2026-09-22

My comment on MCP was always a bad idea? — Hacker News. This article entirely misses the value that MCP brings today. Sure, there's almost no reason to use MCPs if you are running a full-blown terminal agent (Claude Code, Codex, Meta Muse, OpenClaw etc) with unfettered internet access - just let it call APIs directly. If you want to operate something that's less YOLO than that, you'll find yourself wanting: Control over exactly which external services it can access A way to handle authentication

Quote: DeepSeek is training a 2T-parameter model and plans to eventually build an 8T-parameter model. Current DeepSeek models: Flash parameter count of 552 billion Pro: 1.6T (trillion) total parameters with 49B (billion) activated weights per token Mythos / Fable is estimated to be 10T parameter count.

It's not a Qwen3 finetune, it's actually its own fully custom architecture. No Llama.cpp support yet sadly (Also note that this model is NOT post-trained like Qwen3.5/3.6)

The definitive Jev podcast with its lead creator.

The expanded form of a testimony I prepared for Congress.

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 RAND thinks the best AI strategy for the US is to keep all options open:…It’s not clear exactly what is optimal, so the US should spend […]

arXiv:2609.21096v1 Announce Type: new Abstract: In this work, we examine the topology of information flow patterns within attention graphs to effectively distinguish hallucinated from non-hallucinated responses. We analyze the Forman-Ricci curvature to identify structural patterns indicating information bottlenecks in attention graphs. We then introduce a method that captures both semi-local and global information-flow characteristics of attention heads associated with hallucinated responses. We

arXiv:2609.21157v1 Announce Type: new Abstract: Large Language Model (LLM) agents are increasingly being explored for chip design, but most existing approaches operate directly at RTL. We ask whether agents can design better chips by leveraging higher-level abstractions. We compare Direct RTL Design, Agent-based HLS Design, Post-Compiler HLS Refinement, and Post-HLS RTL Refinement, and combine Agent-based HLS Design with Post-HLS RTL Refinement as Agent-based HLS with RTL Refinement (AHRR). We u

arXiv:2609.20974v1 Announce Type: new Abstract: In Mixture-of-Experts language models, the router typically selects and weights experts based on the token's hidden state, utilizing limited contextual information. We propose Attention-Aware Routing (AAR), which augments the router with temporal and spectral features extracted from a sliding window of attention weights that represent a summary of the model's contextual state, disentangled from the hidden state. Keeping the base transformer entirel

Using GPT-5.6, V7 turns scattered company files into context agents can use to complete complex, source-linked work.

Hey everyone! It has been quite a while since the last SupraLabs model - but today we've something special for y'all: Supra2-IMG It's a 100M parameter DiT text-to-image model trained entirely from scratch in under 10 hours on a single H100 on Runpod. It can generate state-of-the-art quality images in 256x256 pixels resolution. Samples: These samples are NOT cherry-picked! Sampling: seed 0, steps 50, cfg 3.0; same settings for every image. If someone here is interested in the prompts, I can give

Sorry for the pretentious name, I know, I know.. It just contains all the pieces I would like to see a AGI model to have, and I can't stand the temptation. Before throwing rocks at me, please take a glance at the Readme, and I hope it will cover your mood a little bit. So, first of all it does work and you can see the sample from the whole training run here: Here is the scaling law graph I have so far, and it looks very promising: The model was built under my deep dissatisfaction so we cannot re

I’ve been working on making small models more capable at agentic coding and work, because most people in the world don’t have the sort of hardware needed to run 3.8-27B, or even 35B-A3B or 9B dense, and I want to extend local agentic coding capability to less privileged users. This quant can be run on a smart phone or older gaming laptop, and can solve real coding problems autonomously in a way I have never seen or measured for this model class. Spark-X2.5-4B is already around best-in-class for

Hundreds of flights were cancelled or delayed on Monday after construction crews in New Jersey accidentally cut a Verizon fiber cable used for air traffic control. FAA Administrator Bryan Bedford said that a circuit failure led to the discovery of the severed fiber cable, ABC News reports. The outage hammered airports across the Northeast, causing […]

The group won't be given leeway to slow down or redirect OpenAI's ongoing mathematical research.

Put together some notes on Jev and the new category of system one aka decision models simonwillison.net/2026/Sep/21/...

We define OmniVChat (Omni Video Chat) as the task of native audio-visual dialogue between a user and an omni model. In OmniVChat, omni models directly and simultaneously receive audio and video from a user and return text. The user's query is embedded in the audio and video, without a separate text question, external captioning, or speech recognition. Direct audio-visual input reduces external latency and computation while preserving perceptual cues. However, research on OmniVChat faces two cons

Agent trajectories record what an agent does and what happens next. Yet standard supervised fine-tuning (SFT) applies loss only to agent-authored action tokens, using environment observations as context but not as prediction targets. We ask whether this convention provides the best initialization for subsequent reinforcement learning. We introduce ActObs, which also supervises the observation tokens already present in each trajectory. Although deployed agents never generate observations, learnin

As corporate AI adoption continues to grow, enterprise-grade LLM agents are being deployed into sensitive contexts such as hiring, healthcare, and finance. In these contexts, compliance with rules specified in an agent's system context is a first-order legal concern. Currently, no evaluation framework systematically measures which LLM models tend to violate compliance rules, especially under pressure from a persistent user, a hurried manager, or circumstances where violation is convenient or att

arXiv:2609.20981v1 Announce Type: new Abstract: Autoregressive (AR) models suffer from local greediness, while diffusion language models (DLMs) often lack the strict causal structure required for reasoning. To combine the advantages and overcome the drawbacks of the dual, we propose Causal Latent Revision (CaLR), a framework that reformulates reasoning as constrained latent optimization. By adopting a causal topology matrix (CTM) from an expert model and implicit differentiation, CaLR performs g

arXiv:2609.21139v1 Announce Type: new Abstract: Replacing attention in a pretrained language model is a compatibility problem: a plausible substitute may alter representations expected by later layers. TinyCeNN-LM introduces a \emph{quality-gated post-training conversion} framework using CeNN-inspired cellular-recurrent layers with bounded local processing, compact recurrent memory, routing, fusion, and accept-or-rollback validation. Three implementations are studied: Integrated Memory, MemoryFu

arXiv:2609.21113v1 Announce Type: new Abstract: Fine-tuning has emerged as a widely adopted approach for adapting LLMs to a variety of downstream tasks. However, how it reshapes their internal mechanisms remains poorly understood. To address this, we investigate how fine-tuning alters internal representations in LLMs, including attention patterns and layer-wise activations, and examine whether these changes are linked to task-relevant components identified by EAP (e.g., attention heads and logit

The headlines right now are full of stories about AI models "escaping their sandboxes" and literally killing all humans, lol. I've even heard several commentators and writers say that AI escaped an "Air gap". But that is SO WRONG. It's actually TOTALLY WRONG. *To be clear, not a single one of these sandboxes was actually air-gapped.* That's a crucial computer science fact. An air gapped sandbox would require *ZERO* cables and network interfaces. It would also require absolute physical isolation.

In summer 2025, when we started building fluado , one of our aspirations was to run outside US Big Tech. But in the big scheme of registering and building a company, it felt "not so important", so we started on Google Cloud because we were familiar with it. It felt icky, but it was one less thing to think about. Now one year later, we realised we had deepened our dependency on US providers: Cloudflare, GCP, Supabase, GitHub, Slack, Google Workspace, etc. At the end of August we set out to move a

OpenAI is working with an independent Advisory Group on Mathematics and Artificial Intelligence to guide the review and communication of emerging AI results.

Meta’s new AI agent Muse has racked up more downloads and daily active users in the U.S. and Canada than ChatGPT did over the same period after its mobile debut, according to new estimates from Appfigures.

Article URL: Comments URL: Points: 103 # Comments: 37

Officials discussed setting up a mechanism for the two countries to notify each other of AI incidents which could threaten national security.

9月19日，第一届中国网络空间安全大会（CCSC 2026）的高水平专题论坛在安徽合肥正式举办。

arXiv:2609.21192v1 Announce Type: new Abstract: Organizations deploying agentic artificial intelligence must determine more than whether a model is trustworthy; they must establish what to validate, control, and observe for a use case to deliver its intended outcome while meeting applicable obligations. This paper proposes AI-GRACE (Agentic Intelligence-Governance, Risk, Assurance, Controls, and Evidence) as a use-case operationalization framework connecting organizational governance with techni

arXiv:2609.21181v1 Announce Type: new Abstract: The Abstraction and Reasoning Corpus and related benchmarks evaluate whether AI models can solve novel reasoning tasks, but often leave unclear whether success reflects inference of the intended underlying rule or reliance on shortcuts. We address this gap by studying test-time task embeddings in Vision ARC (VARC), a model in which a pre-trained backbone is complemented by a trainable embedding representing the transformation rule. In the original

I’m currently a Staff/TL at a FAANG company and moving to a similar TL role elsewhere. In my current role, I’ve built enough trust and context that setting direction and influencing decisions feels natural. My concern is how to rebuild that quickly when joining a new org with zero existing influence. For Staff+ engineers who’ve made a similar move: what concrete things helped you ramp up and establish credibility? Looking for specific actions—e.g. 1:1s with other TLs in week 1, picking an early