2026-08-25

Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now AI is accelerating some types of progress but not others:…A nice METR study lays out where acceleration is showing up…Here’s a little analysis from METR which […]

ToMoE : Converting Dense Large Language Models to Mixture-of-Experts through Dynamic Structural Pruning Large Language Models (LLMs) have demonstrated remarkable abilities in tackling a wide range of complex tasks. However, their huge computational and memory costs raise significant challenges in deploying these models on resource-constrained devices or efficiently serving them. Prior approaches have attempted to alleviate these problems by permanently removing less important model structures, y

Qwen3.8-27B is amazing, but it’s slow. A stronger 35B-A3B Mixture of Experts-coder that can run and solve real codebase issues fast (even on constrained hardware) is a valuable addition to the arsenal. This one is the strongest and most consistent 35B-A3B I’ve benchmarked, on both correctness and speed, in addition to being the fastest to fix out of all the 35B-A3B models when you throw them at real codebases. On top of Ornith-1.5’s fine tune, TielCoder uses a code-weighted imatrix for dynamic q

I just finished building my (relatively) low rent local inference machine: * Epyc 7663 * 256GB ECC DDR4-3200 * 1x RTX 5090 32GB Yeah I realize it's weird to throw a 5090 and 256GB of anything together and call it low end, but relative to ~151GB of weights it is. I'm running UD-Q8_K_XL and getting 23.8-24.6 tokens/sec, with pp ranging from 60 on the first prompt to 385 near the last (no doubt lots of caching) on tasks using 100-128k total context. It was slower with DFlash so I took that out. It

Sounds quite interesting, a big IDE provider optimizing for local AI with their coding harness. Especially that they picked Qwen3.6 over Qwen3.8 because of the thinking needs. Haven't read the full article yet, but sounds really cool.

Is the technique outdated? Yes. Is it still creepy? Also yes.

Early testers are raving about what Instinct can do, but some say the AI assistant’s sweeping access, broad terms and ability to act on users’ behalf come with uncomfortable trade-offs.

Mixture-of-Experts (MoE) architectures significantly expand model capacity without a proportional increase in computational cost. However, optimizing their hyperparameters---particularly the learning rate---at extreme scales of both model size and token budget via sweeping remains computationally prohibitive. In this paper, we propose a compute-efficient, two-step hyperparameter transfer framework that estimates optimal learning rates for training large MoE models by transferring them across sca

全球大厂开始押注的AI科研，终于有了统一标准

定义国产推理算力新范式 赋能万卡级AI推理算力规模化落地

arXiv:2608.20397v1 Announce Type: new Abstract: Agentic large language models (LLMs) on the Model Context Protocol (MCP) re-encode verbose tool schemas every turn, so prefill - quadratic in sequence length - dominates time-to-first-token (TTFT) as the tool registry grows. Nexus's primary lever is to decouple routing from the schema-prefill cost: an INT8 semantic lookaside buffer (SLB) with a calibrated cross-encoder margin gate selects tools by retrieval, and arguments are generated over a compr

arXiv:2608.20400v1 Announce Type: new Abstract: Agentic memory under a fixed budget involves two stages: retention and retrieval. Existing retrieval-centered paradigms implicitly assume necessary evidence survives eviction, but we challenge this by isolating a pre-retrieval failure mode: structurally indirect prerequisite eviction, in which upstream blocks weakly aligned with the query are discarded under budget pressure. We provide an operational definition of this failure, a reproducible deter

The five largest GPU neoclouds now run on very different models. CoreWeave and Nebius report to the SEC; Lambda and Crusoe are private and heading toward IPOs; Groq rebuilt itself as an inference cloud after licensing its LPU technology to NVIDIA. This comparison checks each provider's live rate card, Q2 2026 financials, active and contracted gigawatts, anchor contracts, and SemiAnalysis ClusterMAX tier. Nebius posts the lowest H100 rate and the only published B300 price, Lambda has the cheapest

arXiv:2608.20379v1 Announce Type: new Abstract: Advances in large language models (LLMs) have fueled a wave of research into agency: the ability to reason, plan, and act. This effort has produced agentic frameworks that orchestrate perception, memory, and decision-making around powerful LLM backbones. With the advent of large multimodal models (LMMs), these systems can process and integrate diverse modalities, including images, audio, and video, thereby improving their real-world applicability.

Hybrid-thinking multimodal large language models (MLLMs) allow a single model to alternate between deliberative thinking and latency-efficient non-thinking inference. Although these modes differ in reasoning budget, their delivered responses should satisfy the same user-facing standard. Correctness alone may not characterize this response quality; we therefore evaluate task accuracy and response-pattern failures as complementary outcomes. We study this gap through response-pattern alignment: whe

Should you replace your text-embedding pipeline with a large language model? We answer this with a controlled, cost-aware comparison of ten LLMs across six families and 26 embedding models (118M to 14B parameters) on 37 tasks spanning classification, semantic textual similarity (STS), clustering, pair classification, and retrieval. In aggregate the two paradigms are effectively tied: the best LLM (Gemini 3.1 Pro, 77.6) and the best embedding model (77.2) differ by 0.4 points. Their strengths dif

Modern LLM agents are often improved by modifying prompts, tools, or workflows manually, while the executable scaffold surrounding the model---the harness---is typically treated as a fixed artifact after deployment. This work studies an alternative where the harness is task-specific and continuously evolvable: each task family maintains its own harness, which is hot-swapped across iterations through a fixed task-injection seam and rewritten using environment feedback. We introduce Hierarchical S

My first attempt didn't work. I built on Genie's architecture and the videos looked great, but the controls barely did anything. The effect of a keypress was basically zero. Genie learns its actions unsupervised into 8 codes, and that was too loose a grip for us. So I scrapped it and started again with Dreamer 4. The second attempt: Tokenizer at 40.41 PSNR (Genie's paper reports 35.7) FVD 32.19 end to end 144 frames before it falls apart 1.57B parameters, 9.6M frames, ~$150 Two important learnin

Try it here: Model: Qwen 3.8 27b Q8_X_KL Unsloth Hardware: 3 x RTX3090 Harness: DeepSeek Harness Prompt: /goal I want you to create a **JavaScript + Node.js WebGL project** that renders a highly realistic real-time ocean in the browser. Use **JavaScript only, no TypeScript**. You may use WebGL2, GLSL, and Three.js. The ocean should include realistic waves, vertex displacement, Fresnel reflections, sun highlights, sky/environment reflection, foam/whitecaps, horizon treatment, atmospheric effects,

Article URL: Comments URL: Points: 80 # Comments: 39

Nvidia worker indicted after Jensen Huang scolded Supermicro for AI server smuggling.

At TechCrunch Disrupt 2026, Replit CEO Amjad Masad will share his perspective on the future of programming and Replit's role in developing it.

Hugging Face has reportedly been fielding acquisition offers that would value the company at around $13B. But with the founders' feeling of responsibility to community, doubts arise as to whether a sale will happen.

Recent omni-modal large language models (Omni-LLMs) show great potential as real-time video assistants, which continuously perceive environments and guide users to achieve specific goals. Unlike traditional passive video understanding, interactive assistants should actively combine visual states, user goals, and prior knowledge to provide effective help. Evaluating this is rather challenging, as the model's unpredictable response dynamically changes the user's subsequent actions, which static of

8月24日消息，阿里巴巴达摩院联合中国医科大学附属盛京医院等机构研发出肝癌诊断AI模型DAMO LiON

arXiv:2608.20389v1 Announce Type: new Abstract: A production agent harness must discover and rank, from a growing library of skills, the one most appropriate for a user's task. At small scale this selection happens in context: the LLM planner chooses among skill representations exposed in its system prompt, without an explicit embedding-based retrieval step. We treat this in-context selection as the small-N counterpart to embedding-based skill retrieval at scale, and present a case study of how

arXiv:2608.20378v1 Announce Type: new Abstract: Safety alignment in Large Language Models (LLMs) is often superficial, relying on refusal mechanisms that trigger only at the final stages of generation without erasing the foundational knowledge of harmful concepts acquired during pretraining. This study demonstrates that this architectural disconnect leaves models vulnerable to Semantic Camouflage -- adversarial attacks that wrap harmful intent in benign narrative contexts (e.g., creative writing

In this tutorial, we explore a LabPlot-inspired scientific data analysis workflow in Python while preserving the structure and terminology of LabPlot’s aspect tree, analysis kernels, plotting system, and project model. We build reusable components to import tabular data, compute descriptive statistics, smooth and differentiate signals, perform Fourier analysis and filtering, detect peaks, integrate curves, reduce […] The post Scientific Data Analysis with LabPlot in Python: Signal Processing, Sp

On-policy distillation (OPD) transfers teacher capabilities by supervising trajectories sampled from the student's own policy, yet its generalization behavior remains poorly understood, as most studies evaluate OPD on a single domain and on benchmarks close to the training data. We present a controlled study that varies one generalization factor at a time, from in-domain distribution shifts to cross-domain transfer and the multi-teacher setting. We find that OPD transfers a teacher's reasoning b

I commented on another Qwen 3.8 27B post that I was frustrated getting anything to work. You all gave some great comments. I nuked openwebui and straightened out my llama.cpp docker config. 1 hour of work and I have a model I can chat with, connected to my HomeAssistant server, which I have already updated dashboards with a short prompt and a screenshot (wtf vision built in?) Guess all I needed was the right push. I bought several GPUs in 2023 in impulse purchases for Folding@Home, but have alwa

Given OpenRouter.ai was snapped up by Stripe, who do we think would go after the "GitHib" of AI models? It is a big chunk of change they are looking ($13B). Apple may be a contender to give them a real chip in the AI race, given how they are focused on local AI execution.

Article URL: Comments URL: Points: 144 # Comments: 27

Credit to Twitter Post

As the memory shortage continues to cause trouble for hardware makers, Amazon says it is now being forced to pass on the costs to its consumers.

General Intuition, the startup building a foundation model that trains generalized AI agents how to move through space and time, is in talks to raise at a $6 billion pre-money valuation from new investors including Valor Ventures, Point72 Ventures, and Seven Seven Six.

GrapheneOS, an open source version of Android that prioritizes security and privacy, has detailed its plans for supporting Motorola smartphones. Official support is set to arrive next year, starting with traditional flagships, before rolling out to Motorola's foldable phones and perhaps cheaper models, eventually. In a Mastodon thread, the GrapheneOS Foundation announced that it will […]

framework that folds aggregate human movement into text-based place embeddings. Language models describe what a place is; they miss how it is used. ME-POIs encodes each visit as a contextualized vector and aligns it with one learnable prototype per POI through contrastive learning, then transfers visit distributions from data-rich anchors to the long tail across three spatial scales. Across five map-enrichment tasks on Los Angeles and Houston mobility data, adding ME-POIs improved 34 of 35 model

8月24日，阿里巴巴视频生成大模型Wan3.0正式上线。

Real-world image search queries are multimodal and compositional: ``find this shirt in pink'' specifies an entity to retain, an attribute to modify, and context to ignore. Yet existing re-rankers either compress such multifaceted relevance into an opaque embedding or rely on free-form chain-of-thought that easily omits or hallucinates fine-grained constraints. Drawing on rubric- and checklist-based evaluation from NLP, we recast multimodal image re-ranking as a semantic constraint satisfaction p

The memory layer that decides what's worth remembering Discussion | Link

The AI runtime that remembers, learns, and acts everywhere Discussion | Link