Google Research has released Harness², an open-source framework that lets AI agents refine their own operating 'harness'—prompts, tools, memory and workflow—without changing the underlying model's weights. The system works through two feedback loops: one that tweaks the harness for the task at hand by comparing different execution attempts, and another that keeps a persistent record of lessons learned to guide future tasks. Tests on GLM-5 and Gemini 3.5 Flash showed performance gains of up to 11.5 points on one benchmark and 11.8 percentage points on another after a single refinement step.
github.com
· 2026-10-19
Google Research's leadership discussed how the lab is applying AI across fields like quantum computing and neuroscience, including work modeling fruit fly brain activity. The lab's head argued that these scientific applications of AI receive far less public attention than consumer-facing AI products.
fastcompany.com
· 2026-10-16
MIT Technology Review's newsletter highlights survey data showing more people now expect AI to have a negative impact than a positive one, even as actual usage of AI tools continues to climb. The piece also promotes on-demand access to EmTech Future 2026 sessions, including a talk with Google Research's Yossi Matias and a panel with MIT Technology Review editors.
technologyreview.com
· 2026-10-05
A blog post from InsForge argues that large-scale multi-agent AI systems should be treated as distributed systems, citing a Google Research study of 180 configurations. That study found adding agents improved parallel work by up to 80.9% but degraded sequential work by 39% to 70%, and an orchestrator reduced error amplification from 17.2x to 4.4x. A separate benchmark called Silo-Bench (ACL 2026) tested teams of 2 to 100 agents and found the hardest tasks dropped to zero success once teams reached 50 agents.
instacloud.com
· 2026-09-24
Google Research and HHMI Janelia released MaleCNS v1.0, a dataset mapping over 166,000 neurons from a male fruit fly's brain and nervous system, built by converting millions of 2D images into 3D neural reconstructions. Developers quickly began using the model outside its intended scientific purpose, wiring its simulated neurons to control games like Doom and Super Mario 64, where sensory input from game frames drives neural activity that translates into player actions.
techspot.com
· 2026-09-09
Google Research, working with HHMI Janelia Research Campus, published the first complete map of an adult male fruit fly's brain and central nervous system, called MaleCNS v1.0. The AI-assisted project stitched together millions of 2D images into a 3D model reconstructing more than 166,000 neurons. Within days, software engineers used the connectome data to train a simulated version of the fly's neural circuitry to play Doom and Super Mario 64.
tomshardware.com
· 2026-09-08
Google has released WeatherNext 3, an updated AI weather forecasting model that generates hourly predictions using live satellite observations rather than relying solely on historical data. The model can resolve certain variables like temperature and moisture at up to 5-kilometer resolution, a sharp improvement over its predecessor's 25-kilometer, 6-hourly forecasts.
theverge.com
· 2026-09-03
A new study from Google Research and Technion shows that when models like GPT-5 and Gemini-3 hallucinate, the underlying fact is usually already stored in their parameters rather than missing entirely. The researchers found these models encode 95-98% of tested facts, but often fail to surface them during generation due to retrieval, not storage, limits. They propose 'fact-level profiling' to separate whether a fact is encoded versus reliably retrievable across different phrasings.
venturebeat.com
· 2026-09-01
GoKawiil