kamran hakiman

Studying decentralization and collective action in difficult contexts || currently agora fellow, phd candidate at Pitt.

[kamran hakiman @kamran_hakiman]

@holdenmatt I could be wrong, but it seems that, if skills are always available you can otherwise only soft-constrain your orchestrator from triggering a sub-agent persona.

[kamran hakiman @kamran_hakiman]

@holdenmatt So, instead of the subagent inheriting my orchestator's system prompt, it can get a task-specific persona.

Ironically, this is best composed via SKILL.md (i.e. for certain sub-tasks, use a subagent with persona Y (pass it '.agent/personas/Y/PERSONA.md')

[Daniel @growing_daniel]

Statement from Anthropic sounds so much more like a government than our actual government

[Evan A. Feigenbaum @EvanFeigenbaum]

About half of this platform is making some version of the below argument today. I ain't no Carl von Clausewitz and the super-secret American "strategy" to curtail Chinese power still eludes me but I will say this: If the "true" American strategy is to pivot to Asia to curtail the rise of Chinese power, then picking trade fights with every single ally and partner in Asia, withdrawing from regional standard-setting pacts, taking steps that perversely accelerate Chinese planning and might even shorten Beijing's timelines, canceling a sizable chunk of U.S. foreign assistance to Asia, coercing friends, scaling down diplomatic presence, and loudly asserting "primacy" but withdrawing from providing public goods, all while carving out export control exceptions for China to chips and other technologies and fighting wars in third regions, seems like a rather mysterious way to do it. But as I said, I ain't no Clausewitz.

[kamran hakiman @kamran_hakiman]

@polynoamial This is a great write up. To what degree did it seem the models made non-overlapping logical mistakes? Do you see a layered ‘swiss cheese’ model (multiple models must sign off) as potentially useful?

[kamran hakiman @kamran_hakiman]

The same process occurring in Myanmar, as anti-junta resistance forced new, decentralized behavior and institutions to emerge across the country.

[kamran hakiman @kamran_hakiman]

Excellent article on how war, though always a tragedy, reshapes societal trajectories.

https://t.co/H7wMaleRuh

[kamran hakiman @kamran_hakiman]

@geoffreylitt Being able to 'play' with ideas programmatically (before understanding them) is amazing. Seems like DiSessa's dream is very close.

[Steven N. Durlauf @sndurlauf]

Exceptionally interesting paper by @SoerenHenn and James Robinson. This paper and the associated broader research program seek to understand African social/political/economic phenomena through the lens of African civilization and culture, rather than through that of the West.

Here is the key sentence:

"We provide a microfoundation for this based on basic ontological notions of African society, in particular the salience and importance of the community and its institutionalizations, and we argue that the mechanisms that this generates are common throughout the continent."

[Juan Mateos Garcia @JMateosGarcia]

"A vision of Metascience"
I started reading this essay by @michael_nielsen
+ @kanjun and it's great. I love the idea of using metascience as a source of signals to guide decentralised experiments with the social processes of science.
https://t.co/RAT7liWZB0

[kamran hakiman @kamran_hakiman]

@0thernet @ESYudkowsky @So8res If something is actually powerful, understanding + control must be distributed. Moral onus to stop making nightmare bicycles. https://t.co/JXryz4pFM0

[kamran hakiman @kamran_hakiman]

@0thernet @zocomputer that computer looking cozy on the office couch

[kamran hakiman @kamran_hakiman]

@0thernet @zocomputer hell yes. first ai tool i've used getting seriously at this autonomic vision https://t.co/l2TyaPSAXY

[Luke Davis @lukedavis]

Autonomic Systems will be key to the next phase of AI.

This is not about parameter count and activation functions or transformers vs RNNs, it's the vast machinery of ordinary code wrapping the models, taking their output and bending it back as input for the next go round.

[kamran hakiman @kamran_hakiman]

@thisloops would love low/mid-level caffeinated coffee for the same reason. making decaf + caf in one day becomes work.

[kamran hakiman @kamran_hakiman]

@holdenmatt Have you seen analysis if LLMs actually respond differently to structured tags? If so, is it a product of their training on ingesting code with high fidelity?

[kamran hakiman @kamran_hakiman]

@emollick Claude weirdly hates acknowledging gpt-5. Even after forcing a search, if debugging code which calls a gpt-5 agent, it multiple times insisted on blaming the ‘incorrect information’ that gpt-5 exists.

[kamran hakiman @kamran_hakiman]

@pettertornberg Fascinating. Are you concerned that the models were trained on a certain type of social media behavior, and ported that over despite the changes? Potential difference may be that llm agents respond to institutional novelty differently than humans. (Llm training is more sticky?)

[kamran hakiman @kamran_hakiman]

Beautifully said. Applies equally to international development - it is an engineering discipline.

[François Chollet @fchollet]

Because AI is an engineering discipline and not a scientific field, it's never possible to fully separate the properties of a given approach from those of its specific implementations. The artifact is the method.

[kamran hakiman @kamran_hakiman]

Yes. The real damage, though, is over claims laundering this rhetoric into policy and media discourse.

Worse, often these claims are about theory, not the measurement. Vanishly rare for a study to ‘show’ a theory to be generalizably true across a domain.

[Robin Hanson @robinhanson]

The common phrase "studies show that" should usually be replaced by the phrase "studies suggest".