visarga

23 ms·

511.
by visarga 1y ago
Gemini 2.5 Pro and it was this month, so probably the latest version.
512.
by visarga 1y ago
I recently tried to get Gemini to collect fresh news and show them to me, and instead of using search it hallucinated everything wholesale, titles, abstracts and links. Not just once, multiple times. I am kind of afraid of using Gemini now
513.
by visarga 1y ago
There is a good tool for searching HN comments, it uses semantic search too. I find using it to extract content for LLMs on a topic and then chatting with the content very useful. HN has great signal to noise ratio. https://hacke
514.
by visarga 1y ago
> There may be no moat for any of this. It's what I was thinking looking at the launch video. Is there a moat? No, there is none. The LLM itself is fungible, what matters is the agentic and memory layer on top. That can be reconstr
515.
by visarga 1y ago
Hey, I like your thinking, I wrote about this a bit in another comment here https://news.ycombinator.com/item?id=45663857
516.
by visarga 1y ago
I think this is the natural endpoint - local models doing something like what Atlas and Codex are doing, acting like a firewall between the user and web. You don't need to wade through the crap online yourself, the AI extracts the usef
517.
by visarga 1y ago
I made my own browser extension for that, uses readability and custom extractors to save content, but also summarizes the content before saving. Has a blacklist of sites not to record. Then I made it accessible via MCP as a tool, or I can u
518.
by visarga 1y ago
I hate that. ChatGPT makes itself sound smarter through compression and insider phrasing, and the output is harder to read without adding real clarity. That's why I can't use o3 or GPT-5. The performance layer gets in the way of t
519.
by visarga 1y ago
I think you misread the situation. The move is towards open models, small efficient models, what makes you believe there will be a moat around AI for automation?
520.
by visarga 1y ago
> we still lack schematic-level understanding of what “intelligence” even is or how it works. Not to mention how it interfaces with “consciousness”, and their likely relation to each other I think you can get pretty far starting from beh
521.
by visarga 1y ago
Yes, language is an evolutionary system that colonizes human brains. It doesn't need intelligence, only copying is sufficient for evolution.
522.
by visarga 1y ago
> then the human came up with language to describe that and then encoded the language into the LLM No individual human invented language, we learn it from other people just like AI. I go as far as to say language was the first AGI, we&#x
523.
by visarga 1y ago
> A "skill" could have easily just been a tool The problem skills solve is initial mapping of available information. A tool might hide what information it contains until used, this approach puts a table of contents for docs in
524.
by visarga 1y ago
> and a custom mcp to search indexed mkdocs sites that effectively cover the kinds of things I would include in these skills file Search and this document base pattern are different. In search the model uses a keyword to retrieve results
525.
by visarga 1y ago
I think it has a compact table of contents of all the skills it can call preloaded. It's not RAG, it navigates based on references between files, like a coding agent.
526.
by visarga 1y ago
Sonnet 4.5 was two weeks ago. In the past I never had such issues, but every week my quota ended in 2-3 days. I suspect the Sonnet 4.5 model consumes more usage points than old Sonnet 4.1 I am afraid Claude Pro subscription got 3x less usag
527.
by visarga 1y ago
The forcing argument has merit, it should not be forced, in fact they should say very little about how we do our work. But the "rubber-stamp" framing is wrong, if it were true then you would not be needed at all. It's actuall
528.
by visarga 1y ago
I don't think "replaced" is a good word here.. augmented and expanded. With AI we are expanding our activities, users expect more, competition forces companies to do more. But AI can't be held liable for its actions, tha
529.
by visarga 1y ago
Simulated fire would burn down simulated building
530.
by visarga 1y ago
> I do think GPT5 is doing a better job on just "remembering personal details" and it's compelling. Today I asked GPT5 to extract a transcript of all my messages in the conversation and it hallucinated messages from a prev
531.
by visarga 1y ago
My state space mapping approach, which is reminiscent of expert systems and tabelar RL, only makes sense when you repeat the task in the same environment so you can gradually discover the states and their policies. You can look at execution
532.
by visarga 1y ago
Nice idea, I came up with a similar system. The idea is to map the "state space" of the agent, and describe a number of discrete states. Then assign a policy to each one. Both state space mapping and policy are generated by the ag
533.
by visarga 1y ago
> I am still of the conviction that "reducing employee head count" with AI should start at the top of the org chart. The current iterations of AI already talk like the C-suites That is exactly what it can't do. We need som
534.
by visarga 1y ago
I once spent an hour trying to get an API key for Gemini from AIStudio and Vertex, with assistance from LLMs, and finally gave up and used OpenRouter. It's that great an experience.
535.
by visarga 1y ago
Is "vibe engineering" a correct term for this? It's not vibe based when you scaffold constraints around the agent: automated testing, planning in advance, comprehensive documentation, automated formatting and linting, and man
536.
by visarga 1y ago
Time goes forward, in the future when will you be in a situation you can't access a LLM? Better use LLM as much as possible to learn the skills of controlling agents, scaffolding constrains, docs, breaking problems in such a way that A
537.
by visarga 1y ago
I am working on a World Atlas based approach to computer use agents. If the task and app environment are reused, building an atlas of states and policies might be better than observe-plan-execute. We don't rediscover from scratch how a
538.
by visarga 1y ago
> what we are largely looking at is an emerging means of collaboration. They surpass open source, "out-open source-opensouce" by learning skills everywhere and opening them up for anyone who needs them later.
539.
by visarga 1y ago
1. If I wanted the "style of Van Gogh" I would simply download Van Gogh, why waste time and money on approximative AI. But if I want something Else, then I can use AI. But Gen AI is really the worst infringement tool, for example
540.
by visarga 1y ago
> That's when A.I. starts advancing itself and needs humans in the loop no more. You got to put the environment back in the loop though, it needs a source of discovery and validity feedback for ideas. For math and code is easy, for

More ›