Rich Hickey: Thanks AI

Rich Hickey's "Thanks AI!" is a scathing critique of modern AI and LLMs, arguing they pirate creative works, degrade education, and increase environmental impact. He contends that AI wastes developer time, eliminates entry-level jobs, and diminishes customer support and search quality by filling the internet with "BS." Hickey concludes that AI misleads executives and ultimately floods human communication channels, making interactions suspect.

Read more

AI is forcing us to write good code

Effective development with AI agents and LLMs demands rigorous adherence to "good code" principles, which were often optional for human developers. This includes enforcing 100% code coverage to ensure agents validate their work, structuring codebases with thoughtful, small files for optimal LLM context loading, and providing fast, isolated dev environments for rapid iteration. Furthermore, leveraging automated linters and strong static typing with semantic names is essential to constrain the LLM's search space and enforce code correctness.

Read more

AI employees don't pay taxes

The author contends that the widespread displacement of human workers by AI, including LLMs, poses a critical threat to government tax revenue, which is predominantly derived from human income. This erosion of the tax base could lead to societal collapse, as public services and infrastructure rely on these funds. The text refutes common counter-arguments, such as corporate taxes covering the gap or AI being untaxable, suggesting instead that value flow or energy consumption could be taxed. Ultimately, it advocates for human-in-the-loop systems to preserve the tax base and ensure societal stability, rather than a complete human-to-AI replacement.

Read more

Show HN: Evidex – AI Clinical Search (RAG over PubMed/OpenAlex and SOAP Notes)

Evidex is an AI-powered platform providing medical evidence search capabilities for clinicians.

Read more

Are We Ready to Be Governed by Artificial Intelligence?

AI, including LLMs, is already incrementally governing democratic processes across executive, judicial, and legislative branches. In the executive, AI is used for administrative functions like healthcare coverage decisions, raising ethical concerns about care denial. Judges are leveraging AI to interpret laws and assist in rulings, often without disclosure. Legislatures are employing AI for drafting laws and engaging constituents. The authors emphasize that AI is a power-enhancing technology, and the critical challenge is ensuring its application decentralizes power and serves democratic principles rather than concentrating authority or enabling authoritarianism.

Read more

Deep sequence models tend to memorize geometrically; it is unclear why

Contrasting the associative view of memory in sequence models, this work proposes a geometric view, demonstrating that Transformers synthesize a geometry of atomic facts encoding global relationships, even for non-co-occurring entities, which simplifies complex reasoning. The emergence of this geometry, despite not being driven by typical architectural or optimization pressures, is attributed to a spectral bias, shown via a connection to Node2Vec. This understanding offers avenues to enhance Transformer memory's geometric properties.

ArXiv

Information, complexity, brains and reality (Kolmogorov Manifesto) (2007)

This essay unifies information theory, physics, and neuroscience through the lens of Kolmogorov Complexity. It posits that cognition involves finding algorithms that approach the Solomonoff-Kolmogorov-Chaitin Complexity limit with appropriate tradeoffs. Furthermore, it argues that the universe is an interpreted abstraction—a mental construct based on coherent sensory inputs and interactions—making it a model itself.

ArXiv

Propose, Solve, Verify: Self-Play Through Formal Verification

Addressing the challenge of self-play training for LLMs in code generation, where unit test rewards are unreliable, this work introduces Propose, Solve, Verify (PSV). PSV is a self-play framework that leverages formal verification for robust correctness signals to generate challenging synthetic problems and train a solver via expert iteration. PSV-Verus, trained with PSV, achieves up to 9.6x improvement in pass@1 over inference-only and expert-iteration baselines on three benchmarks, with performance scaling with generated data and iterations, and formal verification and difficulty-aware proposal identified as essential components.

ArXiv

LLMs, LoRA, and Slerp Shape Representational Geometry of Embeddings

This research compares the generalization properties of dense text embeddings from LLM versus non-LLM backbones, and the efficacy of SLERP model-merging in mitigating LoRA-induced over-specialization. Experiments with numerical sequences show LLM-based backbones capture higher-order compositional patterns but are prone to adapter dominance. SLERP merging consistently recovers base-model structure and task gains, yielding superior clustering separability and robustness compared to model souping or unmerged models.

ArXiv

Optimal Software Pipelining and Warp Specialization for Tensor Core GPUs

Optimally leveraging complex GPU architectures, especially with transformations like SWP and WS, is challenging due to reliance on brittle heuristics. This work introduces a novel joint optimization formulation for SWP and WS, solvable by off-the-shelf constraint solvers. The resulting system, Twill, automatically derives optimal, heuristic-free schedules for iterative programs, demonstrating its ability to rediscover and prove optimal Flash Attention schedules on NVIDIA Hopper and Blackwell GPUs.

ArXiv

Show HN: Z80-μLM, a 'Conversational AI' That Fits in 40KB

Z80-μLM is a micro LLM designed for retrocomputing, specifically a Z80 processor with 64KB RAM, achieving a ~40KB .COM binary size. It utilizes quantization-aware training (QAT) with 2-bit weights and 16-bit integer inference for character-by-character autoregressive generation. Input is processed via trigram hash encoding, making it typo-tolerant and word-order invariant for short inputs. While not a general intelligence, it provides personality-driven, terse responses, demonstrating a functional LLM on highly constrained 8-bit hardware.

Repo

Show HN: Kuack – Run Kubernetes jobs in visitor browsers

Kuack is a Virtual Kubelet provider that transforms web browsers into Kubernetes worker nodes, enabling distributed edge computing for WebAssembly (WASM) workloads. It aggregates CPU, memory, and GPU from connected browser agents, allowing K8s to schedule stateless, CPU-intensive jobs to millions of idle devices. This system executes WASM binaries directly in the browser, with results and logs streaming back to K8s.

Repo

Show HN: Interactive plan annotation and sharing for Claude Code

Plannotator is an interactive plan review tool designed to integrate with Claude Code's plan mode. It provides a UI for users to visually annotate and refine AI-generated plans. The system intercepts Claude Code's ExitPlanMode, allowing users to review and mark up plans, then either approve them for implementation or send annotations back to Claude for further changes.

Repo

Show HN: Operon – Reliable Agents Using Biological Motifs and Category Theory

Operon is a research library applying biologically-inspired architectures and category theory to create robust LLM agentic systems. It addresses agent fragility, such as prompt injection and runaway recursion, by leveraging "network motifs" and system topology for safety, rather than solely relying on prompt engineering. Key "organelles" provide functions like adaptive threat defense (Membrane), safe tool execution (Mitochondria), output validation (Chaperone), and LLM integration (Nucleus). The framework also includes biologically-inspired state management, higher-order network topologies (e.g., CFFL, Quorum Sensing), and WAgent for typed wiring validation, ensuring safety emerges from structural design.

Repo

PowerMem – Persistent memory layer for AI agents

PowerMem is an intelligent memory system designed to enable LLMs to persistently remember historical conversations, user preferences, and contextual information. It utilizes a hybrid storage architecture combining vector retrieval, full-text search, and graph databases, integrating the Ebbinghaus forgetting curve theory for dynamic memory management. The system demonstrates significant improvements in accuracy, response speed, and token reduction compared to full-context methods. Key features include LLM-based intelligent memory extraction, multimodal support, multi-agent memory isolation and collaboration, user profile management, and optimized data storage with knowledge graph capabilities.

Repo