📈 New benchmark ranks LLMs on ability to push back on bad prompts

A new benchmark measures how often language models recognize nonsense in a prompt instead of going along with it.

The chart shows three outcomes: green means the model flagged the issue, yellow means partial objection, red means it complied.

Most of the Claude family scored highest on detecting flawed or misleading prompts.

📊 @tech

📺 Casio built a pocket TV back in 1986

In 1986, Casio released the Pocket TV-200, a portable television that ran on 2 AA batteries for about 2–3 hours.

The device used a natural light system. Light passed through a monochrome LCD matrix, and the final image was reflected onto a mirror inside the unit.

Long before smartphones, Casio was already experimenting with handheld screens powered by simple hardware.

📊 @tech

⭕️ Anthropic adds 1-minute memory transfer to Claude

Anthropic introduced a simple way to move long-term memory from another AI into Claude. The update comes as more users switch from ChatGPT to Claude.

Users can export saved context from their previous assistant and paste it into Claude’s memory settings. The process takes about a minute and avoids re-explaining preferences or project details.

Anthropic provides a ready-made prompt to use in ChatGPT or Gemini, then asks users to copy the response into Claude.

📊 @tech

🎙 Claude Code rolls out voice mode to first 5% of users

Voice mode is now live in Claude Code for about 5% of users, with a broader rollout planned in the coming weeks. Access appears via a note on the welcome screen, and can be toggled with /voice.

It works as push-to-talk. Hold space, speak, release. The trannoscript streams directly at the cursor position, so users can mix typing and voice in the same prompt.

Voice mode is included at no extra cost. Trannoscription tokens do not count against rate limits. It is rolling out across Pro, Max, Team, and Enterprise plans.

📊 @tech

🤖 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