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.
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?)
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."
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/
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
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
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"
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'."
> 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
OpenAI’s o3 was big on en dashes—one time it produced a Deep Research result containing >200 of them. I’m not aware of any other LLM using them commonly, though. I’d guess humans use them even less often; I don’t think Apple auto-inserts en dashes, and very few people (myself being one) are pedantic enough to bother.
On the other hand, I don’t think o3 was ever a common choice among people copying from LLMs, so en dashes remain infrequent regardless.
[1] https://unsloth.ai/docs/models/tutorials/minimax-m27
(Unsloth's deepseek-v4 support is still WIP)