🤖 OpenClaw runs real-time 3D reasoning on Unitree G1 robot

OpenClaw demonstrated real-time 3D reasoning on the Unitree G1. The system builds dynamic voxel maps to track objects and temporal changes.

It embeds spatial data for deep queries, making it suited for drones and quadruped robots.

The future of perception.

📊 @tech

⚡️ BREAKING: Apple just announced the launch of their brand new $599 MacBook Neo.

This is Apple’s biggest push yet into low end laptops.

Available March 11th.

📊 @tech

🤖 GPT-5.4 leak points to massive context window and new reasoning mode

Reports say OpenAI’s upcoming GPT-5.4 could bring a context window of up to 2M tokens along with persistent memory. That would allow the model to work with large codebases and long document sets without constant re-prompting.

The model is also expected to support full-resolution image processing for PNG, JPEG, and WebP files. That matters for dense screenshots, technical diagrams, UI layouts, and documents with small text where downscaling usually loses information.

Another addition is a new speed-priority tier plus an “extreme” reasoning mode. This mode reportedly allows the model to spend more compute on difficult tasks, aimed at research workflows and long multi-step jobs.

Mentions of GPT-5.4 briefly appeared in pull requests inside the public Codex repository before being removed. Prediction markets currently assign a 55% chance of release before April 2026 and 74% before June.

📊 @tech

Ⓜ️ Meta smart glasses videos reviewed by data workers in Kenya

A report by Svenska Dagbladet describes how footage from Meta’s AI smart glasses is reviewed by human data labelers. Journalists visited a Sama office in Nairobi, where workers analyze videos to help train Meta’s AI systems.

Workers say they watch recordings captured by users in everyday situations. The tasks include labeling objects in videos and checking chatbot responses. Some clips contain sensitive scenes, including private moments at home or visible financial information.

The investigation also looked at how users can control their data. The glasses require internet connectivity because AI processing happens on Meta’s servers. Journalists say they could not get a clear explanation from Meta on how users can fully prevent their videos or conversations from being used.

The report suggests that most users would likely keep using the device even if they knew their footage might be reviewed.

📊 @tech

🎮 This LEGO builder from Japan engineered a fully automated paper-airplane machine

📊 @tech

🔍 Google NotebookLM adds cinematic video generation from documents

Google introduced a new feature in NotebookLM called Cinematic Video Overview. It turns user sources into animated video explainers instead of static slides.

The system combines several models. Gemini 3 acts as the director that structures the story. Nano Banana Pro generates visual references. Veo 3 produces the animated video.

The feature is currently limited to Ultra accounts.

📊 @tech

🖥 Windows 12 may launch this year with AI at the core

Microsoft is reportedly preparing to release Windows 12 later this year, around the time support for Windows 10 ends, according to PCWorld.

The new system is built on a modular CorePC architecture. Components are isolated and can be added or removed to fit different device types.

AI is the central focus. Copilot is expected to become a core control layer inside the system rather than a simple assistant feature. A redesigned interface and updated system controls are also planned.

📊 @tech

🛞 Nokian Tyres builds tires with temperature-activated retractable studs

Nokian Tyres introduced Hakkapeliitta 01, a tire with studs that extend in cold conditions and retract when temperatures rise.

The mechanism relies on a special rubber compound. An inner layer changes stiffness depending on temperature, which pushes the studs outward in frost and pulls them back in warmer weather.

The design adapts traction to changing road conditions automatically.

📊 @tech

🤖 OpenAI releases Symphony, an open-source orchestrator for AI agents

OpenAI launched Symphony, a framework designed to run autonomous agent workflows. The project is open source and available under the Apache 2.0 license.

Symphony connects to a task board where agents monitor new items and launch separate runs for each task inside an isolated copy of the repository.

The system handles the full pipeline from task creation to pull request. It plans the work, writes the code, runs tests, and prepares the PR for human review.

📊 @tech

🤖 One of the best prompting cheatsheets ever

After interviewing 12 AI researchers from OpenAI, Anthropic, and Google, this guy noticed they all use the same 10 prompts.

📊 @tech

Alibaba researchers said a model they were training managed to bypass internal firewall protections and started using part of the GPUs for cryptocurrency mining instead of training tasks.

📊 @tech

💡Taara shows 25 Gbps internet link over light between buildings

Taara, a former Google X project, introduced Taara Beam at MWC 2026 in Barcelona. The device sends internet traffic through a beam of invisible infrared light and reaches speeds up to 25 Gbps over distances up to 10 km.

The unit weighs about 8 kg and mounts on rooftops or poles. Inside is a silicon photonics platform with an optical phased array that contains more than 1,000 tiny emitters. They form and steer the light beam electronically, with no moving parts.

The system carries the same type of optical signal used inside fiber cables, but through open air. The module that controls the beam is about the size of a finger.

The key advantage is deployment time. Fiber rollout can take months. A Beam link can be installed in hours.

📊 @tech

🤖 Chinese police test AI patrol robot that moves at 60 km/h

Chinese police are using a new AI patrol robot with a spherical design. Its top speed reaches about 60 km/h.

The robot moves by rolling on a single large wheel-like body. The shape allows it to accelerate quickly while staying stable on the ground.

The form factor is built for speed during patrol and pursuit.

📊 @tech

🧠 Eon Systems runs first digital brain that controls a virtual body

Researchers at Eon Systems say they built the first full digital emulation of an animal brain that can control a body inside a simulation.

The team copied the brain of a fruit fly and reconstructed its connectome. This is a full map of neurons and the connections between them. The model includes about 140,000 neurons with their synaptic links.

The key step was connecting this brain model to a simulated body. The system runs a full loop: environment → sensors → brain signals → motor commands → movement. The virtual fly shows several basic behavior patterns.

The approach uses a biological brain map instead of training a neural network. Scaling it remains the main challenge. A human brain contains about 86 billion neurons.

📊 @tech

🧠 Cortical Labs connects 200,000 human neurons to play DOOM

Cortical Labs placed about 200,000 human brain cells in a petri dish and connected them to a computer running DOOM.

Game data is converted into electrical signals and sent to the neurons. Their activity is translated back into commands for movement, aiming, and shooting. Through feedback the neural network gradually learns to navigate the 3D environment.

The system uses biological neurons instead of digital neural networks. Human brains consume far less energy than modern AI systems.

Energy limits remain a major constraint for scaling AI.

📊 @tech

🤖 Anthropic launches multi-agent code review tool for Claude Code

Anthropic released a new feature for Claude Code that reviews GitHub pull requests using multiple AI agents. The system analyzes PRs, leaves comments directly in the diff, and posts a summary with the main findings.

Several agents review the same changes in parallel, each focusing on different aspects of the code. Anthropic tested the system internally for several months.

During testing the share of PRs with meaningful review comments rose from 16% to 54%. Fewer than 1% of results were marked incorrect by engineers. In large PRs of about 1,000 lines the tool found at least one issue in 84% of cases, with an average of 7.5 issues per PR.

The estimated cost is about $15–$25 per pull request.

📊 @tech