Hacker News
1Ousted Savannah police officer gave ICE agent husband access to Flock system
- Savannah police fired Amber Harpe after she gave her ICE agent husband her Flock license plate reader login, which he used for about 983 searches tied to ICE targets. Her termination was upheld and her appeal was withdrawn.
Law enforcement abuse of surveillance tools
Commenters argue every surveillance tool given to police gets abused, that officials naively assume officers will follow rules, and that a Flock officer sharing her login with her DHS agent husband shows the problem.Weak access controls for sensitive systems
Many were surprised Flock uses basic username and password login accessible from anywhere. Commenters noted that health and corporate systems typically restrict access through IP allowlists, VPNs, device certificates, and MDM.Accountability and legal consequences
Users question whether credential sharing violates the CFAA and should be prosecuted, but doubt charges will come since ICE and federal officials are seen as above the law. Chatrie v. US may affect Flock warrant rules.
2Linux 7.3
- Linus Torvalds released Linux kernel 7.3, which opens the 7.4 merge window the next day. This record-breaking release includes 18,765 non-merge commits from 3,021 developers, 807 of them first-time contributors. Notable additions include the failfs filesystem, the fchroot() system call, and completion of the sched_ext subscheduler feature.
Smb and Time Machine support
Commenters debate ksmbd's new opt-in Time Machine support. Some note Apple's SMB stack is buggy and hard to work with, while others point to Apple's published SMB client source. Others question whether an in-kernel SMB server is worth it versus userspace Samba.AI-assisted kernel contributions
Commenters link the record 7.3 contributor count to AI-assisted patches. They note the kernel allows AI use with the submitter responsible for quality, and that the disclosure requirement was reportedly dropped.Apple silicon Linux support
Device tree support for M3 Pro, Max, and Ultra is welcomed as fast progress. Critics counter that GPU and Thunderbolt support is still missing, and that some community work is messy and unstable.
3Someone behind the software
- Good products carry the perspective of their makers, shaped by experience, care, and judgment rather than just requirements and process. Linear values intuition built through customer conversations and learning. Many products now feel messy, growing by addition rather than unified vision, as agent-driven individual workflows dilute character. The author hopes easier software creation lets people follow a clear perspective, and wants to feel the creator behind the software.
Taste and judgment remain human
Commenters argue AI can write code, but product decisions, taste, empathy, and knowing what not to build still need humans. Others doubt AI's taste and note that an AI-built product without a clear vision looks generic.Feature bloat dilutes products
Products grow by adding capabilities rather than rethinking them, becoming "mega-products" without a unified point of view. Commenters cite social media and SaaS examples and warn that focus gets lost.Being your own customer
Linear's advantage may be that its team uses its own product, which helps them make good judgment calls. Others reply that most developers work with non-technical domain experts and can't rely on that.
4Timestral: Mistral's time-series foundation models
- Mistral releases Timestral, its first family of time-series foundation models in four sizes (5M to 209M parameters). The 209M model ranks first on three major forecasting benchmarks, outperforming larger models. It delivers probabilistic, multivariate forecasts, runs efficiently on CPUs, and supports zero-shot forecasting and fine-tuning via Mistral Forge.
Value of covariates and domain knowledge
Commenters debate whether time-series models need external covariates like holidays, gas prices, or weather. Some argue large pretrained nets can infer patterns from history alone, while others say ignoring known drivers wastes signal.Foundation models versus classical methods
Users note ARIMA, Holt-Winters, and per-series XGBoost remain strong and fast. Foundation models are seen as useful for scaling across thousands of series, though skeptics question their real-world advantage.Licensing and benchmark skepticism
Commenters criticize the "Modified" MIT license's $20M revenue restriction and the 0.03 gain being framed as a large improvement. They urge comparisons against naive baselines.
5Hybrid Reasoning Agents: LLM agents that think in latent space, act through text
Latent-space reasoning raises safety and monitorability concerns
Commenters argue that reasoning in latent vectors instead of text removes the main tool for auditing model intent, calling it "the Most Forbidden Technique," though some note CoT was already unfaithful and latent monitoring may be needed.Looped transformers are similar to existing approaches
Participants say looped transformers equal a deeper network with shared weights, a form of continuous chain of thought, and that latent states can be decoded into text with interpreter models.Small local model struggles with looping and reasoning
A user testing the 8B GGUF found it stuck in a repetitive thinking loop on a simple question, while others suspected inference-stack or llama.cpp incompatibility rather than the model itself.
6Wine ported to OpenBSD
- Thomas Frohwein proposes a new OpenBSD port of Wine, the Windows compatibility layer, developed collaboratively after work on %gs/gsbase emulation. Testing of several dozen 64-bit Windows programs and games, including The Witcher 3 Remaster, has mostly succeeded, while 32-bit software does not yet work. He asks for testers and approval.
Wine port technical hurdles
Commenters detail the hard parts of porting Wine to OpenBSD: %gs register handling, SIGUSR1 thread signaling replaced with SIGUSR2, and ReadProcessMemory restrictions that break Electron and Steam.AI-generated code debate
Users argue whether OpenBSD's stance against LLM code is principled or anti-AI bias. Some cite copyright concerns, while others note a Claude-assisted commit and question the port's origins.GPU driver support on OpenBSD
OpenBSD ports Linux open-source AMD and Intel drivers, with no Nvidia support. Commenters say this makes the port usable on AMD and Intel hardware, and that VKD3D works through Vulkan.
7I no longer care to work on any Ruby gems so I've archived all my repositories
- The author is archiving their Ruby gems to show they've stopped working on them. They say Ruby was once designed for human joy and creativity, but Ruby and Rails are now shaped by LLM agents, and they object to the commons being exploited to train AI. They are no longer willing to contribute.
Grief and loss of joy in coding
Many developers who got into software for the love of writing code feel grief as AI takes over, losing the human, collaborative craft. Others find AI liberating, shifting focus to design, problem-solving, and building.Craft versus economics of AI-generated code
Commenters compare AI coding to artisan carpentry versus robotic housebuilding, debating whether craft survives economically. Many fear job losses, wage drops, and rug-pulls, while others see craft persisting as a hobby.Whether LLMs have produced great new software
Some argue that despite years of LLM development, no standout new software has emerged, with bugginess persisting. Others counter that bespoke personal tools are proliferating and that the projects are simply too new.
8MiniMax M3.1 Flash
- MiniMax launched M3.1 Flash, a next-generation model with major gains in agentic tasks, coding, computer use, and cybersecurity, running nearly three times faster at 70% lower cost. It is available now via API, M Plan, and MiniMax Code; model weights and the Pro model follow later.
Benchmarks overstate MiniMax's real-world performance
Users who tested the model as the free stealth release "Space Bunny" found it underwhelming, with many early mistakes and weaker results than DeepSeek V4 and GLM. Commenters argue benchmark scores are increasingly unreliable due to incentives to optimize for them.Cost-performance favors DeepSeek over pricier options
Commenters say DeepSeek Flash offers strong value at very low prices, with some switching from Claude subscriptions. Others note that output costs, verbosity, cache pricing, and quota limits change the real cost comparison.Model announcements rely on selective, cherry-picked comparisons
Commenters say launch charts pick favorable benchmarks and rivals, often skipping the latest frontier models. They recommend independent evaluations like Artificial Analysis and their own testing over marketing claims.
9Dogs seem to agree with humans that 'bouba' is round and 'kiki' is spiky
- Dogs, like humans, appear to link the made-up sound "Bouba" with round shapes and "Kiki" with spiky ones. In a study led by Eleonora Biffi at the University of Parma, 70 dogs looked longer at the rounded shape when "Bouba" played and at the spiky shape when "Kiki" played.
Sound-shape mapping across animals and humans
Commenters argue bouba/kiki reflects real-world sound-shape links (sharp objects make percussive sounds, round ones softer), possibly shared with dogs, birds, and chicks, rather than arbitrary linguistics.Language origins and the step-change debate
Users debate whether language emerged suddenly (~100,000 years ago, per Chomsky-influenced views) or gradually, and whether animals share the foundations of language like understanding basic words.Whether dogs truly understand the effect
Skeptics question if dogs are simply copying human cues (a Clever Hans effect), while others note controlled chick studies and cross-language consistency support a genuine innate association.
10I built a drone with only one motor [video]
Underactuated control in robotics
Commenters praise underactuated robotics, citing BattleBots designs that move with a single motor by timing speed changes to spinning impactors. Others note fighter jets and parafoils as examples, and recommend MIT's Russ Tedrake course.Single-motor drone design
Users say the virtual swashplate concept was done earlier in RC helicopters. Critics point out that a single motor makes the drone spin in flight, while others suggest imaging or scanning uses for the spin.Robotics learning and autonomous threats
Commenters advise learning through obsessive projects and LLM-assisted study. Others express concern about swarms of mini drones being used for harmful purposes.