Do note that I only use LLMs in the ChatUI, I never use agents. I don't believe having a blackbox codebase managed by entities with a half-life of 'delete conversation' or 200k tokens is a responsible idea. In ChatUI, I lay the ground rules, kill assumptions about our working relationship, give it foundational context on the problem and codebase we're working on, explain the problem and then we have a conversation about it and I gradually disclose more logically context as it becomes relevant. So, to directly answer your question, maybe I'm missing out on a ton of upside by not using the absolute best but I'd say familiarizing yourself with a specific model has all the benefits of having a human friend you've grown up with... except your buddy's a savant and would absolutely love to help!
> We’ve been calling some of these “magic words” at work, specific technical terms or references/techniques that you need only mention to get vast improvements in outcome.
First, I almost always try to seed every new project or context/domain with canonical technical specifications or examples I found elsewhere. When I set up this project recently, I linked to a bunch of the official Apple docs for sysctl, and told it to use a specific technique for calling assembly code from Go, that from experience it almost never realizes it can do or knows about (and similarly for sysctl, I knew it kinda sorta knew about it, but not in its entirely): https://github.com/accretional/sysctl/commit/da52438233e5b33...
The other thing I did was tell it to enumerate all the test cases ahead of time rather than to just directly implement them; again this is something where you have to explicitly tell it to go digging for information where it has blind spots and get it to set up properly grounded self-eval in a way that it can test against. I usually tell it to take notes as it works or commit notes to itself that will persist over sessions: https://github.com/accretional/sysctl/blob/main/FINDINGS_2.m...
Once we get back to working on this project we'll just have it implement / validate the rest of the sysctl feature support against the full inventory we had it uncover: https://github.com/accretional/sysctl/blob/main/cmd/darwin-n...
Another thing we do is have it specify an API that it can produce against; then in other projects we have them consume the API via reflection (and our special sauce we've been working on is the ability to discover and integrate against these automatically across thousands of APIs from many providers, which we've got working and can share if you're interested in using it as an early customer): https://github.com/accretional/sysctl/blob/main/proto/sysctl... This isn't the greatest example because it doesn't actually fully specify the sysctl keys yet. But I did have it create a knowledge base trying to cover the 1000+ keys as best as it could, to reference as it continued: https://github.com/accretional/sysctl/tree/main/macos-sysctl...
We have a better example in eg https://github.com/accretional/proto-sqlite/tree/main/lang where we were able to encode the entire sqlite grammar into a grpc interface so that you could eg find the exact structure (and sanitize) of a select statement: https://github.com/accretional/proto-sqlite/blob/main/lang/p... This way integration and discovery becomes a matter of telling it "use reflection against this endpoint to discover the sql interface, then implement against it" and we can model formats/input validation as formal grammars via EBNF (all magic words) vs just adhoc
We also tell it to set up and use a browser automation toolkit/testing and always run it at the end of testing workflows (often in a way that auto-opens screenshots on our local machines + commits them to git) via tools like https://github.com/accretional/chromerpc#headlessbrowser-aut... so that whenever we produce UIs it can evaluate its own output and iterate without direct human intervention. This is another case where the knowledge-discovery problem becomes a problem so we tell the models to use reflection to discover the browser automation apis. That ends up giving us things like this where it records user journeys through sites and creates visualizations without us having to debug them or do them ourselves: https://github.com/accretional/proto-css/tree/main/chrome-te...
It can be frustrating. The AI pretends to be a human, and so a part of my brain expects them to commit and have a "parti pris" like a human, so the exercise is a good reminder of the feedback loop. My mental model is that before the first three or four messages, the model has many finer points of its personality still underdetermined. I'd suggest that as the mechanism for "role-based prompting". And it explains the "savant sleeper agent" thing you describe. You want to get the state in the right attractor on the manifold.
These machines are pretty incredible, but for conversation-driven workflows you really have to be in the driver's seat. A human has a property that the AI does not have, at least under current architectures: we are regulated by the outside world. A bit of a tangent, but I can see how AI psychosis arises from these dynamics.
I'm surprised how little attention this is getting in today's comments. Open-weights means being able to afford multiple runs, space sampling, critique and synthesis.
Last night, sketching some intro biology content emphasizing cross-cutting concerns, Qwen would get sucked into the Next Generation Science Standards attractor. But nudge it by adding just one similar phrase of Chinese, and most runs ran free (outputting English but for headers with parenthesized Chinese). The multi-lingual LLM "no, not *that* region of latent space".
I recommend everybody do this because you don’t need any special data except what you are already using, and the results will be very eye opening: there is WAY more randomness or instability involved than you would otherwise assume. A lot of what you might think is a better prompt technique, or a particularly good or bad outcome, could just as well be random chance or just different behaviors across model version or sizes. And your results can be massively biased by small differences in input. We’ve been calling some of these “magic words” at work, specific technical terms or references/techniques that you need only mention to get vast improvements in outcome.
There’s a skill to it. With agentic loops if you get the model into a self-eval structure where it’s hard to cheat or take shortcuts, and it’s in the right structure or domain that models its training, you’re golden. But it’s hard to find the sweet spots (pro tip, have Opus 4.8 convert PyTorch models into ONNX or quants or get them running on different hardware, I swear it was like I activated some kind of savant-like skillset; meanwhile I can’t for the life of me get it to properly write/test EBNF formalizations of common languages and formats without cheating).
The worst part is that it changes so much so frequently that it’s almost useless to really go digging for this kind of knowledge unless you’re actually the one training the models. I wish this kind of “stability” in output was more emphasized in their training so they’d be predictable. I assume it’s hard to do without overfitting or breaking the explore-exploit loop but also, I would spend so much more on LLMs for batch workloads if they could do them more reliably…