vdelpuerto
vdelpuerto on Hacker News
Not really, although I appreciate the sarcasm. What I'm saying is
that Anthropic and other providers tells you that you can rely on
lower models for certain tasks that are usually daily tasks that
don't require much more than basic understanding. To prepare a
quotation doesn't seem like something that requires Opus 4.7 at
max effort, don't you think? So the question is: what counts as
a task you can delegate to a lower model and still trust the output?
Posted by
vdelpuerto on Thursday May 21, 2026 @08:08AM
from the self-post dept.
By standard I use Opus 4.7 for every task no matter what it is
with Claude. But, because I was reaching almost 90% of the model
use and with two days left before refresh I set Sonnet 4.6 for a
simple task with my Price analyst agent. And it was kind of painful.
It just had to look at an email from a client requesting a quotation,
look for the database of our usual values, scope analysis...
Thanks. I'll take it a look
I was trying to look options outside the box (everything is more context or RAG) and been using this approach for about a month with good results.
https://github.com/VDP89/fscars
honest question here, why there are lot of comments about SEO is dead?
That is the same fight the 2D animators were having with 3D aninmation 30 years ago. The resolution is likely to be the same: the tool wins but the fundamentals stay, and the line between competent and incompetent practitioners moves but does not disappear.
Posted by
vdelpuerto on Tuesday May 19, 2026 @08:54AM
from the github.com dept.
I'm still waiting something more specific or groundbreaking too. Feels like a lot of noise with just the goal to get people to talk about it. And now I realize I am talking about it and about nothing at the same time. Just fugazzi.
through agent Harness the user can get a kind of model personalization as it iterates, you adjust patterns, mistakes, KB and preferences. Its the closest we have to add some kind of personality to our LLM agents. mistakes are the hardest one because you put the rule in AGENTS.md or system prompt and the model still goes back to the wrong pattern. context alone don't fix it
Under the premise of low taxes and Itaipú's great energy generation, Paraguay has an open and political interest in bringing data centers and AI centers to the country looking for capital injection to the economy, but with very poor infrastructure and grid base development to face this. Just looking at recent studies and articles about what consuming energy is doing to the grid in the US, it raises the question of how we can prevent this from happening here.
I wrote something about it trying to look other way around the context or memory data in models. The gravitational pull of information stills very hard to manage. Ive been using "functional scars" about 30 days now and getting good results in repetitive mistakes across sesions.
https://github.com/VDP89/fscars
and not only feelings, but failures with real consequences is what makes us more creative and aware of our actions.
funny thing, local government in Paraguay is importing a model that USA itself recognizes to not fully understand. Paraguay is proposing to accommodate projects whose demand exceeds the available supply
Shortcut works sometimes. But if X is common in training and Y is rare, the model regresses on the next turn even with 'do Y, not X' right there in history. @vanviegen's 'fighting instincts' — you can't trust the model to read the correction. Gate it before the model runs instead of inferring it from context
Annotation queues are a great discovery layer. The enforcement question is separate — once the failure mode is named, what stops the model from producing it again? Evals catch it at runtime, hooks prevent it at the trust boundary. Different layer of the pipeline.
Author here. I run a civil engineering firm in Paraguay and use Claude Code daily in production — 530+ files across three knowledge bases. This essay came from months of watching the same corrections fail across sessions, even after centralizing everything. The five invariants and the "functional scars" pattern came out of cross-referencing 163 findings from my own project logs. Happy to answer questions about the methodology or the hook architecture
Funny — I run LLM agents in production and my problem is the exact opposite. The model can't remember corrections I already made. Fix something in session 4, session 5 it comes back like nothing happened.
Took me like 4 resets to solve it.
The framing of "scary stories" misses something interesting: most of the actual operational fear isn't about consciousness or superintelligence — it's about systems that seem to work until they quietly don't.