New

Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillationapple.com

Apple researchers propose Trajectory‑Shaped Discrete Flow Matching (TS‑DFM), a training‑time distillation method that replaces blind stochastic jumps in discrete flow‑matching with an energy‑based compass to select higher‑quality intermediate tokens. On a 170 M‑parameter language model, the 8‑step student outperforms the 1 024‑step teacher by 32 % perplexity while being 128× faster, beating basel…

CISA's Logging Reference Architecture for OMB M-26-14: What federal agencies should do nextelastic.co

CISA’s Logging Reference Architecture translates OMB M‑26‑14 requirements into eight concrete shifts, from searchable storage tiers to schema‑first design and continuous validation. Agencies should use Elastic’s tools to build a heterogeneous, schema‑first logging architecture and file a plan by 18 Nov 2026 to meet the mandated maturity milestones.

On-Demand Masked Sessions with Twilio Proxy, Voice and Serverlesstwilio.com

A step‑by‑step tutorial showing how to build a Just‑in‑Time masked‑call workflow with Twilio Voice, Proxy, and Sync, using a two‑bounce out‑of‑session pattern to collect a tracking code via IVR, resolve the counterpart’s number, stash it in Sync, and auto‑create a Proxy session on the fly—all deployed as Twilio Serverless Functions.

A Zeroth-Order Paradigm for LLM Preference Alignmentarxiv.org

The paper proposes Comparison-based Preference Optimization (ComPO), a zeroth‑order method that uses comparison oracles to align LLMs without a differentiable loss. Experiments on several LLM families show it improves win rates and mitigates likelihood displacement compared to direct alignment approaches.

Can MiniMax-H3 Reason About the Physical World? An Evaluation of Omni-Modal Generative Modelarxiv.org

This paper introduces a novel evaluation framework to assess the physical world reasoning capabilities of omni-modal generative models like MiniMax-H3. It found that MiniMax-H3 achieved an overall success rate of 41.97% across 517 instances, with significant performance variations depending on the input modalities and reasoning tasks.

GAVEL: Graph World Models for Verified and Efficient Long-Horizon LLM Task Planningarxiv.org

GAVEL augments LLM‑driven robot planners with an explicit graph world model that verifies actions, repairs violations, and reasons over belief distributions, boosting single‑task success from 41 % to 92 % and multi‑task success from 20 % to 93 % on BEHAVIOR‑1K.

Elastic announces GA of cross-project search on Serverless, enabling teams to query across all linked projects without moving a byteelastic.co

Elastic Cloud Serverless now offers General Availability of Cross‑Project Search (CPS), letting you run a single query across up to 100 linked projects (any region, cloud, or solution type) without moving data. Setup is UI‑driven, permissions are evaluated per‑project, and costs are $0.009 / GB retained plus $0.05 / GB egress. CPS works for Observability, Security, Search, vector DB, and ML jobs,…

Transform and route security logs to Microsoft Sentinel tables using Observability Pipelinesdatadoghq.com

Datadog Observability Pipelines now ships pre‑built Microsoft Sentinel Packs that map logs from Palo Alto, Fortinet, Cisco ASA, Cisco Meraki, and ExtraHop into Sentinel’s CommonSecurityLog or Syslog tables. Packs handle field extraction, severity derivation, and device‑action mapping, letting you filter or drop low‑value events before ingest, validate mappings with Live Capture, and reduce per‑GB…

DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Modelsapple.com

DACA‑GRPO adds denoising‑aware credit assignment to GRPO‑style RL trainers for diffusion LLMs. It computes per‑token importance scores from intermediate denoising steps and uses stratified masking to reduce mean‑field bias in likelihood estimates. Plug‑and‑play on three existing GRPO methods, it yields consistent gains on seven downstream tasks (up to +5.6 pp math, +7.4 pp code, +36.3 pp constrai…

How Value Induction Reshapes LLM Behaviourapple.com

Apple researchers fine‑tune LLMs on curated subsets of value‑oriented preference data and measure cross‑value effects, safety, and anthropomorphic language. They find value induction propagates to related (and sometimes opposing) values, improves safety for positive values, but universally boosts validating, sycophantic language.

CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agentsarxiv.org

CERA-MoA proposes a reinforcement‑learning loop where a router and a set of LLM agents are trained together. A “familiarity” estimator reads mid‑layer hidden states to predict each agent’s competence on a query, letting the router activate only a minimal subset of agents that meet a cumulative confidence threshold. The system also feeds targeted training examples to agents based on their evolving…

Agora: Git as Shared Memory for Collective AutoResearcharxiv.org

Agora treats a Git repository as a shared, append‑only memory for autonomous research agents, recording each claim as an immutable commit in a DAG. In a 12‑day run with 13 language‑model workers it generated 1,703 reproducible contributions and closed 62 % of the gap to a trained GPT‑2‑124M, demonstrating that shared research state can accelerate discovery.

From alert to resolution: Manage incidents with Bits Chat in Slackdatadoghq.com

Datadog’s Bits Chat adds a natural‑language interface to Slack, letting responders start investigations, get root‑cause analysis, generate code fixes, and close incidents without leaving the channel. The post walks through a sample e‑commerce outage to show the end‑to‑end workflow.

Build zero-trust AI agents that judge intent, not just syntaxgoogleblog.com

Google's Gemini Enterprise Agent Platform adds three managed runtime controls—Model Armor, Semantic Governance Policies, and Agent Anomaly Detection—to protect LLM agents from prompt injection, social‑engineering refunds, and multi‑turn abuse, moving enforcement out of agent code.

Saving Jet Fuelmarksblogg.com

A step‑by‑step tutorial showing how to use the open‑source scikit‑decide framework together with the OpenAP aircraft performance model to compute fuel‑optimal flight trajectories. The post details the author’s high‑end workstation, installs Python 3.12, scikit‑decide, OpenAP, and DuckDB with several extensions, then explores OpenAP’s aircraft data (e.g., A380‑800 specs and drag polar) and demonst…

"The Secret Life of Circuits" is herecoredump.cx

Unsizing unsized valueshackmd.io

Recreating Voodoo Graphics and a Late-1990s Gaming PC on an FPGAgithub.io

The author built zSST, a SystemVerilog recreation of the 3dfx Voodoo 1 graphics chip, and integrated it with a z486 CPU on a Xilinx KV260 to run DOS games with original 3D rendering at 100 MHz. The post details the fixed‑point pipeline, stage breakdown, and memory subsystem needed to sustain one pixel per clock.

Gemini Live audiosimonwillison.net

Simon Willison shows a minimal JavaScript UI that talks to Google’s Gemini 3.8 Live speech‑to‑speech models via a WebSocket API, using the Web Audio API for capture/playback and no external libraries.

Lucasart's Afterlifewordpress.com

This blog post revisits Lucasart's 1996 game "Afterlife," a unique SimCity-like game where players manage Heaven and Hell, offering strategies for new and returning players. It details specific game mechanics, zoning tactics, and common pitfalls to help optimize soul collection and progress.

‘Now We Can Know Everything and Do Anything,’ Jensen Huang Says at Dreamforcenvidia.com

Nvidia’s Jensen Huang announced Salesforce’s Koa, a CRM‑reasoning LLM built by fine‑tuning Nvidia Nemotron 3 Super on a synthetic, three‑decade‑spanning dataset. Koa uses supervised fine‑tuning plus RL (NeMo RL, Gym, AutoModel), covers 14+ industries, and claims 3× fewer errors on Salesforce’s CRM‑Bench versus leading models. It’s already in internal Slack agents and slated for limited customer p…

The engineering behind the US Strategic Petroleum Reservejohnjwang.com

The US Strategic Petroleum Reserve stores 714 M barrels in ~60 underground salt‑caverns excavated by water dissolution, a cheap, low‑permeability solution that protects against attack and enables rapid drawdown, but requires brine handling and careful cavern monitoring.

txcript: Switching coding agents mid-conversationgithub.com

German Rheinmetall open-sources its Battlesuite connected weapon system protcolgithub.io

Rheinmetall has open‑sourced onboardapi 9.10.0, a C++ middleware that uses DDS (XTypes, XCDR2) for low‑latency, backward‑compatible data exchange between sensors and software. The library includes Java, C#/.NET, and Python wrappers and comprehensive docs.

My temporary PHP fix from 2014 has nearly 20M installs. Today I'm deprecating itjakeasmith.com

A 174‑line PHP shim for `http_build_url()` was published in 2014, ended up with ~20 M installs, and is now deprecated in favor of the PHP‑League URI library and the native URI API in PHP 8.5.

Small Programming Trickswill-keleher.com

A grab‑bag of command‑line, SQL, and JavaScript shortcuts – from fuzzy‑searching shell history with fzf to using `SELECT` without a `FROM` in Postgres, `git log -S`, and keeping an HTTPS agent alive in Node. The author argues that such bite‑size tips boost daily productivity and suggests sharing one per day on a team channel.

1Password's AI patching benchmark is misleadingtrailofbits.com

Bypassing inference bottlenecks: Accelerating complex AI search with Retrieve-for-Trainresearch.google

Retrieve-for-Train addresses LLM inference bottlenecks in complex AI search by using offline reinforcement learning to train a lightweight diffusion model. This enables efficient, single-pass generation of diverse, property-aligned sub-queries, bypassing slow autoregressive reasoning.

Building a Linux GPU Driver for the M4 Mac Mini in One Monthcodyho.dev

Built a clean‑room OpenGL ES 3.0 Linux driver for Apple‑silicon M4/A18 Pro GPUs in ~4 weeks, covering reverse‑engineered firmware ABI, a Rust kernel driver, a custom IR/shader compiler, and user‑space Metal translation; achieved 200 fps Minecraft and WebGL demos, with heavy LLM assistance for debugging and code generation.

WangNet – 1.8 MB, zero-dependency Numberwang adjudication in 11 languagesgithub.com

Numberwang is a tiny character‑level CNN packaged as a 1.8 MB JSON model and a ~100‑line pure‑Python inference script that classifies numeric strings in eleven languages without any external dependencies. It reaches 88.9 % macro‑F1 on a held‑out set, but struggles with unseen arithmetic expressions.