Hacker News: lostmsu

New comment by lostmsu in "SNL Weekend Update: Anthropic CEO Dario Amodei on A.I.'S Threat to Humanity [video]"

lostmsu — Sun, 27 Sep 2026 19:45:20 +0000

Video unavailable

The uploader has not made this video available in your country

New comment by lostmsu in "Unreal Agent"

lostmsu — Sat, 26 Sep 2026 15:56:54 +0000

Don't give the agent any kind of wait or sleep tools and deny any attempt to read status without new progress notifications.

New comment by lostmsu in "A single function Jev-like wrapper for LLMs, including vision models"

lostmsu — Sat, 26 Sep 2026 12:51:59 +0000

Much less code https://news.ycombinator.com/item?id=49769618

New comment by lostmsu in "Tech Needs Humanists More"

lostmsu — Fri, 25 Sep 2026 11:30:19 +0000

It fits in a paragraph, and we all read it. The rest of it is bullshit.

New comment by lostmsu in "Tech Needs Humanists More"

lostmsu — Fri, 25 Sep 2026 11:13:45 +0000

> you wouldn't need to spend years in school

Do you?

New comment by lostmsu in "Mercury 2.5 LLM hits 770 tokens per second"

lostmsu — Thu, 24 Sep 2026 10:56:49 +0000

Inco sucks. I tried their GLM 5.3 Flash and it was quantized to the point of hallucinating Chinese in the middle of English only agentic sessions. Never happened with any other provider.

Radicle: Disclosure of Vulnerability in the Network Protocol

lostmsu — Wed, 23 Sep 2026 15:23:05 +0000

Article URL: https://radicle.dev/2026/09/23/disclosure-of-vulnerability-in-network-protocol

Comments URL: https://news.ycombinator.com/item?id=49817524

Points: 154

# Comments: 58

New comment by lostmsu in "Unreal Agent"

lostmsu — Wed, 23 Sep 2026 13:26:09 +0000

This problem has a few rather trivial solutions.

New comment by lostmsu in "Looking forward to Git 2.56 – and 3.0"

lostmsu — Tue, 22 Sep 2026 11:48:32 +0000

How is this different from branches?

Show HN: SoTA Feed – Every open-weights release from the labs that matter

lostmsu — Mon, 21 Sep 2026 22:19:31 +0000

Article URL: https://sota.borgcloud.ai/

Comments URL: https://news.ycombinator.com/item?id=49794230

Points: 2

# Comments: 0

New comment by lostmsu in "Wall Street Is Growing Skeptical of the Data Center Boom"

lostmsu — Mon, 21 Sep 2026 20:26:00 +0000

What's the output rate and config?

New comment by lostmsu in "Show HN: Mini-AGI – Dynamic continual learning model trained on 8GB VRAM"

lostmsu — Mon, 21 Sep 2026 11:17:26 +0000

It doesn't show any indications of solving catastrophic forgetting.

New comment by lostmsu in "Show HN: Mini-AGI – Dynamic continual learning model trained on 8GB VRAM"

lostmsu — Mon, 21 Sep 2026 11:15:50 +0000

This is slop. 8M parameter dense model with context length 64 that you train on enwik9 in 2h will have 1.15 bpb. This model has 1.8 (bits per byte, lower is better).

New comment by lostmsu in "Introducing System One Models and Jev"

lostmsu — Sat, 19 Sep 2026 19:41:10 +0000

According to https://goodstartlabs.com/research/verification-is-the-bottl... it is only 2.6x cheaper than DeepSeek V4.1 Flash, and they did not test v4.0 Flash, which would have been same price.

New comment by lostmsu in "Show HN: Cactus Needle 3: 8-29MB automation models can match DeepSeek V4 Flash"

lostmsu — Sat, 19 Sep 2026 18:51:51 +0000

> Monarch Hadamard MLP: replaces the dense FFN with three learnable Walsh-Hadamard-initialized Kronecker (Monarch) factor pairs interleaved with per-channel diagonal scales, fixed permutations, a SiLU nonlinearity, and a rank-8 input-conditioned gate, so each token gets a fully mixed nonlinear transform of its d_model channels at O(d√d) parameters and compute instead of the O(d²) a dense 4x-expansion MLP would cost.

Wow, I was just researching W-H in transformers. Did yours seem to work? In my experiments swapping various components for W-H-like transforms caused extreme quality degradation.

UPD. according to the comments here, this model simply does not work at all, so I guess the answer is NO

New comment by lostmsu in "A heap overflow and SSO misconfiguration to compromise OpenAI internal repos"

lostmsu — Sat, 19 Sep 2026 18:27:12 +0000

MS absolutely has a couple of stock-based incentives.

New comment by lostmsu in "Apple M6 Pro Achieves the Highest Single-Core CPU Score in Geekbench 7"

lostmsu — Sat, 19 Sep 2026 17:31:43 +0000

> And there are bugs in any Linux distribution that are way too complex to fix even if theoretically possible. So it doesn’t make any difference.

It does now with LLMs

New comment by lostmsu in "Microsoft exec called AI scraping 'the largest theft of labor in human history'"

lostmsu — Fri, 18 Sep 2026 12:16:23 +0000

Thank God there's the Internet Archive then. https://archive.org/donate I do.

New comment by lostmsu in "Microsoft exec called AI scraping 'the largest theft of labor in human history'"

lostmsu — Fri, 18 Sep 2026 12:07:00 +0000

Good luck with that.

New comment by lostmsu in "Xiaomi Mimo 2.6 live post-training dashboard"

lostmsu — Thu, 17 Sep 2026 15:44:48 +0000

That's posttraining. Pretraining is the expensive part.