The distinction between a procedural graph and a static workflow seems important: self-editing topology can capture reusable strategy, but it also makes regressions harder to localize. I’d be curious whether the refinement loop treats a successful trajectory as sufficient evidence, or uses counterexamples and held-out tasks to avoid encoding a brittle shortcut. A practical evaluation might report graph churn and rollback frequency alongside task success, since a graph that keeps growing could be trading inference cost and auditability for a small gain. The explicit entity–relation–procedure representation also seems like a promising place to attach permissions or provenance to tool calls.
What if open-ended agents are overkill for 99% of problems? Let's just take that premise for a second. Most organizations want to follow "best practices" and train their employees to do so. Hiring a genius and giving him total freedom to complete every task is not what companies usually want for MOST things. They want repeatability, reliability, predictability. Especially if they are regulated.
I'm going to drop a bomb over here: what if agents are the root of all our problems in AI safety, cost, and even adoption by regulated organizations? I really do believe this. Here is what I argue:
I wonder if the graph part is a distraction (data representation / syntax), and once you step back, if/how this relates to the wider concept of 'durable task lists' used in this kind of structured dynamic planning.
Ex: Durable task lists generally use more textual representations, eg, a hierarchical text list that supports named references -- so a graph. Likewise, they're mutable, contain statuses, etc. They're fairly popular and AI models at this point have internalized them at this point.
I mean, sentences themselves are graphs of data and context with connectors of various types. Embeddings within these models help form the links to the underlying semantic concepts.
Directly representing things as a graph would likely reduce some of the 'translation overhead', done correctly.
I'm not so sure on the latter -- that introduces a bunch of tool calls, while the text patches are much closer to the semantic space imo and easy for harnesses
If we rephrase this as instruction following alignment, what is the concern here in practice vs a skill? (Which a model eventually internalizes)
It doesn't detract from the research - as a paper, it shows more crisply the structure is useful. I'm just not sure how necessary the encoding, and given durable tasks, novel the insight. Is there new alpha here somewhere, esp given the similarity?
Maybe I missed in when scanning the paper, but did they compare to a simple self modifying harness, e.g. instructing Pi to update some code or skill docs based on the results?
xnx | 15 hours ago
tomhow | 13 hours ago
tizerluo | 14 hours ago
EGreg | 13 hours ago
I'm going to drop a bomb over here: what if agents are the root of all our problems in AI safety, cost, and even adoption by regulated organizations? I really do believe this. Here is what I argue:
https://safebots.ai/agents.html
https://safebots.ai/kimi.html
And here is my overall thesis:
https://safebots.ai/thesis.html
If you do manage to read (or skim) that, I welcome any questions, comments or rebuttals.
guybedo | 13 hours ago
EGreg | 12 hours ago
conception | 12 hours ago
cyanydeez | 12 hours ago
EGreg | 12 hours ago
horacemorace | 8 hours ago
vatsachak | 13 hours ago
lmeyerov | 11 hours ago
Ex: Durable task lists generally use more textual representations, eg, a hierarchical text list that supports named references -- so a graph. Likewise, they're mutable, contain statuses, etc. They're fairly popular and AI models at this point have internalized them at this point.
NBJack | 10 hours ago
Directly representing things as a graph would likely reduce some of the 'translation overhead', done correctly.
lmeyerov | 8 hours ago
I'm not so sure on the latter -- that introduces a bunch of tool calls, while the text patches are much closer to the semantic space imo and easy for harnesses
If we rephrase this as instruction following alignment, what is the concern here in practice vs a skill? (Which a model eventually internalizes)
It doesn't detract from the research - as a paper, it shows more crisply the structure is useful. I'm just not sure how necessary the encoding, and given durable tasks, novel the insight. Is there new alpha here somewhere, esp given the similarity?
schmuhblaster | 8 hours ago