Hacker News: akg_67

New comment by akg_67 in "Astra for Law"

akg_67 — Fri, 18 Sep 2026 03:22:38 +0000

Context compaction? I noticed llm seem to forget partially or completely the original task when context compaction happens. The problem is more serious with local llm with low context size.

New comment by akg_67 in "My local model setup on an M4 Pro Mac Mini"

akg_67 — Wed, 02 Sep 2026 09:04:01 +0000

The above tests were done with 24k context window. Testing was mostly driven by ChatGPT analyzing oMLX server logs and suggesting changes.

Finally, I settled on Qwen3.6-35B-A3B-4bit with 32,768 context window and 16,384 max tokens.

---

Additional results from Qwen3.6-35B-A3B-4bit (Can't edit previous comment)

Qwen3.6-35B-A3B-4bit, 329.7 PP, 41.3 TG

New comment by akg_67 in "My local model setup on an M4 Pro Mac Mini"

akg_67 — Wed, 02 Sep 2026 09:01:42 +0000

What system? If on *M1 Max 32GB* or weaker, I will be interested in learning more about your setup.

New comment by akg_67 in "My local model setup on an M4 Pro Mac Mini"

akg_67 — Wed, 02 Sep 2026 06:11:20 +0000

Recent performance data on my M1 Max 32GB MacBook using oMLX. I have been working on identifying suitable model and config for my use case and system. Using a refactor and suggest improvements prompt for a specific Django code block using VSCode Cline extension.

---

Qwen3.8-27B-4bit, Prompt Processing (PP) 66.3 tok/s, Token Generation (TG) 11.8 tok/s

Ornith-1.5-35B-A3B-MLX-4bit, PP 379.7, TG 45.8

Ornith-1.5-35B-A3B-MLX-4bit, PP 381.5, TG 46.4

Qwen3.6-35B-A3B-mxfp4, PP 389.6, TG 47.6

Qwen3.6-35B-A3B-OptiQ-4bit, PP 342.6, TG 44.4

---

Qwen3.8-27B-4bit generally runs out of output token before completing the task though excellent partial results.

Ornith-1.5-35B-A3B-MLX-4bit seems to get in the loop often specially with tool calls.

Qwen3.6-35B-A3B-mxfp4 seems to be optimal with speed and quality output.

I am going to test Qwen3.6-35B-A3B-4bit soon with same code block just to check my intuition that any derivatives don't seem to perform better than the originals.

New comment by akg_67 in "Why your local LLM feels dumber than it is"

akg_67 — Sun, 23 Aug 2026 01:43:33 +0000

RAM 32GB, tokens are less than 20, may be about 12-15. Main issue is context window, I have set to 64k but practically beyond 32k, too much swap and slow down.

New comment by akg_67 in "Why your local LLM feels dumber than it is"

akg_67 — Sat, 22 Aug 2026 23:58:23 +0000

Macs fan control set to full blast and lifting the Mac above table for air circulation across bottom.

I use M1 Max with qwen3.8 27b mlx. The gpu temp can reach easily to 95°C as fan doesn’t kick in automatically until 90°C. With Macs fan control at full blast and MBP off table, temp usually hover around 85°C.

New comment by akg_67 in "Qwen 3.8 27B"

akg_67 — Sat, 15 Aug 2026 06:12:38 +0000

Are you using MLX version on M5 Max?

qwen3.8:27b-mlx

New comment by akg_67 in "AI financial advice is surprisingly good, especially if you ask right questions"

akg_67 — Sun, 02 Aug 2026 05:13:23 +0000

Though almost a year old (may be too long in LLM age), just sharing a meta-article on LLM applications in different areas of finance.

Large Language Models in equity markets: applications, techniques, and insights, Frontiers in Artificial Intelligence, A. Jadhav and V. Mirza, 27 August 2025, DOI: 10.3389/frai.2025.1608365.

Personally, I am using LLM, mostly for analyzing and shortlisting companies for in-depth analysis, for investing in Japanese equity market for over a year with relatively decent results.

New comment by akg_67 in "Japan's Genki Robotics to launch first product this year, says CEO"

akg_67 — Wed, 29 Jul 2026 09:01:55 +0000

Andy Rubin, chief executive of Genki Robotics, told Nikkei Asia that the company plans to open up its operating system to outside developers, allowing them to create and sell new robot behaviors through an app-store-like marketplace.

Japan's Genki Robotics to launch first product this year, says CEO

akg_67 — Wed, 29 Jul 2026 09:01:55 +0000

Article URL: https://asia.nikkei.com/editor-s-picks/interview/japan-s-genki-robotics-to-launch-first-product-this-year-says-ceo

Comments URL: https://news.ycombinator.com/item?id=49094986

Points: 2

# Comments: 1

New comment by akg_67 in "7.1 Earthquake in Japan"

akg_67 — Tue, 28 Jul 2026 09:38:37 +0000

NHK TV is showing extensive damage including fallen smokestacks at Nippon Paper Factory facility, bridge road separation, fallen bridge lane, road cracks, fire, water coming out of drainage etc. It was a huge earthquake and will take some time to assess the damage.

New comment by akg_67 in "Americans express unease over SpaceX's influence on retirement savings"

akg_67 — Sat, 20 Jun 2026 02:02:13 +0000

Index licensing might be an issue.

New comment by akg_67 in "Foreign business owners are scrambling to raise capital to stay in Japan"

akg_67 — Mon, 15 Jun 2026 13:39:48 +0000

> As a *Brit living in Japan* (non-resident) I think they should protect themselves at all costs, lest *what happened to my country happen to theirs.*

New comment by akg_67 in "Foreign business owners are scrambling to raise capital to stay in Japan"

akg_67 — Mon, 15 Jun 2026 13:35:02 +0000

Double standards. He is an immigrant in Japan but he doesn’t want immigrants both in his host (Japan) and home (UK) countries. Pretty ironic come to think of. I guess he thinks his type are “good” immigrants, others are not so much.

Mega-cap IPOs: Implications for institutional investors and index managers

akg_67 — Sat, 06 Jun 2026 06:41:30 +0000

Article URL: https://www.ssga.com/nl/en_gb/institutional/insights/mega-cap-ipos-implications-for-institutional-investors-and-index-managers

Comments URL: https://news.ycombinator.com/item?id=48422073

Points: 3

# Comments: 0

New comment by akg_67 in "Samsung chip workers will get an average $340k bonus as AI profits soar"

akg_67 — Fri, 22 May 2026 02:43:59 +0000

And, soon someone will come along claiming unions and strikes don’t work. /s

New comment by akg_67 in "Europe built sovereign clouds to escape US control. Forgot about the processors"

akg_67 — Sat, 16 May 2026 22:43:04 +0000

With the advent of LLMs, data sovereignty is being bypassed. All three Service, Data, and Hardware sovereignty are important and should be the focus. It is not about prioritizing one over the other.

New comment by akg_67 in "Europe built sovereign clouds to escape US control. Forgot about the processors"

akg_67 — Sat, 16 May 2026 13:06:36 +0000

> AFAIK, there is absolutely zero evidence either Intel or AMD CPUs are compromised, even less so that they're somehow remotely accessible by the US government...

The concerns are similar to US supplied fighters having the kill switch or remotely damaging centrifuges in Iran using software virus.

No one knows whether CPUs are compromised similar to no one knew beepers with explosives in Lebanon were compromised by Israel, allegedly during manufacturing. CPUs don't need to be accessed remotely, any compromised person locally will be enough.

These are fascinating cases to show how far state actors will go and how long the compromise can stay dormant.

New comment by akg_67 in "John Ternus to become Apple CEO"

akg_67 — Tue, 21 Apr 2026 21:31:49 +0000

Bold claim based on one month experience. May be living for 10 years and continuing to living here gives different impression.

https://www.reddit.com/r/JapanTravel/comments/xbejr2/apple_m...

New comment by akg_67 in "John Ternus to become Apple CEO"

akg_67 — Tue, 21 Apr 2026 08:35:04 +0000

Do you have a source that supports this claim?

I haven’t come across anyone using Apple Maps while living in Japan, most seem to use Yahoo! Maps or Google Maps.