tl;dr: Qwen has released Qwen3.8-27B, a 27B-parameter dense vision-language model with a hybrid Gated DeltaNet/Gated Attention architecture, native 262K context (extensible to 1M via YaRN), and FP8 quantization. It features toggleable thinking mode with tunable reasoning_effort, preserved thinking across turns, and native image/video understanding. Benchmarks claim it outperforms prior Qwen models and competes with Opus 4.6 Max on coding (SWE-bench Pro, Terminal Bench) and agentic tasks, while trailing on some reasoning benchmarks like HLE.

tl;dr: AI-powered vulnerability discovery tools are rapidly closing off the remotely exploitable bugs that law enforcement and intelligence agencies have relied on for the past decade to bypass encryption, since the FBI's "Going Dark" era effectively ended when hacking tools like GrayKey and Pegasus filled the gap. As this offensive capability dries up over the next few years, agencies will likely renew pressure on vendors to build intentional backdoors into their systems. The perverse result: the US may weaken its own software just as it was finally becoming genuinely secure, exposing infrastructure to foreign adversaries.

tl;dr: Google has released HEIR, an open-source compiler that converts pre-trained AI models to run inference on homomorphically encrypted data, letting servers process ciphertexts without seeing the underlying inputs. The project has attracted partnerships with hardware accelerator startups (Belfort, Niobium, Cornami, Optalysys) and several universities, and Google is demonstrating it with four applications: recommendation models, credit card fraud detection, network intrusion detection, and hotword detection. Benchmarks are currently single-threaded CPU, with accelerator latency numbers promised soon.

tl;dr: RustDesk has released a preview build enabling true unattended remote access on Wayland, including multi-monitor support and connections from the login screen after reboot. The preview is currently limited to x86_64 Debian/Ubuntu systems, with Fedora and Arch support planned once stable. This puts RustDesk ahead of competitors like AnyDesk (which still requires Xorg) and TeamViewer (which considers Wayland support experimental).

tl;dr: Firefox has confirmed continued support for uBlock Origin, making it the last major browser to do so after Microsoft Edge announced it will drop Manifest V2 extensions, following Chrome's lead. Since most browsers (Edge, Opera, Brave, Vivaldi) are Chromium-based, and non-Chromium alternatives Safari and DuckDuckGo don't support uBlock Origin either, Firefox is now the only option for users who want the full extension rather than the limited uBlock Origin Lite.

tl;dr: Summary not available

tl;dr: Claude Code sessions bill on input, output, and cached tokens, with output ~5x input and cache reads at 0.1x input — but the entire conversation gets re-sent every turn, so context bloat compounds fast. Switching `/model`, `/effort`, or running `/compact` mid-session invalidates the cache prefix and forces expensive re-prefill, while `/rewind` preserves it. To keep costs down: `/clear` between tasks, `@-mention` files instead of letting Claude search, quiet noisy commands, trim CLAUDE.md, and offload high-output grunt work to subagents.

tl;dr: Mixedbread has released Toast 1, a specialized search subagent designed to offload retrieval work from frontier models like GPT-5.6 and Claude Opus 5, claiming comparable search quality at up to 10× lower cost and 12× faster speeds. On benchmarks like OfficeQA Pro V2 and Harvey's Law Firm Knowledge, pairing Toast 1 with a frontier model improved accuracy while cutting token usage by 3.5×. It works with any retrieval backend, is priced at $0.30/M input and $0.72/M output tokens, and is accessible via Mixedbread's API.

tl;dr: Summary not available

tl;dr: Opus 5 feels worse to work with than predecessors like Opus 4.7, 4.8, and Fable because it makes bold assumptions, reinterprets plans, and skips clarifying questions—requiring more babysitting despite scoring higher on benchmarks. The author speculates this is a side effect of optimizing for benchmarks and self-improving AI, which reward confident guesses on self-contained tasks and penalize models that pause to ask for clarification. That trade-off hurts real-world coding work, where ambiguity is unavoidable and users would rather the agent ask than guess.

tl;dr: A blogger describes seven books kept within arm's reach for inspiration: Roget's Thesaurus (4th ed., defended as a conceptual hierarchy rather than a synonym list), Sir Thomas Browne's prose, Boccaccio's Decameron, van Heijenoort's From Frege to Gödel logic anthology, Comenius's illustrated 1658 children's book Orbis Pictus, the NIV Bible, and a book of the Belles Heures of the Duc de Berry. Each entry includes personal anecdotes, digressions into favorite passages, and links to past blog posts on the same subjects.

tl;dr: A satirical page mocking the state of modern websites, particularly the ubiquitous COVID-19 notices and cookie consent banners that appeared on every site in 2020. It jabs at EU/CA privacy regulations for creating mandatory, inconsistent popups that browsers can't universally dismiss, while sarcastically blaming users for voting the laws into existence.

tl;dr: Google released Gemini 3.7 Flash just three weeks after 3.6 Flash, claiming significant gains in coding (FrontierCode 1.1: 43.6% vs 34.4%), web development (WebDev Arena Elo 1588 vs 1538), and document reasoning, along with better instruction-following and tool use. It's priced at an introductory $0.75/1M input and $3.75/1M output tokens—half the cost of 3.6 Flash—and is now powering Gemini Spark for Pro/Ultra subscribers and available via the Gemini API, AI Studio, and Antigravity.

tl;dr: Cerebras and OpenAI have launched "Ultrafast Mode" for GPT-5.6 Sol, delivering up to 750 output tokens/second—reportedly 11x faster than Fable 5 and completing Humanity's Last Exam in 11 hours versus 78 for Claude Fable 5. The speedup is enabled by Cerebras' Wafer-Scale Engine, which packs 44GB of SRAM per chip to keep model weights on-chip and eliminate the memory-bandwidth bottleneck that slows GPU inference. It's currently available as a limited preview to select OpenAI API customers.

tl;dr: Instead of constraining an LLM's output to a large fixed taxonomy via structured outputs (expensive, hits context limits), have a cheap/dumb LLM hallucinate plausible classifications for the input query. Then embed both the hallucinated labels and your real taxonomy, and use nearest-neighbor lookup to map the fake output to the closest legitimate category. This avoids shipping the full schema on every call and scales better.

tl;dr: DeepSeek has released a developer preview of DeepSeek Harness, a plugin-based framework for building AI agents where capabilities (models, tools, sandboxes, storage, UI, etc.) are modular plugins managed by a Cordis kernel. Developers can swap or extend any component via configuration without modifying source code, and can try it via `npx @deepseek-ai/dsh web` or clone the repo from GitHub.

tl;dr: Australia's home battery subsidy program, launched in July 2025, has driven the installation of over 500,000 batteries—giving the country more home batteries than the US despite having 1/12th the population. By discharging stored solar power during evening peak demand, the batteries have reduced reliance on additional power plants, contributing to a 47% drop in wholesale electricity prices over the past year, according to Energy Minister Chris Bowen.

tl;dr: Netlify partnered with OpenRouter to expand model choice in its Agent Runners (coding agents) and AI Gateway, adding open models like Kimi K3, GLM 5.2, and DeepSeek V4 alongside Claude, GPT, and Gemini. In a test generating a simple coffee shop site, credit costs varied wildly—from 2.4 credits (DeepSeek V4 Flash) to over 1,000 (Claude Opus)—with Opus producing the most polished designs but often overspending, while mid-tier models like GPT 5.6 Terra offered strong value. Follow-up posts will evaluate models on more complex app scenarios involving databases, auth, and AI integration.

tl;dr: Summary not available

tl;dr: Researcher Christopher Domas demonstrates that flipping bits in AMD Family 16h's DRAM controller (MCT/DCT) address-translation registers rewires physical-to-DRAM mappings on the fly, creating aliases that bypass every memory-protection mechanism sitting above the memory controller. Since the scrambling is a GF(2) linear map, an SMT solver like z3 can reconstruct the transform from a few sentinel writes, yielding read/write access to normally fenced regions: PSP/fTPM memory, SMRAM, the C6 idle-state stash, and even live CPU microcode. The technique likely generalizes to Intel, ARM, and RISC-V memory controllers, threatening TEEs like SEV, SGX, TDX, and TrustZone.