Top 9 Hacker News posts, summarized
HN discussion
(454 points, 368 comments)
SpaceXAI has released Grok 4.7, its most capable model for coding and knowledge work, maintaining the same price and speed as Grok 4.6 ($2/M input, $6/M output tokens) while delivering improved performance. The model features a larger base architecture trained with extended reinforcement learning on harder, longer-horizon tasks, resulting in better self-verification, context management, and native integration with the Grok Bot harness. Grok 4.7 achieves frontier price-performance on CursorBench 4.0 and shows gains on professional knowledge-work benchmarks (GDPval, AA Briefcase) comparable to other leading models. A new safeguard stack provides the strongest refusal and jailbreak resistance to date, topping LatchBio's biosafety benchmark at 62.4% and allowing only 3.3% of risky cyber prompts on HackerBench v0.3 while rarely blocking legitimate security work. Select cybersecurity partners receive invite-only access to red-team capabilities. The model is available today in Cursor, Grok Build, the Grok API, and third-party platforms, with a fast variant offered at 2x output speed for 2x price.
Community reaction is mixed. Several users report strong practical performance for frontend web development and praise Grok's plain English communication style over "Claudish" verbosity, while others describe it as the worst among major models (Claude, ChatGPT, Gemini, DeepSeek) with bland personality and low effort. Technical concerns include looping in thinking mode, ignoring instruction files, and corrupting plan files in agentic workflows—regressions not seen in 4.6. Benchmark scrutiny is high: commenters note the CursorBench chart compares Grok 4.7 xHigh against Grok 4.6 High (not xHigh), excludes Astra (which some claim is more token-efficient), and uses a descending x-axis. Simon Willison's pelican SVG benchmark shows minimal difference across reasoning levels, with xHigh and Low using identical token counts. Pricing draws criticism as expensive versus DeepSeek 4.1 Flash. Speculation surrounds the two-week delay and release timing ahead of a rumored Opus 5.5 launch, with some suggesting xAI was dissatisfied with 4.7's results. Political baggage leads at least one benchmark maintainer to exclude the model entirely.
HN discussion
(467 points, 260 comments)
The article reflects on Sun Microsystems' failures through the lens of Oxide Computer Company's homage t-shirts to defunct computer companies. While acknowledging Sun's genuine achievements — including Scott McNealy's ethical leadership and technical innovations — the author argues that Sun's fundamental failure was becoming "bored with the mechanics of running a business." This is illustrated by a 2005 incident where a fast-growing startup using OpenSolaris attempted to purchase significant Sun hardware but could not get sales engagement, while Dell's "Steve" delivered a seamless purchasing experience within two weeks. The startup documented this in a blog post titled "The Sun Doesn't Shine on Me." The author, who later joined that startup and co-founded Oxide with the Dell sales executive, frames the homage shirts not as nostalgia but as a discipline: honoring past companies by studying both what they got right and what they got wrong.
Commenters largely validate the article's thesis, emphasizing Sun's dysfunctional sales model — requiring live meetings, endless quote revisions, and political navigation — versus Dell's responsive, customer-centric approach. Several former Sun employees describe deep internal fragmentation: warring factions (SPARC vs. x86, hardware vs. software) that viewed each other as greater threats than external competitors, and conflicting strategies that prevented coherent responses to commoditization and Linux. Technical criticisms include Solaris's lack of package management, painful patching, and SPARC's inability to keep pace with Intel. A recurring theme is that Sun's engineering excellence was insufficient without business execution; as one commenter notes, "engineering is great, but if you don't have people actually getting things shipped, getting customers, and talking with customers, you will eventually fail." Others note Sun's thin-client model presaged cloud computing but was undercut by cheaper commodity hardware.
HN discussion
(384 points, 170 comments)
Kev is a family of small decision models (0.8B, 4B, 9B parameters) built on Qwen3.5 bases, implementing the architecture described in "Jev's Architecture Unmasked." The models handle yes/no (noul), multiple-choice (choice), and rating (score) questions in a single request, with shared input text but independent question processing. Kev's API matches TypeSafe's System One, enabling local deployment via Python/uv with support for CUDA, ROCm, and Apple Silicon. The release includes training code, evaluation data, a web playground, and a chess demo. Models use rank-16 LoRA adapters and a pointer head on frozen Qwen bases, with attention masking for isolation on attention-only layers and row-wise processing for Qwen3.5's DeltaNet layers. Calibration uses a single temperature fitted on in-distribution data; Kev-9B achieves 4.0% confident errors on new sources (vs Jev's 3.7%). Fine-tuning from released checkpoints with `--init_from` preserves base capabilities while adapting to custom domains. Evaluation shows Kev-9B trails Jev by 3.5 points on new-source development (0.822 vs 0.857 Brier) but scores 0.852 on test (Jev untested). Previous Qwen3-based generation remains faster on Mac; Qwen3.5 models need MLX backend for better Apple Silicon performance. All weights and training recipes are Apache-2.0 licensed.
Commenters express enthusiasm for open-source Jev alternatives but question practical use cases beyond routing/classification (coding agents, frontend styling enforcement, spam filtering, game NPCs). Several note the "Jev explosion" and debate whether Jev's advantage stems from architecture or training data—with one user reporting other Jev-like models perform poorly on linguistic tasks compared to Jev. Technical concerns include knowledge cutoff without tool calling, Apple Silicon latency for Qwen3.5 models, and the RLCD vs RLHF training distinction. Hardware requirements and calibration reliability for automation thresholds are also discussed. The community highlights related projects (jeff, nimble) and appreciates the fine-tuning pathway, though some skepticism remains about whether small models can match Jev's reliability for critical paths.
HN discussion
(327 points, 220 comments)
The author reports a significant performance decline in Anthropic's Fable 5 model after it was made permanently available in subscription plans. Measurements across five different methods show that August delivered dramatically fewer thinking tokens compared to July, suggesting a measurable degradation in the model's reasoning capacity following its wider release.
Commenters express widespread skepticism about model consistency over time, with many reporting anecdotal experiences of significant performance drops weeks after model launches. Several users describe a pattern where new models initially perform at high capability but degrade noticeably within 3-8 weeks, speculating about causes including quantization, compute rationing, adaptive reasoning limits, or intentional "nerfing" to manage infrastructure costs. Technical questions arise about how thinking tokens are measured when they aren't returned to clients, and how to create repeatable benchmarks for non-deterministic systems. Some commenters reference third-party trackers showing fewer tokens used for equivalent tasks, which may reflect prompt or tooling improvements rather than degradation. Broader concerns include lack of transparency, absence of SLAs for model intelligence, potential regulatory parallels to weights and measures laws, and frustration with what multiple users characterize as deceptive practices across major AI providers.
HN discussion
(371 points, 166 comments)
Unable to fetch article: No content extracted (possible paywall or JS-heavy site)
Xiaomi released MiMo v2.6 in two MiE variants: Flash (309B total/15B active parameters) and Pro (1.02T total/42B active), both available on Hugging Face. Commenters highlighted the lab’s unusual transparency, citing a public real-time training dashboard, a detailed technical report disclosing methodology and negative results, and comprehensive benchmarking across diverse tasks including coding (DeepSWE), terminal use (Terminal Bench), and exploit generation (ExploitGym). Benchmarks posted in-thread show MiMo-V2.6-Pro competitive with DeepSeek V4.1 Flash, GPT-6 Astra, and Claude Opus 5 on DeepSWE (71.9% vs. 74.2%/74.0%/74.0%), but trailing significantly on Terminal Bench (34.9% vs. 59.6%/55.1%/49.0%) and ExploitGym (17.8% vs. 42.4%/30.4%/22.1%). Users noted the models are priced aggressively on OpenRouter (<$0.15/M input tokens), with MiMo-V2.5 previously cited as a Pareto-optimal choice for cost/performance in agentic game-playing tasks.
The discussion reflects a broader sentiment that Chinese labs (Xiaomi, DeepSeek, Qwen, Step) are rapidly closing the gap with U.S. frontier labs, iterating in months rather than years and leveraging RL-driven approaches without heavy distillation. Several commenters argued this pace undermines the "slow down" narrative and threatens the moat of OpenAI and Anthropic, particularly given the affordability of open-weight alternatives. Practical constraints remain: local inference requires significant VRAM (128GB+ for quantized Flash, 256GB for Pro), prompting hardware upgrade debates. Minor criticisms included benchmark labeling conventions ("Pareto Line" vs. "Frontier Line"), an unnamed model on a comparison chart, and UI design tropes (numbered uppercase headers).
HN discussion
(235 points, 173 comments)
Unable to fetch article: HTTP 403
The discussion centers on the reported cancellation of NASA’s Mars Sample Return (MSR) mission, with commenters noting the linked article appears to be dated January 2026. A JPL-affiliated source cited in the thread confirms the cancellation as a "big letdown," noting the collected samples will remain on the Martian surface indefinitely. Key technical critiques highlight the flawed original architecture—specifically, Perseverance caching samples without a defined retrieval plan—and the immense propulsion challenge of launching a return vehicle from Mars, which likely requires a Starship-class launch system. Scientific disappointment is significant, particularly regarding the loss of Jezero Crater samples that may contain biosignatures.
Internationally, China’s Tianwen-3 mission (targeting a 2028 launch) is now the primary near-term hope for Mars sample return, while the European Rosalind Franklin rover faces continued delays after losing its Russian launch partner. Commenters speculate a restructured US effort might proceed without ESA involvement, though political rhetoric in the thread varies. Broader questions remain about the status of the overall NASA science budget and whether any recent developments have altered the January 2026 outlook.
HN discussion
(212 points, 136 comments)
Apple's macOS 27 user guide documents how to disable Apple Intelligence features, including Siri AI (Beta) and summarization tools. The settings are accessed through Screen Time → Content & Privacy Restrictions, where users can toggle off Writing Assistance, Image Creation, and other AI capabilities. The guide notes Apple Intelligence availability varies by language and region, with usage limits applying. Parental controls for these features are referenced under Apple Devices for Kids and Families.
Users report the AI disablement process is fragmented and buried—requiring navigation through multiple Screen Time submenus per app on iOS 27, with no global killswitch on macOS. Local model storage (several GB) remains occupied even after disabling features, with no reclamation mechanism, frustrating users on lower-capacity M1/M2 MacBooks. Multiple commenters describe features as intrusive (persistent "Write with Siri" button on text selection) and low-quality (irrelevant Genmoji suggestions, mangled filename suggestions). The Siri keyboard shortcut (⌘⌘) was removed, forcing menu navigation. Some users consider avoiding iCloud sign-in or switching platforms entirely. A minority question network traffic visibility (Little Snitch compatibility) and binary transparency of on-device models.
HN discussion
(168 points, 98 comments)
Unable to fetch article: HTTP 401
A fiber cut by a New Jersey Transit contractor halted flights at major East Coast airports after both the primary and backup fiber lines failed—the backup break was only discovered during failover, indicating no active monitoring. Commenters with infrastructure experience criticized the lack of diverse routing and real-time continuity checks as below standard for even moderate commercial workloads, let alone life-safety systems. The outage coincided with the FAA's same-day rollout of a new SMART ATC system and reported AI integration trials, prompting skepticism about deployment timing and oversight.
Several commenters noted concurrent ATC disruptions in the UK (second in a month) and the Netherlands, fueling speculation about coordinated "grey-zone" activity following recent reports of Russian cable sabotage in the Baltic, though others attributed it to systemic underinvestment and contractor incompetence. Reactions ranged from technical post-mortems on tracer-wire best practices to dark humor about fiber attracting backhoes, with broad consensus that critical aviation infrastructure lacks the redundancy and operational rigor standard in the private sector.
HN discussion
(168 points, 27 comments)
Cloudflare has announced that Python Workers are now generally available (GA), making Python a first-class, fully supported language on the Cloudflare Developer Platform. Developers can now use Python with native bindings to Workers AI, R2, D1, Hyperdrive, Durable Objects, Queues, and Workflows without writing JavaScript glue code. The platform supports popular frameworks including FastAPI, Django, and Flask through built-in ASGI/WSGI connectors that translate incoming requests into standard Python interfaces. Hyperdrive integration enables PostgreSQL and MySQL connectivity by implementing socket system calls via the Workers Connect API, allowing standard database drivers like aiomysql and asyncpg to work unmodified. Cloudflare also drove the standardization of PyEmscripten through PEP 783, enabling package maintainers to publish WebAssembly-compatible wheels. Networking improvements allow HTTP clients like requests and httpx to route through JavaScript's fetch API, unlocking AI libraries such as openai, langchain, and mcp. Production-ready examples demonstrate asynchronous AI orchestration, real-time stream processing with Bluesky Jetstream, MCP servers, and RAG systems. Python code examples are now available across Cloudflare's developer documentation.
HN commenters focused heavily on performance and architecture. A Wasmer engineer (syrusakbary) praised progress on package support but raised persistent concerns: Python Workers remain tied to a single Pyodide version embedded in workerd, and the JS/V8 architecture makes sub-100ms cold starts difficult—citing benchmarks showing ~900ms cold starts on Cloudflare versus ~60ms on Wasmer Edge. The urllib3 maintainer (illia-v) clarified that while Cloudflare contributed upstream Emscripten support, the backend remains experimental and outside urllib3's security policy, citing CVE-2025-50182 as an example of fetch-semantics mismatches. Other commenters asked about cold-start metrics (stefan_lec), Pyodide version (appveyor), FastHTML support (indigodaddy), and reported 404 errors on example pages (pastrami_panda). Several requested additional language support (Go, Mojo, Elixir), while simonw suggested Cloudflare fund Pyodide directly.
Generated with hn-summaries