0:00 / 1:29ai Swiss AI initiative releases Apertus Mini as an open, sovereign model set
The Swiss AI Initiative, led by EPFL, ETH Zurich, and CSCS with Swisscom as a strategic partner, released Apertus Mini: 16 small language models meant to demonstrate distillation and quantization techniques. The project emphasizes open weights, open data, documented training methods, reproducibility, multilingual coverage across 1,000-plus languages, and EU AI Act-aligned practices such as opt-outs, PII removal, and anti-memorization. The broader significance is not just another model drop, but a European push for sovereign AI infrastructure that can be inspected, rebuilt, and governed outside the dominant US and Chinese AI ecosystems.
Discussion: Mixed — HN liked the open-science and sovereignty goals, especially the promise of open weights, data, code, training methods, and EU AI Act-oriented data handling. But many commenters were skeptical about model quality, pace, and competitiveness, with several citing weak prior Apertus performance, hallucinations in multilingual use, and stronger alternatives such as OLMo, Nemotron, and Chinese open models. (Open weights, open data, and reproducible training are valued even when performance lags, Sovereign AI is seen as increasingly important amid distrust of US and Chinese platforms, Skepticism that a committee-style academic/government effort can keep up with frontier or top open models)
0:00 / 0:26ai Open-weights GLM 5.2 gets close to Opus, but not all the way
Techstackups compared Z.ai’s open-weights GLM-5.2 against Claude Opus 4.8 by asking each agent to build the same 3D platformer from scratch in raw WebGL. Opus finished in about 33 minutes versus about 71 minutes for GLM, and the author judged Opus’s game cleaner, while GLM cost much less: $5.39 billed versus an estimated $21.92 for Opus. The piece argues GLM-5.2 is not a drop-in Opus replacement, especially because it is text-only, but that its MIT-licensed open weights, 1M-token context, and low API pricing make it a serious tool to keep in the stack.
Discussion: Mixed — HN was impressed that an open-weights model can produce a credible raw-WebGL game and many commenters praised GLM 5.2’s cost-to-capability ratio. But the dominant discussion was methodological skepticism: a single greenfield agent run, with different harnesses, was seen as a weak proxy for real coding work. Several users also said Opus still feels better for collaboration, steerability, and taste, while GLM is slower and less multimodal. (skepticism of one-shot and greenfield coding tests, cost-performance appeal of open-weights models, importance of coding harnesses and system prompts)
0:00 / 0:27ai Open-weight AI is catching up, but HN isn’t buying “minimal downside” yet
Andrew Marble argues that moving from Claude and GPT-class proprietary models to open-weight models is starting to look less professionally costly, comparing it to how Linux eventually became a practical default rather than a risky compromise. The trigger is Claude’s identity-verification rollout, plus broader concern about safeguards and trust in closed AI services. He still acknowledges that proprietary models top leaderboards, have better APIs and tooling, and are more socially accepted for confidential work, but says open models are now close enough that the productivity hit may be manageable.
Discussion: Mixed — Commenters were sympathetic to the appeal of open-weight models, especially for privacy, control, resilience, and avoiding Claude’s identity-verification rollout. But the dominant reaction was skeptical: many said the post’s headline overclaims, because the author admits proprietary models still lead and has not yet shown real-world evidence that switching has minimal cost. (Skepticism that the article supports its headline, Open-weight models may be only months behind frontier proprietary models, Real-world coding quality often lags benchmark claims)
0:00 / 0:23ai Claude Code’s ‘extended thinking’ isn’t the raw reasoning log
Patrick McCanna argues that Claude Code’s ctrl+o “extended thinking” output is not the actual reasoning that drove the agent’s actions, but a summary of Claude’s thinking process. The practical warning is that local Claude Code files cannot provide a true reasoning audit trail: users can log inputs, outputs, and actions, but not the hidden reasoning tokens themselves. That matters for anyone treating agent logs as compliance evidence, debugging material, or a security record.
Discussion: Mixed — The discussion is mostly skeptical of hidden reasoning, especially for auditability, security review, and prompt debugging, but many commenters say this is industry-standard behavior rather than an Anthropic-only scandal. Several defend summaries as a trade-secret and anti-distillation measure, while others argue it makes the product less trustworthy and strengthens the case for open models. (Hidden chain-of-thought limits auditability and debugging, Security concerns around prompt injection, tool use, and unseen reasoning, Vendors hide reasoning to prevent distillation and protect competitive advantage)
Moebius is a new lightweight image inpainting model that claims 10B-level quality with only 0.22B parameters, less than 2% of FLUX.1-Fill-Dev’s size, and more than 15× faster total inference. The project uses a reworked latent-diffusion backbone with LλMI blocks plus latent-space distillation from a larger teacher model. If the claims hold up, it points toward practical local and edge-device image editing instead of relying on massive general-purpose models.
Discussion: Mixed — HN was intrigued by the promise of a small, fast, local inpainting specialist, especially after one commenter got it running in-browser via ONNX. But the mood was tempered by hands-on skepticism: several users reported failed demos, 512-by-512 limits, smoother-looking inpainted regions, weak novel-object performance, and doubts that it truly matches 10B-class models. (enthusiasm for local, task-specific AI models, browser and ONNX deployment experiments, quality skepticism versus 10B-model claims)
0:00 / 0:23ai Sakana’s Fugu pitches model orchestration as the next AI product layer
Sakana AI introduced Fugu, presented as an orchestrator that coordinates multiple strong models rather than relying on a single frontier model. The company’s examples show Fugu-Ultra outperforming unnamed frontier baselines on tasks including automated ML training-recipe search, classical Japanese reading-order recovery, Rubik’s Cube solver generation, CAD design, and blindfold chess. The pitch matters because it reframes frontier performance as a routing-and-composition problem, but the HN discussion focused heavily on whether the result is worth the cost, latency, and added abstraction.
Discussion: Mixed — HN was intrigued by the orchestration idea but skeptical of the product value, pricing, latency, and benchmark framing. Several commenters defended Sakana’s research pedigree and the strategic appeal of non-OpenAI/Anthropic alternatives, while hands-on users reported slow responses, tight limits, and quality that did not clearly beat existing frontier-model workflows. (Model orchestration may be useful when models cover each other’s blind spots, Pricing and usage limits drew heavy criticism, especially around $200/month AI tiers, Skepticism that a single Sakana API really avoids vendor dependency)
0:00 / 0:29ai Schmidhuber argues Munich laid the groundwork for today’s AI boom
Jürgen Schmidhuber’s article, introduced by David Ha, argues that several core ingredients of today’s AI systems were published in Munich in 1991, including early forms of transformer-like mechanisms, unsupervised pretraining, neural-network distillation, residual learning, and generative adversarial ideas. The piece frames TU Munich and Schmidhuber’s lab as an under-recognized origin point for techniques now used in large language models and generative AI. The HN discussion treated the history as important but contested, with many commenters separating solid claims like LSTM’s influence from shakier claims around attention and transformers.
Discussion: Mixed — HN readers broadly acknowledged Schmidhuber and the Munich/IDSIA lineage as important to deep learning, especially LSTM and early work on pretraining, distillation, and residual ideas. But many pushed back on the article’s sweeping priority claims, arguing that the modern AI boom was driven as much or more by GPUs, CUDA, AlexNet, deep-learning frameworks, and industrial-scale engineering. The discussion also turned into a broader debate over academic credit, citation norms, and whether academia or private labs deserve more recognition for modern AI. (Munich and IDSIA’s role in deep-learning history, Skepticism about transformer priority claims, Hardware and GPUs as the practical trigger for the AI boom)
0:00 / 0:42ai Tiny Qwen model gets a big boost from fine tuning
A personal-project writeup fine-tuned Qwen 3:0.6B to classify household questions into metadata categories for a RAG chatbot, so retrieval can be narrowed to areas like pool, car, HVAC, or cooking. Prompting the base 600-million-parameter model performed poorly, with about 10% accuracy on 131 tests; fine-tuning with Unsloth and QLoRA raised that to 79%, and changing labels to opaque two-character IDs reportedly improved accuracy to about 92%. The result is a useful local-LLM case study, but it also highlights how much of classification performance can hinge on label design and evaluation setup.
Discussion: Mixed — HN liked the hands-on experiment and the clear results, but many commenters argued the chosen tool was overkill or the wrong architecture for a closed-set classification task. The dominant reaction was constructive skepticism: try BERT-style encoders, embedding-plus-classifier baselines, scikit-learn, constrained decoding, or synthetic data before fine-tuning a decoder LLM. (Fine-tuning small local LLMs can work, but may not be the best fit for simple classification, Encoder models like BERT or ModernBERT are widely seen as better suited to classification, Traditional ML baselines such as n-grams, logistic regression, SVMs, and embedding classifiers were repeatedly recommended)
Alex Kritchevsky’s essay proposes treating a logarithm without a specified base as an abstract quantity, with ordinary base-specific logs emerging when that quantity is measured in units such as bits, nats, or digits. The piece extends the analogy toward vectors and coordinate systems: choosing a base is like choosing a unit or frame, while the underlying logarithmic object is treated as more fundamental. The idea matters less as a new theorem than as a conceptual reframing of why change-of-base works and how logarithms turn multiplicative structure into additive structure.
Discussion: Mixed — HN was engaged and mostly appreciative, but mathematically picky. Many commenters liked the pedagogical framing and used it to discuss torsors, units, slide rules, decibels, and physics; skeptics argued the essay needed stricter type signatures, better terminology, or an actual new result beyond notation. (Logarithms as quantities measured in units like bits, nats, and digits, Connections to torsors, affine spaces, vector bases, and Lie theory, Historical enthusiasm for log tables, slide rules, and hand computation)
0:00 / 1:13software Deno takes on Electron with desktop apps
Deno has published documentation for “deno desktop,” a canary feature in Deno 2.9.0 that turns Deno projects—from a single TypeScript file to frameworks like Next.js, Astro, Remix, Nuxt, SvelteKit, and Vite SSR—into redistributable desktop apps. The pitch is a smaller-by-default webview backend with optional bundled Chromium via CEF, in-process bindings, cross-compilation, DevTools, menus, tray support, notifications, and built-in binary-diff auto-updates. It matters because Deno is positioning itself directly in the crowded space between Electron’s consistency and weight, and Tauri’s smaller system-webview approach.
Discussion: Mixed — HN was broadly intrigued and often positive about Deno Desktop as a serious new Electron/Tauri alternative, especially because it pairs Deno’s TypeScript-first workflow and npm compatibility with desktop packaging. The skepticism centered on familiar web-desktop tradeoffs: CEF versus system webviews, binary size, permissions, native look-and-feel, and whether the “in-process” binding story is as simple as advertised. (Interest in competition with Electron and Tauri, Debate over bundled CEF versus system webviews, Questions about binary size and shared runtimes)
0:00 / 1:08software Mitchell Hashimoto Pledges Another $400,000 to Zig
Mitchell Hashimoto says his family is pledging another $400,000 to the Zig Software Foundation, split as $200,000 per year over two years, bringing their total pledged support to $700,000. He frames the donation as support for Zig’s technical progress, maintainership culture, and independence, while explicitly saying his own pro-AI usage does not prevent him from respecting Zig’s strict no-LLM contribution policy. The post matters because it is a major individual commitment to a systems programming language that is still building out its compiler, ecosystem, and governance norms.
Discussion: Positive — HN’s reaction is strongly supportive, with praise for Hashimoto’s funding, Zig’s technical direction, and the mature tone of the post around disagreement over AI policies. The thread also branches into enthusiasm and skepticism about Ghostty, plus broader reflections on open-source funding and whether small donations can still matter. (Respect for Hashimoto personally backing open source, Approval of Zig’s independence and quality-focused culture, Debate over Zig’s no-LLM contribution policy without much hostility)
0:00 / 0:23software A photo hoarder accidentally discovers years of wigglegrams
The author realized that years of near-duplicate photos in their camera roll could be turned into accidental “wigglegrams”: looping stereo-like GIFs made from slightly different viewpoints. They wrote a script using perceptual hashing to find runs of similar images, measure Hamming distance between hashes, and extract candidate sequences from iCloud or a local photo directory. The result is part photo-art experiment, part practical demonstration of how reverse-image-search-style hashing can surface unexpected structure in a personal archive.
Discussion: Positive — HN largely enjoyed the post as a playful, clever use of perceptual hashing and personal photo archives, with many commenters sharing related stereo-photo tools, cameras, and techniques. The main reservations were practical and sensory: accidental wigglegrams can be chaotic, less convincing than intentional stereo captures, and for some readers triggered motion sickness or migraine symptoms. (Perceptual hashing as a practical way to find similar photo sequences, Nostalgia and technical history around stereo cameras, lenticular prints, Nimslo/Nishika-style multi-lens cameras, and Nintendo 3DS photos, Image alignment, horizontal motion, frame ordering, and stable focal points as key to good wigglegrams)
0:00 / 0:25software Codex logging bug threatened to burn through SSD endurance
A GitHub issue reported that Codex was persistently writing large volumes of TRACE and telemetry logs into local SQLite files under ~/.codex, with one user estimating about 37 TB written over 21 days, or roughly 640 TB per year. The issue pointed to global TRACE-level logging, raw websocket/SSE payload logging, and insert-and-prune churn as likely drivers, raising concerns about SSD endurance on consumer drives. The issue has since been closed after two merged PRs that the reporter says avoided about 85% of the logs in their setup.
Discussion: Negative — HN reaction is sharply critical, using the logging bug as evidence that AI coding tools are being shipped with poor polish, high resource use, and weak operational discipline. A few commenters note Codex is more patchable than Claude Code and that a fix has landed, but the dominant mood is frustration with battery drain, disk churn, memory leaks, and vendor responsiveness. (AI coding tools seen as sloppy or rushed, Concern about local resource abuse: SSD writes, GPU, CPU, RAM, battery, Skepticism toward 'vibe coding' and claims that AI has solved software development)
0:00 / 0:24software JSON-LD for personal sites meets an AI-era backlash
The article explains how personal websites can add JSON-LD structured data using Schema.org, with examples for WebSite, WebPage, Person, ProfilePage, and SoftwareApplication nodes. The pitch is that this helps crawlers understand a site, produce richer previews, disambiguate identity, and potentially improve search presentation. On HN, readers treated it less as a pure tutorial and more as a referendum on whether helping crawlers still helps authors now that search engines increasingly answer questions directly with AI-generated summaries.
Discussion: Mixed — HN found the guide useful as a practical intro, but the discussion quickly turned skeptical about whether structured metadata still benefits site owners in an era of Google AI overviews and LLM scraping. Pragmatic commenters said JSON-LD can help with rich snippets, sitelinks, maps data, reviews, and SaaS FAQ results, while others argued much of it is either ignored, better handled by OpenGraph, or simply feeds platforms that keep users off the original site. (SEO value versus AI-generated search summaries, JSON-LD compared with OpenGraph, Microdata, RDFa, and semantic HTML, Practical advice to follow Google and Bing structured-data documentation)
0:00 / 0:37software A laid-off Blizzard engineer says the software job hunt is broken
A software engineer with about ten years of experience, including seven years at Blizzard, writes that after being laid off in June 2025, the current software job market feels worse than any they have seen. The post criticizes final-round rejections, recruiter silence, automated coding screens, AI proctors, keyword filtering, and the pressure to adopt AI coding tools. It resonated on HN because it ties individual job-search burnout to broader anxiety about layoffs, junior roles disappearing, and AI worsening already impersonal hiring systems.
Discussion: Negative — HN commenters largely agreed that the software hiring market feels unusually punishing, random, and demoralizing, especially after layoffs and amid AI-driven screening and hype. The discussion split on whether refusing AI tools is principled or self-defeating, with several commenters arguing that AI in games is especially fraught because it threatens artists and other creative workers. A recurring thread was career escape: some described leaving tech for trades, while others warned that hiring remains risky and unpredictable even for strong candidates. (software job market pessimism, AI as an accelerant for bad hiring practices, frustration with coding tests and automated screening)
0:00 / 1:19hardware Valve’s Steam Machine arrives, but the price is the story
Valve’s new Steam Machine is launching with a reservation system that randomizes order after a multi-day signup window, an approach commenters largely praised as less bot- and scalper-friendly than a first-come rush. The device is being pitched as a living-room SteamOS gaming PC rather than a locked-down console, with Valve emphasizing that users can install their own apps or even another operating system. The catch, according to the HN discussion, is price: commenters repeatedly cite a $1,049 base package without a controller and argue that puts it in an awkward spot against consoles and self-built PCs.
Discussion: Mixed — HN is broadly supportive of Valve’s open, Linux-friendly approach and likes the randomized reservation system, but the launch price dominates the thread and many commenters think it undercuts the product’s console ambitions. The mood is less backlash than disappointment: people want this to succeed, but doubt the market will accept a $1,000-plus living-room gaming PC with the listed specs. (Praise for an unlocked PC that can run other apps or operating systems, Randomized reservations seen as a fairer anti-scalping mechanism, Concern that the $1,049 base price is too high versus PS5, Xbox, and DIY PCs)
0:00 / 0:39hardware Linux tablet drivers hit a branding wall
Artist and reviewer David Revoy says Gaomon declined to collaborate on Linux FLOSS tablet driver support after deciding the relevant repositories appeared too closely tied to Wacom, the dominant competitor. Revoy argues that names like Libwacom and wacom-hid-descriptors are historical artifacts but now discourage non-Wacom vendors from sharing specs or participating. The result is that Linux support for many drawing tablets may continue to depend on manual device-by-device reverse engineering and the availability of a few specialist contributors.
Discussion: Mixed — HN largely agrees with the author that Wacom-branded Linux tablet infrastructure creates avoidable political and business friction for rival tablet makers. The discussion is sympathetic to better open-source driver support, but split on whether renaming is an obvious fix or a costly volunteer-time sink with no guarantee vendors will help afterward. (Vendor-neutral naming for shared open-source infrastructure, Tension between technical maintenance costs and project politics, Open-source hardware support depending on volunteer labor)
0:00 / 1:28startups When a Startup Job Turns Out to Be Part of the Scheme
A former GenieDB engineer revisits the startup that moved him from the UK to the US after learning that its VC backer, Stuart Frost’s Frost VP, was later sued by the SEC over allegedly excessive incubator fees. After reading arbitration and SEC materials, he concludes GenieDB may have been kept alive at least partly because it generated fees for the fund, even though the company had few customers and never gained serious traction. The piece matters because it reframes a familiar startup failure story as a personal encounter with the incentives and collateral damage of investor fraud.
Discussion: Mixed — HN was sympathetic to the author, but the broader mood was cynical and frustrated about incubators, outsourcing, government grants, and corporate budgeting structures that can reward waste or self-dealing. Many commenters responded with their own stories of jobs, contracts, or projects that seemed to exist for accounting games, budget preservation, or outright fraud rather than building a viable product. (sympathy for employees caught inside fraud-adjacent companies, cynicism about VC, incubator, and grant-funded startup ecosystems, perverse incentives in corporate budgeting and outsourcing)
0:00 / 0:38startups AI lowers the build-versus-buy bar, but not to zero
Brandur Leach argues that AI coding tools have shifted the classic buy-versus-build calculation, but not eliminated it: software still requires human iteration, validation, maintenance, and product judgment. He uses a $400-per-month Jira anecdote to argue that rebuilding common SaaS often fails the labor-cost math, while expensive, high-seat-count tools like Salesforce sit closer to a build decision. The key idea is a “zone of viability” for saleable software: novel enough to be non-trivial to clone, but priced low enough that customers won’t rationally rebuild it with LLM help.
Discussion: Mixed — HN broadly agrees with the author’s core point that AI-assisted coding changes the economics of internal tools, but does not erase build, maintenance, specification, integration, and opportunity costs. Commenters were skeptical of simplistic Jira-replacement anecdotes and some thought the Salesforce comparison was too optimistic, while others shared real examples where LLM-built niche tools did make sense. (LLMs reduce initial implementation cost but not ongoing maintenance, Buy-versus-build still depends on core business value and complexity, Specifications, product judgment, and usage feedback remain hard)
0:00 / 1:21policy Claude adds ID checks for some users
Anthropic’s support page says Claude may prompt users for identity verification for certain capabilities, platform-integrity checks, safety, compliance, or age-related enforcement. The process uses Persona Identities and may require a physical government photo ID plus a live selfie; Anthropic says the data is for verification and legal/safety obligations, not model training, and that Persona stores the ID and selfie under Anthropic’s instructions. The issue matters because access to powerful AI tools is increasingly being tied to real-world identity, raising questions about privacy, due process, international access, and whether centralized model providers are becoming gatekeepers.
Discussion: Negative — HN reaction is overwhelmingly hostile, with many commenters saying they would cancel or avoid Claude rather than upload a government ID and selfie. The dominant concerns are privacy, surveillance, third-party risk around Persona, unclear gating of features, and fears that U.S. policy pressure will fragment access to frontier models. A smaller thread pushes back that the support page and process appear to have existed for months, so the timing may be getting misread amid broader Anthropic controversy. (Privacy and surveillance concerns, Distrust of Persona as verification provider, Unclear which Claude capabilities require verification)
0:00 / 0:24policy Danish privacy activist says masked police raided his home
Danish privacy activist and former police officer Lars Andersen says armed, masked police broke down his door after he published coded references to Prime Minister Mette Frederiksen’s social security and phone numbers and tried to contact her about encryption and surveillance policy. He claims officers immediately cut power, disabled his router, and seized Google Nest cameras, preventing full video of the arrest. The story matters because it sits at the intersection of privacy activism, alleged doxxing of public officials, and the accountability of police during politically sensitive raids.
Discussion: Mixed — HN is broadly uneasy about the reported police tactics—especially cutting power and taking cameras—but far from unified in defending Lars Andersen. Many commenters say his activism has crossed ethical lines, citing alleged GPS tracking of ministers’ cars and doxxing or involving politicians’ families, while others argue his stunts expose hypocrisy in anti-encryption and surveillance politics. (Concern over police disabling recording equipment during a raid, Debate over whether confrontational activism helps or hurts privacy causes, Allegations that Andersen’s tactics include stalking, GPS tracking, or doxxing families)
IPVM reports that a Holiday Hills, Illinois police chief was arrested and charged with official misconduct after prosecutors alleged he used Prairie Grove’s Flock license-plate-reader system and the Illinois LEADS database to track six people he knew personally, including former romantic partners. The article ties that case to other alleged abuses by police chiefs and officers, arguing that Flock’s claim it tracks vehicles rather than people collapses when officers use plates to follow specific individuals. IPVM’s conclusion is that LPRs may have legitimate law-enforcement uses, but routine access should require a warrant, with exigent exceptions for imminent danger.
Discussion: Mixed — The thread leans skeptical of Flock and license-plate-reader surveillance, with many commenters arguing that warrants, subpoenas, access logs, and real penalties are necessary. But there is also a strong countercurrent noting that LPRs can help recover stolen cars or aid investigations, and some push back on the article’s framing and on claims that the Fourth Amendment clearly bars this use. (Warrant requirements versus real-time police access, Abuse by officers tracking romantic partners or rivals, Debate over whether public license-plate scans are Fourth Amendment searches)
0:00 / 0:39policy Canada pitches a hundred-billion-dollar nuclear buildout
Canada’s federal government released a nuclear strategy that aims to start construction on two large reactors by 2035, have five more planned or under development by 2040, deploy a Canadian microreactor to a remote community in the late 2030s, expand CANDU exports, and double uranium exports. Officials said the reactor buildout could cost more than $100 billion, but the strategy does not specify how it would be paid for. The plan matters because Canada says it needs to double grid capacity by 2050 while keeping power low-carbon, and because nuclear exports are being framed as an industrial and geopolitical strategy as allies move away from Russian enriched uranium.
Discussion: Mixed — HN was broadly interested and often supportive of Canada leaning into nuclear, especially given uranium reserves, CANDU expertise, and early Darlington SMR work. But the thread quickly turned into the familiar nuclear-versus-renewables fight, with skeptics arguing the plan is too slow, too expensive, and likely to drown in permitting, subsidies, and megaproject risk. (Canada has real nuclear assets: uranium, CANDU experience, and existing Ontario expertise, Optimism around Darlington’s BWRX-300 small modular reactor project, Concerns that nuclear is uneconomical compared with wind, solar, batteries, and storage)
Beyond All Reason, a free real-time strategy game inspired by Total Annihilation, drew attention for its fully simulated projectiles, explosions, terrain deformation, and large-scale battles with thousands of units. The HN thread treated it as both a serious technical and design achievement and a reminder of how beloved TA remains nearly three decades later. The main caveat from players was not the game itself, but the multiplayer culture around competitive public lobbies, especially large team games.
Discussion: Mixed — HN was impressed by Beyond All Reason’s scale, performance, and lineage from Total Annihilation, with many commenters sharing strong nostalgia for TA and related Spring-engine RTS games. But the dominant discussion turned quickly to multiplayer culture: several players praised the game while warning that 8v8 public lobbies, rigid metas, visible rankings, and team-kick dynamics can make the experience hostile for newcomers or casual players. (Total Annihilation nostalgia, technically impressive large-scale RTS simulation, free-to-play and actively developed)