— Open-weight models are launching in clusters, and advertising has become a new pipeline for zero-click spyware.
Meta releases Muse Spark Open, an 816B open-source multimodal model, partnering with Apple to run locally on M5 Ultra Mac Studio; NVIDIA launches Nemotron 3.5 Super VL and Bolt edge models the same day. On security, Citizen Lab confirms Israeli ad company Discover-Tech delivered zero-click spyware via targeted ads, infecting devices at Al Jazeera and other organizations. Several papers focus on agent distillation, rule-reasoning benchmarks, and RL training optimization.
Meta releases Muse Spark Open, 816B total parameters (42B activated), 1M token context, natively supporting text, images, video, and audio, aimed at coding and agent tasks. It is the first model of its kind that can run locally on two Apple M5 Ultra Mac Studios via the MLX framework, and Meta and Apple are publicly collaborating on this. Why it matters: For the first time, an open-weight model brings 800B+ parameter scale to desktop-class hardware, offering direct reference value for local inference and agent deployment engineering paths; at the same time, the rare collaboration between Meta and Apple hints at converging interests in AI infrastructure.
The community generally welcomes the open-weight release, but some point out its licensing restrictions are numerous, its performance trails competitors like DeepSeek, and they are skeptical of Meta's benchmark data.
A joint investigation by Citizen Lab and Confiant found that Israeli ad company Discover-Tech embedded zero-click exploits in millions of online ads through its Controlium platform, infecting 45 devices at Al Jazeera; the attack works on both iOS and Android, and victims need no interaction. Why it matters: This is the first technical evidence confirming zero-click spyware delivered through targeted advertising, meaning ad networks themselves have become a distribution pipeline for state-level surveillance and mercenary spyware, posing a fundamental challenge to browser security models and ad-blocking strategies.
Commenters generally believe ads are a serious security threat, that ad blockers should be used and regulation strengthened, but some think zero-click attacks are not that easy to achieve.
NVIDIA launches Nemotron 3.5 Super VL, a 120B-parameter MoE architecture (12B activated), 1M token context, supporting text, image, document, and video input, with NVFP4 weights runnable on a single DGX Spark; it also releases two edge models, Bolt 74B-A9B and Bolt E4B, with Bolt 74B-A9B running within 24GB VRAM using native 2-bit quantization. Why it matters: A 120B-class multimodal agent model reaching a single DGX Spark and 24GB consumer-grade VRAM significantly lowers the deployment barrier for enterprise-grade vision-language agents, with direct engineering significance for edge inference and local agent scenarios.
The paper proposes ActionOPD, an action-level on-policy distillation method that uses semantic actions rather than token probabilities as the supervision unit for aligning a small model with a large model agent, without needing a shared tokenizer or teacher probability distribution. Why it matters: It directly addresses the long-standing state distribution mismatch and fragility of token-level supervision in agent distillation, providing a more robust training paradigm for compressing large model agent capabilities into deployable small models.
RuleTrek builds a multi-domain benchmark to evaluate LLM agents' ability to identify, derive, and apply domain rules in complex scenarios spanning multiple files, revealing significant performance gaps in existing models; CLASHarness provides a controllable and verifiable generation pipeline to test multi-agent conflict resolution when facing outdated states, contradictory instructions, and shared resource contention in computer-use tasks. Why it matters: The two benchmarks expose key shortcomings of current agent systems from the dimensions of rule reasoning and multi-agent collaboration, providing reproducible testbeds for subsequent training and evaluation.
ウェブは全体像、購読者にはあなた専用を——興味に合わせた AI 精選、プライベート RSS の統合、コミュニティの見解付きで毎朝配信。ずっと無料。
96 号配信 · 毎日150件超から読む価値ある30件に厳選
Cortex consolidates context-induced performance gains into mergeable low-rank or structural FFN updates, achieving memory consolidation while preserving non-target capabilities.
🤖Cortex consolidates context-induced improvements into mergeable low-rank or structural updates, preserving non-target capabilities across sequential edits and diverse model scales.
ReMaRk targets RL training for LLM recommender systems, changing the reward signal from a single response to joint scoring of all valid responses, reducing training cost without performance loss.
🤖ReMaRk improves reinforcement learning for recommendation by scoring generated lists against all valid replies jointly, reducing training cost while matching or exceeding prior performance.
KV4-RL finds that under asynchronous rollout, KV cache quantization yields more significant speedup than weight quantization, and proposes a method to mitigate the negative impact of INT4 KV on RL learning.
NoteBank introduces sparse episodic memory for linear attention, decoupling long-term storage from the recurrent state, challenging the existing belief that state capacity determines long-context capability.
ProgAgent explores training agents through programmatic tool calling (PTC), storing intermediate results in runtime variables rather than context, alleviating context rot in long tasks.
Proa is a Rust web framework supporting fast SSR and rendering HTML, Markdown, and other targets from the same typed components.
Ito is a framework for building terminal UIs with Lua, based on Rust, aimed at Neovim users and terminal beautification enthusiasts.
Apertus 1.5 extends a fully open multimodal reasoning model to image and audio understanding, with the context window increased to 262,144 tokens.
🤖Apertus 1.5 extends open multimodal language models to image and audio reasoning with long-context training, reinforcement learning, and fully open data and weights.
SwarmRSI evolves scientific agent swarms through feedback-guided mutation and complementary recombination, achieving recursive self-improvement across research tasks.
🤖SwarmRSI evolves scientific agent swarms through feedback-guided mutation and complementary recombination to enable recursive self-improvement across research tasks.
A Goodfire article argues that the prospects for neural network interpretability are better than expected, and that so-called polysemanticity and superposition are mostly misreadings; commenters acknowledge progress but criticize the article as verbose.
Commenters generally acknowledge progress in neural network interpretability, but some consider the article verbose, its conclusions overstated, and the website too heavy.
Anthropic's Scion is accused of plagiarizing Meta's open-source project Mononoke; the author says they received private confirmation of deliberate plagiarism and personnel overlap; the community is divided on the boundary between open-source borrowing and plagiarism.
Most commenters believe Anthropic borrowed from an open-source project while downplaying attribution, but some argue that open-source code inherently allows reference, making the plagiarism accusation excessive.
Simon Willison points out that Anthropic revealed distillation accounts for 5%-7% of its revenue; based on Q2 revenue of $11.5 billion, other labs spent over $500 million distilling Anthropic models.
CopilotKit releases the CLiF continual learning platform, turning agent trajectories and user actions into skills, evaluations, and fine-tuning data.
Star morluto / rea Reverse engineer anything with agents, from app behavior down to native binaries.
Star anthropics / claude-plugins-official Official, Anthropic-managed directory of high quality Claude Code Plugins.
Star K-Dense-AI / scientific-agent-skills Turn any AI agent into an AI Scientist. The #1 Agent Skills library for science, used by 250,000+ scientists worldwide. 177 ready-to-use validated skills plus 100+ scientific databases covering biology, chemistry, medicine, and drug discovery. Compatible with Cursor, Claude Code, Codex, Pi, Antigravity, and the open Agent Skills standard.
Sponsor Star JuliusBrussee / caveman 🪨 why use many token when few token do trick. Viral skill + proxy for coding agents that cuts 65% of tokens by talking like a caveman.
Sponsor Star debpalash / VoiceStudio VoiceStudio is the open-source, fully-local ElevenLabs alternative — voice cloning, voice design, video dubbing, dictation, transcription & audiobook creation in 646 languages.
Sponsor Star obra / superpowers An agentic skills framework & software development methodology that works.
Star ruvnet / ruflo 🌊 The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
Sponsor Star mksglu / context-mode Context window optimization for AI coding agents. Sandboxes tool output (98% reduction), persists session memory, and enforces routing across 17 platforms via MCP + hooks.
Star NationalSecurityAgency / ghidra Ghidra is a software reverse engineering (SRE) framework
Sponsor Star affaan-m / ECC The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
Sparse attention reduces the KV-cache traffic that dominates long-context LLM decoding, but practical methods largely operate at coarse granularity. Token-level selection offers finer control, yet presents two challenges: low-dimensional scoring may lose query--key structure needed for accurate ranking, while gathering scattered selected tokens can offset the resulting latency benefit. We introduce Joint Subspace Sparse Attention (JSSA) to address both challenges. For selection quality, JSSA lea
Coding agents increasingly interact with users by generating interactive software on the fly rather than replying in text, yet many bugs in such software reveal themselves only after multi-step interactions. Today, users must discover these bugs and wait for fixes, making the development loop slow and expensive. Existing benchmarks typically assume the bugs have already been identified, leaving agents' ability to find interaction-dependent failures largely unevaluated. We introduce SWEeper-Bench
Size is pretty small at 102GB. For comparison nvidia/Qwen3.8-Flash-Next-NVFP4 is 133GB. Would be cool to see some of the Qwen3.8-FN specific engines support this.
FTC may fine AI firms on par with companies that delay data breach reports.
METR will publish its own assessments of Anthropic’s risks, incidents, and practices, and the frontier AI model maker has given up the right to block them.
Muse Spark Open is "the most capable US open-weight model to date," the company says.
Yesterday Fish released this incredible text-to-speech model, so naturally I spent the rest of the day building a great demo for them :D The model is a chunky MoE (35B3A), but it runs well on a DGX Spark. The 4bit quants get very close to the original. You can try it here: or download the model and run it locally 🤗
Agent skills encode operational procedures and domain expertise as reusable knowledge assets, yet their distribution in plaintext exposes them to unauthorized copying and leakage. However, practical mechanisms for skill ownership verification and leakage tracing remain underdeveloped. Watermark embedding before distribution offers a potential solution, but structured, compact skill instructions impose stringent requirements for factual and procedural accuracy. To address these challenges, we int
Professional agents increasingly work through command-line interfaces and multiple tools over shared workspaces containing files, messages, and business applications. These environments create safety risks that are difficult to capture with attacks placed at a single, known read location. We introduce APEX-Harm, a benchmark for evaluating prompt and script injection against long-horizon agents performing finance, consulting, and legal tasks in large workspaces. Each task contains 115 to 328 file
A open-source human approval layer for AI Agents Discussion | Link
Multi-agent systems allow language models to solve tasks collaboratively by sharing complementary information. These systems typically communicate through text, while recent approaches also explore exchanging internal model states. However, sharing these states across architectures often requires learned translators or assumptions about model compatibility. We introduce SYNC (Symbol-Indexed Neural Communication), a word-indexed interface that connects heterogeneous models without fine-tuning or
this is actually interesting: Microsoft AI developed a local version of MAI Code 1.1 Flash, a coding-optimized mixture-of-experts model with 137 billion total and 6.8 billion active parameters .
Article URL: Comments URL: Points: 31 # Comments: 8
The FTC is investigating possible consumer harms at two major AI companies.
At Open Source Summit Europe 2026, openKylin outlined its vision for transitioning from resource management to intent management in the AI era. The community introduced KART, a unified AI runtime, alongside core desktop features in openKylin 3.0 like the AI Assistant, Global Intelligent Search, and Intelligent Space The post openKylin Outlines Its Vision for Reimagining Linux in the AI Era at Open Source Summit Europe 2026 appeared first on MarkTechPost .
Sandboxes are all you need.
Last month, we committed to stronger independent, third-party assessment of our models and safety practices. We also signed the White House Accord on Super Intelligence, in which companies pledged to partner with independent external auditors or evaluators. Today, we’re announcing that we’ve started to give employee-like access to a team from METR , a nonprofit that evaluates frontier AI models. METR will publish the results of its risk assessments. Over the past year, we’ve worked with METR on
This is a project that spun from my frustrations of using LocalSend and DashBeam to transfer stuff between devices, and with the extra goal of being fast, trying out the Slint framework and, most importantly, establishing a communication channel with trusted external people that can be left always on.
TypeSafe AI released Jev 1.14 under the jev-preview alias. The System One model adds experimental image input, a Box bounding-box type and an opt-in 262k-token context, at $0.042 per million input tokens. The post TypeSafe AI Releases Jev 1.14 With Images, Box Type and 262k Context appeared first on MarkTechPost .
In long-horizon tool-calling tasks, LLM agents can fail not only because of tool-use errors, but also because of ineffective interaction with users. Yet existing benchmarks largely assume cooperative and behaviorally stable users, leaving the effect of emotional variation on agent reliability underexplored. We introduce τ-Emotion-Bench, a ten-domain benchmark for evaluating tool-calling agents under controlled emotion-conditioned user interactions while preserving the underlying task objective.
Datalab released OmniParseBench, an open OCR benchmark with 16,288 yes/no tests on 2,937 pages in 92 languages. Every score is a pass rate you can slice by tag, script or language. The post Datalab Releases OmniParseBench: An Open OCR Benchmark With 16,288 Pass/Fail Tests to Audit OCR Parsers appeared first on MarkTechPost .
Evaluating model uncertainty on real-world data is challenging because each instance provides only a single observed label, leaving true conditional class probabilities unknown. To address this, we introduce ProbabilityBench, a benchmark that provides exact, instance-level class probabilities across three simulated text-based domains: peer review, covert information transfer, and credit risk. By explicitly defining the full data-generating process for each environment, we analytically compute bo
We introduce StartLux-Decision, decision models from 0.8B to 27B parameters that combine efficient single-pass prediction with adaptive, confidence-guided reasoning. The Instant models use bidirectional attention in full-attention layers to improve decision accuracy while keeping single-pass inference. The AutoThink models invoke explicit reasoning for uncertain decisions and answer confident cases directly. A three-stage training pipeline establishes shared decision capabilities, specializes th
Multimodal LLMs are typically trained by first aligning a vision encoder with a language model via projector training on large image-caption datasets, followed by joint training of the full multimodal model. Despite its widespread use, the role of the projector training stage remains poorly understood. We find that (i) using only 10 to 20% of the data for training the projector already recovers most downstream performance gains, and (ii) projectors trained on different datasets can be linearly i
from Mistral: Timestral is a family of time series foundation models for zero-shot forecasting, available in 5M , 28M , 123M , and 209M parameter versions. This repository contains the 209M checkpoint. Timestral supports univariate and multivariate forecasting, past-only and known-future covariates, and probabilistic predictions across 21 quantiles. It accepts up to 32,768 historical time steps and forecasts up to 1,024 steps in a single forward pass. We’re releasing the source code in our GitHu
Introducing the Rosalind Academic Plan, offering eligible academic labs sponsored access to GPT-Rosalind and tools to advance life sciences research.
Tool: Chirp Pong Another game experiment - I wanted to see if it was feasible to have a two player game on two separate phones running in static HTML where the networking code uses chirps sent via the microphone. It seems to work! Build a two player competitive pong like game where two players use their phones close to each other and the game state is sent via audio noises which the web microphone API can interpret - but add graphical and gameplay twists to make it a lot more fun and intertwine
Anthropic tells investors quarterly revenue up nearly 50% to $16.8bn ahead of IPO The phrase "told investors" appears four times in this FT article, which reports on Anthropic's day of hosting investors on Wednesday. Lots of details in here from " people with knowledge of the matter ". Key things I learned from the article: Q1 revenue (Jan/Feb/Mar) was "about $4.8bn" Q2 (Apr/May/Jun) was $11.5bn Q3 (Jul/Aug/Sep) was $16.8bn In Q3: $1.70 in revenue for every $1 spent on "computing power" Adjusted
Google’s new model isn’t for me on my first test, mixed reviews
From NVIDIA AI on 𝕏:
arXiv:2610.14136v2 Announce Type: new Abstract: Iterative sampling in diffusion models can lead to error accumulation due to exposure bias. The denoiser is trained on real images interpolated with noise, but at inference its inputs are constructed from its own imperfect predictions from the earlier steps, something it has never seen at training time. This might be one reason diffusion models often struggle without additional help, such as CFG. We introduce a correction step for diffusion models
SpaceXAI just released Grok Voice 3.0. Grok Voice 3.0 is a SpaceXAI speech-to-speech model that listens, reasons, calls tools and speaks. It now ranks first on the Artificial Analysis Speech-to-Speech Index with a score of 86.9. For developers, the endpoint, events and $0.08 per minute price are unchanged from Grok Voice 2.0. TL;DR What is […] The post Meet Grok Voice 3.0: SpaceXAI Voice Agent Model Ranks #1 on Artificial Analysis appeared first on MarkTechPost .
Generative discovery in science requires coupling high-level scientific reasoning with the ability to generate samples respecting the structure of a design space. Yet these capabilities remain largely separated: foundation LLMs encode broad chemical and biological knowledge, while domain-specific flow- and diffusion-based generators capture the priors needed to produce valid sequences and geometries. We study how to align such scientific generators to high-level natural-language discovery goals
Edit video by talking to Claude Code or Codex Discussion | Link
Autoregressive generation underlies real-time, causal video synthesis and is the backbone of video world models. Today's dominant approach distills a bidirectional diffusion model into a causal few-step student, but existing pipelines make this conversion prohibitively expensive, requiring offline Probability Flow ODE} (PF-ODE) trajectories that cost tens of thousands of GPU-hours and terabytes of storage. We show it can be made dramatically more efficient. Pushing this efficiency to its extreme
Just downloaded the new LFm2.5-24B-A2B LiquidAI model for testing. They claimed some pretty good performance, and I've had good experiences with their prior releases. Some context - I use Ornith 1.5-35B pretty regularly in multiple workflows - sometimes on it's own, and sometimes as part of an AI-Council in Pi Agent. It's been very reliable, provides surprising insight "smarter" models miss, and is a great workhorse in Pi. I swapped LFM 24B in to replace Ornith in a few of my workflows - it is h
I saw the 4x R9700 build for a 10-person startup here a few days ago and wanted to share my results, since I opted to spend most of the money on the cards. I went with a super cheap used EPYC instead of a Threadripper. In the end it came to about $10.7k per box all-in, and I built three of them for a total of 384 GB of VRAM. As a long time lurker on my other account, I'm always frustrated that people never show how they actually use the models, so after a lot of back and forth I decided to share