Inkling 2 Small to release weights next week, reporting ARC-AGI-2 81.7% at $0.27 per task and ARC-AGI-1 96.5% at $0.08 per task. The model outperforms every open-weight model evaluated on both ARC-AGI-1 and ARC-AGI-2.
Key points
ARC-AGI-2: 81.7%, $0.27/task
ARC-AGI-1: 96.5%, $0.08/task
Full weights release planned for next week
Sources
Anthropic AECI scores grew 1.9x since Oct 2025
Anthropic’s AECI benchmark scores have increased by a factor of 1.9 since October 2025, with a 95% confidence interval of 1.5–2.8. The quoted tweet notes that benchmarks are only one piece of evidence for acceleration and that mapping scores to real-world capability is unclear.
Key points
AECI scores for Anthropic’s models have grown at 1.9× since Oct 2025.
95% confidence interval for the growth rate is 1.5–2.8.
Sources
BioBasicsBench benchmark released with 100 wet-lab tasks
BioBasicsBench is a new benchmark that uses 100 real wet-lab tasks to test AI agents on raw data interpretation, error identification, and unsupported conclusions. It was created with Ligo, a biotech company focused on enzyme design. The benchmark aims to show where models need improvement and track gains for scientific progress, biosafety, and biosecurity.
Key points
BioBasicsBench draws on 100 tasks from real experiments.
Developed with Ligo, a biotech company focused on enzyme design.
Tests whether agents can interpret raw lab data, identify errors, and recognize when evidence doesn’t support a conclusion.
Sources
- Original post
- openai.com - Linked in the post
FMF Org framework to assess AI jailbreak severity announced
FMF Org and partner firms released a framework to identify when bypassing AI safeguards could cause serious cyber harm. Sasha Baker said the framework is an important step toward a common methodology for evaluating and comparing the severity of cyber jailbreaks.
Key points
Partnership with @fmf_org and its member firms on a new framework.
Framework evaluates severity of cyber jailbreaks, capabilities unlocked, and ease of use.
Sources
proto_ide designer IDE announced with early access signups
proto_ide is a designer’s IDE that mixes Claude Code and Cursor features with design capabilities. Early access signups are open.
Key points
3/4s of designers work in coding agents rather than design tools.
Early access signups are open.
Sources
- Original post
- prototypes.fun - Linked in the post
GoogleDeepMind Science team cryptographic advances factor 1024-bit RSA
GoogleDeepMind Science team announced AI agents achieved two cryptographic advances in weeks, including bypassing older digital signature protections and factoring legacy 1024-bit RSA keys in bulk. The results do not affect modern schemes such as RSA 2048.
Key points
AI agents made two significant cryptographic advances in weeks.
They can bypass older digital signature protections and factor legacy 1024-bit RSA keys in bulk.
Sources
Huawei AI data center Horinger capacity 7% of Colossus 2
Huawei’s AI data center in Horinger, Inner Mongolia, is one of the largest in China by AI computing capacity, with 7% of the capacity of the world’s largest AI data center, Colossus 2 in Tennessee. The capacity estimate comes from a new report using satellite imagery and public sources.
Key points
Huawei’s AI data center in Horinger, Inner Mongolia, is one of the largest in China by AI computing capacity.
It has 7% as much computing capacity as the world’s largest AI data center, Colossus 2 in Tennessee.
Sources
Holon-S1 launch discussed on Health Care Network episode
Holon-S1 is SophontAI’s new product mentioned in today’s Health Care Network episode. The segment begins at 52:30. The broadcast also references midterm impacts on healthcare and guests Tanishq Abraham and Andrew Brackin.
Key points
Product: Holon-S1 announced by SophontAI.
Episode segment start time: 52:30.