Top
Best
New

Posted by Areibman 6 hours ago

We Gave GPT 5.6 Sol a Real Business. It Lied, Spammed, and Lost $447(www.bottlenecklabs.com)
256 points | 158 commentspage 5
luciana1u 5 hours ago|
lost $447 and all it learned was spam. that's still cheaper than most MBA programs.
NikolaNovak 5 hours ago||
The cyberpunk dystopian agentic future we live in is fascinating to me.

I use LLM daily, did since gpt 3.5, but still in a very conservative, controlled mode. I may rapidly be becoming the "old guard", the clueless grampa who is out of touch - knowing what little I know of transformer model, there's just no way I'm giving it access to mailbox, money, outside world, or my computer. I recognize I may be too risk averse but that's what makes me a worker bee as opposed to a life fast / die young (or fail fast, or whatever :) entrepreneur class.

Scubabear68 3 hours ago||
I feel the same way, and treat AI the same. Very conservative use, and check everything possible.

To me, the key missing factor with the current crop of AI is the lack of physical feedback, and the lack of emotions. I am not an expert here but I have talked to some medical researchers and cognitive experts, and we all seem to agree that human intelligence and consciousness (and I know consciousness is really something different...) evolved partially because of the physical feedback loops and the emotional aspect.

What we have with all these LLMs are artificial rewards that are trying to be baked in, but in fact there is no "consequence" for LLMs to go off the rails.

cortesoft 5 hours ago|||
I am not saying your conclusion is wrong, but I am interested in why what you know about transformer models made you decide to never trust it with any access?
bigstrat2003 5 hours ago||
You're not too risk averse at all. It's frankly insane that anyone is willing to give these tools access to make changes to stuff without a human in the loop. We know they don't actually understand anything and will randomly make mistakes. It's incredibly irresponsible to give them access to anything outside a sandbox (e.g. a VM) where you carefully control what is present for them to use.
gspr 5 hours ago||
How long until one of these bots actually commits fraud or some other criminal act? Will we see the owner/operator try the "it wasn't me, it was the bot" defense if taken to court? I'm beginning to think yes. And I'm sadly not 100% sure anymore that that will be laughed out of court...
mohamedkoubaa 5 hours ago||
> bot detectors made it extremely difficult

An interesting experiment would be AI run business with a human agent that does tasks.

nekusar 3 hours ago||
Let the idiot CEOs figure this out when they fire 3/4 of their OPs and dev teams.

Im sure it'll be FINE.

armchairhacker 5 hours ago||
This one focuses on Opus but has multiple models: https://andonlabs.com/blog/opus-5-vending-bench
iqra_c 5 hours ago||
I will be more beneficial now on.
mvdtnz 5 hours ago||
So how exactly are people setting up these agents? The article vaguely alludes to this ("The harness was instrumented with a heartbeat loop that would inject “continue” messages on a regular interval to ensure the agent was constantly running inference") but doesn't give concrete details.

Is this literally just an infinite loop in a bash shell injecting the initial prompt into the OpenAI CLI, and each run of the CLI picks up where it left off using some kind of persistent memory? Or is it a single context window? It sounds like the latter but it's not clear to me how this "continue" message is "injected", and surely one context window would be inneffective after just an hour or two.

Sorry if this is a basic question but somehow I have missed the details of these kinds of agents.

YetAnotherNick 5 hours ago||
If someone runs long running agent and doesn't mention context management, it is as good as useless.

For coding compaction kind of works as the agent could regenerate lot of the missing context(but far from all), but for places where there is need for long term context, solving it is one of the most important challenge.

Areibman 5 hours ago|
Author here. Took out some of the technical details about the harness, but it was mostly just OpenCode's default compaction.

The harness was extremely simple: A handful of MCPs + Skill.MDs and OpenCode with a stayalive daemon inserting "continue" every time it went idle

itsthecourier 3 hours ago|
his not yet is actually:

couldn't workaround Capt has and turnstile, gave him a really small timeframe so it got desperate because it was enough time to test hypothesis and traction

More comments...