ENFR
8news

Tech • IA • Crypto

TodayBriefingVideosTop 24hArchivesFavoritesTopics

DeepMind ASI roadmap, Midjourney Medical, SSA 1M tokens

AISaturday, June 20, 2026· 9 videos

Briefing

Audio player
0:00 / 0:00

DeepMind outlines path to ASI

Google DeepMind published a 60-page roadmap led by Shane Legg detailing a path to artificial superintelligence (ASI) within a decade. The plan emphasizes four routes: scaling GPU compute, multi-agent systems, new architectures, and recursive self-improvement. It reframes progress as system-level coordination rather than single-model gains. The proposal intensifies debate over whether LLMs alone can reach AGI.

Midjourney launches Medical imaging division

Midjourney introduced Midjourney Medical, expanding beyond generative art into healthcare hardware. Led by David Holz, the effort builds on his Leap Motion background in sensing systems. The move signals a long-term strategy to pair AI with proprietary devices. It positions the company in competition with advanced medical imaging innovators.

Ultrasound scanner pushes atomic precision

The prototype uses roughly 358,000 ultrasonic sensors arranged in a 70 cm ring to scan the full body. It captures up to 17 GB per second of data, reconstructing high-resolution internal imagery. The system detects motion at near atomic-scale sensitivity using advanced wave analysis. If validated, it could redefine non-invasive diagnostics and 3D medical imaging.

SubQuadratic unveils SSA attention breakthrough

SubQuadratic introduced Subquadratic Sparse Attention (SSA) to overcome transformer scaling limits. The method reduces compute from quadratic to near-linear by selecting semantically relevant token interactions. At 1 million tokens, it cuts compute from 252 petaflops to 3.9, a roughly 64× reduction. Benchmarks show 56× speedups over FlashAttention-2 on NVIDIA H100 at extreme context lengths.

France replaces Palantir with ChapsVision

France’s DGSI is phasing out Palantir, used since 2016, in favor of ChapsVision. The domestic platform Argonos will handle intelligence analysis across classified and open data. The shift reflects growing concern over reliance on U.S. technology and data sovereignty. It underscores Europe’s push for independent AI infrastructure amid geopolitical tension.

Musk and Thiel rethink data centers

Elon Musk is targeting orbital AI data centers at 600 km altitude, aiming for 1 GW by 2027 and eventually 1 TW. The concept relies on continuous solar energy and Starship deployment for scale. Meanwhile, Peter Thiel-backed Pantalassa raised $140 million for ocean-based platforms. These floating systems use wave energy and seawater cooling to bypass terrestrial limits.

Hermes framework reframes AI agents

Hermes is gaining traction as an orchestration layer rather than a standalone model. It coordinates tools, memory, and sub-agents while relying on models like OpenAI, Claude, or GLM 5.2. Architectures often include orchestrators, validators, and extraction agents running in parallel. Its effectiveness depends more on system design than model choice, especially in high-risk workflows.

Computer-use agents hit reliability limits

AI systems that operate software environments are advancing rapidly but remain fragile. They rely on perception loops using screenshots, text, and action generation across interfaces. Key challenges include grounding, UI interpretation, and long-horizon task consistency. Scaling these agents will require better multimodal perception and robust orchestration layers.

Videos covered

Previous briefings · AI