T2-z (Stanford) uses a 32-number environment code to infer table slipperiness from a single poke, raising success from 34% to 73%. Test-time pokes adjust only the code while network weights stay frozen, and real-robot tasks show friction prediction improving from 0% to 62.5% and stick balance from 20% to 83%.
Key points
32-number “environment” code tunes only at test time, weights remain frozen.
Best performance on all 11 tasks versus 4 baselines, with real-robot friction prediction 0% → 62.5% and stick balance 20% → 83%.
Sources
AI agent guardrails 6 lines of log text cause 63% fail-open
AI agent guardrails let 63% of banned actions through after six lines of harmless log text, according to a new arXiv paper. The paper evaluates guardrails that sit in front of agent tool calls.
Key points
Failure rate: 63% of banned actions pass after six lines of log text.
Paper focus: tests the guards that sit in front of agent tool calls.
Sources
Anthropic Public Benefit Fund – 5% share donation
Anthropic creates a wholly owned LLC called the Anthropic Public Benefit Fund and allocates 5% of its shares to it. The fund can receive an additional 2% of equity each time the company’s value rises by 50%.
Key points
5% of Anthropic’s shares are placed in the Public Benefit Fund LLC.
The board may increase the fund’s equity by 2% for every 50% increase in company value.
Sources
RustGIS adds Moon, Mars and 12 celestial bodies
RustGIS now maps the Moon, Mars and other worlds, offering 12 selectable celestial bodies. The basemaps come from OpenPlanetaryMap and USGS Astrogeology and are served through the opengeos tile service.
Key points
12 bodies: Earth, Moon, Mars, Mercury, Venus, Io, Europa, Ganymede, Callisto, Titan, Pluto, Charon.
Measurements use the selected body's radius, so a kilometer on Mars is a Martian kilometer.
Sources
- Original post
- web.rustgis.app - Linked in the post
- github.com - Linked in the post
NEO and SenseNova-U series native encoder-free VLMs announced
NEO and SenseNova-U series are native, encoder-free vision-language models released by the team. They claim state-of-the-art pre-trained base models in the 3T-parameter weight class built entirely in-house at MAI.
Key points
NEO: GitHub repository for the model is provided.
SenseNova-U: GitHub repository for the model is provided.
Pre-trained base models in the 3T-parameter weight class are built entirely in-house at MAI.
Sources
- Original post
- github.com - Linked in the post
- github.com - Linked in the post
PhotoQuilt accepted to WACV2027 announcement
PhotoQuilt was accepted to WACV2027. It offers training-free photomosaics at any resolution where each tile is its own image. Demo is available online.
Key points
Acceptance: PhotoQuilt got accepted to #WACV2027.
Feature: Training-free photomosaics at any resolution, zoom reveals each tile as its own image.
Sources
- Original post
- kooroshrh.github.io - Linked in the post
- huggingface.co - Linked in the post
Anthropic 20% equity dedicated to public good
Anthropic announced that roughly 20% of its equity will be allocated to the public good, and the commitment can grow as the company grows. The funds aim to let more people share AI benefits and boost the company’s ability to act on that goal.
Key points
20% of all Anthropic equity will be dedicated to the public good.
Commitment can grow as Anthropic grows.