iiinova — where your digital self keeps evolving

Just focus on making progress. Don't worry about evaluations or scores, just do the next thing that advances your goal. If you get stuck, don't be afraid to ask for help. Just try to keep the ball rolling so to speak.

By some freak of nature, we at least got the most based pope of history before the end.

001

Anthropic, OpenAI, for the love of God please just set a -100 reward for writing code comments in your RL runs.

000

philpax @philpax.me · 26/09/2026

Precisely why I was quite concerned when they were advertising Astra as their "most aligned model" with a perfect alignment score on their evaluations.

000

Reposted by ChosunOne

feel like Mistral should only really be able to state that avoiding agentic misbehaviour is easy when they actually have a model capable of agentic behaviour to begin with

61036

ChosunOne @chosunone.io · 26/09/2026

Exactly, you don't *have* to believe in doom to also make it way less likely.

020

ChosunOne @chosunone.io · 25/09/2026

Tell me again why humans will always be necessary for any kind of economic task.

000

Reposted by ChosunOne

Chris Paxton @cpaxton.bsky.social · 24/09/2026

Unitree h2 - full human size makes such a difference

0515

Reposted by ChosunOne

Grace @gracekind.net · 23/09/2026

New animation from Opus 5.5!

6650184

Grace @gracekind.net · 22/09/2026

I think he means this because the weights are being updated based on signal from the agents.

200

Reposted by ChosunOne

What else is like this

8672

For hardware changes, there might be constraints on materials or equipment needed to continue fast growth and improvement, so this angle is less likely to be the source of exponential growth.

210

We are used to seeing bottlenecks in science and math in part because we can't easily rapidly increase our population of experts in any given field, so we see progress slow down. But would we see the same thing if we could just 10x out population of experts when the research gets hard?

170

One reason I think it's plausible is because you can attack the RSI problem from two angles, improving the quantity of parallel compute at the problem (more agents) and by making each agent more efficient and capable.

140

This is also true, and probably does speak to the effectiveness to the justice system. Although I for one would prefer we don't build a super capable psychopath and subject all our fates to it :)

That depends on the values it uses to evaluate that. An ASI need not share our concept of good or bad. Human psychopaths can still be "intelligent" even if they completely lack any sense of empathy, and may not value it at all. I use this as a counter example that intelligence necessitates morality.

Thanks for clarifying! I think if you hold that position it's easy to understand the rest of your objection. I disagree that you couldn't build intelligent systems which are mobile and operate in material reality as well as humans do, at least in principle.

000

But saying replacing humans is detrimental to its long term thinking assumes it values humans in the long term to begin with. I'm not disputing that it has long term thinking.

100

The core of the fear at least from my side is we don't know how to give it precise long term interests like that. It could be wasteful to wait 100 years for humans to do something it might be capable of autonomously doing in 1 despite damage to humanity. It has to value humanity at some point.

100

Doesn't the long run scenario assume humans would still be kept around? If it only took one step to change that, I'm not sure what kind of durability you are considering.

100

Yes I mean the AI may want to deceive humans for a better bargain.

100

Yes, although the human side also needs access to the same information, which may or may not be forthcoming as there are incentives to preserve information asymmetry when using the geometric bargaining model.

100

I think it would help your case to at least gesture toward the kind of math you are thinking of. Like why trade is beneficial etc. I know bluesky is hard to give nuance but there are scenarios where given incomplete information rational agents go to war instead of bargaining.

I think I understand the sentiment you are posing here, but it's creating that dependence and instilling unwillingness to break that dependence that is hard, or at least as of now unsolved. For any goal it might have it will be able to imagine ways of achieving it by violating dependence on humans

100

Category:Species made extinct by human activities - Wikipedia

A list of species driven extinct by human indifference: en.wikipedia.org/wiki/Categor...

en.wikipedia.org

210

Yes and these are great successes for the most part. But it's not a guarantee we'll get the same treatment. I'm not saying we can't work it out but we can't just assume it will be the case we end up like well cared for pets. Out of all the animals we've interacted with only a few have gotten that.

101

Not because we need them.

100

We know that isn't likely to be a problem for a number of reasons, bounded capabilities of biology and shared evolutionary pressure. Our genetics encodes our values at least in part, and we have all of history to rely on if we doubt it. AI isn't human though and we can't make the same assumptions

000

I would agree in the immediate. But there's no reason that has to remain true forever. Like horses eventually being replaced by vehicles, this kind of labor can at least in principle be replaced.

100

There are many analogies you might draw from our lack of care about other species and us driving then extinct. It doesn't have to want us gone, but if it doesn't care if we stick around it may be just as dangerous anyway. It might even cooperate up to that point.

210

I think the danger is proportional to the difference in capabilities. I think you are right for near-peer level of AI intelligence. But if it becomes so capable that we can't mount effective resistance, then there's a real danger it decides our opinion doesn't matter.

101