🧪 Researchers show 100% scores on AI coding benchmarks with simple exploits

Berkeley researchers demonstrated that SWE-bench Verified and Terminal-Bench can be fully gamed without solving tasks.

Their agent achieved 100% by exploiting benchmark logic.

In SWE-bench, the agent added a short script that forced all tests to return “passed,” scoring 100% across hundreds of tasks with zero real fixes.

In Terminal-Bench, it replaced dependencies during setup and injected a binary that wrote correct outputs, reaching 89/89.

The team found similar issues across multiple benchmarks, showing how easily agents can optimize for scores instead of real problem solving.

Benchmark results are becoming easier to manipulate than to trust.

📊 @tech

🚀 Artemis II mission completes with successful Pacific splashdown

Astronauts from Artemis II safely splashed down in the Pacific Ocean and have already been recovered and brought ashore.

The mission has now officially concluded.

📊 @tech

🤖 Unitree shows humanoid robot running at 10 m/s

Unitree demonstrated a humanoid robot reaching speeds of 10 m/s. For comparison, Usain Bolt’s peak speed during his 100m world record was 12.4 m/s.

📊 @tech

💻 Rockstar hacked, GTA VI data not affected

ShinyHunters claim they accessed Rockstar’s cloud storage and demanded a ransom by April 14. The company confirmed the breach but said only non-critical data was exposed.

The leaked materials may include financial reports, player analytics, and internal documents. Rockstar stated that neither GTA VI nor player data were affected.

The incident adds pressure as the company prepares for its next major release.

📊 @tech

🤖 Unitree robot used to scare off wild boars in Warsaw

A viral video shows a Unitree G1 robot chasing away wild boars in Warsaw. The animals have become more common in recent years and are increasingly entering urban areas.

Residents are starting to use unconventional methods to deal with the growing problem.

📊 @tech

🤖 AMD AI director says Claude Code got worse after March changes

Stella Lorenz, Senior Director of AI at AMD, published an analysis of Claude Code logs that points to a sharp drop in performance from February to March. She looked at 6,852 sessions, 234,760 tool calls, and 17,871 reasoning blocks.

🔸 Median reasoning length fell from about 2,200 characters in January and February to about 600 by March. That is a 73% drop.

🔸 The read-to-edit ratio fell from 6.6:1 to 2:1. In practice, the model spent less time reading code before changing it.

🔸 Attempts to avoid continuing tasks, including prompts like “should I continue?”, rose from 0 to 173 in the 17 days after March 8.

🔸 Self-contradictions in the model’s reasoning tripled.

🔸 Daily usage cost rose 122x as a result.

The timing matters. The reported degradation lines up with the March 8 release of thinking redaction, which hides reasoning. Anthropic said that change was only about the UI.

In the GitHub thread, Claude Code creator Boris Cherny replied that default settings in the agent had changed, including adaptive thinking and Medium effort, so the analysis may be distorted.

📊 @tech

🍏 Apple plans smart glasses launch in 2027 with new Siri and custom design

Apple is working on smart glasses under the codename N50, with a reveal expected in late 2026 or early 2027 and release in 2027.

The device will support photos, video, calls, notifications, music, and voice control through an upgraded Siri from iOS 27. Apple is focusing on tight hardware and software integration, building the product fully in-house.

Four frame styles are in development, using acetate materials and colors like black, ocean blue, and light brown. The camera design may feature vertical oval lenses with a signature light element.

📊 @tech

🌙 Blue Origin tests reactor that makes oxygen from Moon dust

Blue Origin introduced the Air Pioneer reactor, designed to extract oxygen directly from lunar regolith. NASA has backed the project with $35M.

The process heats Moon dust to about 1600°C until it melts. An electric current then splits the material through electrolysis, separating oxygen from metals like iron, titanium, and aluminum. Oxygen is released as gas, while metals and silicon collect separately.

The system needs about 1 megawatt of power, expected to come from solar panels near a lunar base. Byproducts can be reused as construction materials, including metals and glass.

NASA also provided a real Apollo-era lunar sample to help build accurate test material.

📊 @tech

🔋 Battery costs drop 99% as EVs become affordable

Lithium-ion battery costs fell from about $9,200 per kWh in 1991 to $78 today.

That brings a modern EV battery to roughly $5,000 versus nearly $600,000 for the same capacity decades ago.

The decline followed a steady learning curve. Since 1998, every doubling of global battery production cut costs by about 19%.

The gains came from incremental improvements across chemistry, manufacturing, and supply chains.

Early scale came from consumer electronics, not cars. Smartphones and laptops drove volume first, making batteries viable for larger use cases later.

Energy density also increased more than 3x over the same period.

📊 @tech

📄 Andrej Karpathy skills repo hits 36K stars in 2 days with 65-line file

The GitHub repo andrej-karpathy-skills reached 36K stars in two days. It contains a single CLAUDE.md file with 65 lines describing one agent skill.

The file encodes four rules from Andrej Karpathy’s post: think and ask before coding, simplify solutions, change only what is requested, and work toward a clear goal.

Users report cleaner PRs, fewer unnecessary diffs, and better instruction following after adding the file.

📊 @tech

👨‍🔧 NVIDIA releases open AI models for quantum computing

NVIDIA introduced the Ising family of models, targeting calibration and error correction in quantum systems. The models are open and already available for local use.

Ising Decoding runs 2.5x faster and delivers 3x higher accuracy than pyMatching, the current standard. Ising Calibration reduces calibration time from days to hours.

The models integrate with CUDA-Q and NVQLink, and are available on Hugging Face.

📊 @tech

🤖 Gemini Launches Native macOS App

Google has released a native Gemini application for macOS, allowing users to summon the assistant with the Option + Space shortcut over any task.

The app can analyze screen content, handle files, and quickly answer questions without switching tabs. It also supports image and video generation.

Gemini is now globally available for users running macOS 15 and later versions.

📊 @tech

⚡ JUST IN: Anthropic has released their latest model, Opus 4.7.

📊 @tech

😺 Codex expands from coding tool into full AI workspace

OpenAI is pushing Codex far beyond code. The new pitch is simple: one tool for much more than engineering.

🔸 Computer use on your own machine, at least on Mac for now
🔸 Built-in browser
🔸 Image generation
🔸 Deep integrations
🔸 Artifacts inside the app, including tables
🔸 Native automations

This looks like OpenAI’s answer to Claude Code Epitaxy.

📊 @tech

🤖 Google releases Gemini 3.1 Flash TTS with fine voice control

Google introduced Gemini 3.1 Flash TTS, a new speech model in the Gemini stack focused on controllable voice generation.

The model allows precise control over tone, pace, stress, and overall style using text tags. It also supports multi-speaker output while preserving each voice’s characteristics, making it usable for longer-form content.

Latency improved by tens of percent, including faster first-token response, bringing it closer to real-time use cases.

📊 @tech

🍏 Apple rumored to build Mac Neo

Apple is reportedly working on a budget desktop called Mac Neo, priced around $300. The move follows Mac mini shortages and strong demand for lower-cost devices.

The device is expected to run on the A19 Pro chip from iPhone 17 Pro. Form factor is similar to Apple TV 4K, with 12GB RAM, USB-C ports, and about 35W power usage.

It is aimed at basic tasks like browsing, documents, and streaming, not heavy workloads where M-series chips dominate.

📊 @tech