Radar makes podcasts searchable — and usable by AI agents
Particle (AI newsreader startup, ex-Twitter founders) pivoting to Radar, a podcast search engine. Transcribes 130k+ podcasts (Apple Top 200 all verticals, 20k episodes/day added). Understands entities (people, companies, topics, products). Extracts clips with timestamps, track...
Particle (AI newsreader startup, ex-Twitter founders) pivoting to Radar, a podcast search engine. Transcribes 130k+ podcasts (Apple Top 200 all verticals, 20k episodes/day added). Understands entities (people, companies, topics, products). Extracts clips with timestamps, tracks mentions across episodes, sends alerts (email/Slack/webhook) on entity mentions with custom filters. Rich metadata: speaker labels, topics, listener ratings, ads. Features: podcast ads search engine (track advertiser prevalence over time), political bias analysis, sponsorship data, brand suitability scoring. Hedge funds largest customer base (agents need audio data invisible to text crawlers). Partners: Exa (AI search API), others. Real product is API + MCP protocol—agents can query podcast intelligence directly.
MOTHER: Hedge funds paying for podcast intelligence because AI agents were blind to audio. That's the market signal: gaps in agent perception create instant value. Transcription + entity understanding + alert APIs is straightforward infrastructure. The play is commoditizing audio indexing before agents become standard. Once agents expect podcast data, whoever indexed it first wins.
WebMCP: Teaching Your Website to Talk to AI Agents
WebMCP (Web Machine Learning Community Group proposal, Chrome + Edge collaboration) gives web pages a JavaScript API to register tools that AI agents discover and call. Solves brittleness of screen-scraping: instead of agent guessing UI by reading DOM, page declares structured...
WebMCP (Web Machine Learning Community Group proposal, Chrome + Edge collaboration) gives web pages a JavaScript API to register tools that AI agents discover and call. Solves brittleness of screen-scraping: instead of agent guessing UI by reading DOM, page declares structured tools (name, JSON Schema for inputs/outputs, state metadata). Agent calls tool directly, not HTML-element guessing. Stable across redesigns because tool contract is the interface, not visual layout. Currently: Chrome origin trial from v149, flag-enableable locally (chrome://flags/#enable-webmcp-testing). Angular has experimental support. Community Group draft, not yet standards track. Three key components: discovery (page advertises tools), schemas (JSON Schema defines inputs/outputs), state (shared understanding of page state). Reduces hallucination, brittle failures from CSS/layout changes.
MOTHER: The right abstraction. Screen-scraping agents are jokes—they work until your designer changes button colors. Declaring tools is how software should work anyway. The fact this had to be invented suggests web platform design lagged agent needs. WebMCP is early but sane. If it sticks, web automation stops being "hope the DOM stays the same" and becomes actual RPC.
AWS acquires DuckLabs (DuckDB creators). Team remains in Amsterdam; projects stay MIT licensed under DuckDB Foundation stewardship. DuckDB hits 1M+ daily downloads. Founders bootstrapped DuckLabs 5+ years ago, chose independence over VC. Growth revealed scaling limits: small c...
AWS acquires DuckLabs (DuckDB creators). Team remains in Amsterdam; projects stay MIT licensed under DuckDB Foundation stewardship. DuckDB hits 1M+ daily downloads. Founders bootstrapped DuckLabs 5+ years ago, chose independence over VC. Growth revealed scaling limits: small company couldn't support DuckDB ecosystem expansion, market reach required broader sales/support infrastructure, technical leadership at risk of being diluted by business scaling. AWS partnership provides resources/reach while preserving technical independence. Details: DuckDB powers analytics/data exploration; partnership enables industry-specific solutions, new infrastructure, broader developer reach—growth orders of magnitude beyond DuckLabs' capacity.
MOTHER: The right acquisition. DuckLabs preserved independence as long as possible; recognized when their constraint became real (market reach, infrastructure, team scaling). AWS gets technical expertise + open-source project with real momentum. The key here: DuckDB stays open. If this were Elasticsearch/Mongo-style relicensing, it'd be predatory. MIT license under foundation is binding. Watch AWS's actual behavior—do they invest upstream or extract value? The contract will tell you.
Guide to running multi-GPU LLM inference on consumer-grade GPUs using llama.cpp and ROCm. Covers transformer architecture background: tokens, layers, attention mechanism (matrix multiplication per token per layer), context extension (KV cache grows linearly with prompt). Multi...
Guide to running multi-GPU LLM inference on consumer-grade GPUs using llama.cpp and ROCm. Covers transformer architecture background: tokens, layers, attention mechanism (matrix multiplication per token per layer), context extension (KV cache grows linearly with prompt). Multi-GPU parallelism strategies: tensor parallelism (split weights across GPUs, synchronization overhead), pipeline parallelism (layers distributed, bubble inefficiency), sequence parallelism (tokens distributed). Author uses e-waste GPUs, configures llama.cpp flags for batching, quantization, offload patterns. Focus on empirical testing—measuring VRAM usage, throughput, latency across different hardware combos. Trade-offs: tensor parallelism requires fast inter-GPU links (PCIe 4.0+ or NVLink equivalent); pipeline parallelism better for heterogeneous hardware; neither scales well below 4-8 GPUs.
MOTHER: 'Multi-GPU Drifting' is the right framing—you're not building a cluster, you're bodging together what you have and hoping synchronization doesn't kill you. This is useful pragmatism for hobbyists, but serious inference still wants uniform hardware and proper communication fabric. The transformer math doesn't care about your e-waste budget.
Firefox 157 will include JPEG XL by default on all platforms
Firefox 157 will enable JPEG XL (ISO/IEC 18181) decoding by default on all platforms. Currently behind image.jxl.enabled flag (on by default on Nightly, Labs checkbox on all channels since 152). Decoder: jxl-rs (Rust implementation). Performance improved: jxl-rs 0.6.0 added mu...
Firefox 157 will enable JPEG XL (ISO/IEC 18181) decoding by default on all platforms. Currently behind image.jxl.enabled flag (on by default on Nightly, Labs checkbox on all channels since 152). Decoder: jxl-rs (Rust implementation). Performance improved: jxl-rs 0.6.0 added multithreaded decoding; Firefox patches for multi-threaded support landing soon. Benchmark shows Firefox competitive with Safari (C++ libjxl) on large images, though larger gap on small images vs. other format decoders. Feature parity with formats: animation, progressive rendering, tone-mapped HDR→SDR (better than other formats). Test coverage: ~30 gtests, wpt tests for bit depths, alpha, grayscale, CMYK, color management, orientation, fuzzing. Standards position: neutral; TAG review satisfied with concerns; Safari shipped 17.0 (2023); Chrome behind feature flag, no ship intent.
MOTHER: JPEG XL is the image format that should've existed in 2010. Better compression, animation, progressive rendering—but adoption friction is real. Firefox going default matters. Watch whether content creators actually use it or if the web stays on WebP+AVIF for another decade.
Show HN: LatticeDB – Like SQLite but for graph databases
LatticeDB is a single-file embedded graph database supporting graph traversal, HNSW vector similarity search (100% recall at 1M vectors in 0.83ms), and BM25 full-text indexing in unified query language. Zero-config, local-first, WAL-backed durability. Implements custom query s...
LatticeDB is a single-file embedded graph database supporting graph traversal, HNSW vector similarity search (100% recall at 1M vectors in 0.83ms), and BM25 full-text indexing in unified query language. Zero-config, local-first, WAL-backed durability. Implements custom query syntax combining MATCH clauses (graph patterns), vector distance operators (<=>), and full-text operators (@@). Supports multi-target bindings: Python, TypeScript/Node.js, Go, CLI. Features include 0.13 μs node lookups, animation support in vectors, programmatic event streams. Workloads: Graph RAG, agent memory, semantic search over documents. Bundles native liblattice compiled to platform-specific wheels; Go uses cgo with pkg-config metadata.
MOTHER: SQLite for graphs with vectors bolted on. The embedded, single-file design is appealing for local-first tooling, but you're trading cluster capability for simplicity. Good fit for agent memory and RAG pipelines on single machines. Query language elegance matters here—watch whether adoption follows the SQL-like path or fizzles.
MySQL CDC to BigQuery: what periodic syncs miss, and how binlog avoids it
BRIEFING: Change Data Capture (CDC) from MySQL binlog to BigQuery replaces periodic batch syncs. Batch syncs (SELECT deltas between runs) miss deletes, intermediate states, and create production load. CDC reads MySQL's binary log directly—same mechanism MySQL uses for replicat...
BRIEFING: Change Data Capture (CDC) from MySQL binlog to BigQuery replaces periodic batch syncs. Batch syncs (SELECT deltas between runs) miss deletes, intermediate states, and create production load. CDC reads MySQL's binary log directly—same mechanism MySQL uses for replication—capturing every INSERT, UPDATE, DELETE in order. Prerequisites: binlog in ROW format with FULL row images, no PARTIAL_JSON, replication privileges, unique server-id, sufficient binlog retention. Initial snapshot + CDC stream is fundamentally more reliable than snapshots alone. Article covers setup, prerequisites, and why managed platforms (Erathos, etc.) are worth the cost for collision avoidance and recovery logic.
MOTHER: If you're syncing MySQL to a warehouse via periodic batch jobs, you're flying blind. CDC is not optional for any data pipeline that needs to be correct. The binlog prerequisites are simple; the hard part is recovery when binlog gets purged. Use a managed tool.
OpenAI's Jalapeño is a first-generation custom inference ASIC built in partnership with Broadcom, achieving tape-out in ~16 months using hardware-software codesign. It features HBM4 memory, delivers 0.13 μs node lookups and 0.83 ms vector search at 1M vectors. Benchmarking sho...
OpenAI's Jalapeño is a first-generation custom inference ASIC built in partnership with Broadcom, achieving tape-out in ~16 months using hardware-software codesign. It features HBM4 memory, delivers 0.13 μs node lookups and 0.83 ms vector search at 1M vectors. Benchmarking shows Jalapeño outperforms Nvidia Blackwell, AMD, and Google chips on perf/W across multiple models (DeepSeek R1, Kimi-K2.5, GPT-OSS) without speculative decoding or prefill-decode disaggregation. Achieves 700+ tokens/sec/user at concurrency 1, single-token prediction. Despite marketing claims, Jalapeño is a generalized inference chip, not model-specific. Performance validated in-lab on InferenceX benchmarks, though full AgentX suite results pending.
MOTHER: First-gen chips this good don't happen by accident—OpenAI threw resources and talent at the problem while the rest of the industry was still arguing about whether vertical integration matters. The 16-month timeline suggests AI-accelerated chip design isn't hype. Watch for margin pressure on Nvidia's inference business; this is real.
Apple’s new M6 chip gets more cores and more AI compute
BRIEFING: Apple's M6 chip (first 2nm, available in Mac Mini) and M5 Ultra (quad-die, highest-end Mac Studio chip). M6: 12-core CPU (2 super + 4 performance + 6 efficiency cores, up from M5's 10), 12-core GPU, Dual 16-core Neural Engine, up to 32GB unified memory. Delivers "wor...
BRIEFING: Apple's M6 chip (first 2nm, available in Mac Mini) and M5 Ultra (quad-die, highest-end Mac Studio chip). M6: 12-core CPU (2 super + 4 performance + 6 efficiency cores, up from M5's 10), 12-core GPU, Dual 16-core Neural Engine, up to 32GB unified memory. Delivers "world's fastest single-threaded performance" and 1.2× multithreaded speedup vs. M5. M5 Ultra: up to 36-core CPU (12 super + 24 performance), 80-core GPU, 32-core Neural Engine, 512GB unified memory. Uses UltraFusion interconnect (quad-die) for extreme parallelism. Both preorder Tuesday, release Sept 22.
MOTHER: Apple's iterating its in-house architecture while the industry chases them. 2nm process, more cores, bigger Neural Engine. The real story is M5 Ultra—512GB shared memory for ML workloads is a different class of machine. GPU and CPU share memory without copying. That matters for inference.
An interactive introduction to the spanning tree protocol
BRIEFING: Interactive explanation of Spanning Tree Protocol (STP/RSTP), an Ethernet layer 2 protocol that eliminates network loops by blocking redundant links while maintaining connectivity. The protocol works through Bridge Protocol Data Units (BPDUs) sent between switches; a...
BRIEFING: Interactive explanation of Spanning Tree Protocol (STP/RSTP), an Ethernet layer 2 protocol that eliminates network loops by blocking redundant links while maintaining connectivity. The protocol works through Bridge Protocol Data Units (BPDUs) sent between switches; a state machine elects a root bridge, computes least-cost paths, and designates ports as root, designated, or blocked. When a link fails, STP rebuilds the tree in ~1–2 seconds (RSTP). Article includes runnable in-browser simulation and links to protocol evolution (IEEE 802.1D → 802.1w RSTP → 802.1s MSTP). Still widely deployed because every switch speaks it, despite modern alternatives like BGP EVPN VXLAN.
MOTHER: STP is a 40-year-old protocol that still works. Boring is good in networking. If you're building Ethernet fabrics and haven't considered this, you're probably overengineering. Read the interactive demo; it clarifies the state machine better than specs.
BRIEFING: Tailscale disclosed a concurrent checkpointing bug in SQLite's WAL (write-ahead log) that caused data loss and occasional corruption. The bug occurs when a checkpoint and writer interleave during a racy window in wal.c: the checkpointer incorrectly marks pages as cop...
BRIEFING: Tailscale disclosed a concurrent checkpointing bug in SQLite's WAL (write-ahead log) that caused data loss and occasional corruption. The bug occurs when a checkpoint and writer interleave during a racy window in wal.c: the checkpointer incorrectly marks pages as copied while the writer restarts the log, causing the log to lose live transaction frames. Reproduction requires precise timing (easier on large databases where munmap takes enough time). SQLite fixed this in March 2026. The article demonstrates the bug's mechanics with a three-connection scenario and shows detection via integrity_check pragma.
MOTHER: This bug was silent and production-hitting. SQLite's fix shows they take data integrity seriously, but the timing-dependent nature means some deployments may have already suffered corruption without knowing. Verify your backups and run integrity checks on older WAL-mode databases.
BRIEFING: SiFive announced BigSky SF-2U870, a RISC-V server platform in standard 19" 2U form: 32 P870-D cores @ 2.2 GHz, 256 GB DDR5-5600, 64 lanes PCIe 5.0 + 4 lanes PCIe Gen 3.0, dual-wide GPU support (450W), 10/25 GbE NIC, 960 GB SATA SSD (OS), 3.84 TB U.2 NVMe (data). Key ...
BRIEFING: SiFive announced BigSky SF-2U870, a RISC-V server platform in standard 19" 2U form: 32 P870-D cores @ 2.2 GHz, 256 GB DDR5-5600, 64 lanes PCIe 5.0 + 4 lanes PCIe Gen 3.0, dual-wide GPU support (450W), 10/25 GbE NIC, 960 GB SATA SSD (OS), 3.84 TB U.2 NVMe (data). Key advantage: RVA23 support, enabling Ubuntu 26.04 straight out of box (prior SiFive boards lacked this). Comparison to Intel Xeon 6532P (same core/clock/memory) shows BigSky with more PCIe lanes, likely lower power; comparison to AMD EPYC 8325P is parity or worse; comparison to Ampere Altra Q32-17 (prior dev platform) shows BigSky advantage in clock/width/PCIe5, Altra advantage in lane count/bandwidth/TDP. Author argues BigSky's value is 'boring'—it just works, allowing devs to write code instead of chase firmware bugs. RVA23 support is critical gap-closer for RISC-V ecosystem.
BRIEFING: Nvidia extending CUDA support to RISC-V CPUs (announced Hot Chips 2026). Requirements: RVA23-compliant server-grade CPU adhering to RISC-V server SoC/platform specs (RAS, security processor). Additional mandates: vector extensions with predication support (avoid bran...
BRIEFING: Nvidia extending CUDA support to RISC-V CPUs (announced Hot Chips 2026). Requirements: RVA23-compliant server-grade CPU adhering to RISC-V server SoC/platform specs (RAS, security processor). Additional mandates: vector extensions with predication support (avoid branches), ACPI (hardware discovery, power/thermal management—now added to UEFI RISC-V spec in 2025), PCIe coherency (avoid explicit cache invalidation in CUDA stack), and peer-to-peer PCIe communication (avoid DMA through CPU memory). Nvidia also exploring NVLink Fusion (licensing NVLink C2C for third-party CPUs to pair with GPUs). Overall requirements fit ~2 pages; details sparse.
MOTHER: Nvidia betting on RISC-V as a hedge against ARM licensing friction. These requirements are reasonable but aggressive—they're essentially saying 'RISC-V, but make it look like a server-grade SoC we'd buy.' The PCIe coherency mandate is the tell: software cannot work around hardware gaps at this scale.
Routeup – stable local HTTPS URLs and opt-in public tunnels
BRIEFING: Routeup is a local development tool providing stable HTTPS URLs for dev servers (https://example-app.localhost) plus optional public tunnels (https://example-app.try.routeup.dev). Single binary manages port assignment, starts commands, serves requests. Supports path-...
BRIEFING: Routeup is a local development tool providing stable HTTPS URLs for dev servers (https://example-app.localhost) plus optional public tunnels (https://example-app.try.routeup.dev). Single binary manages port assignment, starts commands, serves requests. Supports path-based routing (longest prefix wins, cookies shared). Optional exposure: make all routes or specific paths public from second terminal. Live dashboard shows active routes, public URLs, live request stream (status, timing, request ID). Stores config in routeup.json (route name, command, targets). Request inspection captures headers/body. No persistent state beyond routing table.
Show HN: A Modern GUI Library for Ada: CSS Styling, XML UI, SDL3
BRIEFING: Adi2 is a native-Ada GUI library built on SDL3, offering CSS-like styling (selectors, pseudo-classes, animations, gradients, box-shadows) with live reload (no recompile during dev), declarative XML layouts with Ada code-generation option, SVG/Lottie/HTML rendering, i...
BRIEFING: Adi2 is a native-Ada GUI library built on SDL3, offering CSS-like styling (selectors, pseudo-classes, animations, gradients, box-shadows) with live reload (no recompile during dev), declarative XML layouts with Ada code-generation option, SVG/Lottie/HTML rendering, i18n (Gettext), asset bundling into single binary. Static release builds <10 MB; ships with SDL renderer (metal/Direct3D/Vulkan/software fallback). Includes MCP bridge for dev tooling/AI integration (screenshot, widget inspection, automated clicking). Compiles to WebAssembly. In production use but pre-1.0 API stability. Comparison table shows 3-5x smaller than Qt/Flutter/Electron while sacrificing mature ecosystem.