2026 tech layoffs reach 45,000 in March

Global tech layoffs reached 45,000 in March 2026, with approximately 20% directly linked to AI integration and organizational restructuring. Major firms like Block and WiseTech Global are citing the increased productivity of generative AI and LLMs in software engineering as a primary driver for workforce reductions. This shift marks a transition toward AI-first strategies, automating tasks ranging from code maintenance to customer service workflows across global tech hubs.

Read more

AI didn't simplify software engineering: It just made bad engineering easier

LLMs have significantly lowered the barrier to code generation, but they do not replace the core engineering disciplines of architecture, reasoning, and system alignment. The author argues that rapid AI-driven output often accelerates "spec drift," where implementation diverges from specifications and tests, ultimately increasing technical debt and unmanageable complexity. While LLMs are powerful tools for exploration and drafting, professional software development still requires human expertise to manage the lifecycle of complex systems and ensure long-term reliability.

Read more

GitAgent – An open standard that turns any Git repo into an AI agent

GitAgent is an open, framework-agnostic standard for defining and versioning AI agents natively within Git repositories. It treats agent configurations, prompts, and memory as code, enabling human-in-the-loop workflows via PRs and stateless compute with Git-based state persistence. The CLI supports exporting definitions to various runtimes, including Claude Code and OpenAI, while providing built-in tools for CI/CD and regulatory compliance auditing.

Read more

AI Gets Wrong Woman Jailed for Six Months, Life Ruined

A facial recognition system in South Dakota generated a false positive that led to the wrongful arrest and five-month incarceration of a woman 1,200 miles away in Tennessee. This case highlights the systemic risks of using biometric AI as a primary source for law enforcement without sufficient human-in-the-loop verification or corroborating evidence. It serves as a critical example of how algorithmic errors in large-scale datasets can bypass constitutional protections and result in severe real-world harm.

Read more

Postgres with Builtin File Systems

db9 is a serverless PostgreSQL database and cloud filesystem purpose-built for AI agents to manage memory, RAG context, and task artifacts. It features native vector search, auto-embeddings, and outbound HTTP calls directly from SQL, alongside environment branching and distributed cron jobs. The platform provides a CLI-first interface and built-in type generation, enabling agents to autonomously provision and query unified structured and unstructured data.

Read more

Can RL Improve Generalization of LLM Agents? An Empirical Study

A systematic study of RFT generalization for LLM agents reveals that while models scale well across task difficulty, cross-environment transfer is limited by shifts in semantic priors and observation/action interfaces. Sequential multi-environment training demonstrates strong downstream gains with minimal forgetting, while mixture training optimizes overall performance across diverse environments.

ArXiv

Researchers improve lower bounds for some Ramsey numbers using AlphaEvolve

Improved lower bounds for several classical Ramsey numbers, including R(3, 13) and R(4, 15), were achieved using AlphaEvolve, an LLM-based code mutation agent. This work highlights AlphaEvolve as a single meta-algorithm capable of generating search algorithms for these results, contrasting with traditional bespoke computational methods. The agent also successfully recovered known exact lower bounds and matched other best known bounds.

ArXiv

Terraforming Mars: Mass, Forcing, and Industrial Throughput Constraints

An evaluation of Mars terraforming feasibility indicates that human-relevant atmospheric pressures require exaton-class inventories. Accessible CO2 provides minimal warming ($\lesssim 10$ K), and achieving habitable temperatures demands significant IR opacity and multi-$10^2$ TW to PW-class power over centuries to millennia. While regional paraterraforming is plausible on near-term industrial scales, global transformation necessitates massive exogenous volatile supply or much larger discovered inventories, alongside sustained high-authority climate control.

ArXiv

Stellar engines and Dyson bubbles can be stable

Analysis of ultra-large space structures shows that passive stability for stellar engines and Dyson bubbles is contingent on non-uniform mass distributions or dense cloud configurations rather than uniform geometries. These stability parameters serve as critical constraints for modeling and identifying potential technosignatures in SETI research.

ArXiv

Benchmarking Language Modeling for Lossless Compression of Full-Fidelity Audio

Trilobyte is a byte-level tokenization schema that enables autoregressive LMs to perform lossless audio compression on high-resolution 16-bit and 24-bit waveforms. By reducing vocabulary scaling from $O(2^b)$ to $O(1)$, it achieves SOTA compression at 8-bit and 16-bit, consistently outperforming FLAC across diverse domains. While effective, compression gains relative to traditional codecs become more modest at higher bit depths.

ArXiv

Han – A Korean programming language written in Rust

Han is a statically-typed, general-purpose programming language written in Rust that uses Korean keywords and identifiers. It features a dual-execution model, providing a tree-walking interpreter for instant execution and a compiler backend that generates LLVM IR for native binary production via clang. The toolchain includes an LSP server and REPL, supporting modern language constructs such as closures, structs, pattern matching, and generics.

Repo

Claudetop – htop for Claude Code sessions (see your AI spend in real-time)

Claudetop is a real-time cost and token monitoring tool for Claude Code designed to provide granular visibility into session spending and cache efficiency. It tracks live burn rates, compares pricing across different models, and issues smart alerts for budget milestones or context window saturation. The tool includes session analytics, tagging for project-based cost tracking, and a plugin system to integrate Git status or CI/CD metrics directly into the terminal prompt.

Repo

Data-anim – Animate HTML with just data attributes

data-anim is a lightweight (<3KB gzipped) library for declarative HTML animations using data attributes, featuring 30+ animations and anti-FOUC protection. It provides a low-overhead solution for developers building interactive interfaces for AI and LLM projects.

Repo

AgentArmor – open-source 8-layer security framework for AI agents

AgentArmor is an open-source security framework providing 8-layer defense-in-depth for agentic AI applications, aligning with the OWASP Top 10 for Agentic Security. It secures the entire agentic lifecycle through features like prompt injection detection, action plan validation, and PII redaction. Recent updates include the OpenClaw Identity Guard for encrypted agent storage and an MCP Server Scanner for evaluating tool and server risks before connection.

Repo

OpenClaw-superpowers – Self-modifying skill library for OpenClaw agents

OpenClaw-superpowers is a plug-and-play skill library for OpenClaw agents, enabling them to be self-modifying and teach themselves new behaviors. Unlike session-based AI tools, agents can dynamically create and integrate new skills during conversation via create-skill, which become immediately live and persistent. This framework supports 24/7 operation, long-running tasks, native cron scheduling, and advanced capabilities like task handoff, memory persistence, and self-recovery. It also includes critical security skills for vetting new behaviors and guarding against prompt injection or dangerous actions.

Repo