visarga

11 ms·

211.
by visarga 5mo ago
> ? Claude, ChatGPT, etc are heinously expensive for tiny benefits lmao Unfortunately local inference is inefficient, 100s of times more inefficient than cloud. When you answer one request at a time you still have to fetch all active wei
212.
by visarga 5mo ago
> I don’t think “oh we can use Qwen now?” Would exactly devastate the US You'd be surprised how useful it can be to fine tune it in enterprise.
213.
by visarga 5mo ago
It can't happen. For one - if it did happen it would mean all domains reach singularity at once, but we know the capability curve is jagged. Each domain advances at its own speed. Second - the more you make progress, the harder it gets
214.
by visarga 5mo ago
Yes, that is how I see it too. What I would add is - intent testing - collect user messages, and check them against executed work from time to time. Every ask must be implemented and tested, every code must be justified by a user message.
215.
by visarga 5mo ago
> I do not want to rediscover for the hundredth time that in fact all this time an agent took shortcuts for acceptance tests I rely upon and didn’t catch. Or once again get the agent to understand why and what I want it to do after its c
216.
by visarga 5mo ago
You talk as if problem solving is a supervised (imitation) learning problem. No, it is a reinforcement learning problem, models learn by solving problems and getting rated. They generate their own training data. Optimal budget allocation is
217.
by visarga 5mo ago
Yes, and? Are you challenging I wrote those 30 words myself?
218.
by visarga 5mo ago
I usually type 5000 words researching for a 500 word output. It's not "write me an article on X", it's 99% my own ideas, but worded and structured and polished a bit. But I don't post them here. They are on my blog.
219.
by visarga 5mo ago
Beautiful idea, an autoencoder must represent everything without hiding if is to recover the original data closely. So it trains a model to verbalize embeddings well. This reveals what we want to know about the model (such as when it thinks
220.
by visarga 5mo ago
Large LLMs on MacBook produce tokens at an acceptable speed but the problem is reading context. Not incremental reading like when you have a chat session, because they use KV cache, but large size reading, like when you paste a big file. It
221.
by visarga 5mo ago
> And the solution? More AI, unfortunately. I think the solution to using AI in coding is more testing, which unlocks even more AI.
222.
by visarga 5mo ago
> BUT the benefit now is you can basically prototype for free. But.. so can your competitors. And that changes the value proposition.
223.
by visarga 5mo ago
I think copyright is far for being the most important aspect related to AI, it's geopolitical and economical. And even if it was the most important, there is only a case to be made for 1. that copy used to train models and 2. rare or i
224.
by visarga 5mo ago
How does that work? Is it a kind of infringement without substantial similarity?
225.
by visarga 5mo ago
Even the case for copyright infringement is weak. LLMs are not copying machines, we already have copying machines at much lower price, almost zero, and perfect fidelity and much faster than generating it probabilistically. So it makes no ec
226.
by visarga 5mo ago
I did this too, ablating all the components in my coding agent harness. The insight from my meta-optimization loops was "have judge agents review the plan and implementation". One of my own insights here is that you need to collec
227.
by visarga 5mo ago
wow, that brings back memories from my first encounter with Apple
228.
by visarga 5mo ago
That makes the bite less damaging - if everyone hax "Co-authored-by AI" in their commits less shame for it, just a normal fact of life now, not a sign of low quality.
229.
by visarga 5mo ago
When classifying resumes it is better to use the LLM as a feature extractor, think of 10-20 features you base your decision on, and extract them by LLM. The LLM only needs to do lower level task of question answering. Then you fit a classic
230.
by visarga 5mo ago
Yes, I too think it's authored by AI, but can you indicate where it is wrong?
231.
by visarga 5mo ago
Good research, but man do I feel the LLM vibe shining through. That sustained information density...
232.
by visarga 6mo ago
Rather than talking about consciousness which we can't even define or observe in others directly, why not focus on something more concrete - cost. A process or pattern that pays its costs, or gains to offset its costs. Why cost? Becaus
233.
by visarga 6mo ago
> Weird, I thought AI was going to create so much economic surplus that we wouldn’t know what to do with it. What happened? The surplus is converted in new structure and becomes baseline. Even if a company does not change, their competit
234.
by visarga 6mo ago
I think it is in the interest of chip makers to make sure we all get local models
235.
by visarga 6mo ago
> the new tool suddenly gets ragged pulled from under your feet If that happened at this point, it would be after societal collapse.
236.
by visarga 6mo ago
Yes, I think they unlock a whole new level of capability when they have a r/w file system (memory), code execution and the web.
237.
by visarga 6mo ago
Yes, I have a theory - that higher efficiency becomes structural necessity. We just can't revert to earlier inefficient ways. Like mitochondria merging with the primitive cell - now they can't be apart.
238.
by visarga 6mo ago
yes, $200/mo is a serious subscription, we are owed something, and I won't feel ashamed for saying that especially when you are told using the subagent for code review "claude -p" is now billed on API on top of $200 sub
239.
by visarga 6mo ago
No, I want a little monkey doing tricks. /s
240.
by visarga 6mo ago
There are ways to quantize or compress KV cache down.

More ›