Y
HN Search
Hacker News Search
new
|
comments
|
top
|
jobs
0x457
searching PlanetScale…
1.
▲
2.
▲
3.
▲
4.
▲
5.
▲
6.
▲
40 ms
·
601.
▲
by
0x457
1y ago
Pretty sure, almost every Mac user is using emdash. I know I do when I'm macOS or iOS.
602.
▲
by
0x457
1y ago
I had a weird bug in elixir code and agent kept adding more and more logging (it could read loads from running application). Any way, sometimes it would say something "The issue is 100% fix because error is no longer on Line 563, howev
603.
▲
by
0x457
1y ago
Because while Nginx always has access to .well-known, thing that validates on issuer side might not. I use DNS challenge to issue certificates for domains that resolve to IPs in my overlay network. The issue is that supporting dns-01 is jus
604.
▲
by
0x457
1y ago
You're looking at a summary for chunks of code that are relevant to the given library. If you type what specifically you need documentation for and adjust output token count, it will give LLM relevant fragments. It lets you emulate RAG
605.
▲
by
0x457
1y ago
Not sure how are you doxxing yourself, what stopping me from YOLOing my life savings into this short after reading a few comments in this thread?
606.
▲
by
0x457
1y ago
That's how Kiro works and how the agent that I'm working on works. Kiro creates tasks from spec document, and you can revise tasks either by prompting or editing tasks.md file. My thing works a bit differently, but essentially the
607.
▲
by
0x457
1y ago
Yes, I can, and I do, I'm pointing out that compressing an entire conversation history into a single message is so lossy that I might as well start a new session. Yes, I can also tell any agent to commit more often, but that's aga
608.
▲
by
0x457
1y ago
Yes, you can prompt, I find it confusing why that + worktrees (i.e. agent working on its own working copy of the repo) is the standard/
609.
▲
by
0x457
1y ago
That's CANBUS not OBD2, and it only works on some cars because not moronic manufacturer prevent it. Try doing it, on a European car you will fail.
610.
▲
by
0x457
1y ago
Well, "porn" has a very loose definition and varies country by country. IMO it's either nudity (but remember free the nipple movement when you try to define nudity) and/or sexual act usually involving penetration. Anyway
611.
▲
by
0x457
1y ago
I get the "good" result with phi-4 and gemma-3n in RAG scenario - i.e. it only used context provided to answer and couldn't answer questions if context lacked the answer without hallucination.
612.
▲
by
0x457
1y ago
I always find it wild that none of these tools use VCS - completed logical unit of work, make a commit, drop entire context related to that commit, while referencing said commit, continue onto the next stage, rinse and repeat. Claud always
613.
▲
by
0x457
1y ago
They do, but if you look at the graphs...what is the point of the large context window if accuracy drops off waaaaay before context window is maxed?
614.
▲
by
0x457
1y ago
Better if you see it for yourself. Setup GitHub MCP and enable all tools. It will start using wrong tools at wrong time, overuse it. Add languageserver-mcp, and it suddenly will start trying to use it for file edits and create a huge mess i
615.
▲
by
0x457
1y ago
It gives me elementary school oral report. The same level of acting and script.
616.
▲
by
0x457
1y ago
The fact that you don't understand that humans absorb a lot of knowledge just by existing is concerning (if I cared). > The vast majority of humans (>99.99999%) can tell that those things are different. Citation needed.
617.
▲
by
0x457
1y ago
I don't think YouTube comment section ever contain useful information regardless of what the video/stream is about.
618.
▲
by
0x457
1y ago
I mean...where do you think context is stored?
619.
▲
by
0x457
1y ago
> Looking at real things while living life is categorically different from viewing millions of artworks made by other artists I don't see a difference.
620.
▲
by
0x457
1y ago
A blind monkey smashing a keyboard can produce better PR and PR reviews than GitHub copilot. I don't get how they managed to make copilot so bad.
621.
▲
by
0x457
1y ago
Entirely different product. Claude Code runs on my machine, Jules runs on some sandboxed ubuntu vm in GCP without any input (beyond high level user story promp)
622.
▲
by
0x457
1y ago
Do you need an invitation? I'm just using my Amazon Q Dev account that I pay $20 a month for. Works fine with Kiro.
623.
▲
by
0x457
1y ago
Well, first Jules came before Gemini CLI. Second, that's okay, as long as it can verify its work (i.e. run tests) it will eventually figure out what to do. Its sandbox is very limited and prevents proper grounding IMO. However, if thei
624.
▲
by
0x457
1y ago
Jules is web only from my understanding, similar to OpenAI's Codex (web version...) You give it a task and it produces a PR. While gemini-cli is more like pair programming with AI.
625.
▲
by
0x457
1y ago
Are you telling me artists never get inspired by something they saw living their life? Only allowed to be inspired when specifically looking at other art?
626.
▲
by
0x457
1y ago
When people say "modern workstation" in context of LLM, they usually mean its consumer(pro-sumer?) grade hardware on a single machine. As opposed to racks of GPUs that you can even buy as a mere mortal (min order size) It doesn&#x
627.
▲
by
0x457
1y ago
Yes, but I thought we're talking about category difference. Proper RLHF surely boosts "predicted next token until it couldn't" to feel more like "actually recalled".
628.
▲
by
0x457
1y ago
Only if you don't count artists with able eye(s) living their life?
629.
▲
by
0x457
1y ago
The context is "you don't need an entire terminal multiplexer such as Tmux when all you need is pseudo-tty". Bringing another terminal multiplexer isn't an alternative. In my agent, I use pseudo-tty without any tmuxes&#x
630.
▲
by
0x457
1y ago
yes
More ›