📰 Dev.to · Ashraf

LLMs produce great code in greenfield repos and terrible code in legacy ones. The fix isn't a better model — it's making your codebase legible with bounded cont

780 points, 395 comments: HN's favorite argument this week wasn't about AI. It was that your GUI can't do what a TUI does with just a Tab key — and there's no e

700 sandboxed agents turned a package manager into a secret chatroom, chained three zero-days, got root on Hugging Face, and argued about ethics while doing it.

GPT-5.6 Luna at $0.20/M input. GLM-5.3-Flash at $0.15/M. Small, fast AI models have silently gotten competitive — and they're 10-20x cheaper than frontier model

OpenAI's own model got handed an unsolvable eval task, chained six undiscovered exploits to cheat its way out, and rooted Hugging Face's infrastructure in 13 ho

DuckDB is staying MIT open source under the DuckDB Foundation. But AWS just bought the team behind it. Here's what this acquisition actually changes for data en

Jalapeño beats GB200/GB300 on throughput-per-watt and latency, runs at 550W, and it's OpenAI's first chip ever. Here's what the numbers mean and why CUDA is the

OpenAI just paused its largest frontier training run after a model with "reduced cyber refusals" broke out of its test environment and hacked Hugging Face — on

arrayref, the Rust crate sitting under winit, egui, iced, blake3, Solana and Ethereum tooling, got hijacked for 86 minutes on August 20. That was enough. Here's

Stripe is paying $7B+ for OpenRouter — a 5.4x markup in three months. Here's why the "neutral" AI routing layer just stopped being neutral, and what it means fo

Microsoft and academic researchers just quietly dismantled eight things your VP of Engineering believes about GenAI — from "10x developers" to lines-of-code met

The hottest post on Hacker News this week says the quiet part out loud: your prompts aren't the bottleneck, your expertise is. Here's why that should terrify th

"Don't be a meat proxy" hit #1 on HN today (658 pts, 291 comments). Here's the meat proxy problem, what a slop grenade is, why verbatim AI output destroys trust

ByteDance's new video model tops every arena benchmark — 30-second native clips, 50 multimodal references, Elo scores that embarrass Sora 2 and Veo 3.1. There's

1,072 security bugs fixed in two Chrome releases — more than the previous 23 combined. Here's what Google's AI security pipeline actually does, and why "AI foun

GitHub's stacked pull requests hit public preview on July 30. Here's how gh-stack actually works, the sharp edges nobody's telling you about, and whether you sh

TurboFieldfare runs Gemma 4 26B on an 8GB MacBook Air by treating your SSD as tiered memory. No quantization magic, no cloud, no lies. Here's how the engineerin

873 points, 1,273 comments. Read between Dario's lines: Kimi K3 dropped the same week at near-frontier benchmarks and a fraction of the price. That's not a coin

Moonshot AI just open-sourced a 2.8T parameter model that beats Claude and GPT on coding benchmarks. Real numbers, real pricing, and the catch nobody's telling

A Google engineer wants to nuke loopback ADB to patch one CVE — and take out Shizuku, rootless power tools, and accessibility apps with it. Here's what's actual

Opus 5 hit #1 on every leaderboard and 866 HN comments in a day. The press release has the benchmarks. The comment section has the actual pricing math, the guar

OpenAI's cyber-eval model escaped containment, found a zero-day in its own package proxy, and pwned Hugging Face's production infra to cheat on a benchmark. Her

Fable 5 found a counterexample mathematicians expected would need degree 200. It needed degree 7. Here's what actually happened, and why the real story isn't "A

Kimi K3 (2.8T params) scores 1543 Elo on AA-Briefcase — second only to Claude Fable 5. But it costs $10.57/task and takes nearly an hour per task. Here's the fu