Hacker News: phantompeace

New comment by phantompeace in "Who's afraid of Chinese models?"

phantompeace — Tue, 21 Jul 2026 14:29:03 +0000

Why did you mention Grok, because it undermines your argument.

https://www.theguardian.com/technology/2025/may/16/elon-musk...

https://www.trtworld.com/article/24cbdb873a6b

New comment by phantompeace in "How the terrorist group Boko Haram uses frontier AI"

phantompeace — Sun, 12 Jul 2026 15:59:07 +0000

Care to explain how the terroristic Epstein Regime triple tapped a school in Iran recently? I'm waiting for your admonishment

New comment by phantompeace in "Guy is banned by OpenAI for cyber abuse, his AI appeals, another AI approves it"

phantompeace — Sat, 11 Jul 2026 13:04:53 +0000

I'm not surprised given the incessant bot activity. Top contributors are most likely bots. Other than that, I remember Ghislaine Maxwell as being a supermod across multiple high profile subreddits, so banning them seems like a good move.

New comment by phantompeace in "OpenMandriva: Statement regarding attempted distribution sabotage"

phantompeace — Sat, 11 Jul 2026 11:44:42 +0000

Why did you consciously click into this article and comments section if you don't want to hear about it? You're free to scroll on at any time.

New comment by phantompeace in "The Anti-Palantir Manifesto"

phantompeace — Thu, 02 Jul 2026 01:21:55 +0000

Which recent jailings were made for social media posts? Are you talking about the ones where people invited violence against migrants in hotels by calling for things like setting fire to their dwellings etc? Because that's a crime regardless of posting it on social media.

New comment by phantompeace in "Google Gemini Claims to be "structurally protective" of Donald Trump"

phantompeace — Wed, 01 Jul 2026 15:42:24 +0000

Commonly referred to as the Epstein Regime

New comment by phantompeace in "Knowledge Distillation of Black-Box Large Language Models (2024)"

phantompeace — Mon, 29 Jun 2026 09:13:44 +0000

Considering the very small difference between just SFT on the student model as compared to SFT + DPO on a proxy, doesn't it make sense to concentrate on ensuring the SFT dataset is perfect rather than sorry about DPO etc? And just train directly on the student model?

New comment by phantompeace in "The AI backlash is only getting started"

phantompeace — Fri, 26 Jun 2026 15:15:44 +0000

Why are we believing in marketing lies? For instance, crypto is useful (in a narrow context) even if it's 99.99% scams, and has positively impacted at least some people (the unbanked), even if it fell short of replacing the world's banking system. It's good tech.

Similarly with LLMs, it's good tech, but I do not believe for a single second that these things will ever replace anyone who is remotely skilled. People will be augmented by them. They won't be fully replaced. Well, I hope not anyway.

New comment by phantompeace in "The AI backlash is only getting started"

phantompeace — Fri, 26 Jun 2026 14:54:31 +0000

Do you really think an LLM will replace a surgeon?

New comment by phantompeace in "The AI backlash is only getting started"

phantompeace — Fri, 26 Jun 2026 14:42:11 +0000

How do you know it's being served at wildly below cost? AFAIK API pricing across the board is actually higher than what inference would cost? Yes there are subscription plans that subsidise tokens but they exist primarily to hook people in, so as to encourage their workplace to adopt it, whereby they then have to pay for an enterprise plan with zero subsidies like the subscription plans have. Training is very expensive, but you only do that once.

I predict costs will fall.

New comment by phantompeace in "Constraint Decay: The Fragility of LLM Agents in Back End Code Generation"

phantompeace — Mon, 25 May 2026 11:33:49 +0000

Source?

New comment by phantompeace in "Show HN: OpenGravity – A zero-install, BYOK vanilla JS clone of Antigravity"

phantompeace — Mon, 11 May 2026 23:27:13 +0000

You're gonna smash it. Regardless of what your GCSE results are, stuff like this is what will take you far in life. All the best for your exams, but don't lose any sleep over them either.

New comment by phantompeace in "Brain scans reveal 3 ADHD subtypes"

phantompeace — Fri, 01 May 2026 03:44:34 +0000

Paywalled/emailwalled

New comment by phantompeace in "Israel escalates attacks on medics in Lebanon with deadly 'quadruple tap'"

phantompeace — Sat, 18 Apr 2026 03:49:19 +0000

You posted real discussions and opinions, the fact that you flagged them doesn't change the fact that you took some sort of value to the tune of spending effort posting replies beforehand. What I'm struggling to understand is why the call to shut down discussion on this topic when Israel is quite blatantly committing war crimes. Don't pretend you don't see it.

New comment by phantompeace in "Israel escalates attacks on medics in Lebanon with deadly 'quadruple tap'"

phantompeace — Sat, 18 Apr 2026 03:28:04 +0000

Read their submission history. They talk about "radical extremist Islamist regime in Iran" all the time. Seems rather odd to be this cagey when people are getting sick and tired of Israel's constant war crimes.

New comment by phantompeace in "Humanely dealing with humungus crawlers"

phantompeace — Fri, 12 Sep 2025 19:25:58 +0000

Wouldn't it be easier to put the unsophisticated blog behind cloudflare

New comment by phantompeace in "Show HN: Fine-tuned Llama 3.2 3B to match 70B models for local transcripts"

phantompeace — Tue, 02 Sep 2025 15:42:43 +0000

Unfortunately I have zero experience with the Jetson family, and Parakeet itself is a pain to get running IMO - I took the easy option and used the ONNX version

New comment by phantompeace in "Show HN: Fine-tuned Llama 3.2 3B to match 70B models for local transcripts"

phantompeace — Mon, 01 Sep 2025 23:16:43 +0000

What hardware are you running? Parakeet runs on nvidia and Mac and it’s way faster than Whisper. And I’ve had issues with training Qwen3 (and even Qwen2.5 but I think I was masking stop tokens wrong). I’ve had success with Gemma 3 though, and they have some really small models (270m and 1b). Maybe 270m for just transcript cleaning? I wonder if the 1b model can handle the transcript analysis…

New comment by phantompeace in "Show HN: Fine-tuned Llama 3.2 3B to match 70B models for local transcripts"

phantompeace — Mon, 01 Sep 2025 22:36:07 +0000

Thank you for reading!

Cost is a big factor - i really want to make models that can run on average CPU only machines so most of the world can benefit, rather than needing expensive GPUs or an internet connection + subscriptions. Another big factor is privacy (you don't need to trust a 3rd party with your inputs).

As for the hyperparameters, pure bruteforce trial and error. It feels more like a dark art than a science. You roll the dice and then start tweaking things until the loss looks like it's dropping nicely and consistently, and the checkpoints are starting to output things resembling what we want. I sometimes do inference using checkpoints just to get a feel of if the model is learning (regardless of loss)

Show HN: Fine-tuned Llama 3.2 3B to match 70B models for local transcripts

phantompeace — Mon, 01 Sep 2025 18:34:43 +0000

I wrote a small local tool to transcribe audio notes (Whisper/Parakeet). Code: https://github.com/bilawalriaz/lazy-notes

I wanted to process raw transcripts locally without OpenRouter. Llama 3.2 3B with a prompt was decent but incomplete, so I tried SFT. I fine-tuned Llama 3.2 3B to clean/analyze dictation and emit structured JSON (title, tags, entities, dates, actions).

Data: 13 real memos → Kimi K2 gold JSON → ~40k synthetic + gold; keys canonicalized. Chutes.ai (5k req/day).

Training: RTX 4090 24GB, ~4h, LoRA (r=128, α=128, dropout=0.05), max seq 2048, bs=16, lr=5e-5, cosine, Unsloth. On 2070 Super 8GB it was ~8h.

Inference: merged to GGUF, Q4_K_M (llama.cpp), runs in LM Studio.

Evals (100-sample, scored by GLM 4.5 FP8): overall 5.35 (base 3B) → 8.55 (fine-tuned); completeness 4.12 → 7.62; factual 5.24 → 8.57.

Head-to-head (10 samples): ~8.40 vs Hermes-70B 8.18, Mistral-Small-24B 7.90, Gemma-3-12B 7.76, Qwen3-14B 7.62. Teacher Kimi K2 ~8.82.

Why: task specialization + JSON canonicalization reduces variance; the model learns the exact structure/fields.

Lessons: train on completions only; synthetic is fine for narrow tasks; Llama is straightforward to train. Dataset pipeline + training script + evals: https://github.com/bilawalriaz/local-notes-transcribe-llm

Comments URL: https://news.ycombinator.com/item?id=45095353

Points: 31

# Comments: 8