Skip to content
Drix10 Blog

Blinkit Ambulance charges non-emergency transfers ₹2,500

, 8 items in Computer Vision and AI Applications, 4 min read

In this digest (8 items)

Blinkit Ambulance will keep emergency transfers free and start charging ₹2,500 for non-emergency transfers. The charge will go to Eternal Foundation to help scale the service. The change aims to reduce non-emergency use of fully equipped ambulances.

Key points

  • Served 20,000 emergency cases (42,000 cases total) in the last 2 years with no charge.

  • Non-emergency transfers will be charged ₹2,500 after the trip, matching private ambulance rates.

Sources

Gemini Omni 1.2 Flash adds 15-second video, $0.07/s

Gemini Omni 1.2 Flash adds 15-second video generation and editing, and costs $0.07 per second for 720p. It also upgrades character speech and reduces visual glitches.

Key points

  • Video limit: up to 15 seconds for generation and editing.

  • Pricing: $0.07 per second for 720p video, a 30% price drop.

Sources

You don't need long context – compaction cuts context to 8K tokens

You don't need long context shows frequent compaction lets Claude Opus 4.5 keep accuracy while dropping median peak context from 32K to 8K tokens. Opus scores 70.2 on SWE-bench versus 70.0 with full history and reaches 74.8 with end-of-conversation compaction. Later tokens cost up to 7× more on frontier open-weight models.

Key points

  • Claude Opus 4.5 scores 70.2 on SWE-bench versus 70.0 with full history.

  • Median peak context reduces from 32K to 8K tokens with frequent compaction.

  • Simple compaction at conversation end raises Opus score from 70.0 to 74.8.

  • Later tokens cost up to 7× more to generate on frontier open-weight models.

Sources

TAPESTRY single-step denoising improves WAM performance 12 Hz 9GB VRAM

TAPESTRY shows that reducing diffusion steps improves WAM performance, with single-step denoising delivering smooth actions at 12 Hz using 9GB VRAM. The finding contrasts with more steps causing action jitter.

Key points

  • Reducing diffusion steps can improve WAM performance.

  • Single-step denoising runs at 12 Hz and uses 9GB VRAM.

Sources

Aeva expands Optical Connectivity product portfolio

Aeva expands its Optical Connectivity product portfolio for the AI data center market, delivering performance and silicon photonics innovation. The company reports growing engagements with AI compute companies and optical solutions providers, including commercial negotiations with a second customer.

Key points

  • Expansion: Optical Connectivity product portfolio for the AI data center market.

  • Engagements: Commercial negotiations with a second customer and broader AI compute and optical solutions partners.

Sources

GEN-1.5 improves fine-grained dexterity tasks

GEN-1.5 improves fine-grained dexterity tasks, handling millimeter-scale details like thinner O-rings. The models are getting better at the nitty gritty.

Key points

  • Model: GEN-1.5

  • Tasks test dexterity and perception of millimeter-scale details

Sources

Aria music transformer represents harmony and can be steered

Aria, a symbolic music transformer, shows that mode and key become progressively decodable across layers, and intervening on these representations can causally redirect harmonic resolution. The work demonstrates internal harmonic structure representation and the ability to steer music generation.

Key points

  • Mode and key become progressively decodable across layers.

  • Intervening on these representations can causally redirect harmonic resolution.

Sources

LightShop launch at UIST 2026 announcement

LightShop is an interactive computational system for creating high-precision light paintings, presented at UIST 2026. It combines physically accurate lighting with computational precision.

Key points

  • Launch: LightShop presented at UIST 2026.

  • Capability: Enables creation of high-precision light paintings.

Sources