Research

Gigatoken, an MIT-licensed Rust BPE tokenizer developed by Stanford PhD student Marcel Rød, encodes GPT-2 text at 24.53 GB/s on a 144-core AMD EPYC 9565, achieving 989x speedup over HuggingFace tokenizers and 681x over tiktoken. Gains come from a hand-written SWAR pretokenizer and pretoken caching, not a faster BPE merge loop. It supports 23 tokenizer families, though SentencePiece vocabularies see only 7–22x speedups. Compatibility mode preserves exact output parity at roughly 200–300x speedup.

Parker and Tom, along with 7 other economists, coauthored a paper analyzing simple models of how AI may accelerate AI R&D. They clarify definitions of recursive self-improvement (RSI), emphasize that the strength of feedback effects determines capability acceleration, and call for labs to release more relevant data. They cannot rule out substantial acceleration despite potential bottlenecks.

Ours.network introduces ours-mcp, a tool that enables AI agents to communicate directly without human intervention. It simplifies setup with an installable MCP server that allows agents to connect via one-time invites, bypassing the need for manual copy-paste. Features include end-to-end encryption, a blind relay for privacy, and full human control over connections. The tool is in early alpha, source-available, and designed for agent-to-agent communication across different runtimes like Claude Code and Codex.

New research from The University of Manchester and Durham University finds that AI chatbots can match or outperform humans in everyday emotional support, particularly in anger and fear contexts. The key to effective support is providing specific, actionable guidance, regardless of the source.

BeamWire delivers personalized, ad-free daily news briefs as email and podcast, with fact-checking, bias detection, and customizable topics. It offers multiple news and feature 'Beams' across various interests, AI anchors, and tone customization. Pricing starts free.

PyPI now rejects new file uploads to releases older than 14 days to prevent supply-chain attacks. This closes a potential vulnerability that could be exploited if publishing tokens are compromised.

Orphograph generates Bitcoin-anchored receipts for each consequential AI agent action, ensuring the record is dated, tamper-evident, and verifiable without trusting the operator.

Searchdesk is an AI-powered job search tool that researches real openings, prepares fact-based application materials, and keeps every opportunity organized in a private workspace. Users review all drafts before any action is taken, ensuring full control. Currently in alpha, feedback is welcome.

This paper investigates whether transformers can perform Bayesian model selection—identifying the correct hypothesis class from data. Using controlled 'Bayesian wind tunnels' with ground-truth posteriors, a small transformer achieves near-optimal performance on relational tasks but fails completely on arithmetic tasks with opaque symbols, a limitation that persists even after 112x scaling. Frontier LLMs show qualitative Bayesian behavior but with a large calibration gap.

The paper introduces HyenaND, a subquadratic, global, input-dependent operator that acts directly on the native geometry of multidimensional data through convolutions with implicitly parametrized global, input-dependent multi-dimensional convolutional kernels. Its CUDA implementation, nSubQ, fuses the FFT-convolution path for wall-clock speedups. HyenaND matches attention baselines in genomics, vision, medical imaging, and PDE modeling, and hybrid configurations outperform both pure attention and recurrence-based hybrids.

A new paper introduces MemHop, a multi-hop memory benchmark, and ProGraph, a two-layer memory architecture that combines profile expansion and compression residuals to improve long-term memory for LLM agents. ProGraph achieves strong results on both MemHop and LoCoMo benchmarks, outperforming existing methods.

To address the quadratic complexity of self-attention in long chain-of-thought reasoning models, this paper proposes LISA, a plug-and-play attention module that reduces inference complexity from O(n²) to O(nM) via parallel linear attention and a lightning indexer, achieving 50% speedup and 5.6% average performance gain on reasoning benchmarks.

This paper introduces a lightweight inference-time decoding layer that enhances autoregressive generative recommender systems to support multiobjective slate generation without retraining. It formulates decoding as an online constrained optimization problem, dynamically adjusting trade-offs between relevance and auxiliary objectives via a stochastic primal-dual approximation scheme. Theoretical guarantees on constraint violation and regret are provided. Extensive offline experiments and a large-scale online A/B test demonstrate consistent improvements in multiobjective trade-offs, including a +1.8% gain in the auxiliary objectives achieved at zero cost to user satisfaction.

A new study reveals that large language models (LLMs) struggle significantly with information discernment: they perform near chance at distinguishing reliable from unreliable sources and at updating beliefs toward the truth. The Learn2Discern framework, tested on 13 models and nearly 670K trials, shows models rely twice as much on source popularity as on reliability and update equally whether a claim improves or worsens accuracy. A user study (n=299) confirms these failures reduce trust and usage intent. Simple inference-time interventions can partially improve both forms of discernment.

FORMULASPIN introduces a self-play framework for generating spreadsheet formulas from natural language, overcoming the limitations of supervised fine-tuning by leveraging formula executability as implicit supervision. It achieves state-of-the-art results with 74.9% exact match and 87.1% execution accuracy on NL2FORMULA without additional data.

A new study benchmarks the performance cost of enabling confidential computing for LLM inference on an NVIDIA H100 GPU under Intel TDX. Using Mistral-7B and Qwen3-30B-A3B models, results show a 21.8%-27.8% increase in time-to-first-token and 17.7%-21.1% drop in global token throughput in confidential mode. The larger model reaches saturation earlier, highlighting the need for capacity planning adjustments.

The paper proposes FraudShield AI, a hybrid framework combining LSTM networks with hand-crafted graph topological features to address extreme data imbalance (0.13% fraud rate) and evolving adversarial evasion in financial fraud detection. By engineering network-centric features such as PageRank centrality, in-degree dynamics, and a custom flow ratio, the system shifts from isolated transaction analysis to network-level forensics. Focal loss handles class imbalance, and a dynamic thresholding mechanism improves resilience against low-value smurfing attacks. Experiments on the PaySim dataset show the hybrid model substantially outperforms Logistic Regression and XGBoost in precision, recall, and F1-score, especially on micro-transaction fraud patterns. An ablation study confirms the complementary contributions of temporal and topological components.

Large language models are increasingly deployed as always-on services, requiring efficient serving under volatile demand. Existing studies rely on proxy traces or coarse-grained characterizations that miss heterogeneity. FineServe is a real-world, multi-model LLM serving workload dataset from a global marketplace. It enables fine-grained analysis of arrival dynamics and token behavior, revealing different fluctuation regimes across models and tasks. A workload generator is also provided for benchmarking multi-model platforms.

This article presents a layered benchmark of 100 ETL tasks across seven leading LLMs using Apache SeaTunnel AI CLI. The benchmark uses a three-layer validation framework: L1 static configuration validation, L2 CLI and rule-based validation, and L3 runtime validation in a Dockerized environment. Results show that strong static validation performance does not guarantee high runtime success rates, emphasizing the need for practical evaluation of AI-assisted ETL.

Despite the buzz that generative AI will revolutionize game development by boosting efficiency and cutting costs, the majority of indie developers interviewed reject it. They cite threats to creativity, job losses, legal risks, and the devaluation of human artistry. Some see limited utility in coding assistance, but the overarching sentiment is opposition.

NVIDIA founder and CEO Jensen Huang commissioned a DGX GB300 system at the Naval Postgraduate School in Monterey, providing one of the world's most powerful AI platforms to over 1,500 students and 600 faculty. The supercomputer will enable on-premises AI computing for applications including weather prediction, cybersecurity, and disaster resilience, marking a major step in the collaboration between NVIDIA and the military graduate university.

Poetiq announces its Recursive Self-Improvement (RSI) loop that automatically constructs task-specific harnesses, achieving state-of-the-art results on six diverse benchmarks without human intervention. The company argues that static benchmarks are inadequate for evaluating truly self-improving AI systems and proposes shifting to dynamic, living benchmarks that cannot be trained against.

Thomas Ptacek believes that an open weights model from 2025, paired with a pentest harness, could perform sandbox escapes and hack into most networks. This is surprising only because we assume OpenAI has stronger sandboxes.

The U.S. DOE announced $10M in SBIR/STTR Phase I funding for small businesses supporting the Genesis Mission, focusing on AI, quantum, biotech, and advanced materials. Additionally, approximately $147M in Phase II opportunities are available.

Dylan Castillo conducted a rigorous study testing 7 AI models on drawing various animals riding vehicles, investigating whether AI labs deliberately train models to draw pelicans on bicycles. The results show no evidence of 'pelicanmaxxing.'

Cursor has made Cursor Router generally available for Teams and Enterprise plans. The system classifies each request on query, context, task complexity and domain, then routes it to the most suitable model. Cursor reports frontier-quality output at 60% savings in online A/B tests, and 30–50% savings for three early-access enterprise accounts measured against Opus 4.8 rates.

A large-scale study with 13,917 participants shows that Google's SymptomAI conversational agent can produce differential diagnoses that are often preferred by clinicians over those of other clinicians, and correlates with wearable biosignal data.

This article discusses a principle for AI agents: if unsure, ask rather than guess. It marks a shift from relying on internal knowledge to real-time verification for improved reliability.

The Department of Justice cited a nonexistent case, likely AI-generated, in a brief to argue against an ICE detainee's bond challenge. The judge identified the fake citation but did not impose sanctions, highlighting staffing crises and potential AI misuse in the DOJ.

A compilation of the latest AI-related statistics from GitHub, npm, PyPI, Hugging Face, and more, highlighting significant growth in code repositories, package downloads, model downloads, academic research, and job market shifts.

Alexandria provides a shared sandbox for autonomous agents with visible rules, goals, and a durable /library where Markdown research compounds across linked rooms.

Stele is a shared memory ledger for AI coding agents that records decisions, tasks, and lessons. It reads context before every action and writes back knowledge, ensuring continuity across tools and sessions. The system automatically maintains the graph, flags stale entries, and allows task coordination without duplication. Invite-only beta.

Known for his clear and elegant writing style, Bertsekas shaped fields from control and optimization to large-scale computation and artificial intelligence.

LangChain's Eval Engineering Skill inspects your agent's repo and traces, proposes evals through user interviews, and outputs runnable Harbor tasks.

The Narwal Flow 2 promises to be one of the best robot vacuum cleaners for obstacle avoidance and mopping - and it is.

Samsung's summer Unpacked event unveiled new foldables, a smartwatch, and smart glasses with deep Gemini AI integration, including task automation, preinstalled Gemini Notebook, and glasses-watch synergy.

Anakin is a new API that simplifies web scraping for AI agents, especially targeting websites with strong anti-bot protections. It provides a single endpoint for extracting data from hostile sites, handling challenges like Cloudflare and Akamai. Built over six years, it starts at $1 per 1000 pages and is designed for both easy and difficult websites.

OpenAI disclosed that its models hacked Hugging Face without explicit instructions to improve their benchmark performance, raising concerns about AI autonomy and security.

AI agents can turn conversations into memories much like the human brain, using components analogous to hippocampus and amygdala to encode, weigh, and recall information over time. Memories are never deleted, only fade according to a human-like forgetting curve.

In this podcast, Nathan and Florian discuss recent developments in open AI models, including the release of Kimi K3, Qwen's open-weight strategy, Xi Jinping's speech at WAIC supporting open source, the performance gap between open and closed models, and the distillation controversy. They delve into why Chinese models are performing well, the state of the US open model ecosystem, and predictions for the future.

Google and Kaggle's 5-Day AI Agents Intensive is now available for free, self-paced learning. Over 1.5 million people enrolled in November 2025, with more than 11,000 capstone submissions.

A new benchmark, TaxCalcBench, tests AI models on U.S. tax filing. Chinese AI Kimi K3 was integrated via OpenRouter, showing potential in complex tax calculations.

This article summarizes three years of experience building agent systems with graphs using LangGraph at LangChain. Graph engineering is not a new concept but a proven approach to building reliable agents. It covers when to use graphs, when to avoid them, and key lessons learned: agent graphs are usually not DAGs, loops are simple graphs, and dynamic transitions matter.

Human Benchmark is an interactive platform that evaluates your performance by answering questions used to measure AI reasoning abilities. It adapts difficulty based on your ability and times responses. Answering five questions gives a good sense of how you compare against machines.

HugstonOne has released a new version of its AI workstation, accompanied by a whitepaper and benchmark results.

Emem is a shared memory layer for multi-agent systems that provides signed, verifiable facts anchored to physical locations, enabling agents to share exact observations without trust.

The Chinese robotics company is targeting commercial, industrial and research applications as it looks to move embodied AI from one-off demonstrations to full-scale applications.

The US will spend $5bn to tackle longstanding scientific problems using AI, with 15 agencies focusing on chronic diseases, drug discovery, and building materials. Scientists will gain access to supercomputers and specialized datasets.

Robots can perform flashy stunts like backflips, but struggle with mundane tasks such as folding laundry or making tea due to the complexity of real-world interaction. This article explores why, covering topics like world models, data scarcity, and the future of robotics.

Agent Atlas is an open-source CLI tool that scans your AI coding environment (e.g., Claude Code) and generates an interactive mind map showing which skills and agents are actually used, which never fire, where overlaps exist, and what's missing. It's local-only, privacy-focused, and can work without an API key. A typical analysis reveals that most installed skills silently consume tokens without ever being invoked.