China's Kepler Robotics shared a video of its Forerunner K2 humanoid with next-gen hands With an 11-DOF, each 'NimbleMaster' hand can handle a 15 kg payload It also has a 25-element tactile sensing matrix per finger to better grasp different objects

OpenAI added native image generation to GPT-4o and Sora The move creates an integrated system, where ChatGPT understands conversation context and creates/edits visuals with precision Rolling out to Free, Plus, Pro, and Team users

Big Figure update: our humanoids can now walk naturally like humans! We introduced an end-to-end neural network for sim-to-real actions—tapping RL and a high-fidelity physics simulator to enable Figure 02 to walk with a smoother gait and increased speed

Alibaba's Qwen team dropped two new AI models: —Qwen2.5-Omni-7B for processing across text, audio, image, and video in real-time —QVQ-Max to analyze and reason about visual information across images and videos, including blueprints and geometry problems

Ideogram released v3.0 of its image generation model The model outperforms Imagen 3, Flux Pro 1.1, and Recraft V3—excelling at complex layouts, logos, and typography Also supports style references to control generations with access to free users

Reve emerged from stealth with Reve Image 1.0 (in preview) The model ranks #1 in Image Arena, outperforming Imagen 3, Midjourney v6.1, and Recraft V3 It shows impressive prompt adherence and long text rendering—with natural language editing!

Microsoft 365 introduced two new thinking-capable AI agents to help users with workplace tasks: —Researcher: To research reports using different knowledge sources, including the web —Analyst: To extract insights from data, much like data analysts

Perplexity launched answer modes The feature enhances AI searches on specific verticals (like travel and shopping) with entities like images, videos, and cards Each mode will support commercial transactions, allowing users to act right where they are!

Tencent released Hunyuan T1, a reasoning AI that uses the industry's first Transformer-Mamba architecture The hybrid model matches or surpasses DeepSeek R1 and OpenAI’s o1 & GPT 4.5, with 2x faster speed Priced at $0.14 and $0.55 per million I/O tokens

DeepSeek released an updated version of its V3 model —V3-0324 is a 641GB model but capable of running on high-end PCs —Promises a major boost in performance across the board —Uses MoE, activating only 37B params/token —Open-source under MIT

South Korean researchers developed a gummy-like robot using water droplets coated with Teflon particles PB can shape-shift, pick up debris, move across water, and merge with another similar robot

We're hiring for hundreds of roles @Figure_robot: > AI Engineers (many) > Staff Security Engineer > HMI Design Lead > System Integration & Test (many) > Legal (many) > Manufacturing (many) Apply here:

That's it for this week's AI and Robotics breakdown. I share the latest research every week, so follow me @adcock_brett for more. If you found this valuable, consider a like/retweet to spread the word.

🎯 Holds breath in sniper mode... 🎮 X-Air Pro's clutch features: 💨 Breathable mesh for sweaty 1v5 moments 🔫 Perfect height for that winning headshot ⚡ 5D armrests for precise aim control 😅 "Why did I solo queue" comfort mode 🎯 For those who main AWP and live dangerously! 🪑✨ Aim better: 🎪 Headshot deal: Snipe $20 off! #AndaSeat #CSGO #Valorant #ESports #Gaming

How does that compare to F.03?

Fingers are always hard for AI. It’s like it’s thinking “why are there so many of them?” Maybe there’s a better design.

with Figure v2 at 25 kg per hand, what is your target kg per hand longterm? and can they lift items together? looking to plant lots of trees 🌲 🌲