← Oct 1Friday, October 2, 2026 · all daysOct 3 → · today »
1.Pi 1.0(earendil.com)
1646 points by sergiotapia 12 days ago | 575 comments | permalink
tl;dr: Earendil released Pi 1.0, a minimal, extensible agent harness/coding CLI that now adds Codemode (with MCP and non-LLM model support), virtual model extensions, deferred tool loading, Anthropic cache warming, mid-conversation system messages, and a full-screen TUI by default. Alongside it, they're shipping Pi Durable, an experimental MIT-licensed package for building long-running agentic applications beyond the terminal. Both are installable via shell script or npm.
HN Discussion:
  • ↑Enthusiastic users praise Pi's minimalism for enabling local models and flexible use cases
  • ↑Pi's hackability and vendor-agnostic design make it ideal for building custom agents and harnesses
  • ↓Skepticism about the minimalism claim given feature creep like bundled cache warming
  • •Curious users ask why they should switch from existing tools like Claude Code, Codex, or OpenCode
  • ↓Recommends alternative even more minimal tools over Pi
2.Several vulnerabilities have been discovered in the Linux kernel(lwn.net)
552 points by luispa 12 days ago | 394 comments | permalink
tl;dr: Debian Security Advisory DSA-6528-1 patches an enormous batch of Linux kernel vulnerabilities—well over a thousand CVEs—that could enable privilege escalation, denial of service, or information leaks. Fixes are available in version 6.12.111-1 for the stable distribution (trixie), and users are advised to upgrade immediately.
HN Discussion:
  • ↓CVE count is a misleading metric since kernel assigns CVEs to any bugfix
  • ↑AI is accelerating vulnerability discovery, exposing fragility of computing infrastructure
  • ~AI will introduce vulnerabilities at similar rates to humans while finding more, worsening the problem
  • ↑More vulnerabilities found is beneficial and will strengthen project quality long-term
  • ↓Advisory lacks critical details like remote vs local exploitability
3.Clef: Open-weight decision models, and new RL fine-tuning platform(blog.cloudflare.com)
620 points by jasondavies 12 days ago | 215 comments | permalink
tl;dr: Cloudflare released Clef and Clef-flash, two open-weight (Apache 2.0) decision models hosted on Workers AI that produce structured, probabilistic classifications rather than generated text. Built on Qwen with a non-autoregressive scoring head, they're Jev-API compatible, support vision and a 64k context, and claim competitive benchmark performance with lower latency. Cloudflare is also launching an RL fine-tuning service—initially hands-on via their FDE team, later self-serve—leveraging AI Gateway, Workers AI, and Containers.
HN Discussion:
  • ↑Impressed that Clef surpasses Jev using Jev's own paradigm and benchmarks in short time
  • ~Open weights without published data/training pipeline isn't truly open source
  • ↓Clef's pricing is significantly more expensive than Jev, undermining its value proposition
  • ↓Real-world testing shows Clef is slower and less accurate than Jev, contradicting claims
  • ~Smaller decider-4B variant performs surprisingly well and deserved benchmark inclusion
4.DeepSeek Harness Desktop for macOS and Windows(deepseek.com)
385 points by Kuyawa 12 days ago | 207 comments | permalink
tl;dr: DeepSeek has released Harness, an open-source agent framework available as a desktop app for macOS/Windows or web UI, built on Cordis's plugin-based architecture. It handles tasks like file organization, coding, research, and scheduled background jobs, with a "Creator mode" for building new plugins via chat. Developers can launch it instantly with `npx @deepseek-ai/dsh web` or clone the repo from GitHub.
HN Discussion:
  • ↑Cordis plugin architecture is the truly exciting innovation, potentially the 'emacs of harnesses'
  • ↑Enthusiastic user praise for the desktop app's performance and plugin creation capabilities
  • ↑Official packaging is needed to combat third-party impostor sites with potential security risks
  • •Benchmarks and comparisons with competing harnesses (Juggler, OpenCode) are needed for context
  • ↓Criticism of poor webpage performance and accessibility issues undermining the launch
5.SvelteKit 3(svelte.dev)
396 points by sampsn 12 days ago | 171 comments | permalink
tl;dr: SvelteKit 3.0 ships as an incremental release focused on polish and type safety, with an `npx sv migrate sveltekit-3` command to automate upgrades. Breaking changes include moving config to `vite.config.ts`, replacing the `$lib` alias with the standard `#lib` subpath import, improved environment variables, less boilerplate for service workers, and better error handling. Remote functions for type-safe client-server communication are the team's next priority but aren't ready yet, as they depend on the still-experimental Async Svelte.
HN Discussion:
  • ↑Enthusiastic praise for SvelteKit as a production-ready framework with excellent DX
  • ↑Svelte preferred over React/Next.js for personal and professional projects
  • ↓Criticism that Svelte constantly reinvents itself with breaking changes
  • •Questions about LLM/vibe-coding compatibility with Svelte compared to React
  • ↓Dislike of SvelteKit's routing conventions and file naming
6.Pi Durable(earendil.com)
490 points by paulsmith 12 days ago | 68 comments | permalink
tl;dr: Earendil released Pi 1.0 alongside Pi Durable, an experimental TypeScript framework for building long-running, crash-resilient agentic applications that can run anywhere with a JavaScript runtime. It provides durable storage, checkpointed tasks, pluggable extensions/tools, concurrent conversations with forking, multiplayer client support, and automatic compaction—letting agents survive process death and resume from checkpoints. Pi Durable complements rather than replaces the Pi coding agent, and ships with examples including a small coding agent and a TUI-based vacation planner.
HN Discussion:
  • ↑Enthusiasm for durable agent harnesses as an important innovation space
  • ~Questions design choices like lack of branching conversation trees
  • ↓Criticism that sandboxing and security policy aren't treated as first-class concerns
  • ~Skepticism about whether the added complexity is justified over simpler setups
  • ~Desire for broader durable state beyond JSON documents, e.g. outbox pattern for external stores
7.Ask HN: Who is hiring? (October 2026)
260 points by whoishiring 12 days ago | 269 comments | permalink
tl;dr: The monthly Hacker News "Who is hiring?" thread for October 2026, where companies post job openings directly (no recruiters or job boards). Posts should specify location, indicate REMOTE or ONSITE status, and describe the company if not well-known. Several third-party search tools are linked for filtering listings, along with the companion "Who wants to be hired?" thread.
HN Discussion:
  • ↑Companies posting job listings following the thread's format conventions
  • •Request to include 4DWW tag for Four Day Work Week visibility
  • •Remote-first positions dominate the hiring landscape across Europe and North America
  • •AI and ML roles are heavily represented in current hiring needs
8.Git 3.0's upcoming SHA-256 default will be a costly mistake(blog.gitbutler.com)
549 points by chmaynard 12 days ago | 521 comments | permalink
tl;dr: Git 3.0 will default new repositories to SHA-256 instead of SHA-1, forcing an ecosystem-wide migration that breaks interoperability between repos, submodules, tooling, libraries, and existing URLs/signatures — with no clean path for mixing formats. The author argues this is solving a largely theoretical problem, since real-world attacks exploit social engineering and compromised maintainers, not hash collisions. A better approach would be adding a secondary independent content hash (like SHA-256 or BLAKE3) into signed object headers, preserving SHA-1 for addressing while providing stronger verification where it matters.
HN Discussion:
  • ↓The article contains factual errors and misrepresents the real security risks of SHA-1 collision attacks
  • ↓Git's docs and actual implementation already address many of the author's concerns, including bijective mapping and submodule compatibility
  • ↓The article's screenshots and framing of the forge/UI problem are misleading or strawmen
  • ↑The SHA-256 transition is well-argued against and resembles IPv6's problematic rollout; crypto agility via signatures is preferable
  • ~The transition is driven by compliance/politics rather than real security needs, or Git should make the modes more interoperable
9.StreetComplete on iOS is now in public beta(github.com)
619 points by Snowly 12 days ago | 170 comments | permalink
tl;dr: StreetComplete, the OpenStreetMap editing app, is being ported to iOS using Kotlin Multiplatform and Compose Multiplatform, allowing most of the existing Kotlin codebase to be reused rather than rewritten (as would be needed with Flutter). The migration—estimated at roughly one man-year of work—is about 50% complete after full-time work in early 2024, and the maintainer is seeking contributors and sponsors to finish the remaining work tracked on a public project board.
HN Discussion:
  • ↑Gratitude for government and NGO funding that made the iOS port possible
  • ↑Enthusiasm for the project and excitement about iOS availability enabling broader participation
  • ↑Personal testimonials about how StreetComplete lowered the barrier to OSM contribution
  • •Criticism of the OSM community's pedantic edit-reverting behavior discouraging new contributors
  • ~Comparison with alternatives like EveryDoor and noting bugs/missing features in current app
10.RIP, vector database(turbopuffer.com)
375 points by razin 12 days ago | 106 comments | permalink
tl;dr: Turbopuffer is rearchitecting its storage in v3 to stop using the ANN vector index as the primary key, which has caused storage amplification (duplicating docs for multi-vector cases), write amplification (rebalancing cascades), and suboptimal block sizes for non-vector queries. The new design treats ANN as just another secondary index, enabling faster text, regex, vector, and SQL-style queries. V3 currently passes 100% of CI, with performance tuning and public benchmarks coming before production rollout.
HN Discussion:
  • ↑Draws historical parallel to Postgres/MySQL indexing tradeoffs, validating the design shift
  • ~Skeptical of vector DB hype cycle, preferring to wait for mature SQL databases to absorb features
  • ↑Confirms the approach by sharing similar successful experiences building on SQLite or alternative engines
  • ↑Other vendors (LanceDB, TopK) already treat ANN as secondary index, validating turbopuffer's direction
  • •Wants concrete evidence like p99 benchmarks or questions dashboard progress before accepting claims
11.Various Projects Find Hidden SDR Capabilities in ESP32 Microcontrollers(rtl-sdr.com)
287 points by nkw 12 days ago | 66 comments | permalink
tl;dr: Researchers have discovered an undocumented feature in several ESP32 chips that allows firmware to bypass the fixed WiFi/Bluetooth radio and capture raw IQ baseband samples, effectively turning them into SDRs covering 2.2–2.7 GHz (and 4.8–6.0 GHz on the ESP32-C5) at up to 80 MS/s. Bandwidth limits mean most chips can only export snapshots for spectrum analysis, though the ESP32-S31 can stream continuously over Gigabit Ethernet. Independent projects have leveraged the finding for phase-coherent direction finding (ESPARGOS), a general-purpose SDR with an FPGA frontend, and a 5.8 GHz FPV video receiver (C5VRX).
HN Discussion:
  • ↑Excitement about potential applications like V2X data collection and ham radio use cases
  • ~Concern that Espressif may patch out this undocumented feature due to regulatory pressure
  • ↑Technical speculation on extracting IQ data via new interfaces like Gigabit Ethernet or PSRAM
  • •Questions about signal quality metrics like phase noise and DC offset not covered in article
  • •Curiosity about extending capabilities such as sampling harmonics or building LORA-like systems
12.Cloudflare K2: serverless event streams(blog.cloudflare.com)
282 points by elffjs 12 days ago | 111 comments | permalink
tl;dr: Cloudflare launched K2 in public beta, a serverless event streaming service built as a durable log on top of R2 object storage, offering fan-out consumption, long-term retention, and scaling independent of compute. Unlike Kafka, K2 offloads replication/consensus to R2, trading ~1s p99 produce latency for simpler, cheaper operation at edge scale. Pricing is expected at $0.04/GB produced and consumed, with a roadmap including multi-GB/s throughput, key-based ordering, and Kafka client compatibility.
HN Discussion:
  • ↑Object storage as a foundational substrate is an exciting trend for building simpler systems
  • ↓Consumer-side pricing is too expensive, especially for fan-out use cases
  • ↓Cloudflare's rapid product release pace raises concerns about security and quality
  • ~Questioning how K2 meaningfully differs from existing S3-based solutions like WarpStream and AutoMQ
  • ↑Making individual streams cheap and simple addresses real Kafka complexity pain points
13.How to speed up the Rust compiler in September 2026(nnethercote.github.io)
272 points by trickypr 12 days ago | 163 comments | permalink
tl;dr: Over the past two months, the Rust compiler saw a mean 4.57% wall-time reduction across benchmarks, with 555 of 629 measurements improving. Major contributors include an LLVM 23 upgrade, PGO for Clippy (up to 18% faster), rustdoc optimizations, dataflow analysis changes yielding ~30% speedups on cranelift-codegen, and ongoing work to mitigate regressions from the newly-enabled Polonius borrow checker and next-gen trait solver on Nightly. The author is also starting a new role at Hexcat focused on compiler performance.
HN Discussion:
  • ↑Sharing additional personal optimization work or technical contributions complementing the article's progress
  • ↑Appreciation for the measurable gains and corporate funding that enabled them
  • ↓Rust compile times remain too slow for modern AI-agent workflows, pushing users to Go
  • •Curious questions about why the compiler is slow in the first place
  • ↑Pointing to future/related speedup efforts like parallel frontend and TPDE
14.Gemini 4 Argon(blog.google)
1684 points by bradleyg223 13 days ago | 1173 comments | permalink
tl;dr: Google announced Gemini 4 Argon, a frontier model with a 1M output token limit, priced at $2/M input and $10/M output tokens, initially rolling out to trusted cyber defenders before wider release. It claims state-of-the-art results on benchmarks like DeepSWE v1.1 (77.9%), LVBench (91.7%), and CWE-bench v1 (68%), and is already being used internally at Google for tasks like C++ to Rust migrations and data center memory optimization. Google is also releasing it without cyber guardrails to trusted defenders and emphasizing new safeguards against misuse, prompt injection, and misalignment.
HN Discussion:
  • ↑Personal anecdote showcasing impressive Gemini capabilities reinforces the model's power claims
  • •The leapfrogging competition disproves winner-takes-all AI theories, suggesting a distributed market
  • ↓Frustration that the model isn't actually available to paying subscribers
  • ↑The C++ to Rust migration at scale is the truly significant announcement here
  • ~Google should open-source Gemini weights to undermine competitors
15.OpenDLSS: A Vulkan Reimplementation of Nvidia's DLSS 5 Neural Rendering Network(github.com)
267 points by sagacity 14 days ago | 122 comments | permalink
tl;dr: OpenDLSS is a Vulkan reimplementation of NVIDIA's DLSS 5 Neural Rendering network (a 71-block Swin/ViT model running FP8 on tensor cores), bit-exact against the original down to all 75 block boundaries. It includes a WebGPU port that achieves the same exact outputs in-browser without tensor cores or FP8, demonstrating that the determinism lives in the spec rather than the hardware. Users must supply their own weights; the repo ships no NVIDIA code, weights, or instructions for obtaining them, and benchmarks on an RTX 4070 SUPER show 7.8 ms at 1080p.
HN Discussion:
  • ↑Impressed by the technical achievement of bit-exact reproduction across hardware
  • ↓Questions the practical utility given weights aren't provided and may be game-specific
  • ~Concerned about performance cost at 1080p being too high for real-world use
  • •Curious about technical details like inputs (z-buffer) and training data
  • ↓Skeptical due to LLM-generated README suggesting lack of author effort
16.Book of Shapes – Collection of minimal, generative and customizable SVG-patterns(bookofshapes.com)
279 points by eustoria 14 days ago | 25 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↑Simple appreciation for the beauty and quality of the collection
  • ↑Shares related personal projects inspired by similar generative design work
  • ↓Questions originality, suggesting content may be derived from another source
  • ↑Finds practical use cases such as enhancing presentations or design work
  • •Questions the target audience and practical applications of the tool
17.Micron CEO Says Memory Supply Will Be Much Tighter in 2027 and 2028 Than in 2026(techpowerup.com)
384 points by speckx 12 days ago | 437 comments | permalink
tl;dr: Micron CEO Sanjay Mehrotra says memory and storage supply will be significantly tighter in fiscal 2027 and 2028 than in 2026, with no clear line of sight to supply-demand balance even accounting for planned DRAM cleanroom expansions. Over 75% of Micron's 2027 output is already committed to customers, and its new Idaho ID2 fab won't produce wafers until late 2028. Last quarter, DRAM prices rose in the high teens percent and NAND jumped roughly 30%.
HN Discussion:
  • ↓High prices will invite new competition that undermines Micron's long-term market control
  • ↓CEO's claims are self-serving marketing to drive demand and should be distrusted
  • ↓Micron is engaging in deliberate market manipulation by abandoning consumer DRAM
  • ~Hope Chinese competitors like CXMT will enter and relieve the painful consumer pricing
  • ↑Tight supply will force welcome memory optimization and a return to leaner software
18.Cops Can Bypass iPhone's Automatic Reboot to Get into Locked Phones(404media.co)
282 points by speckx 12 days ago | 220 comments | permalink
tl;dr: Magnet Forensics, maker of the GrayKey phone-cracking tool used by law enforcement, has developed new products ("GrayKey Preserve" and an "Evidence Preservation Mode") that bypass Apple's inactivity reboot feature. That iOS feature, introduced in late 2024, automatically reboots iPhones after 72 hours of being locked, reverting them to a state that's harder to crack. The workaround helps police preserve seized devices in a more accessible state while waiting for warrants or working through backlogs.
HN Discussion:
  • ↑Police may be searching phones before obtaining warrants, raising legal concerns
  • •Technical explanation of how the bypass likely works by extracting AFU keybags
  • ↑Users should add extra encryption layers to protect sensitive data on phones
  • ↓Article's framing is biased by implying the reboot feature only hinders police rather than all attackers
  • •Apple's implementation of the reboot timer is flawed and gives false sense of security
19.Returning from vacation? The government can search your phone without a warrant(arstechnica.com)
403 points by rbanffy 12 days ago | 369 comments | permalink
tl;dr: Under the long-standing "border exception" to the Fourth Amendment, CBP can search and copy data from travelers' phones without a warrant at ports of entry. A Florida immigrant rights organizer is suing after his phone was seized returning from vacation, alleging CBP targeted him for his political activism. CBP says such searches are rare, affecting only 0.0013% of the 419 million travelers processed in FY2025.
HN Discussion:
  • ↑Border agents wield excessive power globally, often exploited by law enforcement for easier data collection
  • ~The core problem is lack of transparency and accountability, not just the warrant exception itself
  • ↑The Fourth Amendment text is clear and this practice plainly violates it
  • ↑Individuals should take technical measures (GrapheneOS, data sovereignty, wiping, dual profiles) to protect themselves
  • ↑CBP's cited statistics are misleading as search rates have actually increased over time
20.LinkedIn Larpmaxxing(hereticpleb.vercel.app)
317 points by BurnerBurner 14 days ago | 249 comments | permalink
tl;dr: The author rants about LinkedIn's flood of nearly identical "YOLO computer vision" projects (hand tracking, pothole detection, etc.) that are performative, trivial, and never actually deployed anywhere useful. To prove the point, they built their own YOLO-based detector for these slop posts in about 90 minutes using Roboflow and Ultralytics, concluding that LinkedIn rewards looking impressive over actual skill, since nobody dares criticize for fear of hurting their job prospects.
HN Discussion:
  • ↑LinkedIn's technical hypocrisy: obfuscates against scrapers while praising AI everywhere
  • •LinkedIn poisons Google search results with login-walled pages
  • ↓Larpmaxxing actually works when strategically targeted at specific decision-makers
  • ↓If you hate LinkedIn just leave; it still has genuine networking value
  • ↑LinkedIn is full of performative self-promotion slop, confirming the article's thesis