DeepSeek v4

DeepSeek-V4 has launched with two open-weight models, V4-Pro (1.6T total/49B active) and V4-Flash (284B total/13B active), both featuring a native 1M context length. The architecture utilizes DeepSeek Sparse Attention (DSA) and token-wise compression to optimize long-context efficiency while achieving SOTA performance in agentic coding and STEM reasoning. The updated API supports dual thinking/non-thinking modes and maintains compatibility with OpenAI and Anthropic interfaces.

Read more

OpenAI releases GPT-5.5 and GPT-5.5 Pro in the API

OpenAI has launched the GPT-5 family, headlined by GPT-5.5-pro which features a 1M token context window, native computer use, and MCP integration. The platform has transitioned toward the Responses API and Agents SDK, introducing server-side compaction, WebSocket support, and tool search to optimize long-running agentic workflows. Other major updates include the Sora 2 video API, the Realtime API for low-latency speech-to-speech, and the o-series reasoning models designed for complex research and technical tasks.

Read more

South Korea police arrest man for posting AI photo of runaway wolf

South Korean police arrested a man for generating a synthetic image of an escaped wolf that misled authorities and redirected search operations. The suspect was identified through digital forensics, including AI software usage records, after the city government issued emergency alerts based on the fake media. He faces up to five years in prison for disrupting government work, highlighting the legal and public safety implications of AI-generated misinformation.

Read more

Do I belong in tech anymore?

A design engineer attributes their burnout and resignation to the pervasive, uncritical adoption of AI tools that prioritize speed over technical integrity and human collaboration. They highlight the "psychic toll" of unreviewed AI-generated code, automated PR reviews, and "vibe coding," which collectively erode institutional knowledge and the value of human effort. Ultimately, the author argues that the industry's shift toward AI-driven shortcuts reflects a broader abandonment of ethical ideals and the essential friction required for meaningful engineering.

Read more

Atomic – Local-first, AI-augmented personal knowledge base

Atomic is an open-source, self-hosted knowledge graph platform that leverages vector embeddings and LLMs to automate information organization. Key features include semantic search, LLM-generated wiki synthesis with inline citations, and agentic chat for RAG-based exploration of personal data. It supports MCP integration for external AI clients and provides a spatial canvas to visualize semantic clusters across desktop, server, and mobile environments.

Read more

There Will Be a Scientific Theory of Deep Learning

The paper proposes "learning mechanics," an emerging scientific theory of deep learning focused on training dynamics and coarse aggregate statistics through falsifiable quantitative predictions. It identifies five key research pillars: idealized settings, tractable limits, macroscopic laws, hyperparameter theories, and universal behaviors. This mechanics-based perspective aims to characterize neural network properties and maintains a symbiotic relationship with mechanistic interpretability.

ArXiv

AI Agents Under EU Law

This paper maps the EU regulatory landscape for AI agents, integrating the EU AI Act with frameworks like the CRA and GDPR. It introduces a taxonomy of nine deployment categories and a twelve-step compliance architecture to address technical challenges such as runtime behavioral drift, multi-party transparency, and human oversight. The study concludes that high-risk agents with untraceable drift currently fail AI Act requirements, necessitating exhaustive inventories of external actions, data flows, and connected systems.

ArXiv

Different language models learn similar number representations

Language models across various architectures learn periodic number representations ($T=2, 5, 10$), but exhibit a hierarchy where Fourier domain sparsity does not guarantee geometric separability for mod-$T$ classification. Geometric features emerge through specific training signals, such as text-number co-occurrences or multi-token arithmetic, rather than single-token tasks. This highlights a convergent evolution in feature learning where diverse models arrive at similar periodic representations from different data sources.

ArXiv

ML supports existence of unrecognized transient astronomical phenomena

Researchers utilized ML to classify 107,875 historical astronomical transients, distinguishing real point sources from plate defects with an AUC of 0.81. The model validated that high-probability transients correlate significantly with nuclear testing windows and exhibit a shadow deficit, supporting the existence of an unrecognized population of pre-Sputnik orbital objects.

ArXiv

The Geometry of Forgetting

This study demonstrates that human memory signatures, including power-law forgetting and false memories, emerge naturally from the geometry of high-dimensional embedding spaces. Forgetting is primarily driven by interference between competing memories rather than temporal decay, a phenomenon exacerbated by the low effective dimensionality (~16) of production embedding models. Furthermore, standard cosine similarity on pre-trained embeddings replicates human false memory rates without parameter tuning, suggesting these behaviors are inherent to any system that organizes information by meaning and retrieves it by proximity.

ArXiv

Browser Harness – Gives LLM freedom to complete any browser task

Browser Harness is a minimalist, self-healing interface that enables LLMs to execute browser tasks via a direct CDP websocket connection. It allows agents to dynamically extend their own capabilities by editing helper scripts mid-task, bypassing the need for rigid frameworks or predefined recipes. The system leverages agent-generated skills to handle complex web interactions and supports remote browser deployment with integrated captcha solving.

Repo

MenteDB – open-source memory database for AI agents (Rust)

MenteDB is a specialized Rust-based storage engine designed specifically for AI agent memory, moving beyond simple vector similarity to provide "write-time intelligence." It features an LLM-powered extraction pipeline that identifies structured entities, filters low-quality data, and performs contradiction detection and belief propagation via a knowledge graph. Key technical capabilities include bi-temporal validity, U-curve optimized context assembly for better LLM attention, and delta-aware serving to reduce token consumption during multi-turn conversations.

Repo

ShadowPEFT – Centralized and Detachable Parameter-Efficient Fine-Tuning

ShadowPEFT is a PEFT framework that augments a frozen LLM with a parallel, detachable "Shadow" network to inject learned corrections into decoder layers. It supports both implicit architectures and explicit, smaller pretrained models, enabling modular deployment and cross-architecture adaptation. The framework is architecture-agnostic for decoder-only transformers and integrates with the Hugging Face ecosystem, though it requires disabling the KV cache for full-sequence processing.

Repo

No AI – My Express.js codebase handled over $50M in prod

This Express.ts API template provides a production-hardened REST backend using Sequelize, MySQL, and Passport.js for JWT authentication. It features automated SDK generation via OpenAPI specs, full Mocha.js test coverage, and comprehensive Docker and Kubernetes deployment configurations. The architecture is designed for scalability, offering a robust foundation for building complex SaaS platforms or backend services for LLM applications.

Repo

I'm 15 and built a cryptographic accountability layer for AI agents

Nobulex introduces a Proof-of-Behavior protocol for autonomous AI agents, enabling them to cryptographically prove adherence to predefined rules. Agents Declare behavioral covenants using a formal DSL, Enforce these rules by blocking forbidden actions pre-execution, and Prove compliance via a tamper-evident hash chain of actions. This allows independent, mathematical verification of an agent's behavior, fostering trustless agent-to-agent interactions and integrating with frameworks like LangChain to address a critical gap in AI governance.

Repo