2026
10,564 articles · 132 days · 10 categories · avg 80 per day
Xiaohongshu speeds up its sparse-attention indexer: DoSA halves TTFT
Xiaohongshu's Dots team proposes DoSA, which prunes about 97% of candidates in the indexer, keeps the same Top-K and cuts time to first token by half.
A Company's Immune System and 7,522 Hours Saved: What Actually Worked When Rolling Out AI at Alpina
Deploying AI at a 300+ person company hinges not on choosing a language model, but on overcoming organizational inertia and employees' fear of losing their…
AstraZeneca to pay insitro up to $1.5 billion for AI-driven MASH drug discovery
AstraZeneca and insitro signed a deal worth up to $1.5 billion: the Virtual Human platform will search for drugs for MASH and cardiomyopathies. It is one of the…
RoboCasa-Shift: a benchmark shows where VLA robots break in the kitchen
WorldStream Builds an Endless 3D World and Remembers Every Player Edit
Qwen-Audio-3.1-ASR: 30 Languages, 16 Dialects, and Diarization in a Single Model
A Team of AI Researchers Found a Linear Attention Design That Beat GDN2 at 1.3B
DiffResearch: A Diffusion LLM Makes Deep Research About 6x Faster
The open-source DiffResearch framework swaps an autoregressive model for a diffusion one and speeds up deep research roughly 6x. Its dFast mode adds another 1.99x…
Of 1 598 RL Papers, 37 Made It Into Frameworks: An Audit of Seven Open-Source Stacks
An audit of commits in seven open-source RL post-training frameworks: of 1 598 methods, only 37 reached code, while paper volume grew fivefold against the same 5–7…
Researchers teach LLMs to adapt to dialect: the CAP method and 10 model families
Agent4Sci: Claude Opus Beats Records on 20 Scientific Benchmarks by 10.4%
MultiAgent-OPD Trains a Whole Team of Agents End to End and Scores 38.68% on Seven QA Benchmarks
DuoLM Speeds Up LLM Recommendations 60x: 26,400 Scores per Second on a Single H100
Together AI Launches Post-Training: RL and SFT on B200 at $9 per GPU-Hour
Together AI has opened Post-Training: RL and SFT on dedicated GPUs, LoRA and full weights, and a one-line migration from Tinker. B200 costs $9 per GPU-hour for the…
UniPat AI to Put 100 Million Yuan Into Fellows Program for Scientists
Chinese startup UniPat AI is launching UniPat Fellows: 100 million yuan over one year, 500,000 yuan in API credits per researcher, plus data, experts and compute.
USTC Releases "Machine Scientist 2.2": AI Moves From Running Experiments to Creating Substances
USTC has unveiled its Machine Scientist 2.2 series. The claim: the system doesn't just run experiments but autonomously creates new materials.
Google Picks Intel Xeon 6 Processors for Training Neural Networks
Safety or Corporate Interest: Why Is Anthropic Holding Back the Mythos Release?
Amazon Values Its Chip Business at $50 Billion and Prepares to Sell Externally
Google and Intel Expand Their AI Infrastructure Partnership
Claude Cowork's File Autopilot: Three Examples of Hours Saved
Claude Cowork turns hours of routine work into minutes. With direct access to local files, a single prompt lets Cowork build a plan and execute the task — from…
Black Forest Labs: A 70-Person Team Takes On Silicon Valley's Giants
Black Forest Labs, known for its image generation success, is widening its scope. Its next target is bringing AI into the physical world and robotics.
The 'Verify Before You Refuse' Rule: How to Stop an AI Agent From Giving Up at the First Error
The Complete Guide to ModelScope: From Model Discovery to Export
OSGym: A New Framework for Training AI Agents at Scale on the Cheap
Meta Unveils Muse Spark: A Multimodal AI Model With Visual Reasoning
OpenVLHarness Gives Multimodal Models Memory and Tools: +9.7–12.6 Points
OpenVLHarness, a harness with tools and persistent multimodal memory, lifted the average score of eight MLLMs by 9.7–12.6 points on 22 benchmarks, using at most 37%…
Stateful Harness Lifts World Model Accuracy on Hidden State from 32% to 92%
Researchers split rule execution from video rendering: Stateful Harness with a frozen MiniMax-H3 gets the correct outcome in 92% of WorldState-Bench pairs, versus…
Hybrid LLMs: Recurrent Layers Suppress Attention, and Staged Training Helps
Researchers describe Competing Learning: in hybrid models, recurrent layers learn retrieval faster and suppress attention. They tested staged training on models up…
Halcyon beats LlamaGen and JiT-T2I with 18–275x less pretraining compute
MemT-1.0 builds memory into the model itself: 66x smaller storage at 128K
Artificial Analysis launches HealthBench: GPT-6 Astra leads with 66.0%
