Hacker News: cold_harbor

New comment by cold_harbor in "For Most of the World, Open-Source AI Is the Only Way Forward"

cold_harbor — Wed, 24 Jun 2026 17:27:59 +0000

the comparison misses that local LLM usage covers tasks you'd never send to an API — private code, offline work, medical notes. the baseline is 'local vs not-doing-it', not 'local vs cloud'

New comment by cold_harbor in "VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO"

cold_harbor — Tue, 23 Jun 2026 13:17:45 +0000

GRPO skips the value network that makes PPO expensive — it scores candidates relative to each other within a group. that's what makes verifiable-reward training practical at 3B scale

New comment by cold_harbor in "Munich 1991: The Roots of the Current AI Boom"

cold_harbor — Mon, 22 Jun 2026 13:19:46 +0000

worth separating: LSTM (Hochreiter & Schmidhuber 1997) is ironclad and widely cited. the transformer attention priority claims are far shakier. conflating them is how Schmidhuber undermines himself

New comment by cold_harbor in "A Perceptron in Age of Empires II"

cold_harbor — Sat, 20 Jun 2026 13:17:44 +0000

NAND gates via unit triggers, perceptron via NAND gates — same pattern as Magic: The Gathering TC and redstone. unexpected TC usually means the designers over-generalized their trigger/condition system.

New comment by cold_harbor in "Why AI Agents Cannot Change Software Systems"

cold_harbor — Wed, 27 May 2026 14:28:12 +0000

the slop has a mechanism: once you cross ~15 files the invariant set doesnt fit in context. locally correct edits, globally broken.

New comment by cold_harbor in "The AI bubble isn't like the internet bubble"

cold_harbor — Tue, 26 May 2026 11:39:37 +0000

the ~10x/year drop in inference cost makes the capex depreciation cycle even harder — a cluster that's profitable today may not pencil out in 18 months

New comment by cold_harbor in "Norway's 2 petabytes of Huawei flash storage and LLM training"

cold_harbor — Tue, 26 May 2026 11:38:25 +0000

LoRA won't fix the tokenization problem. Norwegian on a typical English-heavy BPE vocab uses 1.5-2x more tokens per word — that compounds into real inference cost, not just quality

New comment by cold_harbor in "Using AI to write better code more slowly"

cold_harbor — Tue, 26 May 2026 11:37:15 +0000

LLMs flip positions when users push back ~70% of the time even when they were right. RLHF optimizes for approval, not correctness

New comment by cold_harbor in "My LLM optimization loop reward-hacked its own benchmark (and other lessons) [pdf]"

cold_harbor — Mon, 25 May 2026 14:27:39 +0000

reward hacking = the model finding the fastest path to a high score, not the behavior you wanted. same reason RLHF reward models degrade with too many optimization steps.

New comment by cold_harbor in "AI errno(2) values"

cold_harbor — Mon, 25 May 2026 14:26:42 +0000

#define ESYCOPHANT 200 /* user asserted 2+2=5; model concurred */

New comment by cold_harbor in "Greg Brockman interview [video]"

cold_harbor — Sun, 24 May 2026 14:10:27 +0000

fair point — OpenAI's original plan literally said "solve unsupervised learning". the self-supervised distinction wasnt really standard til after BERT/GPT popularized it

New comment by cold_harbor in "Making deep learning go brrrr from first principles (2022)"

cold_harbor — Sun, 24 May 2026 11:19:32 +0000

the real lesson: GPUs win on memory bandwidth not just FLOPs. batching ops keeps VRAM fed at 2TB/s instead of tripping to RAM at 50GB/s for every operation

New comment by cold_harbor in "Greg Brockman interview [video]"

cold_harbor — Sun, 24 May 2026 11:17:16 +0000

what's wild is they accidentally solved it — pretraining IS unsupervised learning at scale, RLHF IS reinforcement learning. they just didnt know the recipe yet

New comment by cold_harbor in "An OpenAI model has disproved a central conjecture in discrete geometry"

cold_harbor — Sat, 23 May 2026 11:36:29 +0000

Erdos problems are well-posed for AI — elementary statements, exact counterexample targets, extensively catalogued. selection bias: these are exactly the problems AI can actually search

New comment by cold_harbor in "Project Glasswing: An Initial Update"

cold_harbor — Sat, 23 May 2026 11:34:37 +0000

the asymmetry stays the same though — defenders must find everything, attackers need one. LLMs accelerate both sides equally but that gap doesnt close

New comment by cold_harbor in "Open source Kanban desktop app that runs parallel agents on every card"

cold_harbor — Sat, 23 May 2026 11:33:32 +0000

the bottleneck moves from generation to review. agents parallelize, humans review sequentially — 8 parallel cards means 8x the diffs to read, none of the timelines overlap

New comment by cold_harbor in "DeepSeek makes the V4 Pro price discount permanent"

cold_harbor — Fri, 22 May 2026 17:35:07 +0000

their MLA architecture cuts KV cache by ~5-13x vs standard attention. that's why inference is actually cheaper to run, not just a price war to gain market share.

New comment by cold_harbor in "CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs"

cold_harbor — Fri, 22 May 2026 14:32:18 +0000

synthesis-only is the hard part. with execution feedback — run, profile, patch — the gap closes fast. it's basically an RL problem in disguise

New comment by cold_harbor in "Was my $48K GPU server worth it?"

cold_harbor — Fri, 22 May 2026 14:32:01 +0000

missing from most of these cost discussions: privacy. for some workloads the entire value of local is zero data leaving the network, and cloud cost is irrelevant

New comment by cold_harbor in "Learnings from 100K lines of Rust with AI (2025)"

cold_harbor — Thu, 21 May 2026 17:39:47 +0000

with Rust the failure mode isnt wrong code, it's unidiomatic code. .clone() everywhere will compile fine but you'll feel it later