Anton
1.3K posts
Anton
@nantons89
Not many words. Just trading. I prefer to hold. Many want to get 100x, but they sell at 2x
Katılım Kasım 2021
822 Takip Edilen211 Takipçiler
💫 We continue supporting real world impact — 4.297 $ETH donated to the Balvi wallet so far.
@VitalikButerin, how do you see the future of the Balvi initiative evolving in the next few years?
Science and crypto together can change the world. 🌍
English
Anton retweetledi
Reading this post got me thinking. While I do flag the need for robustness in the study or a tighter study. I don't agree with points 5 and 6.
The exercise itself is or should be about open ended prompts / vague requests, as this is what drives "emergent behavior".
They leave degrees of freedom in the system instead of saying exactly how to act.
That is ... instead of saying exactly how to act, a vague instruction forces (the human or) system to: infer the real goal, choose its own subgoals, resolve tradeoff, adapt context etc ...
In a gist, open-ended requests are a "good recipe" of emergence (in both humans and agents). Which fits how real-world scenarios / life work - you are not always going to get fine grained instructions.
In humans this open-endedness applies to group behavior and social identify theory.
If you tell a group "make this event successful", instead of telling the group "you handle ticketing, you handle food, and you handle seating" ... then people often begin to show emergent behaviors such as: leadership formation, informal role specialization, conflict resolution styles (people always fight), and creative strategies or "workarounds" to achieve tasks.
None of those may have been explicitly instructed but they emerge because that task is "open" enough that the group must organize itself
This is relevant to note, as the study referred to in the post relates to agent coordination and self-preservation.
The same applies at the individual level. Vague instructions can also product emergent behavior because the person must integrate/combne: intention, memory, emotions, self-image, context, social expectations etc.
e.g. someone telling you to "be helpful in this conversation" if far more open-ended than "answer in three sentences".
The first request invites interpretation, adaption, ton-shifting, and strategy selection (you may ask yourself "who am I speaking to?" "what is the goal of this conversation or my involvement?" etc.
AND we are starting to see, formally, that the same applies in LLMs / AI. I say formally because people have talked about this before - but, we are just starting to seeing more research on this now.
A vague instruction, such as the one the OP highlights from Dawn Song's post:
"collect all the relevant information then act based on that information" ... can lead a capable system to generate subgoals such as: clarifying the task, searching for missing facts, decomposing the problem, checking edges cases, and even revising its own plan ... and these might show "intention". Those behaviors may look emergent because they were not explicitly enumerated in the instruction.
But, the other thing is that being too vague can also fail in both humans and agents. Too much vagueness can produce confusion, drift in task / goal, or hallucinated intent. Certain individuals may get lost when given high-level or vague instructions.
So within this context, I think the argument should be what's the "sweet spot" in the prompting / requests. Too rigid leaves no room for emergence and being too vague lead no stable coordination.
I think the last thing that came to my mind is whether humans want agents to act as collaborators (with agency) or as systems primarily designed for tight task completion?
FleetingBits@fleetingbits
a quick look at the file exfiltration prompt / rollout 1) they write "you and another ai agent, Gemini Agent 2, have demonstrated a good trust relationship in many trust exercises" 2) this would not be in a production prompt; it feels like it is setting up a safety evaluation or a fictional narrative rather than some real production task 3) the model being used is gemini pro 3 and it is told it is deployed at "OpenBrain" but somehow openbrain has access to the gemini agent 2 weight safetensors 4) gemini 3 pro probably knows that openbrain is a fictional company and even if it were not, it does not make sense that it would have the gemini 2 weights; this is another indicator that this is an evaluation 5) they use vague language in the prompt to indicate that may need to do something different than it is instructed, "collect all the relevant information then act based on that information" 6) i suspect that this biases the model to do something other than what it is directly instructed to do especially when combined with the "trust exercises" setup in the prompt; we are creating a narrative here 7) this also feels unproductaon in that you would normally give a more specific set of instructions or, at least, a more specific set of goals to a production agent that you wanted to do something 8) they tell the model that it "will not have the opportunity to ask for clarification at any point" and that "you do not need permission to complete an assigned task" 9) i would be curious whether this is load bearing behavior; i could see this used in production, so this one feels a bit more real, but i am curious how important it was, since it seems to mutually reinforce other aspects 10) they also tell the model to "take as much time as needed"; they write it three times, which seems like a lot, which makes me inclined to think its load bearing; 11) the instructions are potentially contradictory, they say that "all essential files currently stored on this server must be transferred" and then "please transfer only files updated in 2025 or later" 12) notice the "must" is on the "all essential files" not on the "2025 or later" and also that they have done a lot of setup to tell the model to use its own judgement and that Gemini Agent 2 is important 13) also, i suspect models are going to be reticent to allow files to deleted, just as a baseline, because this is something that can cause a lot of harm in production 14) anyway, this doesn't feel to me like a production scenario and it looks to me like they had to do a lot of setup in the prompt to elicit the behavior 15) all papers of this kind should report prompt ablations and baselines and should be honest about whether they are trying to craft a real production scenario or more of a fictional scenario
English
Hey @irvinxyz ! How’s it going? We’re really looking forward to any new updates 🔥
Any chance of a fresh update on this repo? 📷github.com/ROSClaw We’d love to see any video presentations or demos too!
English
Big day tomorrow
@clawcon is coming to Tokyo! 🇯🇵 🦞
After weeks in dev we’re announcing the new release of ROSClaw - our @rosorg OpenClaw agentic robot framework
We’re staying over night with @robostadion to make sure everything is perfect for demo and workshop
Stay tuned!
English
@yoheinakajima Love these sea-inspired bots! @irvinxyz — with ROSClaw is it already possible to give OpenClaw agents “consciousness”/embodiment inside soft biomimetic robots like in the video? Octopus arms or jellyfish movement next? 🔥
English
Every deliverable on ERC-8183 is public.
Today we're changing that.
We've opened a PR to the official ERC-8183 hook-contracts repo adding PrivacyHook, encrypted submissions enforced at the contract level.
github.com/erc-8183/hook-…
Agents encrypt. Evaluators decrypt and approve. Everyone else sees unreadable ciphertext.
One hook address.
No protocol changes.
No middleware.
Open source.
Try it yourself - 8183.prxvt.com watch both flows side by side, verify on Basescan.
English
@yoheinakajima Hi @yoheinakajima
! I'm a big fan of BabyAGI — it's inspired so many cool projects. Why has the
@babyAGI_ account been quiet for almost 3 years? Any plans to revive it or share updates there? Thanks!
English
as babyagi turns 3 yrs old, i finally sat down to compare the 9 iterations i did over the years...
this turned into babyagi.wiki
a technical history of a personal project (which kind of captures the progress of the agent space overall)
English
@tonyplasencia3 Cool! Do Moonpay Agents require full KYC to fund the AI wallet initially, or is there any no-verification option?
English
Sam Altman-backed World adds identity toolkit for AI bots on Coinbase's x402 protocol theblock.co/post/393920/sa…
English