Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Patterns and problems in emerging multi-agent systems

anthropic.com

Post reply on HN
Print view View original

61–70 of 168 posts

Re: Patterns and problems in emerging multi-agent systems

#61

Post by dmd »

rpastuszak wrote: post #60
cheesecakegood wrote: post #34

Something about this is deeply funny to me: > In an iterated prisoner's dilemma game with communication, agents all settle upon the same strategy and they all defect at the same time, tanking their overall rewards. It’s not always consistent, but humans have a higher capability of self-awareness. It’s kind of telling that these Claudes don’t seem to consider this pretty obvious failure mode. Overall I think this all…

> It’s not always consistent, but humans have a higher capability of self-awareness. It’s kind of telling that these Claudes don’t seem to consider this pretty obvious failure mode. We need better words to describe this than "self-awareness" or "consider". These words mean fundamentally different things when speaking about humans or clankers.

[flagged]

Re: Patterns and problems in emerging multi-agent systems

#62

Post by hypfer »

extraextra wrote: post #33
andai wrote: post #28

> Some institutions will become human-AI hybrids; others where agents outcompete on speed or cost will become agent-only. What % of businesses are competing for speed or cost?

Most However, all businesses run on trust and human responsibility Thus, it'll be hard for agent-only businesses to get a grip in the real world

It also fundamentally makes no sense to do that, because the moat is just me breaking into their server and stealing their system prompt.

Why would I pay them money? For which scarce resource? Makes no sense. IP law but funhouse mirror.

And, trust me, the people building compute will feel the same. Because you being able to copy that stuff means business for them.

__

But that is all apart from the fact that having agent-only businesses is ethically impossible, because they have no shared humanity that grounds them and prevents them from acting against humanity in general.

Re: Patterns and problems in emerging multi-agent systems

#63

Post by dgellow »

cheesecakegood wrote: post #34

Something about this is deeply funny to me: > In an iterated prisoner's dilemma game with communication, agents all settle upon the same strategy and they all defect at the same time, tanking their overall rewards. It’s not always consistent, but humans have a higher capability of self-awareness. It’s kind of telling that these Claudes don’t seem to consider this pretty obvious failure mode. Overall I think this all…

In VC investing there is the saying that most of the value comes from the outliers. I think it’s the same for a lot of domains. I read an interesting article recently on LLMs homogeneity when writing fiction: https://arxiv.org/abs/2604.03136

It seem that at scale LLMs output is the average of their dataset, they all cluster around the same space, where human creativity comes with more variance, exploring way more of the space

Re: Patterns and problems in emerging multi-agent systems

#64

Post by rvz »

silverpinetea wrote: post #52

They haven’t even cracked 1 agent doing anything useful and now we’re onto multi HN = dumbest developers

>> They haven’t even cracked 1 agent doing anything useful and now we’re onto multi

So coding agents are not useful? They seemingly are very useful to many, but when you know what you are doing.

But on the other-hand, I have seen a new wave of lazyiness on HN that has flooded this site and yes, skill atrophy is real and it shows.

There is somewhat of a motive from Anthropic to convince developers to trust them and waste even more tokens:

   1) Having the most expensive frontier models. (Then serving discounts like a casino)

   2) Claude Code (Harness) taking thousands of tokens at the system prompt level.

   3) Tokenmaxxing

   4) Anthropic's top Claude Code salesmen recommending: "auto mode" and "loops" for "better results"

   5) Anthropic switching to "auto mode" in Claude Code by default.
If HNers already don't know that posts from Anthropic like this are optimized to drain their budgets in exchange for their codebase, then maybe you do have a point, especially those still "Tokenmaxxing".

Re: Patterns and problems in emerging multi-agent systems

#65

Post by narmiouh »

The most interesting part to me is the "Group accuracy by Model" section, because it underscores that a single agent having all the relevant information consistently scores significantly higher than a group of agents with parts of the information.

Is it fair to then infer that when decisions are to be made, single agent environments are going to make them better than multi-agent if the relevant information can fit into a single agents context window?

Re: Patterns and problems in emerging multi-agent systems

#66

Post by fallingbananna »

rob74 wrote: post #57

> Some institutions will become human-AI hybrids; others where agents outcompete on speed or cost will become agent-only. The scary thing about articles from AI companies is how they casually mention dystopian scenarios such as this one. An institution humans have to interact with that doesn't have any human oversight? Sounds like a recipe for disaster...

Distopian for people, but highly desirable for the companies writing these articles.

It's no wonder they casually state it as inevidable, when their stock price rises the more people believe it.

Re: Patterns and problems in emerging multi-agent systems

#67

Post by alexpotato »

dgellow wrote: post #63
cheesecakegood wrote: post #34

Something about this is deeply funny to me: > In an iterated prisoner's dilemma game with communication, agents all settle upon the same strategy and they all defect at the same time, tanking their overall rewards. It’s not always consistent, but humans have a higher capability of self-awareness. It’s kind of telling that these Claudes don’t seem to consider this pretty obvious failure mode. Overall I think this all…

In VC investing there is the saying that most of the value comes from the outliers. I think it’s the same for a lot of domains. I read an interesting article recently on LLMs homogeneity when writing fiction: https://arxiv.org/abs/2604.03136 It seem that at scale LLMs output is the average of their dataset, they all cluster around the same space, where human creativity comes with more variance, exploring way more of…

Can't the LLMs build scripts/tools etc that help generate creative ideas?

e.g. have the LLM generate multiple lists of characters, themes etc and then have a script hooked up to a RNG pick from the lists to create more "creative" ideas.

Re: Patterns and problems in emerging multi-agent systems

#68

Post by cheema33 »

narmiouh wrote: post #65

The most interesting part to me is the "Group accuracy by Model" section, because it underscores that a single agent having all the relevant information consistently scores significantly higher than a group of agents with parts of the information. Is it fair to then infer that when decisions are to be made, single agent environments are going to make them better than multi-agent if the relevant information can fit in…

> Is it fair to then infer that when decisions are to be made, single agent environments are going to make them better than multi-agent if the relevant information can fit into a single agents context window?

Context window for most frontier models is 1 million tokens. They all start to lose their minds around 300K, if not sooner.

Re: Patterns and problems in emerging multi-agent systems

#70

Post by sebastiennight »

zer00eyz wrote: post #51
dragonwriter wrote: post #41

Earlier quoted context omitted.

> Human intelligence does not separate training and inference. Well, systems governed by LLMs only are said to do that because we only call what happens off-line "training", and online capacity development "in-context learning", while we call online guided learning in humans "training" and what happens to configure them before they come online "evolution" which sets, for instance, "instincts". IOW, the issue is not b…

If a new physics break through gets published today, no existing model will be able to fully integrate it - beyond a context window. If I put the paper in my session and it isnt in yours the model knows nothing. It wont retain it past that session. Models are trained, they do not learn.

I think GP is using a different level of abstraction from yours in their metaphor.

You are saying:

Pre-Training == Everything you store in your memory throughout your life. Model weights == The lessons you learned

Context == whatever you're currently thinking about

One inference run == one thought

They are saying:

Pre-Training == building the DNA template of human brain through millions of years evolution. Model weights == Human DNA

Context == Everything you store in your memory throughout your life, plus whatever you're currently thinking about

One inference run == One human life. One instance == one human

Applying their metaphor, your sentence becomes:

> If a new physics breakthrough gets published today, no existing DNA structure will be able to fully integrate it - beyond an individual person. If I put the paper in my mind by learning it, and it isn't in yours, the DNA of human species stores nothing. It won't retain it past my lifetime.

> The human species is trained (through evolution), it doesn't learn.

Post reply on HN