HackerTrans
トップ新着トレンドコメント過去質問紹介求人

thomasm6m6

no profile record

投稿

Systematic Reward Hacking and Prime Sprints

primeintellect.ai
3 ポイント·投稿者 thomasm6m6·2 か月前·0 コメント

Evaluating Gemini Robotics Policies in a Veo World Simulator

veo-robotics.github.io
2 ポイント·投稿者 thomasm6m6·7 か月前·0 コメント

コメント

thomasm6m6
·2 か月前·議論
FYI, llama.cpp (which antirez/ds4 is inspired by) supports system ram. E.g. [1] is a good guide for running a similar-sized model with 128gb ram and a 3090-sized GPU.

[1] https://unsloth.ai/docs/models/tutorials/minimax-m27

(Unsloth's deepseek-v4 support is still WIP)
thomasm6m6
·3 か月前·議論
And OpenCode just added `opencode export --sanitize` for PII redaction

https://github.com/anomalyco/opencode/releases/tag/v1.4.4
thomasm6m6
·3 か月前·議論
There's https://github.com/badlogic/pi-share-hf by the creator of pi-coding-agent, to redact session data and publish on Huggingface. You can find others of the same idea for Claude Code/Codex on Github, though of varying redaction quality. Or have your LLM fork pi-share-hf to work for your preferred coding agent.

Clem Delangue (HF CEO) tweeted about this[1] and mentioned https://traces.com/ for exporting Claude sessions

Edit: It looks like HF now supports importing your agent's session directory directly[2] (I hope they're redacting PII?)

[1] https://x.com/ClementDelangue/status/2041189872556269697

[2] https://huggingface.co/changelog/agent-trace-viewer
thomasm6m6
·6 か月前·議論
post title should be dated. earliest archive.org snapshot is from 2018:

https://web.archive.org/web/20180501000000*/https://www.medi...
thomasm6m6
·7 か月前·議論
script is here: https://github.com/vtomnet/hn35/blob/main/generate.py

the prompt indeed began with "We are working on a fun project to create a humorous imagining of what the Hacker News front page might look like in 10 years."
thomasm6m6
·7 か月前·議論
> A 100% Rust kernel is now upstream in Linux 7.4

the future predicted by gemini is already coming true; from 3 hours ago <https://news.ycombinator.com/item?id=46213585>:

"Rust ... is now a core part of the kernel and is here to stay"
thomasm6m6
·7 か月前·議論
re: image gen, have you seen the more recent models? gemini-3-pro-image (aka nano banana pro) in particular is stunningly good at just about everything. examples: https://vtom.net/banana/
thomasm6m6
·7 か月前·議論
oh, didn't realize it required login. it's an implementation of the "llm simulates a world as you go" idea. screenshots and twitter for context:

https://sw.vtom.net/tmp/worldsim1.png

https://sw.vtom.net/tmp/worldsim2.png

https://x.com/NousResearch/status/1788283681511035011

https://x.com/karan4d/status/1768836844207378463
thomasm6m6
·7 か月前·議論
thanks! here's the script for reference: https://github.com/vtomnet/hn35/blob/main/generate.py
thomasm6m6
·7 か月前·議論
wow that's cool. didn't expect an llm could make that complex of a ui. it's giving me flashbacks to simulating reddit threads with llama in early 2023
thomasm6m6
·7 か月前·議論
Now I'm curious to try something more real-time. gemini wouldn't work since it's so slow, but gpt-oss-120b on cerebras could be a good fit with careful prompting. might do this after finals

also worth linking https://worldsim.nousresearch.com/console
thomasm6m6
·7 か月前·議論
Ah that one was generated with an earlier prompt, where I asked it to use the original comment count from TFA (mostly as a suggestion, I don't expect it would get the exact number). Then I realized that was too many and it would end up repeating tropes for the other submissions' comments, so reduced it to a random comment count from 20-100
thomasm6m6
·7 か月前·議論
good point. some of the comments actually do have >1 children, e.g. a few here (https://sw.vtom.net/hn35/item.html?id=90098555), but it'd be better if more random
thomasm6m6
·7 か月前·議論
Here it is: https://sw.vtom.net/hn35/news.html

I downloaded the original article page, had claude extract the submission info to json, then wrote a script (by hand ;) to run feed each submission title to gemini-3-pro and ask it for an article webpage and then for a random number of comments.

I was impressed by some of the things gemini came up with (or found buried in its latent space?). Highlights:

"You’re probably reading this via your NeuralLink summary anyway, so I’ll try to keep the entropy high enough to bypass the summarizer filters."

"This submission has been flagged by the Auto-Reviewer v7.0 due to high similarity with "Running DOOM on a Mitochondria" (2034)."

"Zig v1.0 still hasn't released (ETA 2036)"

The unprompted one-shot leetcode, youtube, and github clones

Nature: "Content truncated due to insufficient Social Credit Score or subscription status" / "Buy Article PDF - $89.00 USD" / "Log in with WorldCoin ID"

"Gemini Cloud Services (formerly Bard Enterprise, formerly Duet AI, formerly Google Brain Cloud, formerly Project Magfi)"

Github Copilot attempts social engineering to pwn the `sudo` repo

It made a Win10 "emulator" that goes only as far as displaying a "Windows Defender is out of date" alert message

"dang_autonomous_agent: We detached this subthread from https://news.ycombinator.com/item?id=8675309 because it was devolving into a flame war about the definition of 'deprecation'."
thomasm6m6
·8 か月前·議論
https://github.com/google-gemini/gemini-cli/blob/release/v0....

> For Google AI Ultra subscribers and paid Gemini and Vertex API key holders, Gemini 3 Pro is already available and ready to enable. For everyone else, we're gradually expanding access through a waitlist.

So, not available yet. Tried with a free API key and I did not have access. I do have access on a paid API key, but I'd rather not use that with an agent. The rate limits page the docs link to currently has no info on gemini-3-pro rate limits. Seems to me this is really only for users on the $200/mo subscription plan. Somewhat odd, given the model is already GA in every other coding agent as I understand
thomasm6m6
·9 か月前·議論
2.5 Flash Lite replaced 2.0 Flash Lite which replaced 1.5 Flash 8B, so one might suspect 2.5 Flash Lite is well under 50B
thomasm6m6
·10 か月前·議論
On iOS at least, there is a “Sort replies > latest” option which is strictly chronological