Daily Tech News - 2026-07-03

Today's tech landscape buzzes with innovations pushing the boundaries of AI, optimizing data infrastructure, and redefining software development paradigms. In the AI realm, Google DeepMind unveiled DiffusionGemma 26B-A4B, its first open text-diffusion model, showcasing advancements in generative AI. Efficiency and infrastructure remain key, with Manticore achieving 14x faster embeddings and AMD accelerating large-scale LLM inference on new GPUs for tasks like YOLO26 object detection. The critical shift in AI infrastructure to rack and POD scale, coupled with robust multi-tiered persistent memory architectures, underpins advanced AI agents. Developers are also seeing a surge in tools for LLM observability (OpenLIT SDK, VictoriaMetrics) and security (Nebula AI-powered penetration testing), while new approaches to local LLM deployment and context model optimization are emerging. Even traditional development gets an AI-native twist with "Git for Agents" (Oak) and efficient Python compilers for LLM wikis.

In data management, real-time capabilities and scalability remain paramount. Debezium emerges as a key player for efficient Change Data Capture (CDC), streamlining real-time database replication without legacy complexities. The intricacies of distributed databases are being explored to handle modern data volumes, moving beyond traditional monolithic approaches. PostgreSQL users receive crucial advice on strict memory overcommit to avoid OOM killer issues, and keyset pagination is highlighted as a superior method for navigating large datasets efficiently.

Beyond these core areas, cloud infrastructure discussions highlight how data sovereignty is reshaping design, moving beyond geographical concerns to legal jurisdiction, and prompting explorations into alternatives to dominant providers like AWS. Developer tools are also advancing, with ContextCodeCache in Rust for source file mapping and Red Hat's Quarkus 3.33 delivering stability and performance for enterprise Java. Finally, a philosophical return to "Software, from First Principles" underscores the foundational complexity of our digital world, complemented by practical lessons on OS memory management from FreeBSD. Today's tech landscape buzzes with innovations pushing the boundaries of AI, optimizing data infrastructure, and redefining software development paradigms. In the AI realm, Google DeepMind unveiled DiffusionGemma 26B-A4B, its first open text-diffusion model, showcasing advancements in generative AI. Efficiency and infrastructure remain key, with Manticore achieving 14x faster embeddings and AMD accelerating large-scale LLM inference on new GPUs for tasks like YOLO26 object detection. The critical shift in AI infrastructure to rack and POD scale, coupled with robust multi-tiered persistent memory architectures, underpins advanced AI agents. Developers are also seeing a surge in tools for LLM observability (OpenLIT SDK, VictoriaMetrics) and security (Nebula AI-powered penetration testing), while new approaches to local LLM deployment and context model optimization are emerging. Even traditional development gets an AI-native twist with "Git for Agents" (Oak) and efficient Python compilers for LLM wikis.

Featured Articles

14× faster embeddings: how we rebuilt the ONNX path in Manticore

When we shipped Auto Embeddings — the feature that turns any text column into a vector automatically, with no separate model service to run — the most common piece of feedback was about speed. The pre...

Keywords: architectures onnx, manticore embeddings, manticore embed, onnx runtime, documents onnx, cpu onnx, onnx models, faster onnx, onnx efficiency, manticore search
Source: manticoresearch.com

Software, from First Principles

When you use your smartphone or laptop, it is easy to forget the sheer absurdity of the physical reality underneath. Behind that seamless experience lies centuries of effort from physicists, chemists,...

Keywords: makes computing, computers fundamentally, devices virtually, tremendous computing, hardware smarter, computer behavior, spend computing, smartphones, computing, smartphone
Source: fazamhd.com

What is Debezium? CDC architecture and how it works

A few years ago, setting up real-time database replication meant buying an expensive enterprise software license or writing brittle polling scripts that dragged your production servers to a crawl. Tod...

Keywords: cdc pipelines, cdc pipeline, replication tool, databases debezium, debezium databases, database replication, cdc process, infrastructure debezium, data infrastructure, specialized replication
Source: fivetran.com

Observability for LLM Apps and Agents: OpenLIT SDK + VictoriaMetrics observability stack

Blog /
Observability for LLM Apps and Agents: OpenLIT SDK + VictoriaMetrics observability stack This post is a joint effort by the OpenLIT and VictoriaMetrics teams. OpenLIT brings the OTel-native...
Keywords: observability tools, agent telemetry, observability hosted, hosted observability, observability opentelemetry, telemetry openlit, agents openlit, sdk openllmetry, observability setups, llm telemetry
Source: victoriametrics.com

Building a GPU-Resident YOLO26 Object Detection Pipeline on the AMD Radeon™ AI PRO R9700 GPU

Building a GPU-Resident YOLO26 Object Detection Pipeline on the AMD Radeon™ AI PRO R9700 GPU# Modern AMD GPUs include a dedicated hardware block for video processing called the Video Core Next (VCN) e...

Keywords: yolo26 detector, detectors yolo11, gpu rocm, architecture rocdecode, gpu decoding, amd rocm, inference gpu, amd gpus, amd gpu, predictions gpu
Source: rocm.blogs.amd.com

Jamesob's guide to running SOTA LLMs locally

Note: nothing in this README aside from the tables was written by AI. Have $2k burning a hole in your pocket and want some local, state-of-the-art machine intelligence? How about$ 40k? If Dario and Al...

Keywords: docker containers, rtx pro, good docker, pix gpus, docker, docker container, gpus, gpus communicate, model docker, rtx
Source: github.com

PostgreSQL and the OOM Killer: Why You Must Use Strict Memory Overcommit

Our team members built and operated five managed PostgreSQL services over the past 15 years. Across all of them, one configuration has remained constant: strict memory overcommit. In this blog post, w...

Keywords: memory postgresql, overcommit_memory important, memory overcommit, overcommit_memory, overcommit_memory strict, configure_memory_overcommit, configure_memory_overcommit strict, vm overcommit_memory, allocates memory, def configure_memory_overcommit
Source: ubicloud.com

Beyond the Chip: How AI's Next Battleground Moves to Rack and POD Scale

Beyond the Chip: How AI's Next Battleground Moves to Rack and POD Scale Inside NVIDIA, CSPs, and ASIC Designers' Strategies AI computing demand is growing exponentially, far outpacing Moore’s Law, whi...

Keywords: hardware ecosystem, hardware portfolio, chip performance, architectures nvidia, hardware, ai racks, ai chip, chip development, chips designing, chip ai
Source: insights.trendforce.com

LLM Wikis Are Over-Engineered — I Replaced Mine With a Pure Python Compiler

TL;DR

I built a pure Python pipeline that compiles a folder of raw, messy text notes into a linked, linted markdown wiki. No LLM calls, no embeddings, no external APIs, standard library only. - The...
Keywords: pipeline corpus, recompiles corpus, benchmarked pipeline, efficiently corpus, compiled_wiki attention_mechanism, output compiled_wiki, library pipeline, test_human_notes_preserved_across_recompile ok, test_human_notes_preserved_across_recompile, ok test_human_notes_preserved_across_recompile
Source: towardsdatascience.com

Red Hat build of Quarkus 3.33: Stability and performance advancements for enterprise Java

Red Hat build of Quarkus 3.33 has reached general availability, establishing a new Long-Term Support (LTS) stable baseline for teams that value predictability over chasing every upstream feature drop....

Keywords: quarkus update, latest quarkus, update quarkus, build quarkus, quarkus 33, quarkus redhat, new quarkus, quarkus resources, new features, run quarkus
Source: developers.redhat.com

Oak: Git for Agents

You wait minutes to read one file. Git makes you clone the entire history before you can open a single path. On a big repo that's minutes of dead time, with tokens spent watching a progress bar, befor...

Keywords: git oak, git fastest, benchmarks oak, oak benchmarks, comparing git, usable git, build git, git clone, git equivalent, git easy
Source: oak.space

What is a distributed database? Types, components, and use cases

What is a distributed database? Key components, types, and use cases Scaling a database once meant buying a bigger server. Today, the volume and velocity of enterprise data make that approach impossib...

Keywords: distributed database, distributed databases, databases distributed, database distributed, centralized databases, centralized database, distributed sql, databases spread, data distributed, distributed data
Source: fivetran.com

Anatomy of Persistent Memory's 3 Layers: Comparing ContextNest, Mem0 and Zep

Designing production-grade AI agents requires building a robust, multi-tiered persistent memory architecture. A common pitfall is expecting a single memory database or context retrieval tool to handle...

Keywords: agent memory, contextnest memory, memory paradigms, memory tiers, memory engine, persistent memory, mem0 contextnest, memory treating, mem0 manages, memory architectures
Source: promptowl.ai

FreeBSD Ate My RAM

Last month I posted about my journey migrating my site server from an old Ubuntu server to FreeBSD. Some people on Hacker News noticed that, when I showed the fastfetch result, I said I was confused w...

Keywords: free memory, ram usage, memory use, memory usage, freebsd cached, memory free, freebsd usage, memory difference, manages ram, ram provide
Source: crocidb.com

DiffusionGemma 26B-A4B: Google’s First Open Text-Diffusion Model

Almost every large language model you have used generates text one token at a time, left to right. DiffusionGemma breaks that pattern. It is the first open-weight model from Google DeepMind to use dis...

Keywords: text diffusion, diffusiongemma optimizes, context diffusion, google diffusiongemma, diffusiongemma weights, diffusion decoding, diffusiongemma uses, diffusiongemma posts, diffusiongemma support, diffusiongemma run
Source: codersera.com

Accelerating Large-Scale LLM Inference on AMD Instinct MI350X/MI355X with Eagle3 and AMD Quark

Accelerating Large-Scale LLM Inference on AMD Instinct MI350X/MI355X with Eagle3 and AMD Quark# Large language model (LLM) inference is increasingly constrained by autoregressive decoding. Even when p...

Keywords: speculative decoding, throughput speculative, decoding promising, optimized decode, decoding speedup, inference speedups, expensive decoding, decoding draft, num_speculative_tokens draft_tensor_parallel_size, decoding practical
Source: rocm.blogs.amd.com

Show HN: ContextCodeCache in Rust

Tool that scans a project and generates a ContextCodeCache - a .ccc directory holding a compact, machine-readable map of every source file: its constants, functions (with return types and doc summarie...

Keywords: path ccc, linux ccc, ccc path, ccc directory, file ccc, contextcodecache ccc, install ccc, ccc binary, ccc install, use ccc
Source: github.com

Upcoming change: Lakebase Autoscaling update notifications post 3 days in advance

What's coming? Learn about features and behavioral changes in upcoming Databricks releases. Upcoming change: Lakebase Autoscaling update notifications post 3 days in advance To deliver Postgres versio...

Keywords: models databricks, cloud databricks, databricks model, databricks plans, upcoming databricks, managed databricks, databricks service, databricks managed, beta databricks, organization databricks
Source: docs.databricks.com

How data sovereignty is changing cloud native infrastructure design

The core issue isn’t where your server sits. It’s who can be compelled to hand over what’s on it. For years, cloud providers treated sovereignty as a geography problem. Pick a region. Choose a country...

Keywords: cloud infrastructure, eu cloud, cloud procurement, cloud providers, sovereign infrastructure, cloud act, cloud vendors, infrastructure policy, underlying infrastructure, sector cloud
Source: cncf.io

Stop Paginating With OFFSET: Keyset Pagination and the Deep-Page Cliff

Stop Paginating With OFFSET: Keyset Pagination and the Deep-Page Cliff LIMIT 20 OFFSET 40 is how almost everyone paginates, and on the first few pages it is perfectly fine. The problem is what OFFSET...

Keywords: pagination offset, paginating offset, keyset pagination, pagination needsoffset, pagination workloads, pagination deep, shallow pagination, pagination feels, pagination win, pagination idea
Source: devops-daily.com

Long Context vs. Short Context Model: When Does a Long Context Model Win?

Introduction 1.1 The marketing claim, and the question it skips Each new generation of encoder models comes with a bigger context window. BERT and MiniLM gave us 512 tokens. Then ModernBERT arrived...

Keywords: encoder bert, encoder models, encoder uses, encoder modernbert, encoders embedding, modernbert encoder, encoder embedding, encoders context, encoder model, encoders
Source: towardsdatascience.com

Understanding PyTorch’s Test Infrastructure

Featured projects TL;DR

PyTorch tests are often generated at import time, so CI failures may show device/dtype-specific names that differ from the source template.
For local debugging, pytest -k a...
Keywords: instantiate_device_type_tests ci, example pytorch_testing_device_only_for, instantiate_device_type_tests dtypes, pytorch_testing_device_only_for, debugging pytorch, debug pytorch, imported instantiate_device_type_tests, types pytorch_test_with_slow, pytorch tests, file instantiate_device_type_tests
Source: pytorch.org

Nebula AI-Powered Penetration Testing Platform Automates Vulnerability Assessments

A new open-source security tool is bringing large language models directly into the penetration tester’s terminal. Nebula, developed by BerylliumSec, integrates state-of-the-art AI models into the com...

Keywords: security tool, ai backends, security tooling, automate vulnerability, execution cloud, openai models, ai terminal, openai api, vulnerability discovery, openai
Source: cybersecuritynews.com

Show HN: Mcpsnoop – Wireshark for MCP (transparent proxy and live TUI)

Wireshark for MCP. A transparent proxy that shows every real tool call between your AI client and your MCP servers, live in your terminal. The official MCP Inspector connects as its own client. It nev...

Keywords: wireshark mcp, client mcpsnoop, client mcp, mcpsnoop server, server mcpsnoop, run mcpsnoop, mcpsnoop http, mcpsnoop run, command mcpsnoop, wireshark
Source: github.com

Beyond the AWS Default: The Cloudflare Alternative

I was recently looking at a client project and hit that familiar wall. The standard move, the one everyone expects you to make, was to spin up an AWS environment. It is the “safe” choice. It is what t...

Keywords: simplicity aws, aws environment, aws powerful, aws, philosophy aws, saying aws, complexity vpc, aws focuses, choosing infrastructure, cloud debate
Source: alan.is

🗞️ China claims a new milestone in locally trained AI, as Meituan rolls out LongCat-2.0.

🗞️ China claims a new milestone in locally trained AI, as Meituan rolls out LongCat-2.0. China’s LongCat-2.0 milestone; OpenAI inference costs cut; Claude Science & Sonnet-5 system card; AI power user...

Keywords: ai spending, cost ai, china ai, bigger ai, ai tokens, ai market, ai token, ai use, ai power, ai changes
Source: rohan-paul.com

Claude, please stop trying to memorize random crap

Agentics: Memorizing Session Transcripts Isn't Useful Keep track of artifacts, not scratch. Alt title: Claude, please stop trying to memorize random crap We have found zero performance benefit on SWE...

Keywords: agentics memorizing, memorizing session, transcripts agent, session transcripts, agent session, transcript agent, transcript sessions, session transcript, transcript useful, agent context
Source: 12gramsofcarbon.com

I tested AI-managed VPS without any Linux experience – here’s how it went

I tested AI-managed VPS without any Linux experience – here’s how it went Hosting a website or application on a virtual private server (VPS) has long been associated with needing to “know Linux.” It’s...

Keywords: assistant vps, vps management, managing vps, manage vps, vps exploring, understand vps, integrated vps, managed vps, hostinger ai, vps team
Source: hostinger.com

60% Fable cost cut by converting code to images and having the model OCR it

Cut Claude Code's input tokens by rendering bulky context as images — the same system prompt, tool docs, and history, in a fraction of the tokens. An image's token cost is fixed by its pixel dimension...

Keywords: vs pxpipe, pxpipe vs, image tokens, pxpipe images, image costs, pxpipe gist, tokens images, fewer tokens, pxpipe compresses, tokens image
Source: github.com

Commodore 64 Basic for PostgreSQL

LOAD "PL/CBMBASIC",8,1: Commodore 64 BASIC for PostgreSQL If you are of a certain age, the words 38911 BASIC BYTES FREE will do something to you that no amount of therapy can undo. You remember the bl...

Keywords: commodore interpreter, programmed c64, commodore 64, cbmbasic commodore, c64, postgresql pl, rom compiled, basic postgresql, microsoft commodore, pl cbmbasic
Source: thombrown.blogspot.com

It Still Can't Do My Job: Four Years of Moving Goalposts (2022–2026)

It Still Can't Do My Job I started keeping notes in December 2022, mostly to document why the panic was overblown. The notes turned into this. The quotes in orange boxes are real. You can look them up...

Keywords: created chatgpt, chatgpt substantially, trick chatgpt, answers chatgpt, chatgpt, overblown notes, chatgpt launches, snake game, game write, write snake
Source: publicznyprofil.github.io

Show HN: I made a tool that prevents websites from tracking you

Cleaner pages. Fewer trackers. Local-only. GetBlocked! is a lightweight, local-only Chrome extension that helps reduce common third-party website tracking by blocking known tracker requests, cleaning...

Keywords: block tracker, tracker requests, tracker domains, analytics trackers, tracker domain, trackers local, tracking blocking, website tracking, blocking chrome, trackers ad
Source: github.com

Construction document data extraction: From drawings, RFIs, and submittals to structured data

Construction document data extraction: From drawings, RFIs, and submittals to structured data Table of contents Construction teams run on data — project numbers, drawing references, specification sect...

Keywords: construction documents, construction document, construction extraction, document structured, document data, construction data, structured document, construction workflows, making documents, construction workflow
Source: nutrient.io

How to build a WordPress plugin with AI (Cursor + WordPress Studio)

Earlier this year, we showed you how to use Claude Code with WordPress Studio to build your next plugin idea. While Claude Code is a great option, if you’re not comfortable using the terminal, install...

Keywords: plugins ai, building plugin, build plugins, code editor, wordpress studio, plugin idea, coding editor, studio wordpress, build plugin, plugin build
Source: wordpress.com

Valve open source the Steam Machine e-ink screen so you can make your own

While Valve will not be making and providing their own e-ink display for the Steam Machine, they have opened it up so anyone can now do it. Valve originally teased it with the first lot of reviewers t...

Keywords: ink display, example ink, machine ink, display steam, inkterface, providing ink, images steam, steam machine, inkterface good, plan ink
Source: gamingonlinux.com

Vercel Sandbox now supports FUSE-based filesystems

Vercel Sandbox now supports FUSE, letting you mount remote storage and custom filesystems inside a running Sandbox. Use it to attach S3 buckets, network filesystems, or any other FUSE-compatible drive...

Keywords: s3_bucket_name mount_dir, run s3fs, mountpoint s3, s3fs inside, mount s3, s3fs, env s3_bucket_name, mountpoint amazon, s3 amazonaws, attach s3
Source: vercel.com

Show HN: Updated my landing page with Fable (retro pixel style)

Guildly gives you a small team that plans, builds, and ships together. You set the direction and approve what matters. FREE · MAC / LINUX / WINDOWS Real problems for the people building today. Just yo...

Keywords: team guildly, guildly, dms guildly, guildly turns, guildly keeps, team ai, ai employees, guildly gives, board team, autopilot team
Source: tryguildly.com

Ubuntu 26.04 LTS Restores Video and Audio Thumbnails: What You Need to Know

Ubuntu 26.04 LTS Restores Video and Audio Thumbnails: What You Need to Know If you have recently installed Ubuntu 26.04 LTS and found that video and audio files lack image thumbnails in the file manag...

Keywords: missing thumbnail, thumbnail packages, thumbnails packages, thumbnailer packages, thumbnailer gst, install thumbnailers, audio thumbnails, thumbnailer restart, thumbnailers newer, files lack
Source: serverhost.com

Connect Your ChatGPT Subscription to exe.dev

We're pleased to announce that you can connect your ChatGPT subscription to exe.dev and use our coding agent Shelley with the OpenAI models you're already paying for! To enable this, go to exe.dev/int...

Keywords: openai integrations, openai integration, integration openai, subscription openai, models openai, openai models, dev openai, openai, llm openai, subscription exe
Source: blog.exe.dev

Quarkus SPDX SBOMs: JSON-LD, Licenses, Real Checks

Quarkus SPDX SBOMs: JSON-LD, Licenses, Real Checks Build a small Quarkus service, generate an SPDX 3.0.1 SBOM, inspect the JSON-LD graph, and see exactly where NTIA and CISA checks still fail. I have...

Keywords: sboms dependency, dependency sboms, spdx quarkus, quarkus spdx, sbom spdx, spdx sbom, sbom packaging, quarkiverse spdx, spdx sboms, spdxdocs quarkus
Source: the-main-thread.com

The Free Market Lie: Why Switzerland Has 25 Gbit Internet and America Doesn't

The Free Market Lie: Why Switzerland Has 25 Gbit Internet and America Doesn't You may have heard about 25 Gbit symmetrical internet in Switzerland. This is often cited as the fastest dedicated (non-sh...

Keywords: internet switzerland, switzerland fiber, swisscom infrastructure, broadband prices, speed internet, infrastructure swisscom, switzerland differ, internet speeds, fiber internet, switzerland fast
Source: stefan.schueller.net

More Orb Sizes

<p>You can now pick the size of the <a href="/news/agents-in-orbs">orbs</a> used to run Amp agents remotely:</p> <ul> <li><code>a0.tiny</code>: 1 CPU, 2GB memory, 40GB disk ($0.10/hour)</li> <li><code...

Keywords: size project, amp agents, href projects, large code, orbs amp, 16 cpus, cpu 2gb, run amp, cpus 16gb, agents remotely
Source: ampcode.com