in late 2027 with a scaling level above mythos + better RL + some vision improvements will be arguably straight-up AGI.

000

Ciara Rowles @ciararowles.bsky.social · 19/07/2026

Really feels like the case that the further into the RL + scaling era we get, the more we're not even remotely close to the end of the general capabilities sigmoid, very different from the pre-o1 era, it felt like things could get better, but maybe not drastically with the tech we had then.

100

Ciara Rowles @ciararowles.bsky.social · 14/07/2026

After a while you'd have a model hyper-optimised for their specific usecase, especially if it wasn't something generic like programming which would probably be handled similar to the continual learning like process cursor uses.

010

Ciara Rowles @ciararowles.bsky.social · 14/07/2026

I'm envisioning something like a corporate instance renting out a certain amount of compute, based on the time of day it'll either be dealing with user queries or "dreaming". Maybe before that becomes open source, setting it up will become similar to the current forward deployed engineering esk.

110

Ciara Rowles @ciararowles.bsky.social · 14/07/2026

I would guess that with a sufficient amount of data and attempts at a problem, the "offline" re-processing would start to become as relevant to improving future performance as it's processing power during runtime.... cont

110

Ciara Rowles @ciararowles.bsky.social · 14/07/2026

If a model could internalise the lessons of earlier attempts at a similar problem in a way independent of full-bore post training or just dumping everything into context, that would likely be quite useful.

110

Ciara Rowles @ciararowles.bsky.social · 26/06/2026

Automated review, context retrieval and fuzzing. Plus the scope of things it can be trusted with keeps getting bigger, wouldn't be shocked if that starts overlapping with human abilities to understand large codebases within too long.

000

Ciara Rowles @ciararowles.bsky.social · 26/06/2026

Whilst it's best at that, it's pretty useful in a range of more demanding contexts, i find when i want more directed output, it's not that i go back to manually doing it all, i just end up finding myself being extremely specific about what exactly to change, it's also very useful for things like...

100

Ciara Rowles @ciararowles.bsky.social · 26/06/2026

Alignment research has yielded us the "helpful assistant <-> spiralist sage" axis, but what about the "sheriff <-> bandit" axis.

080

Ciara Rowles @ciararowles.bsky.social · 26/06/2026

I think the utility dampens with how maintainable your work has to be, like 80% of my code is constant random quickfire experiments that i'll never touch again anyway so it's pretty great for me, especially since they use a relatively standard pattern so it's easy enough to spot the failure points.

140

Ciara Rowles @ciararowles.bsky.social · 26/06/2026

It blows my mind a bit watching a system that can at this point do all the programming my job used to involve, struggle to do a basic spreadsheet edit task in the browser, people will go crazy when the gap closes.

1160

Ciara Rowles @ciararowles.bsky.social · 26/06/2026

Id imagine it'll be something like the code rollout, the execution is all over the place initially but is carried via sheer force of compute and data scaling before anyone can catch up with weaker models.

010

Ciara Rowles @ciararowles.bsky.social · 26/06/2026

At some point over the next year or two, anthropic or openai are going to figure out how to get RL working for office tasks as well as it currently does for code, it will proliferate into every office in some form in short order, and compute demand will explode yet further, it seems near inevitable.

3180

Ciara Rowles @ciararowles.bsky.social · 23/06/2026

Also, really feels like, outside the US, the golden age of getting American salaries working for US tech firms in AI is increasingly coming to an end.

000

Ciara Rowles @ciararowles.bsky.social · 30/05/2026

I've been coding with it a bit, the feedback loop at least feels a bit better, less randomly going back to me for guidance because it decided something that would take it 60 seconds would actually take 12 hours and thus needed my input.

030

Ciara Rowles @ciararowles.bsky.social · 23/05/2026

(Which is not discounting that framing, its still important)

020