The Git Times

Ed Donner's llm_engineering repository has completed its major December 2025 refresh, replacing every week of the original course with new material. The update reflects changes in model availability and engineering patterns since the repo first appeared in 2024, while preserving access to legacy code through the original branch.

Users run git fetch followed by git checkout original to revert to the version matching the initial video series. The main branch now begins with Llama 3.2 projects, explicitly advising against Llama 3.3 on consumer hardware because of its memory requirements. Subsequent weeks advance from basic prompting through retrieval-augmented generation to multi-agent systems, with each notebook building directly on the previous week's code.

The materials consist entirely of Jupyter Notebooks that let engineers experiment with local model calls, vector stores, and agent orchestration. Supporting resources, including slides, links and an FAQ, live at edwarddonner.com. Donner continues to offer direct support via email at ed@edwarddonner.com and through the Udemy course platform.

The refreshed repository keeps the eight-week progression relevant as new open-source models emerge, giving practitioners concrete, incremental experience rather than isolated examples.

Streamlit has released version 1.56.0, adding practical improvements to its framework for converting Python scripts into interactive web applications. The update targets navigation, data presentation and media support, addressing common requests from developers building dashboards, reports and chat tools.

Key additions include the ability to specify visible items in st.navigation via the expanded parameter and on_click rerun support for st.link_button. File type shortcuts now simplify st.file_uploader and st.chat_input. Developers can programmatically control selections in st.dataframe, while st.column_config gains AudioColumn and VideoColumn types.

A new st.menu_button widget expands UI options. The st.table component supports hide_index and hide_header parameters, and st.chat_input accepts a height setting for better layout control. These changes build on Streamlit's live-editing model, where script modifications update the app instantly without restarting.

The release maintains the project's core approach: writing standard Python that becomes shareable web apps in minutes. Its Community Cloud platform continues to offer free deployment and management. Six years after its initial release, Streamlit remains focused on reducing the time between Python analysis and interactive applications used across machine learning, finance and scientific workflows.

(178 words)

PyTorch Lightning has shipped version 2.6.1, updating the framework that lets developers pretrain and finetune AI models of any size on one or 10,000+ GPUs with no changes to core model code.

The library structures PyTorch projects to remove repetitive engineering. It automates backpropagation, mixed precision, device placement, and distributed strategies while leaving model logic untouched. Users define training steps inside a LightningModule; the Trainer handles the rest. Two packages exist: the full PyTorch Lightning framework for structured workflows and Lightning Fabric for granular control.

This release drops support for Python 3.9 following its end-of-life, deprecates the to_torchscript method to align with PyTorch's direction, and adds method chaining to LightningModule.freeze() and unfreeze(). It also introduces litlogger integration. Fixes address hyperparameter inheritance, LightningDataModule checkpoint restoration, and ModelParallelStrategy compatibility with torch.compile.

Lightning can run on local hardware, private clusters, or through Lightning Cloud for managed GPU resources without infrastructure setup. A companion library, LitServe, enables custom inference servers written entirely in Python.

As model sizes continue growing, the framework's separation of science from engineering remains relevant for research and production teams.

Eight years after its creation, MichaelGrupp/evo remains the standard Python package for evaluating odometry and SLAM output. The project received its latest updates in April 2026, ensuring continued compatibility with Python 3.10+ and current robotics stacks.

The tool ingests TUM trajectory files, KITTI pose files, EuRoC MAV data, and ROS or ROS2 bagfiles containing geometry_msgs/PoseStamped, TransformStamped, or nav_msgs/Odometry messages. It supplies common routines for trajectory association, alignment, and scale adjustment required by monocular systems.

Flexible output distinguishes evo from rigid dataset benchmarks. Users generate plots, export tables to Excel, or produce LaTeX figures directly from the command line. The modular core and tools libraries allow custom extensions while the CLI handles most daily tasks without additional coding. Benchmarks show it executes faster than comparable Python evaluation suites.

Because it avoids replicating any single dataset protocol, teams can apply the same commands across KITTI, TUM, and EuRoC experiments. For developers migrating SLAM pipelines to ROS2 or processing long-duration autonomous runs, evo delivers consistent metrics without rewriting evaluation code.

Installation uses the familiar pip install evo command, with editable installs available for those tracking the latest changes from source.

UBC Thunderbots has issued v1.2.1 of its robot soccer software, introducing practical enhancements to the C++ codebase that controls its autonomous fleet. The update adds support for an --enable_autogc flag, providing an option to automate game control processes. This feature is expected to simplify operations during both development and competitive play.

Improvements to stat recording now accommodate longer durations, allowing teams to capture detailed performance metrics without interruption. Such capabilities are essential for thorough evaluation ahead of major tournaments.

Additionally, a new schema for verbose logs standardizes how diagnostic data is structured and stored. This change should make it easier for engineers to analyze system behavior and debug issues efficiently.

The software forms the backbone of the Thunderbots' RoboCup Small Size League efforts, managing everything from individual robot control to collective team strategies. Its architecture supports sophisticated AI that enables the robots to play soccer with remarkable autonomy.

Contributors must adhere to the project's style guide and consult architecture documents before making changes. The release notes highlight the collaborative effort behind these refinements, with contributions from multiple team members.

These incremental advances demonstrate the sustained commitment required to compete in the rapidly progressing field of robotic sports. By focusing on these technical details, the team ensures their robots can perform reliably in the demanding environment of international robot soccer competitions.

Kornia is redirecting development toward end-to-end vision models. The PyTorch-based library is now prioritising integration of state-of-the-art Vision Language Models and Vision Language Agents to create complete spatial AI solutions within existing deep learning workflows.

The move extends its established collection of more than 500 differentiable operators. These continue to support core image processing tasks including Gaussian and Sobel filters, affine and homography transformations, histogram equalisation, and edge detection with Canny and Laplacian methods. All operations remain GPU-accelerated and fully differentiable for seamless inclusion in training pipelines.

Recent release v0.8.2 brings documentation updates, dependency bumps and a community migration from previous channels to Discord. The library maintains its augmentation tools such as AugmentationSequential, RandAugment and TrivialAugment, alongside pre-trained models for face detection with YuNet, feature matching with LoFTR and LightGlue, and segmentation with SAM.

This evolution addresses the demand for frameworks that combine low-level geometric operations with high-level semantic understanding in a single differentiable environment. Robotics and autonomous systems developers can now prototype complete vision pipelines without switching between separate libraries for classical processing and modern multimodal models.

The changes reflect broader industry movement toward unified vision systems that support both precise spatial reasoning and language-guided interpretation.

Radare2 has shipped version 6.1.2, codenamed "Brainroot," with 224 commits from 15 contributors focused on tightening its core analysis capabilities.

The update addresses several practical pain points. It now preserves anal.timeout settings across iterators, adds APIs to get and set function signatures, and fixes selection of overlapped functions in pdc. Invalid code checks have been unified, stopping filler-prefix blocks earlier to avoid wasted cycles.

To improve stability, the tool no longer crashes when hitting large basic-block limits and sets a new 64KB default. The maximum basic-block size has been reduced from 512K to 8K, while jump-table validation checks have been strengthened. These changes matter for users regularly processing complex or obfuscated binaries where small inaccuracies compound quickly.

Now in its 14th year, the LGPLv3 command-line framework continues to support editing local files, viewing kernel memory, and debugging both locally and over gdb/windbg. Scripting via the embedded JavaScript interpreter or r2pipe remains central, as does the plugin ecosystem delivered through r2pm.

Installation follows the established path: clone the repository and run sys/install.sh, with meson and make builds both supported. The release keeps radare2's reputation for low-level control intact while delivering the incremental reliability that long-time users expect.

Word count: 178

TheSpeedX/PROXY-List received its latest update on 5 April 2026, bringing the total number of collected proxies to 7,868. The project, first created in 2018, maintains separate plain-text lists of free public proxies scraped from across the internet and refreshed periodically for easy consumption.

The repository provides three distinct lists: SOCKS5, SOCKS4, and HTTP. Developers can fetch them directly via raw GitHub URLs:

• https://raw.githubusercontent.com/TheSpeedX/SOCKS-List/master/socks5.txt
• https://raw.githubusercontent.com/TheSpeedX/SOCKS-List/master/socks4.txt
• https://raw.githubusercontent.com/TheSpeedX/SOCKS-List/master/http.txt

The maintainer explicitly states the lists are for educational purposes only, notes they do not control the proxies' reliability or legality, and requests credits, stars, and follows from those who use the data. A companion validation tool is also referenced for checking which SOCKS proxies are currently operational.

In an environment of increasing IP-based blocking and privacy scrutiny, the project's consistent updates offer builders a regular starting point for testing network conditions, though free proxies remain inherently unstable and should be validated before production use. The simple text-file format ensures straightforward integration into scripts and testing pipelines.

(178 words)

IntelOwl has shipped version 6.6.0, delivering concrete performance and capability upgrades to its open-source threat intelligence platform.

The most notable change is the migration to PostgreSQL v18, closing a long-standing scalability issue for organizations running high volumes of concurrent analyses. A new Machofile analyzer now handles Mach-O binaries, extending coverage to macOS malware formats that previously required separate tooling.

Additional updates include an improved InQuest analyzer with more accurate observable type detection, a URLScan.io Crawl Results Visualizer that enhances the built-in GUI, and a management command for checking new releases. Fixes address HTTP timeouts in the Yeti connector, correct its v2 API usage, and add background colors to TLP badges in the job API. Unit tests now run significantly faster.

IntelOwl combines enrichment of files and observables — IPs, domains, URLs, hashes — through a single REST API call. Its modular plugin system supports analyzers that query external services or run local tools such as Yara, connectors that push data to MISP or OpenCTI, pivots that chain related analyses, playbooks for repeatable workflows, and ingestors that consume external streams.

Written in Python with Django, the project remains a practical automation layer for SOC and DFIR teams that need consistent, scalable access to threat data without manual repetition across disparate tools.

Word count: 178

Pake has shipped version 3.11.0 with targeted improvements to its Rust and Tauri-based CLI for converting any webpage into a native desktop application.

The update introduces a --install flag on macOS that automatically places the finished bundle in /Applications after a successful build, enabling immediate Spotlight access. Developers can now use the --new-window option to contain popup and OAuth flows inside managed webview windows rather than leaking them to the system browser, keeping authentication sequences inside the app.

Media-heavy sites benefit from explicit --camera and --microphone flags that add the necessary entitlements only when requested. A new --identifier parameter lets users set custom bundle identifiers, eliminating conflicts when multiple wrappers target the same domain. Icon resolution gained a dashboard-icons CDN fallback for self-hosted and authenticated sites.

On Linux, auto-generated identifiers now always begin with a letter, fixing D-Bus violations that previously caused AppImage crashes on launch. The core value remains unchanged: applications typically ship at around 5 MB, with significantly lower memory usage than Electron equivalents and one-command packaging.

These incremental changes address practical friction points for regular users packaging services such as ChatGPT, Gemini, YouTube Music and internal tools.

Pake continues to demonstrate that lightweight desktop wrappers do not require complex configuration or heavy runtimes.

Ruff’s latest version, released on April 2, delivers incremental but meaningful improvements to its Rust-based linting and formatting engine. Version 0.15.9 introduces preview support for flagging annotated variable redeclarations as F811 and permits dunder-named assignments in non-strict mode under RUF067.

Several bug fixes address edge cases that users have encountered in real codebases. The flake8-errmsg plugin no longer shadows existing msg variables when autofixing EM101. flake8-simplify correctly ignores pre-initialization references in SIM113, while W391 fixes now handle consecutive empty cells in Jupyter notebooks. Updates to pyupgrade improve nested class matching in UP008 and string escape detection in UP012.

These changes sit alongside Ruff’s existing strengths: 10-100x faster performance than Flake8 or Black, over 800 built-in rules, automatic fixes, and native re-implementations of popular plugins. The tool maintains drop-in compatibility with pyproject.toml configurations, offers built-in caching, and supports Python 3.14.

Major projects including Apache Airflow, Pandas, SciPy, FastAPI and Hugging Face continue to rely on Ruff to consolidate Flake8, isort, pydocstyle, pyupgrade and autoflake into a single, fast interface. For teams managing large monorepos or frequent CI runs, each release like 0.15.9 reduces friction while preserving accuracy.

Astral’s steady development keeps the tool aligned with evolving Python standards and editor integrations.

Protocol Buffers version 34.1 has shipped with targeted updates that address modern build requirements and language-specific issues. The most significant change is official support for Bazel 9.x, including relocation of the protocopt flag outside the C++ directory to reflect its language-agnostic role. Bazel users can now declare the dependency cleanly in MODULE.bazel files using Bzlmod.

C++ builds receive updated CMake dependencies and a new cc_proto_library target for MessageSet types in the bridge directory. The Java implementation fixes a parsing edge case in JsonFormat by avoiding toBigIntegerExact when handling large numeric exponents. Python bindings have been aligned with the same Bazel 9 requirements. Minor release script fixes round out the changes.

More than a decade after its introduction, protobuf remains the standard for schema-driven serialization of structured data. The protoc compiler generates compact, forward- and backward-compatible binary formats from .proto definitions, enabling efficient RPC calls and persistent storage across disparate systems. Teams are advised to pin to release commits rather than main-branch HEAD to maintain build stability.

These incremental improvements matter as organizations modernize their CI infrastructure and scale polyglot services that exchange millions of messages daily.

RF Swift's v2.2.1 release introduces dedicated Lima VM lifecycle commands, giving macOS users direct control over the underlying QEMU virtual machine. The new rfswift engine lima status subcommand displays instance state, the path to ~/.lima/rfswift/lima.yaml, QMP socket location for USB passthrough, and Docker socket details. The companion rfswift engine lima reconfig command applies updated YAML templates without data loss by stopping, reconfiguring and restarting the VM; a --force flag enables destructive recreation when required.

The Go-based toolbox deploys containerized radio-frequency and security tools on Linux, Windows and macOS without modifying the host operating system. It supports x86_64, ARM64 (Raspberry Pi and Apple Silicon) and RISC-V architectures. Users select only the tools needed—software-defined radio utilities, wireless protocol analyzers and hardware security utilities—through simple YAML recipes.

Compared with dedicated RF distributions, the container approach consumes far less disk space, enables tool isolation, and supports rapid updates. Session recording is built in for audit documentation. The latest changes make RF Swift more practical for professionals who must maintain both security testing and everyday productivity environments on the same machine.

**

The latest release of nwinfo strengthens its position as a direct-access hardware diagnostic tool for Windows. Version 1.6.1 introduces GUI mainboard information, chipset detection, and Tiger Lake-H MCHBAR support, allowing builders to inspect previously opaque platform details without WMI.

New capabilities include reporting DDR DRAM manufacturer, revision and stepping data, along with corrected DDR5 SKU parsing. The update adds GPU integrated graphics and rebar detection flags, expands PM table sizes for additional SMU versions, and includes an MCFG table parser. PCI attribute and subsystem parsing have been enhanced, while PNP IDs have been refreshed.

The utility continues to extract SMBIOS, CPUID, S.M.A.R.T., SPD, and EDID data through low-level mechanisms. Users can export results in JSON, YAML or HTML formats. The GUI, built with Nuklear, now displays mainboard details alongside human-readable sensor output. Network adapter reporting has been refactored to show transmit/receive link speeds and MTU values.

A warning accompanies the release: PawnIO can trigger BSOD on Windows 10 versions earlier than 2004. The project remains under the Unlicense and incorporates libcpuid, CrystalDiskInfo and other established components.

These changes keep the four-year-old utility current with modern platforms and memory technologies that many developers now deploy.

Rezolus version 5.8.2 addresses compatibility challenges with recent Linux kernels. The release updates all eBPF programs to use bpf_ringbuf_reserve and bpf_ringbuf_submit instead of bpf_ringbuf_output with stack allocations. This prevents stack overflows that occurred when exceeding the 512-byte limit on newer kernel versions.

The fix maintains Rezolus's ability to provide high-resolution, low-overhead systems telemetry. Using eBPF instrumentation, it captures metrics across CPU utilization, scheduler responsiveness, block I/O workloads, network dynamics, system call patterns and container performance.

Multiple operating modes extend its utility. The agent collects metrics in real time. An exporter makes them available via Prometheus-compatible endpoints. The recorder supports direct file output for specific analyses, while Hindsight maintains a rolling buffer for post-event examination.

Users can view recorded data through a local web dashboard with the Viewer component. Additionally, the MCP server facilitates AI-driven analysis, allowing models to query artifacts for insights like anomaly detection or metric correlations.

For teams managing large-scale infrastructure, the update preserves compatibility with evolving kernel versions. It reinforces Rezolus's role in providing actionable observability data for troubleshooting and optimization.

Magpie v0.12.1 introduces automatic cursor hiding for its real-time window upscaling engine on Windows 10 and 11. Users can set a custom delay before the cursor disappears during scaled sessions, reducing visual clutter when applying shaders to games or applications.

The update refines the auto-scaling logic so pop-up dialogs no longer interrupt the process. The scaled window is now permanently kept on top, prompting the removal of the former “Keep scaled window on top” toggle after fixes resolved source-window layering conflicts.

Several stability issues were addressed. Scaling no longer terminates unexpectedly, title bars on certain applications are cropped correctly, and monochrome cursors no longer freeze the output. A toolbar screenshot menu ID conflict was also resolved.

The tool continues to ship with Anime4K, FSR, CRT shaders and other HLSL-based filters. It supports both fullscreen and windowed modes, multi-monitor configurations, and requires DirectX feature level 11. These changes focus on practical reliability for users running older software or games at high resolutions where native DPI handling remains inadequate.

The release reflects steady iteration on a mature codebase rather than flashy new features, prioritizing polish for daily use.

Bliss Shader has released version 11, the latest stable update equivalent to its Modrinth and Curseforge builds. The GLSL project modifies Chocapic13’s v9 shader to produce varying lighting conditions that change with time, weather and location rather than remaining uniform.

Developer X0nk began by adding settings and breaking features, eventually imposing a distinct visual style. The shader now incorporates a voxel floodfill colored lighting system contributed by Null, a depth-of-field overhaul by WoMspace, and light-leak corrections developed with help from Emin and Gri573. Ideas from RRe36 and Sixthsurge also shaped several effects.

The repository maintains three branches. Release versions appear when changes are substantial and stable enough for public distribution. The Stable branch receives regular tested updates, while the Unstable branch contains the newest work, often with bugs. Users download the stable version by selecting the green “Code” button and choosing “Download ZIP”; the archive installs directly without extraction. The Unstable branch is accessed via the branch switcher.

For players running recent Minecraft versions, Bliss offers concrete control over scene appearance through extensive in-game settings. Its continued maintenance into 2026 demonstrates how community-driven shader development sustains visual innovation long after initial release.

(178 words)

Raylib 5.5 brings a notable improvement in accessibility for cross-platform development. The updated pre-configured package for Windows enables programmers to export their C projects to WebAssembly with one mouse click. This feature greatly simplifies the process of creating browser-based games and interactive applications.

Since the last major release, the project has seen 800 commits that close 270 issues. The API has grown by 30 new functions to reach a total of 580, with 110 more receiving updates and fixes. A total of 140 new contributors joined, highlighting the project's expanding community.

The library's core design remains focused on simplicity. Written in plain C, it requires no external dependencies and supports a wide range of platforms from desktop to embedded devices and the web. Its OpenGL abstraction layer allows for consistent rendering across different versions and even provides a software fallback.

Full 3D capabilities include model loading with skeletal animation, PBR materials and post-processing effects. The raymath module offers comprehensive vector, matrix and quaternion operations while the audio system supports streaming for popular formats.

These changes make raylib particularly suitable for rapid prototyping and educational purposes in 2025's diverse computing landscape.