visarga

12 ms·

241.
by visarga 6mo ago
> pay more, get more Users on $200 plan complaining, already at max level of subscription, I don't think a $200 subscription should make you feel like you are getting unfair advantage. Like restricting claude -p to API ... after I p
242.
by visarga 6mo ago
And they don't use our version of CC, or with our settings. They have flags for internal use only.
243.
by visarga 6mo ago
from 1h to 5 minutes, was in the news recently
244.
by visarga 6mo ago
I want to differentiate 2 kinds of harnesses 1. openclaw like - using the LLM endpoint on subscription billing, different prompts than claude code 2. using claude cli with -p, in headless mode The second runs through their code and prompts,
245.
by visarga 6mo ago
I might consider switching to codex from claude pro 20x but I need the post tool use, pre file write and post user message hooks. Waiting on codex to deliver. - pre file write -> block editing code files without a task and plan of work -
246.
by visarga 6mo ago
I made one similar harness, mine does lightweight sandboxing with Seatbelt on Mac and Bubblewrap on Linux. I initially used Docker too, but abandoned it. I like how these 2 sandboxes allow me to make all the file system r/o except the
247.
by visarga 6mo ago
> Whether that student walks out the door five years later as an independent thinker or a competent prompt engineer is, institutionally speaking, irrelevant. I think this is a simplification, of course Bob relied on AI but they also used
248.
by visarga 6mo ago
Yes, judges should not just look for bugs, they should also validate intent follow, but that can only happen when intent was preserved. I chose to save the user messages as a compromise, they are probably 10 or 100x smaller than full sessio
249.
by visarga 6mo ago
Humans are also unreliable, we are competing for scarce attention, platforms decide what gets visibility and we cater to their algorithms. You could say humans are prompted by feed ranking AI - what and how to publish.
250.
by visarga 6mo ago
My own approach also has intent sitting at the top: intent justifies plan justifies code justifies tests. And the other way around, tests satisfy code, satisfy plan, satisfy intent. These threads bottom up and top down are validated by judg
251.
by visarga 6mo ago
I built a harness where my plans and code are reviewed with 'claude -p' but most work is interactive, now it has been wrecked. I relied and integrated with Anthropic to get burned. I'm not even maxing out my plan, never surpa
252.
by visarga 6mo ago
> In their minds, financial success is the ultimate yardstick. In a loopy recursive way, it is. Cost gates what we can do and become. Paying back your costs to extend your runway is the working principle behind biology, economy and techn
253.
by visarga 6mo ago
I run Claude Code from Zed. Very nice experience.
254.
by visarga 7mo ago
Claude Code says thank you for reporting, I bet they will scan this chat to see what bugs they need to fix asap.
255.
by visarga 7mo ago
ARIMA and ARMA models
256.
by visarga 7mo ago
The cognitive benefit of writing comes from externalizing and evaluating ideas under friction. LLM conversation provides more friction per unit time than solo drafting because you're constantly reacting to a semi-competent interlocutor
257.
by visarga 7mo ago
> it certainly feels against the spirit of what I intended when distributing my works You can own the works, but not the vibes. If everyone owned the vibes we would all be infringing others. In my view abstractions should not be protect
258.
by visarga 7mo ago
I've been saying LLMs are more open than open source for some time...
259.
by visarga 7mo ago
ARC is trying to isolate a unitary intelligence signal, so it strips away coordination, specialization, and division of labor. But that also means it removes one of the dominant mechanisms by which intelligence actually scales in the real w
260.
by visarga 7mo ago
It only tests puzzle solving, intelligence is cost compression that powers itself.
261.
by visarga 7mo ago
That is because platforms both enable us and exploit us, they exploit both those who create/comment and those who read. They perform a necessary function but extract the value from it for their own good.
262.
by visarga 7mo ago
But across a sequence you still have to load most of them.
263.
by visarga 7mo ago
I did something similar, markdown and code agents for memory, multiple feeds for intake, also my own browsing and claude cli messages get indexed.
264.
by visarga 7mo ago
> there is nothing about the training process of these models that would encourage them to make the output of any layer apart from (n-1) meaningful as the input of layer n There is something that does exactly that - the residual connecti
265.
by visarga 7mo ago
I did a similar system myself, then I run evals on it and found that the planning ceremony is mostly useless, claude can deal with simple prose, item lists, checkbox todos, anything works. The agent won't be a better coder for how you
266.
by visarga 7mo ago
Workflow matters too, how you organize your docs, work tasks, reviews. If you do it all by hand you spend a lot of time manually enforcing a process that can be automated. I think task files with checkable gates are a very interesting anima
267.
by visarga 7mo ago
You can differentiate by context, one sees the work session, the other sees just the code. Same model, but different perspectives. Or by model, there are at least 7 decent models between the top 3 providers.
268.
by visarga 7mo ago
An ensemble can spot more bugs / fixes than a single model. I run claude, codex and gemini in parallel for reviews.
269.
by visarga 7mo ago
If AI produces surplus where does it go? Not talking about investment backed datacenter buildout and AI labs. Talking about the results of AI work... I think AI outcomes distribute to contexts where it is used, and produce a change in how w
270.
by visarga 7mo ago
I think you need coding style guide files in each repo, including preferred patterns & code examples. Then you will see less and less of that.

More ›