Hacker News
1OpenCode Hacked: User credentials leaked
- OpenCode reported a security breach caused by a large-scale distillation attack that affected 1.4% of its Console users, who were notified by email. The incident is ongoing, and OpenCode is working with Anthropic and others, though it cannot share more details yet.
Scope of the impact
Commenters agree the leak affected only OpenCode Console, a hosted OpenRouter-like service where BYOK keys were stored server-side, not the open-source CLI or desktop app. Some suspect local .env files were also read.Poor communication
Many criticize OpenCode for confusing naming and an unclear announcement that blurs Console with the coding agent. Users want a detailed root-cause analysis, and the team is seen as having suffered brand damage.Key revocation design
Proposals for an invalidate_api_key endpoint so leaked keys can be revoked quickly. Others warn it could enable denial-of-service abuse and suggest protecting it with auth.
2DeepSeek-v4.1-Flash-1017
- DeepSeek has announced a new V4.1 Flash model in an official news update.
Practical advantages over frontier models
Users say DeepSeek isn't top on benchmarks but "feels" better: fast tokens/s and low TTFT, visible chain-of-thought, no cyber-related blocking, strong cost value, and code that stays closer to what a developer would write.Vendor lock-in and regulatory capture
Commenters argue OpenAI and Anthropic push regulation to stall open and Chinese models, and that frontier labs' revenue depends on compute rather than models, making their position fragile.Model speed and review quality
One tester found DeepSeek v4.1 Flash fast but slower to deliver useful reviews than GPT 6.1 Sol, which found more relevant issues and planned better fixes. Others say results are non-deterministic.
3DeepSeek: DeepSeek v4.1 Flash 1017
- DeepSeek V4.1 Flash 1017 is a sparse mixture-of-experts model with a 1.0M-token context window, priced at $0.08 input and $0.50 output per million tokens. It is aimed at coding, terminal, and computer-use agents, and at long-horizon and long-context tasks. Providers on OpenRouter range from about 52 to 295 tokens per second, with uptime of roughly 99.9%.
Open model diversity
A commenter argues that regulatory hurdles pushed by OpenAI and Anthropic would reduce the diversity needed for solving complex problems. They see a diverse ecosystem, including open models like this one, as the best long-term approach.Regulation and foreign open models
A reply asks how regulatory hurdles would help against foreign open models, noting that China would not cooperate with such regulation.
4You could have invented transformers from memory
Clarity for newcomers
Readers praised the writing but said terms like d_k, M, "tokens," and "similarity" go undefined and the notation is dense. Others replied that the terms are defined in a "Here are the terms" section.Whether the memory analogy explains LLMs
Commenters explained attention as a key-value lookup, with keys and values as learned projections of context tokens. Critics argued the analogy is weak for training and that context tokens aren't really memories.Engineering and neuroscience perspectives
Electrical and communications engineers found concepts like interference-to-signal ratio and capacity familiar, and urged applying channel-capacity intuition. Others pointed to hippocampal pattern separation research and 3Blue1Brown's "compression is intelligence" video.
5How the Claude Code team does code review
- Claude Code's team now has Claude write about 90 percent of its code, so review has shifted from manual line-by-line reading to layered automation. Machines handle most checks, including self-review, AI pull request review, security scans, and CI, while humans focus on risky changes. Merged pull requests more than doubled from June to September, and reverts fell over 40 percent.
Code comprehension and trust
Commenters worry that relying on Claude to explain and debug a huge codebase, instead of humans keeping a mental model, removes recovery options when the system fails. Some argue humans have always lost this context.Claude-style writing
Readers say the article's prose has recognizable LLM patterns, and they note that heavy exposure to AI text may be shaping how people write. Some non-native speakers say they can't tell the difference.Anthropic's profitability and costs
Critics say the company is unprofitable and that its token-heavy review process ignores cost. Defenders argue the spending is an experiment and that AI review can still be cheaper than human review.
6Solar surpasses nuclear power globally for first time
- Solar and wind now outpace nuclear globally. WNISR 2026 reports solar surpassed nuclear generation in 2025, while nuclear capacity has stayed flat for 25 years, with slow growth, long construction delays, and limited SMR progress.
Solar overtakes nuclear in output
Commenters note solar's rapid growth, with generation roughly matching nuclear's 2025 output. Wind also briefly topped nuclear monthly, and solar is expected to pass hydro in a few years.Nuclear stagnation and competitiveness
Some blame regulation, overbuilding, and R&D spending for nuclear's decline. Others argue it is uncompetitive on cost and build time compared with solar and batteries.Storage and grid reliability
Commenters say long-duration storage is needed to cover evening and low-wind periods. Home batteries, transmission, and overbuilding are proposed, though some doubt these fully solve intermittency.
7I love software rendering
- The author loves software rendering, where the CPU draws every pixel, and prefers its sharp, gritty look over the softer 3D-accelerated versions of classic DOS games like Quake, Doom, and Blood. Running these on a fast modern PC allows resolutions far beyond what 1990s hardware could handle, though the heaviest games still demand powerful CPUs.
Aesthetic not exclusive to software rendering
Commenters argue the look the author loves, such as nearest-neighbor textures and no anti-aliasing, can be reproduced on GPUs. Software rendering is not the essential ingredient, and GPUs are far more efficient.Pixelated textures and filtering
Many note the article is really about nearest-neighbor texture sampling rather than software rendering. Some prefer sharp pixels over bilinear filtering, while others criticize TAA and DLSS artifacts.Historical CPU limits and performance
Discussion covers how early CPUs struggled with sound, rendering, and resolution, why sound cards and 3dfx helped, and how modern CPUs could run old engines far faster.
8Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console
- PS5LM turns a jailbroken PS5 into a private home model server. It runs GGUF language models and decision models on the console's GPU and serves them over the network through OpenAI and Anthropic-compatible APIs, with no PC, cloud, or account required. Users can load models from a TV interface, watch requests stream, and monitor console temperature and load.
Takedown resilience of project hosting
Commenters debate whether the project should leave GitHub for a custom domain. Replies argue git's distributed design makes the code resilient, that GitHub's social features are the real draw, and that DMCA takedowns are hard to avoid, with torrents, magnet links, and non-DMCA git hosts as suggested fallbacks.Value and quality of running LLMs locally
Users see a cheap, always-on use for a PS5 running a 2-bit Qwen 27B: monitoring, log review, and background tasks at about 70 cents a day. Quality is seen as well below full-precision models and slower than Claude or Codex, though some find it useful for codebase reading.Repurposing locked-down hardware
Hackers turning consoles into general-purpose machines is seen as a long tradition, with the PS3 cluster and BC-250 boards cited. Commenters note that rising RAM prices make used PS5s good value, though they may drive up prices.
9Apple ML Research: Muse Spark Open with MLX on Mac Studio
- MLX on Apple silicon lets researchers run frontier-class models locally. Using four M5 Ultra Mac Studios with RDMA over Thunderbolt, Muse Spark Open (816B parameters) generates over 36 tokens per second. M5 Ultra outperforms M3 Ultra, with up to 4.6x faster time to first token and 1.6x faster generation.
Apple's shovel-selling strategy
Commenters argue Apple profits by selling hardware for local AI rather than training models, unlike Nvidia. Critics counter that its AI efforts, like Siri, have lagged and its CUDA and datacenter gaps are real.Meta partnership and benchmark skepticism
Users question the benchmarks, which compare Meta's model at 4-bit against Qwen at BF16. Others debate whether Apple and Meta are really partners, given their history of rivalry.Cost and practicality of local hardware
A 4-node Mac Studio cluster costs tens of thousands of dollars, so many say API access is cheaper. Some see a niche for privacy-sensitive or regulated users, and others doubt 8K context is enough.
10Butlerian Jihad
Writing and thought
Commenters argue LLM text mimics form but lacks intent, and even light AI editing subtly erodes one's own writing. Others see AI-driven conformity to the median, echoing Heidegger's worry that technology trains people to think like machines.Who the Butlerian Jihad targeted
Many argue Herbert's Jihad was aimed at technocrats who used machines to enslave people, not at machines going rogue. Others note the Dune Encyclopedia and Brian Herbert's sequels give different accounts of its causes.Plot armor in Dune
Some call the setting contrived, arguing that spice, shields, and knife fighting only make sense in a world built to sustain the Jihad. Others reply that this is ordinary world-building that serves Herbert's point about human agency.