Creamsicle47

I'm pretty sure thinking is practiced by attempting to argue the opposite viewpoint yourself. Here, you're telling me that you "practice" your thinking by having it actually create the argument for you.

I agree that AI is a useful tool, but it's arguably a very easy tool that people can offload their thought process to - which you should not be doing.

> avoid being subverted in this way

In the instances where they were able to detect it. How much do you think they pour over individual traces nowadays to search for these things?

Even then, the actual development of a harness for interesting problems is a difficult task. It requires more skill & experience than people realize, and gets harder as you scale things up.

Can you please point me to any project based around spec-driven development with agents which has worked well long-term, ideally one that isn't another agentic devleopment tool?

I come from a land down under

> Roblox reports user counts are down

Please don't tell me you value your shitty 401k more than keeping kids off of this horrible game

These edge-cases are why the simplest answer works best: just scale your model to be good enough that it can work through the problem on its own

Agents in Paris?

Technically true, but this is a very inefficient optimization strategy and usually indicates some structural issue with your setup.

When you know enough about the JSON structure to understand how to set a decoding order, then you can simply break the problem into smaller parts so it's more manageable.

Otherwise, you have these solutions: - GEPA-style prompt optimization for simple agents - If working with large models, you can just have them do step-by-step problem solving & relying on their CoT for accuracy. - A small model can simply be trained to do what you want

And if those don't work, I would be extremely surprised.

As someone also considering getting one, what would make it worth it in your opinion?

It's because LLMs provide an efficient way of optimizing a complex, non-linear search space when there's a way to measure progress. The less verifiable something becomes, the harder it is to optimize (but not impossible).

What the other guy means is the model is forced to decode potentially out of distribution and subsequent parts of the sequence will be sampled from a higher-entropy distribution.

i.e. it might be that a particular model may need some additional conditioning, or perhaps adjusting the order of the json output affects it to produce lower-quality results.

You have to remember that the LLM will be expected to generate 10s of thousands of tokens on a single turn, where things like quantization noise lead to cascading error further on. So when you can, you really want to stay within the model's distribution.

The most sane comment I've seen in this thread

It comes down to the blast radius of a failure happening in your system. Some stateless frontend won't require much beyond visual inspection, but if you introduce some client-server architecture or require payment processing, you must consider the cost of a security intrusion when auditing code changes.

A human acting on inappropriate thoughts can be held accountable in a physical way. An agent producing harmful or dangerous actions will get written off as just a “trajectory that didn’t work correctly”