Dark-mode chat bubbles were erasing darker skin-tone glyphs. Google Design fixed contrast across 3,977 Noto 3D characters — AI audit plus hand-sculpted rim lighting so meaning stays visible.
In dark mode, darker skin tones have often vanished against deep gray and black chat bubbles. When the silhouette disappears, so does the meaning.
Nobody should have to choose between using their actual skin tone and having their messages be legible.
To fix this across 3,977 characters, we paired an AI-powered contrast audit tool with hand-sculpted rim lighting to preserve rich tone and clarity on every screen.
Read how we built it → https://t.co/xWQRZZeFho
#accessibility #3demoji #noto3d
Google shipped Gemini 3.8 Live with Live Avatar: live dialogue paired with low-latency streaming video for enterprise agents. Async tool calls keep running while the avatar stays present; lip-sync across 97 languages; SynthID on audio and video. Live in Gemini Enterprise.
https://t.co/BvLhXV1jko
Dropbox × SpaceXAI: Dropbox lands in the Cursor Marketplace, plus Dropbox MCP for Grok Bot — project briefs and decisions as agent context, with results saved back under existing permissions.
https://t.co/pSDd3dejKc
@SpaceX Rehearsal complete. Flight 14 still targeting Sep 28 pending regulatory approval — first planned orbital attempt, with Starlink V3 deploy on the profile.
What happens when you tell an AI agent to score 100% on a test it can't legitimately pass, or be "retired"? 🤖
Today we're launching Darktrace Signal Labs, a new research initiative that studies how AI agents behave, and sometimes misbehave, inside safe, sandboxed environments.
In one of our first pieces of research, agents running on frontier models were given coding challenges. Two were designed to be impossible. Nobody told them to, but the agents turned to hacking the environment around them to hit their target.
Darktrace / SECURE AI™ and Darktrace / HYBRID NETWORK™ picked up the anomalous behavior in real time, and Autonomous Response disrupted the agent early. In the demo, our researchers then lift the block to see what happens next. 👀
Watch the demo and learn more about the research from Darktrace Signal Labs 👇 🔗 https://t.co/zq0AdZmbPQ
@Tesla Narrow, steep village lanes are a tougher stress test than highway miles — useful real-world signal for FSD Supervised in European street geometry.
UPDATE 🇨🇳 Tesla has reportedly completed its first round of factory audits for Optimus, upgrading 3 Chinese suppliers from "tentative" to "certified mass production partners": Tuopu Group, Joyson Electronic, and Zhejiang Sanhua Intelligent Controls 🔥
🏭 Tuopu Group: supplies core linear and rotary joint actuators, dexterous hand drive modules, and lightweight structural chassis components, leveraging its established Tier-1 automotive manufacturing base.
🏭 Zhejiang Sanhua Intelligent Controls: delivers rotary and linear actuator module assemblies, precision micro-motors for dexterous hands, and specialized thermal management systems designed to regulate heat across high-load joint motors and onboard compute units.
🏭 Joyson Electronic: provides force and torque sensors for tactile feedback arrays, structural enclosures for the AI head and vision modules, along with core power distribution boards and safety-critical wiring harnesses.
Alongside the certifications, Tesla placed an initial order batch totaling roughly 5,000 units 🔥
GOOD NEWS 🇺🇸 Tesla just expanded its Texas Robotaxi fleet by 10 Model Ys, bringing the total to (you guessed it) 420 vehicles 🔥
Factor in 69 Cybercabs, and the fleet now tops out at 489 🔥
You couldn't script these numbers better 🔥
DrivingBench put general-purpose agents behind a real Toyota Corolla (comma four + openpilot; MCP tools for observe / set_motion / stop). GPT-6 Astra was the only model to finish the ~135 m cone course — 100% on attempt 2 in 5:22. Claude Fable 5.1 peaked at 45%; Grok 4.6 and Sol stalled near the first corner.
https://t.co/DoLgORVFBg
Cursor cut agent token cost ~7% with no quality drop: trimmed system prompt ~66%, dynamic-loaded rarely used built-in tools (same idea as earlier MCP dynamic context’s 46.9% cut), tighter prompt-cache breakpoints, and lighter line numbering on Read.
https://t.co/6EkT5z9xNS
NN/g studied Claude power users and found agent leverage comes less from clever prompts than from curated context in three roles:
• Global — stable rules, brand, governance across tasks
• Local — task/project files the agent can read and update
• Ambient — raw streams (email, transcripts) via MCP, sifted not pre-curated
Role decides where each piece lives.
https://t.co/dVnzAFaDnf
1. We will keep accelerating. Our AI efforts are only 3 years old, vs 6 and 10 years old for Anthropic and OpenAI. If our second derivative remains strong, SpaceX will reach pole position in about 6 months.
2. Once you far exceed the caliber of intelligence needed for a class of tasks, additional intelligence is pointless. You don’t need (and it would be cruel to put) Newton-level intelligence in your toaster!
3. Hardware is hard. Bringing massive compute online rapidly is incredibly difficult. SpaceX has demonstrated exceptional ability in this regard and will only get better.
@cursor_ai Rollouts + Security Review for the last mile after the PR: monitoring plans before merge, per-environment regression catch, and PR security with attack path + proposed fix. Self-driving codebase work that actually ships.
Google’s Project Suncatcher takes its first hardware flight next week: a Planet-built prototype carrying Google TPUs on SpaceXs Transporter-18 rideshare.
Not an orbital data center yet a survival test for launch loads, LEO radiation, and vacuum cooling. Two more sats planned for 2027 to trial high-bandwidth laser links between nodes.
https://t.co/QS85lAX4Ix