2026-08-12

For more than 150 years, the Riemann hypothesis has stood as one of the major unsolved problems in mathematics. Anthropic hasn't solved it — but the company's models made more progress than you might expect.

Interesting examples in the link Paper here: This is your prompt to go out and give us 10mil rows of Opus 5 traces on hf before they fix this workaround

I wanted to find out whether a huge text-only MoE could be given basic vision without retraining the language model itself. The short answer is yes. I froze DeepSeek V4 Flash and a 417M-parameter MoonViT image encoder, then trained a 40.1M-parameter connector between them on 100,000 image-text examples. The completed NVFP4 model loaded across four B200s in a custom SGLang stack and answered real image prompts. It has basic vision, although this 100K-example pilot is not yet a production-quality

Introducing Muse Glimmer Meta are back in the open weights game! Muse Glimmer is a brand new 30B model under a clean Apache 2.0 license (a step up from the janky Llama licenses of old). They claim to have optimized it for exactly the kind of things I'm looking for in a local model: End-to-end Agentic Task Completion. Muse Glimmer achieves strong success rates on full-task benchmarks including DeepSearch QA, MCP-Atlas, 𝛕-Bench and SWE-Bench, which measure its ability to work within scaffolds, wri

Large language model (LLM) inference serving is increasingly constrained by memory rather than compute. As long-context and long-form reasoning workloads become more prevalent, the key-value (KV) cache dominates both memory footprint and memory traffic during LLM token generation, i.e., decode. In particular, HBM capacity has become a scarce and costly resource that heavily limits inference batch size and system throughput. This paper presents OasisKV, a memory-centric LLM inference system desig

Small language models are often the only option for deployment under tight latency, cost, and on-premises constraints, but they are rarely trained from scratch: a compressed model is usually recovered through knowledge distillation (KD). This recovery step largely decides the final quality, yet it is expensive. We present a practitioner's study of how to make distillation training efficient, organised around two systems contributions. First, we show that offline KD (caching the teacher's top-K l

With the rapid advancement of large language models (LLMs), harnesses have become essential infrastructure for deploying agents across a wide range of domains. The fast-evolving harness ecosystem has also made rigorous capability evaluation increasingly important. However, efficiently building an end-to-end, systematic, and comprehensive evaluation pipeline remains a significant challenge. To address this challenge, we introduce A^2E (Agent Auditing Engine), an end-to-end evaluation engine desig

Phone and Bluetooth signals could turn roadside cameras into far richer tracking tools.

Pharma is suddenly paying for Bio × AI tools, and Chai is leading the pack with four deals closed this summer. Cofounder Matt McPartlon and Product leader Neil Patil explain why.

Zoom screen-sharing bug let people fully take over other devices on a call.

A few things right off the bat: it reasons very efficiently. Like Grok 4.5 levels of efficient thinking it quantizes very well. My first few tests with iq3_xxs were better than Qwen/Gemma behaved at that size its knowledge depth is amazing. It beats Qwen3.6 27B on no-tools trivia. in OpenCode it is a much more efficient agent than 27B. Both models accomplish their tasks but Muse-Glimmer got there faster every time I'll say that it's worse at most things coding, probably being closer to Gemma4-31

AMIE promotional video

OpenAI and AWS are making Daybreak cybersecurity capabilities available through Amazon Bedrock to support enterprise security workflows.

Imagine if your everyday laptop could run an AI model smart enough to take care of 90% of your work—totally private, lightning fast, and completely free of monthly fees. That is the exact tipping point of history where the AI bubble bursts completely. I bet this is the nightmare for OpenAI and Anthropic. Not everybody has resources to host big open weight models, but everybody can host small ones like Qwen 3.6. Are we reaching there soon 🔜

AI code review that fixes what it finds Discussion | Link

Hey guys, we just released Unsloth Desktop today! It's an open-source desktop app that works on Mac, Windows, and Linux: We're pretty excited about this release because it's the first (we're like 99% sure) desktop app that enables you to train AND run models on your local device. Here are some of the key features: You can run and train MLX, diffusion image/video models, audio models and GGUFs Connect Claude Code and Codex to local LLMs 50% more accurate with self-healing tool calls and sandboxed

a small win for american open models - Glimmer runs on a fits on a single RTX 3090!

arXiv:2608.07524v1 Announce Type: new Abstract: Training deep learning models on variable long sequences poses significant computational challenges. Existing methods force a difficult trade-off between efficiency and ease-of-use. Simple approaches use static configurations that cause workload imbalance low efficiency, while complex methods introduces significant complexity and code change for new models. To break this trade-off, we introduce Data-Centric Parallel (DCP). Its core principle is to

全球AI安全实战化测评，中国方案DoGNAVY位列前三

Large Language Models (LLMs) have driven rapid progress in autonomous agents, yet standard evaluations remain confined to static task solving. An emerging frontier is harness evolution---the agent's capacity to autonomously optimize its own operating harness. However, systematically benchmarking this capability remains challenging, as existing evaluations fail to isolate harness improvements from base model strength, prevent task-specific overfitting, or capture long-horizon iterative research.

Learning-based memory systems for self-evolving LLM agents face two tightly coupled challenges. First, trajectory-indexed utilities grow with the interaction history, thereby dispersing limited feedback over an ever-expanding state space. Second, because trajectory-level rewards are jointly assigned to co-retrieved memories, irrelevant experiences may receive misleading utility updates and consequently enter the memory-reward trap. To address these challenges, we introduce Reduced-Order Memory R

Brad Lightcap, OpenAI's special projects lead and the company's former COO, announced his departure after an eight-year stint at the AI lab. In an internal memo he later posted to X, Lightcap told colleagues he'd be starting "something new." "Over the last few months, I've been focused on the next horizon and what would stand […]

Why passkey apps treat Windows differently than other operating systems.

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. These startups are chasing the next big thing in LLMs Nine years after Google researchers introduced the transformer, this family of neural networks has become the engine inside every major large…

arXiv:2608.07528v1 Announce Type: new Abstract: Linear probes detect corrupted context in language models with near-perfect accuracy, yet this does not translate into reliable failure prediction. The result is a dissociation with direct implications for deployment monitoring. Across multi-hop arithmetic chains, probes that detect corruption turn out to be uninformative about final answer correctness; models forced into structured confidence formats collapse to two values with indistinguishable e

arXiv:2608.07538v1 Announce Type: new Abstract: As LLM agents move from decision support to autonomous procurement, firms need to know whether delegated negotiators create value, divide it predictably, and avoid money-losing contracts. We study this in a canonical supply chain bargaining problem: a buyer with private demand information negotiates a quantity-payment contract with an uninformed seller. We benchmark nine LLMs from OpenAI, Google, and Alibaba against a validated Perfect Bayesian Equ

arXiv:2608.07542v1 Announce Type: new Abstract: Autonomous research loops driven by large language models can run machine-learning experiments at scale but tend to drift toward local refinements of whichever metric they optimise rather than testing the hypotheses that motivate the experiments. We address this structurally and present an AI Scientist for studying generalisation in quadruped robot navigation policies in simulation. Building on the autoresearch paradigm of Karpathy, our loop adds t

Hey everyone, Today we release Luth-2-0.8B and Luth2-2-2B , two non-reasoning models that set a new state of the art for French across a wide variety of tasks for their size 🚀 A few notable scores on French benchmarks compared to models 〜3 times their size: - Luth-2-2B scores 69.67 vs Gemma-4-E2B-it at 65.17 on Multi-IF. - Luth-2-0.8B scores 72.92 vs granite-4.0-h-micro at 55.60 on MGSM-Rev2. - Luth-2-2B scores 81.52 vs Gemma-4-E2B-it at 81.24 on Math-500. Luth-2 builds on our previous work with

arXiv:2608.07474v1 Announce Type: new Abstract: Prior work showed that human-in-the-loop oversight becomes structurally untenable in high-loss domains when AI output velocity V exceeds human cognitive capacity C_max. The operative constraint, however, is not V alone but V x L, where L denotes per-item cognitive load. L consists of triage, judgment, and response, which respond asymmetrically to AI capability improvement. Triage cost does not decline as models become more capable, because semantic

On-policy distillation (OPD) provides dense teacher supervision on student-generated trajectories, but standard reverse-KL training can assign insufficient probability to other plausible continuations. Teacher entropy alone does not reveal whether uncertainty is concentrated among a few plausible next tokens or dispersed over a long probability tail, nor whether the student already represents those candidates well. Moreover, local teacher probabilities may not predict downstream success. We intr

Heeeey all! I just completed some fun tests with Muse Glimmer, I thought I'd let you know. In fact, the summary below was written by Muse itself! I ran a 2× DGX Spark cluster and got Meta's day-old Muse Glimmer 30B running the day after release — then pushed its context from the trained 131K all the way to 1M with YaRN, verifying retrieval at every rung. Sharing config + results since the "131,072+" hint in the model card turned out to be very real. Setup Hardware: 2× NVIDIA DGX Spark (GB10, 128

Zuck argues for releasing more open-weight models and invites governments to work with AI makers to test safety..who's I have yet to figure.

Interpretability is often treated as a tax on capability: language models are trained as opaque systems, then explained after the fact, with methods whose reliability is difficult to establish. In this work, we challenge this premise. Rather than reverse-engineering a model, we make interpretability a constraint of the training pipeline, optimized alongside the language modeling objective. Across three orders of magnitude of compute, on both autoregressive and diffusion language models, interpre

Hard prompt compression reduces long-context inference cost by independently scoring tokens, sentences, or chunks and retaining the highest-scoring units under a budget. We identify a structural failure in this procedure: independent selection can split dependent evidence pairs, retaining one member while deleting the other. When retained text contains an answer but deleted text defines the entity needed to interpret it, we call the result referential dangling. At a compression ratio of 0.30, Be

The Delta flight crew switched off the aircraft's legitimate Wi-Fi network for around 30 minutes due to the incident, according to a spokesperson.

In a new alert, the FBI said cybercriminals are targeting adults and minors in an attempt to steal their personal and intimate pictures in extortion campaigns.

Thirteen authors joined a little-known Google AI study in 2022. Now they're facing backlash.

One of OpenAI's longest-serving executives is headed out the door, although the longtime COO told staff that he was "excited to help you all advance the mission from a different vantage point."

arXiv:2608.07530v1 Announce Type: new Abstract: SHACL is a core technology for validating the conformance of RDF knowledge graphs (KGs). Yet, authoring SHACL shapes requires technical expertise that most domain experts lack. Translating natural language requirements into SHACL (NL2SHACL) would lower this barrier. However, there is no dedicated benchmark for NL2SHACL, and evaluating generated shapes requires methods beyond string comparison, as semantically equivalent shapes can differ in seriali

arXiv:2608.07473v1 Announce Type: new Abstract: Evaluative AI (EAI) has been recently proposed as a way to support human decision-making, not by producing a single recommendation, but by presenting competing hypotheses together with evidence for and against each. In this position paper, we advocate (computational) argumentation as a particularly suitable paradigm to provide a formal, computable foundation for forms of EAI that are explainable and contestable, setting the ground for a long-term r

Hey all, I have about 8YOE. I've been with 4 different start ups, I've more or less passed the first recruiter phase 95%+ more of the time since they were more of a vibe check and I'm a decent interviewee. This was about 4-6 years ago when I was on the market for my 2nd + 3rd job, so this was also when I had about 2-4 YOE. I've been on the job market for past month or so as someone with 8YOE and I have a hard time passing even the first recruiter round, which is more or less a vibe check and non

arXiv:2608.07540v1 Announce Type: new Abstract: AI systems increasingly operate between flexible input representations and formal objects used by downstream tools. A key challenge is recognizing when an unfamiliar formulation denotes a known formal object. We study this challenge through theorem recognition: given an equivalence-preserving transformation of a theorem condition, a model must recover the theorem identity associated with the standard statement. We introduce TREAT, a benchmark for e

We introduce Sci-VBench, a comprehensive benchmark for evaluating knowledge- and reasoning-intensive video generation across scientific domains. It contains 1,253 expert-annotated examples spanning 60 subjects across four core disciplines: Natural Science, Healthcare, Humanities & Social Sciences, and Engineering. Each example requires models to generate temporally rich videos that demand scientific reasoning and knowledge-grounded synthesis, going beyond surface-level visual plausibility. We fu

arXiv:2608.07533v1 Announce Type: new Abstract: An embodied agent is an intelligent entity that interacts with its environment through a physical body. Currently, the evaluation of embodied agents primarily relies on two paradigms: (1) manually annotated Visual Question Answering (VQA) pairs and (2) high-level task completion metrics, such as success in navigation or manipulation. The former is labor-intensive and subject to variability in annotation quality. The latter may obscure critical vuln

Audio reasoning is essential for machine understanding of the acoustic world. Reinforcement learning with verifiable rewards can elicit such reasoning, yet existing reward designs are complementary in their limitations: outcome-based rewards supervise only the final answer and let the model reach it without attending to the audio, whereas process-based rewards score the reasoning itself but rely on coarse, hand-crafted, and fixed criteria that neither adapt to each question nor stay grounded in

The malicious use of generative artificial intelligence to create highly realistic deepfake videos raises serious ethical concerns and poses substantial challenges to AI safety. However, existing deepfake video benchmarks provide limited coverage of recent synthesis methods and generally lack reliable fine-grained textual annotations. Meanwhile, conventional detectors and multimodal large language models (MLLMs), whether operating as a single model or relying on a single analytical perspective,

Vision-Language Models (VLMs) should answer from concrete image evidence rather than language priors, dataset shortcuts, or irrelevant visual context. Existing perception-aware post-training methods encourage image use through global perturbations or attention proxies, but they do not test whether a sampled answer causally depends on the local evidence that supports it. We propose Counterfactual Evidence Disentanglement (CED), a training-time evidence audit for VLM grounding. For each response,