I think i’ve only once flat out told one it was wrong about a specific assertion I quoted and it immediately was able to find its way to what I knew to be the correct claim.
I just wonder what would happen if i was in fact mistaken and I told it confidently it was wrong without elaborating
The best sort of methodology I’ve found to coerce Claude or whatever (we are strongly encouraged to use it, because you know, tech these days) is (for a single agent) to define a process that includes proving its work and citing sources. For agentic flow, you basically just assign a contrarian role in particular domains to some of the agents - ideally all of this is also hooked into an MCP server that includes deterministic utilities to improve accuracy and solution arrival speed.
It’s basically just a shitty, brute-forced, massively over complicated Monte Carlo algorithm that’s wildly inefficient in terms of energy usage and infrastructural cost, that also happens to be turning our economy into a highly flammable house of cards.
Can you tell what my opinion of all this bullshit is, despite knowing how to do all of this crap reasonably well? 😛
strongly encouraged for me means my entire bonus is tied to skilling up with LLMs.
“cite-or-stop” with journaling, role definitions, subagents to control context windows, and shifting as much of the work as possible to deterministic scripts has yieled pretty high-quality results.
Expensive af though. Like the self-review skill I maintain before engineers put their code up for human review will find 30-40 things on average, where only a small handful are false positives, can burn 10% of a $20/mo plan’s utilization in a single run on a moderately sized PR. The default one my company setup in our source control usually finds 1-5 things and only 0-2 are of any value.
Oh yeah it 100% explodes your token usage by establishing a contentious paradigm. But that’s what the C-suites want, and they don’t understand (or care to understand) the subtleties of how this development methodology works. So we’ll all (somewhat maliciously) follow orders until they change their minds when they see the absolutely ludicrous LLM bill 🤪
I think that’s a good approach. Personally I find LLMs quite fascinating but they’re deeply flawed. They can barely be used in production environments, especially unsupervised. The workflows regarding LLMs are very esoteric with specific prompting techniques etc and while all LLMs have similar flaws each model and model version behaves differently. It’s super weird and unreliable. Like one big workaround that has so much investment that it keeps improving every month but still stays shitty at it’s base.