2026-09-11

Original Source from DeepSeek WeChat Official Account: Today we're officially releasing the DeepSeek V4.1 Flash model. It is the smallest model in our brand-new model architecture series, with native multimodal visual understanding. The new architecture was designed with these goals in mind: a higher capability ceiling, faster inference, greater throughput, and scalability to larger-parameter models. Asymmetric architecture: big intelligence at low cost DeepSeek V4.1 Flash is a 552B-parameter Mo

GPT‑Live‑1 brings natural, full-duplex voice conversations to the API, with stronger instruction following, custom voices, and telephony support.

Hey y'all! We've released a new model in our lineup: GigaChat-3.5 Reasoning. It's a 432B-A28B MoE with Gated DeltaNet for long-context efficiency. We trained domain experts (code, math, general, etc.) with CISPO and then distilled them into a single model via on-policy distillation. In our evals the resulting model lands close to DeepSeek V4 Flash Preview while using 37% fewer tokens in its reasoning traces. Weights are on Hugging Face under MIT: . You can also try it at giga.chat — pick the rea

The downside of uncensoring a model is that it is known to potentially damage it, but CyberTiel is an even more capable software engineer than its censored TielCoder base, while allowing offensive security research. This was achieved by quantizing with an improved imatrix, baked from a curated corpus of cybersecurity- and agentic software engineering work. In short, the small damage from abliteration on a full precision model is negligible under Q4 quantization, and the weights that the model ne

On July 22, 2026, a transmission line fault in Ashburn, Virginia—the heart of the world’s largest data center cluster—knocked more than 3 gigawatts of load off the grid in seconds. And it wasn’t the first time. Two years earlier, a single failed surge arrester dropped roughly 60 Virginia facilities and 1,500 megawatts at once. No…

arXiv:2609.09565v1 Announce Type: new Abstract: Agentic graph learning (AGL) has recently achieved promising results on graph reasoning tasks, where an agent powered by a large language model (LLM) sequentially samples the graph as evidence to support its final prediction. Existing methods either employ a single agent or orchestrate multiple role-based agents to reason and learn over the entire graph, but both essentially rely on a shared reasoning policy across different graph regions, which ca

轻量版LingBot-World 2.0

arXiv:2609.09395v1 Announce Type: new Abstract: Language models act through tools, yet practical agents face libraries containing thousands of interfaces. We introduce the tool menu as the short, ordered subset of available tools shown to an agent before execution. The agent can call only tools in this menu. Multi-step tasks require the final action and the prerequisite tools that create its inputs in a usable order. Current constructors rank tools by request relevance, which can surface the fin

While research on recursive self-improvement (RSI) has predominantly automated model training pipelines, reliable autonomous development demands a missing pillar: post-hoc monitoring and auditing to understand what models learn and ensure safe alignment. Mechanistic interpretability tools are essential to bridge this gap, among which Sparse Autoencoders (SAEs) serve as a cornerstone by isolating interpretable features for model inspection and steering. In this paper, we introduce SAEScientist-Be

Large language models often answer structurally unanswerable questions, such as computing cot(-540°) or evaluating (1).startswith("1"), instead of abstaining. We ask whether this failure reflects missing recognition or failed routing from recognition to abstention. Across instruction-tuned models from 1.7B to 70B parameters, a single linear direction in the hidden state separates answerable from structurally impossible math and code prompts, showing that models represent impossibility before gen

AI research agents combine prior knowledge, public sources, and experimental feedback to produce useful results. The Discovery Certification Protocol (DCP) turns claims about these results into executable recovery and feedback tests. Gate 1 validates useful improvement on sealed evaluation. Gate 2 gives matched agents the registered starting information and observed Web content while withholding the target research history. Every valid method reaching the numerical target supplies a recovery wit

People keep on getting confused about this, so I looked at the safetensors on hf. The title should have been "Deepseek V4.1 Flash is 748B total/552B base, not 284B or 305B or 485B or 522B" The model is not 284B. The original Deepseek V4 Flash is 284B, but not the V4.1 Flash model The model is not 305B, despite what some people claim "So: ~305B real backbone + 203B engram = 508B total" This is incorrect. The model is not 485B, even though Huggingface lists the model as 485B, but that's because th

Today, we're releasing a demo of WeWorm, the first zero-click worm to spread through WeChat calls across iOS and Android. [...] The victim does not need to answer the call, or interact with their phone at all. Even if they do answer, they hear nothing, and the exploit still succeeds. [...] Working with AI, our team found the bug and wrote the first remote code execution (RCE) exploit in about two days. Building the worm took one more week. A worm at this scale used to be the kind of thing that t

OpenAI has decided to fully shut down a protein design project I'm working on for a client. Needless to say, open weight models are the only way forward.

so i saw that openui.com released OUI-1, a model fine-tuned on DiffusionGemma. the training dataset uses OpenUI-Lang, a custom DSL (domain-specific language), instead of plain HTML, Markdown, or React code. what makes it interesting is that you can already get a regular LLM to use OpenUI-Lang through a system prompt, but that eats up a lot of the context window. my thinking is that fine-tuning a model on the DSL could reduce that overhead and leave more room for the actual conversation, without

Github Repo. Blog post. 💡 TL;DR (from the Github Readme) Spend less without making the agent do less useful work. SoL-Pi is a standalone extension for Pi that packages four reusable efficiency mechanisms discovered through scaled auto-research loops. It reduces repeated model turns, context replay, oversized observations, and unnecessary long-log reading while preserving the work and evidence an agent needs to finish a task. SoL-Pi installs on top of an unmodified Pi release. Every mechanism is

A new report released Thursday by Anthropic alleges persistent distillation attacks by China-based AI companies, which have escalated in recent months as competition in the space has intensified.

Anthropic produced an extensive collection of case studies covering the ways its current AI models have been misused.

“The vast majority of cases we find are people who are entitled to claim for something, claiming for that thing,” the researcher told TechCrunch.

Multimodal Large Language Models (MLLMs) perform strongly on general visual understanding tasks such as visual question answering, yet they often struggle with a basic comparative skill: identifying what has changed between two similar images. We introduce VDiff-Bench, a challenging multiple-choice benchmark for fine-grained Image Difference Identification. VDiff-Bench contains 1,756 four-way questions over image pairs and covers 10 change categories: position, motion, regional image color, over

arXiv:2609.09203v1 Announce Type: new Abstract: Existing benchmarks for autonomous AI scientists evaluate only final outputs---generated code, hypotheses, or papers---yet discard the reasoning process by which those outputs were obtained. This makes it impossible to audit scientific methodology, diagnose failure modes, or distinguish systematic reasoning from fortunate guessing. We present \textbf{OpenDiscoveryTrace}, a public dataset of 558 complete AI scientific agent trajectories that capture

Full-duplex spoken language models can listen and speak simultaneously, enabling them to handle interruptions and backchannels in human conversation. However, text generation, speech synthesis, and audio playback proceed asynchronously. As a result, what a model believes it has said may not match what has actually been played to the user. We refer to the problem of recovering from an interruption while remaining aware of the model's realized speech as anchor interruption. To address this problem

Hoping to see smartest medium size models soon & later with all available optimizations/architectures/etc.,. Thanks Deepseek! Ex 1: 30-50B MOE + 10-15B Engram + DeepSeek-V4.1-Flash type KVCache Ex 2: 15-30B Dense + 10-15B Engram + DeepSeek-V4.1-Flash type KVCache EDIT : Updated Engram to 10-15B from 50B

LandingAI has shipped Agentic Document Extraction Gen2, a rebuild of its document stack on the DPT-3 model family. Chunks are retired in favor of a document, page and block tree. DPT-3 Pro grounds to the line, DPT-3 Verity grounds to the word with a confidence score, and Parse billing now counts output characters instead of flat pages. Gen1 code will not run against Gen2 endpoints. The post LandingAI Releases Agentic Document Extraction Gen2 with DPT-3 Pro and DPT-3 Verity appeared first on Mark

Compressing large-scale neural networks is essential for deploying models on resource-constrained devices. Most existing methods adopt weight pruning or low-bit quantization individually, often resulting in suboptimal compression rates to preserve acceptable performance drops. We introduce a unified framework for simultaneous pruning and low-bit quantization via Bayesian variational learning (\method), which achieves higher compression rates than prior baselines while maintaining comparable perf

A transformer can make an attribute linearly decodable in its residual stream at a depth where that attribute does not yet influence the output. This gap between where information is readable and where it is used has been shown for attributes stated directly in the input. We ask whether it also holds for an attribute the model must infer gradually over a conversation, namely how expert its dialogue partner is. Using ExpertCollab, a corpus of multi-turn research-planning dialogues between model-p

World models are increasingly used as policy-in-the-loop imagination environments, where reliable rollouts require fine-grained controllability with respect to low-level robot actions. A key obstacle to scaling such models in robotics is that actions are not a universal language in pixel space: changes in visual environment, camera view, robot placement, or embodiment alter how the same numerical action manifests visually, leading to conflicting supervision under mixed training and brittle gener

I was not aware that the harness makes such a big difference. DeepSeek V4.1 Flash

Some quick notes on a truly weird week.

InquiryIQ, a previously unreported prototype, tested a model from xAI, maker of Grok, to surface associates, social accounts, and other information about people identified through Clearview.

This week on “Uncanny Valley,” we dig into a former Anthropic researcher’s AI doomsday warning, the latest upgrades from Apple’s event, and the census report that claimed Trump won the 2020 election.

A new feature coming to Slack will allow you to build interactive reports, polls, dashboards, presentations, microsites, and other tools directly inside a chat. With Slackforce Surfaces, you can describe to Slackbot what you need, and it will use AI to gather information from relevant conversations and connected apps, like Google Drive or Salesforce, to […]

Meta's newest app Muse is off to a slower start than the company's other apps, like Meta AI or Threads.

The platform was linked to an FBI investigation into a dark web marketplace last week.

App support is slim right now, but Google says more are coming.

Native is now the future of mobile at Shopify Shopify are moving from React Native back to separate Swift and Kotlin codebases for their native apps, for the exact reason you would expect: We decided to switch from native to React Native in 2020 for three reasons: Stop building the same features twice Allow developers to work across the stack Spend less time chasing feature parity and more time shipping value [...] Native still means building and maintaining software on two platforms, that cost

"Bankruptcy cannot become the new land grab for AI.”

arXiv:2609.09428v1 Announce Type: new Abstract: Evaluating the quality of explanations produced by explainable AI (XAI) methods remains challenging because existing approaches often rely on subjective human judgment, limiting reproducibility, scalability, and comparability between studies. We examine whether LLMs can serve as a reproducible and scalable mechanism to make comparative assessments of the quality of XAI explanations. We introduce XAI-Arena, an LLM-as-a-judge framework for scalable,

Meta has launched its new Muse assistant, marking the company's first real foray into AI-powered productivity tools. The company says its AI agent can "take the busywork off your plate" by helping you with online shopping, emails, trip-planning, and more. I decided to try out the new tool and see how well it performed - […]

Recent advances in wearable sensing enable continuous monitoring of physiological and behavioral signals, yet existing benchmarks rarely evaluate whether AI systems can reason over a real user's longitudinal wearable record. We introduce WearableQA, a benchmark comprising 4,084 10-option multiple-choice questions constructed from the wearable time series, blood biomarkers, and demographics of 200 real users, each with up to 500 days of daily measurements. WearableQA preserves authentic wearable

9月10日，阿里巴巴集团旗下高德正式发布全球首个3D原生城市世界模型ABot-Earth 0.7。

大晓机器人联合南洋理工大学 S-Lab、上海人工智能实验室发布全新人–场景交互重建研究 HSImul3R

arXiv:2609.09578v1 Announce Type: new Abstract: Urban planning is a real-world spatial optimization problem that requires selecting feasible actions from large candidate spaces under practical objectives such as cost and service quality. Existing optimization and reinforcement learning methods are effective for fixed formulations, but often depend on task-specific representations and constraint handling. We propose \emph{CityPlanner}, a sandbox-agent framework for executable urban planning. City

We present Cadence, an error-bounded lossy compressor for numeric time series pairing a 330M-parameter time-series foundation model (Google TimesFM-3) with an adaptive arithmetic coder, guaranteeing |x_t-x_t|leτ on every sample. One negative result constrains the design space: for lossless coding a foundation model is worth nothing, because bits saved are logarithmic in predictor accuracy, Δb=log_2(MAE_{old}/MAE_{new}). So the 1.51times advantage TimesFM-3 holds over a 32-tap linear predictor bu

Tool: .blend URL Viewer I'm continuing to have a lot of fun with GPT-6 Astra and Blender (see my TIL ). As a big fan of the Imperial Fabergé Easter eggs , I've always thought it would be fun to make some new ones that celebrate popular culture. Yesterday I decided to try out the new ChatGPT Images 2.5 by running this prompt : Generate a photo of a faberge egg that's themed after the TV show Pluribus - research first It gave me this - honestly not bad for a first attempt! Then, just to see what w

Humans need to study only a handful of well-written textbooks to master a discipline and attempt its hardest problems. We argue that an ideal self-evolution method should share the same property, that is autonomously learning from raw training material for transferable problem-solving capability. However, we still lack a direct measurement for it. We introduce StudyBench, a controlled physics benchmark that directly measures how efficiently a self-evolution method converts training material into

9月9日，以“JoyAI · 跃迁物理世界”为主题的JDDiscovery-2026京东全球科技探索者大会在北京举行

A few years ago a 100GB was considered a very large language model. What do we call under 100GB models now? Tiny models? haha