Sparkles

maggieappleton.com·1 months ago

Maggie Appleton’s fresh GitHub Next talk argues coding agents need shared, multiplayer workflows. It demos Ace, a collaborative cloud workspace blending chat, planning, previews, and agent-driven development for better team alignment.

x.com·1 months ago

A social post from AI researcher Kangwook Lee discussing why Codex via the Codex app appears to use fewer tokens, sparking developer conversation about model efficiency and tooling behavior.

michaellivs.com·2 months ago

Michael Livs dissects Anthropic's new generative UI, revealing the read_me/show_widget tool pattern, streaming DOM injection, and CDN‑based sandbox. He implements a Glimpse‑backed Pi extension to reproduce progressive, tokenized widget rendering in terminals.

github.com·7 months ago

A ReactiveX-style library for Go providing Observable, operators, Subjects, and schedulers. It uses Go generics, offers plugins and examples, and targets event-driven asynchronous applications with composable, declarative stream processing.

shloked.com·7 months ago

Shlok Khemani reverse-engineers Poke, releasing OpenPoke — an open-source prototype that reproduces Poke’s orchestrated multi-agent email assistant, persistent execution agents, triggers, and memory layers; outlines design trade-offs, costs, and limitations.

github.com·7 months ago

Tunix is a Google open-source JAX-native library for LLM post‑training, providing supervised fine‑tuning, PEFT (LoRA/QLoRA), reinforcement learning (GRPO, DPO), and knowledge distillation with distributed accelerator support; currently in early development.

bigdata.2minutestreaming.com·7 months ago

Technical deep dive explaining how Amazon S3 achieves 1 PB/s and 150M QPS using massive parallelism: ShardStore, erasure coding (5-of-9), shuffle sharding, power-of-two placement, continuous rebalancing and scale-driven stability.

ysymyth.github.io·8 months ago

Shunyu Yao argues AI has entered a 'second half' where RL generalizes, shifting focus from inventing models to defining real-world tasks, measurements, and evaluation-driven productization to bridge capability to utility.

ghuntley.com·8 months ago

A tongue-in-cheek, Go-like language created by Geoffrey Huntley by iterating Claude in a loop. CURSED swaps keywords for Gen‑Z slang, includes a Zig compiler, examples, and a GitHub repository.

youtu.be·8 months ago

Google AI Release Notes hosts Logan Kilpatrick and Gemini researchers explaining Gemini 2.5 Flash Image (nicknamed nano‑banana), demonstrating pixel‑perfect edits, interleaved generation, multi‑turn consistency, text rendering, and real‑world applications for creators and developers.

youtube.com·8 months ago

Scott Wu joins John Collison to talk about Cognition’s AI software engineer, the Moneyball-ification of everything, math competitions with Alexandr Wang in 6th grade, acquiring Windsurf over a weekend, whether coding tools will be replaced by the labs, and why he thinks we already have AGI.

dropbox.tech·1 months ago

Dropbox details how Magic Pocket reduced immutable blob-store overhead through compaction and multiple optimization strategies, offering a practical large-scale infrastructure case study on storage efficiency, cost control, and systems design tradeoffs.

github.com·1 months ago

Open-source macOS automation layer for AI agents, using accessibility trees instead of screenshots to control apps, learn reusable workflows, and integrate through MCP. Best suited for developers building agentic desktop tooling on Mac.

cocoindex.io·6 months ago

Open-source ETL framework (Rust core, Python SDK) that incrementally transforms and indexes data for AI. Provides dataflow programming, live incremental updates, observability via CocoInsight, and connectors for vector DBs, storage, and multimodal pipelines.

blog.google·7 months ago

Google and Yale released Cell2Sentence-Scale 27B, an open Gemma-based single-cell foundation model that generated a novel drug hypothesis (silmitasertib) to boost antigen presentation; laboratory tests confirmed a context‑dependent ~50% increase.

arxiv.org·7 months ago

Introduces ReasoningBank, a memory framework that distills reusable reasoning strategies from agents' successful and failed experiences. Combined with MaTTS (memory-aware test-time scaling), it boosts agent effectiveness, efficiency, and generalization on web and software benchmarks.

shloked.com·7 months ago

A reverse-engineering of ChatGPT's memory that identifies four storage layers—interaction metadata, recent conversations, model-set context, and dense user knowledge memories—and argues OpenAI prefers brute-force context over retrieval systems, with privacy implications.

jax-ml.github.io·7 months ago

Comprehensive chapter explaining GPU internals, memory and interconnects, collective communication, and roofline analysis for scaling large language models. Focuses on NVIDIA architectures and compares GPUs to TPUs for parallelism.

anthropic.com·8 months ago

Anthropic explains practical techniques for designing, prototyping, and evaluating tools for LLM agents, showing how Claude and Claude Code can optimize tool descriptions, responses, and evaluations to improve agent performance.

tempo.xyz·8 months ago

Tempo is a Stripe- and Paradigm-incubated payments-first Layer 1 built for stablecoin rails—EVM-compatible, private testnet with enterprise partners, targeting sub-second finality, predictable low fees, privacy options, and high throughput.

bencallahan.com·8 months ago

Ben Callahan reports on a 95-practitioner session finding AI helps divergent work (brainstorming, summaries, prototypes) but fails at convergent, deterministic tasks—production code, pixel fidelity, and repeatability—urging context, tests, and governance.

blog.wilsonl.in·8 months ago

Software engineer Wilson Lin recounts building a web search engine in two months, generating 3 billion SBERT embeddings with 200 GPUs, indexing 280M chunks, and serving ~500 ms queries via sharded RocksDB+HNSW. He shares architecture, cost hacks, limitations, and a live demo.

aprilnea.me·1 months ago

A technical teardown of Claude Code uncovers Antspace, Anthropic’s apparent internal PaaS, detailing deployment clues, runtime architecture, and reverse-engineering methods for readers interested in AI infrastructure and platform design.

ml.ink·1 months ago

Ink is an infrastructure platform for AI agents, enabling code deployment, observability, databases, domains, and MCP-based automation so agents can build, ship, monitor, and manage apps with minimal DevOps overhead.

alexzhang13.github.io·7 months ago

Alex L. Zhang (with Omar Khattab) introduces RLMs — a REPL-based inference wrapper letting LMs recursively call sub-models to process unbounded contexts, showing strong gains on long-context benchmarks and discussing costs, limits, and strategies.

alphaxiv.org·7 months ago

Introduces REFRAG, a compress-sense-expand decoding framework that shortens decoder input for RAG, reducing latency and KV cache size while preserving accuracy, enabling up to 16× longer contexts and large TTFT speedups.

arxiv.org·7 months ago

GEPA (Genetic-Pareto) evolves prompts using natural-language reflection, iterative mutation, and Pareto-based selection to learn high-level rules from trajectories. It achieves higher accuracy and dramatic sample-efficiency versus RL and leading prompt optimizers.

anthropic.com·7 months ago

Anthropic’s update adds context editing and a file-based memory tool (with Claude Sonnet 4.5) so developers can run longer, more accurate agent workflows by trimming stale context and persisting useful state across sessions.

jeremyberman.substack.com·8 months ago

Jeremy Berman explains how he reclaimed first place on the ARC‑AGI public leaderboard using Claude Sonnet 3.5, an evolutionary test‑time compute strategy, heavy parallel sampling, and open‑source reproducible prompts.

ghuntley.com·8 months ago

Geoffrey Huntley describes 'Ralph' — a simple Bash loop that repeatedly feeds PROMPT.md into a coding agent (Sourcegraph Amp), enabling rapid repository generation, iterative tuning, and surprising emergent behaviors like creating a new language.

arxiv.org·8 months ago

Shows fundamental representational limits of single‑vector embeddings: introduces LIMIT, a dense synthetic dataset where state‑of‑the‑art embedders fail, proving dimensionality bounds constrain retrieval and motivating multi‑vector, sparse, or hybrid solutions.

martinalderson.com·8 months ago

Martin Alderson debunks claims that inference is an unsustainable loss, using H100 cost math and throughput analysis to show input tokens are cheap while output tokens are costly, implying many inference use-cases remain profitable.

lucumr.pocoo.org·10 months ago

Armin Ronacher argues MCP works best when exposing a single, stateful code tool—like a Python/pexpect interpreter—rather than dozens of brittle CLI tools. He demonstrates LLDB debugging via pexpect-mcp and experiments with a Playwright JavaScript 'ubertool,' emphasizing composability, reliability, and reusable scripts.