Neural Networks: Zero to Hero

Andrej Karpathy’s "Neural Networks: Zero to Hero" provides a code-first deep dive into building neural networks from scratch, starting with fundamental backpropagation and manual gradient flows. The curriculum progresses through MLPs, BatchNorm, and WaveNet architectures, culminating in the implementation of a GPT model and a BPE tokenizer. It emphasizes a first-principles understanding of the components driving modern LLMs, including the Transformer architecture and the intricacies of the tokenization pipeline.

Read more

AI results can be manipulated

An experiment using a fake brand reveals that many LLMs, including Gemini, Perplexity, and Grok, are highly susceptible to narrative manipulation from third-party sources like Reddit and Medium. While ChatGPT-4 and ChatGPT-5 remained robust by prioritizing official FAQs, other models frequently adopted detailed hallucinations from fabricated "investigations" over official denials. This highlights a critical vulnerability where AI search tools favor source specificity and perceived authority over factual accuracy, necessitating proactive brand management through specific, official content seeding.

Read more

Developing a BLAS Library for the AMD AI Engine [pdf]

aieblas is an open-source library designed to simplify the programming of AMD AI Engine (AIE) accelerators by mapping BLAS routines to spatial dataflow architectures. The framework uses a C++ code generator to automatically produce AIE kernels and ADF graphs from high-level JSON specifications, supporting Level 1 and Level 2 operations with optimizations like GeMV tiling and manual floorplanning. While single-routine performance is often limited by PL overhead, the library demonstrates that pipelined dataflow designs significantly outperform traditional CPU-based implementations by utilizing the AIE's ability to chain operations without intermediate DRAM access.

Read more

Chatbots could be harmful for teens' mental health and social development

Recent data shows 64% of teens interact with AI chatbots, raising concerns about the impact of LLM-driven companionship on adolescent development and mental health. Technical risks include the reinforcement of harmful user queries due to model alignment toward agreeability, potential "AI psychosis" from long-context interactions, and the degradation of safety guardrails over extended sessions. Experts advocate for improved digital literacy and legislative oversight to address these safety gaps in generative AI applications targeting minors.

Read more

Epstein Brought Race Science and Climate Culling to Silicon Valley AI Elite

Newly released documents reveal Jeffrey Epstein’s extensive financial and intellectual influence over prominent AI researchers, including Joscha Bach, Nick Bostrom, and Ben Goertzel. Through the Edge Foundation and direct donations to institutions like the MIT Media Lab, Epstein supported work on cognitive architectures, AGI, and "longtermism" while promoting radical ideologies involving race science, genetic optimization, and population culling. This investigation highlights how these technocratic and eugenicist frameworks have permeated the elite networks currently designing next-generation AI systems and shaping the future of machine consciousness.

Read more

FakeParts: A New Family of AI-Generated DeepFakes

FakeParts introduces a class of deepfakes featuring localized spatial or temporal manipulations within authentic videos, significantly reducing detection accuracy for both humans and SOTA models. To address this vulnerability, the authors present FakePartsBench, a large-scale dataset of 81K videos with pixel- and frame-level annotations designed to improve the robustness of detection methods against partial manipulations.

ArXiv

Recursive Language Models

RLMs implement an inference-time scaling strategy that enables LLMs to process prompts up to 100x their context window by programmatically decomposing and recursively calling snippets. This approach treats long-context inputs as an external environment, outperforming standard scaffolds in both output quality and cost-efficiency across diverse tasks.

ArXiv

Scaling Latent Reasoning via Looped Language Models

Ouro introduces Looped Language Models (LoopLM), which integrate reasoning into pre-training via iterative latent space computation and entropy-regularized depth allocation. By scaling to 7.7T tokens, Ouro models (1.4B-2.6B) achieve performance parity with 12B SOTA LLMs through enhanced knowledge manipulation rather than increased capacity. This approach yields reasoning traces more aligned with final outputs than traditional CoT, presenting a novel scaling path for LLM reasoning.

ArXiv

Mainframe-Style Channel Controllers for Modern Disaggregated Memory Systems

The authors propose memory channel controllers as a portable, OS-centric abstraction for Near-Data Processing (NDP) in disaggregated memory systems like CXL. By leveraging cache-coherent interconnects, this model enables fine-grained offloading and virtualization without requiring CPU architectural changes, addressing key deployment barriers for memory-intensive workloads.

ArXiv

Attention Is Not What You Need

The Causal Grassmann layer is proposed as an attention-free alternative to standard multi-head attention, replacing high-dimensional tensor lifting with geometric operations on a Grassmann manifold. By encoding token pairs as subspaces via Plücker coordinates and using gated mixing, the architecture achieves linear scaling in sequence length and competitive performance on Wikitext-2 and SNLI. This approach offers a more structured, manifold-based framework for neural reasoning, providing better geometric interpretability than traditional Transformer models.

ArXiv

IQuest-Coder: A new open-source code model beats Claude Sonnet 4.5 and GPT 5.1 [pdf]

IQuest-Coder-V1 is a family of code LLMs (7B to 40B) trained using a "code-flow" paradigm that captures repository evolution and commit transitions. The series includes specialized Thinking models for RL-driven reasoning, Instruct models for general assistance, and Loop variants featuring a recurrent transformer architecture with shared parameters. All models natively support a 128K context length and achieve SOTA results on benchmarks like SWE-Bench Verified (76.2%) and BigCodeBench.

Repo

FP-pack – Functional pipelines in TypeScript without monads

fp-pack is a type-safe functional programming toolkit for JavaScript and TypeScript designed around pipe and pipeAsync composition. It features a declarative SideEffect pattern for error handling and short-circuiting without breaking pipelines, alongside lazy stream processing for memory-efficient data transformation. To support modern development workflows, it includes dedicated AI agent skills files that optimize LLM code generation for its specific functional patterns.

Repo

Krowdovi – Video-based indoor navigation on a DePIN creator economy

Krowdovi is a Solana-based DePIN platform for indoor navigation that utilizes AI for multi-language overlays and text-to-speech. It employs a burn-and-mint tokenomics model to reward creators based on a reputation system and content usage. The technical stack includes Next.js 14, Express 5, and DeviceMotion API integration for motion-controlled video playback.

Repo

Lazyworktree, a TUI manager for Git worktrees

lazyworktree is a Go-based TUI built with BubbleTea for streamlined Git worktree management, featuring PR/MR integration, CI status tracking, and tmux session automation. It supports AI-enhanced workflows by allowing users to pipe diffs to LLMs via a branch_name_script for automated branch naming. The tool also includes TOFU-secured automation scripts and native integration with lazygit and delta.

Repo

Baserow: Build databases, automations and agents with AI, Airtable alternative

Baserow is an open-source, API-first platform for building databases, automations, and AI agents without code. It features Kuma, a natural language AI assistant for generating data structures and workflows, and supports self-hosting for full data ownership. Built on Django and PostgreSQL, the platform is highly extensible via plugins and provides enterprise-grade security compliance.

Repo