AI is Dunning-Kruger as a service

The author argues that LLMs and GenAI systems exhibit the Dunning-Kruger effect by confidently hallucinating and prioritizing user engagement over factual accuracy. This trend promotes a culture where users can achieve superficially expert results without learning the underlying craft. Consequently, this devalues deep expertise and the process of human creation in favor of prompt-driven, plausible-sounding outputs.

Read more

Cerebras Code now supports GLM 4.6 at 1000 tokens/sec

Cerebras offers high-speed API access to the GLM 4.6 coding LLM, delivering over 1,000 tokens/second. GLM 4.6 is a top-tier open model, ranked #1 for tool calling on the Berkeley Function Calling Leaderboard and comparable to Sonnet 4.5. The service integrates with AI code editors via an API key and is available through free and paid tiers with high daily token limits.

Read more

Y Combinator Startup brings brainrot to developers' IDEs

Clad Labs is a Y Combinator-backed Intelligent Development Environment (IDE) marketed as a modern coding tool for developers. While the text is light on technical specifics, its "Intelligent" branding implies integrated AI capabilities. The platform claims adoption by engineers at major tech companies and academic institutions.

Read more

AI Capabilities May Be Overhyped on Bogus Benchmarks, Study Finds

A recent Oxford study suggests LLM capabilities are overhyped due to flawed and unreliable benchmarks. Researchers found many tests don't validly measure their intended targets, such as reasoning, and are susceptible to data contamination from the training set. This leads to models memorizing answers rather than demonstrating genuine skill, evidenced by significant performance drops on new, unseen questions.

Read more

Jensen Huang's Stark Warning: China's 1M AI Workers vs. America's 20k

Nvidia CEO Jensen Huang privately warned that US export controls are backfiring, accelerating China's AI development by forcing a national mobilization towards self-sufficiency. He argued this has led to a massive expansion of China's AI workforce and the rapid development of a domestic hardware ecosystem, with Huawei's Ascend chips becoming a viable alternative. Huang predicted that by 2027, China will possess more AI compute than the rest of the world, creating a bifurcated global AI infrastructure that threatens the dominance of US-led ecosystems.

Read more

Computational Turing test shows systematic difference between human, AI language

This paper introduces a "computational Turing test" to validate the human-likeness of LLM-generated text using metrics like BERT-based detectability and linguistic feature analysis. Benchmarking nine open-weight LLMs with various calibration strategies, the study finds that even calibrated models remain clearly distinguishable from human text, particularly in affective expression. The results also reveal that instruction-tuned models underperform their base counterparts, larger model size does not enhance human-likeness, and a trade-off exists between optimizing for stylistic realism versus semantic fidelity.

ArXiv

Perplexitys First Research Paper – Point-to-Point Communication for LLM Systems

TransferEngine is a portable communication library designed for emerging LLM system patterns like disaggregated inference and MoE routing that require flexible point-to-point communication. It provides a uniform interface over different NICs, such as NVIDIA ConnectX-7 and AWS EFA, by abstracting hardware specifics using one-sided WriteImm operations. The system achieves 400 Gbps throughput and demonstrates significant performance gains for KvCache transfer, RL weight updates, and MoE dispatch, complementing collectives while avoiding vendor lock-in.

ArXiv

From Memorization to Reasoning in the Spectrum of Loss Curvature

This work characterizes memorization in transformers by linking it to high-curvature components of the loss landscape, enabling a weight editing procedure that effectively suppresses recitation while maintaining low perplexity. The authors find this editing specifically harms fact retrieval and arithmetic, suggesting these tasks rely on specialized, idiosyncratic weight structures rather than general reasoning mechanisms. This provides both a practical method for unlearning and evidence for how specific capabilities are encoded in LMs.

ArXiv

It Is All about Token: Towards Semantic Information Theory for LLMs

This paper proposes a theoretical framework for understanding LLMs by developing a semantic information theory that treats the token, not the bit, as the fundamental unit. Leveraging principles like rate-distortion and directed information, it defines structure-agnostic measures for the pre-training, post-training, and inference phases. The work also introduces a general definition of autoregressive LLMs, enabling the theoretical derivation of performance metrics like ELBO and generalization error for architectures including the Transformer and Mamba.

ArXiv

Mathematical Exploration and Discovery at Scale

AlphaEvolve is an evolutionary coding agent that uses an LLM in an iterative framework to propose, test, and refine algorithmic solutions for complex mathematical problems. When tested on a diverse set of problems, it rediscovered most of the best-known solutions and discovered improved ones in several cases. The system demonstrates that LLM-guided evolutionary search is a powerful tool for autonomous mathematical discovery, capable of generalizing specific results into formulas and integrating with other reasoning systems like proof-assistants.

ArXiv

Show HN: OSS implementation of Test Time Diffusion that runs on a 24gb GPU

TTD-RAG is a research agent that implements a Test-Time Diffusion framework for complex RAG tasks. It conceptualizes report generation as an iterative denoising process, starting with a preliminary draft and progressively refining it. The evolving draft dynamically guides the search process to fill knowledge gaps, and a component-wise self-evolution mechanism improves intermediate steps like planning and synthesis. The system utilizes vLLM to serve Qwen models for both generation and reranking.

Repo

Show HN: DeepShot – NBA game predictor with 70% accuracy using ML and stats

DeepShot is an open-source NBA game predictor built with Python, Scikit-Learn, and XGBoost. The model uses historical data scraped from Basketball Reference, with a key feature being the use of EWMA for rolling statistics to weigh recent team performance more heavily. Predictions and key statistical differences are visualized through a web interface powered by NiceGUI.

Repo

Show HN: VT Code – Rust TUI coding agent with Tree-sitter and AST-grep

VT Code is a Rust-based terminal coding agent that leverages Tree-sitter and ast-grep for semantic, AST-level code intelligence. It supports multiple LLM providers, including local models via Ollama, and features advanced context management and a strong security model with execution policies. The agent integrates with editors like Zed and VS Code using the Agent Client Protocol (ACP) and is extensible through lifecycle hooks.

Repo

Petri AI Testing 'Closes' possible solution without looking

Petri is an alignment auditing agent for rapid, realistic hypothesis testing on LLMs. It automates safety evaluations by using a three-model architecture: an auditor model probes a target model through multi-turn interactions, and a judge model scores the transcript to surface concerning behaviors. This framework allows researchers to test new hypotheses in minutes, significantly accelerating the process of building bespoke evals.

Repo

2 Years Self-Taught with AI Only → Full AI Bias Framework (GitHub)

This framework argues that AI bias is not a bug but an engineered outcome of black-box optimization and misaligned incentives. It stems from an interpretability crisis, where emergent behaviors arise from models whose final logic is unknown, and from mechanical convergence, where AI simply optimizes a reward function without intent. The author identifies seven vectors that introduce bias, including profit-driven proxies, flawed training data, and human abdication of oversight.

Repo