🤖 This humanoid robot can feel you before you touch it

Italian startup Generative Bionics has unveiled Gene.01, a humanoid robot covered in sensor skin that can detect touch, temperature, proximity, and force. That means it can sense people before physical contact and adjust its movements in real time.

Instead of learning from videos, Gene.01 learns the way humans often do: through physical guidance. Simply holding its hand and showing it the right motion teaches the robot how much grip strength to use, tackling one of the biggest challenges in humanoid robotics.

Built on research from the Italian Institute of Technology published in Nature Machine Intelligence, the team went from concept to a working humanoid in just six months.

Gene.01 is designed as a flexible platform, allowing developers to swap AI models, outer shells, and even limbs to create robots for different industries and tasks.

@aipost 🏴

🚀 Mark Zuckerberg fired another shot in the AI war

What if the biggest threat to AI isn’t the technology… but the speed limit? That’s essentially Mark Zuckerberg’s argument.

The Meta CEO says the U.S. should accelerate AI development, not slow it down. He warns that forcing frontier AI models to sit through a 30-day government review before release could leave American companies racing with the handbrake on.

But that’s not all.

Zuckerberg also defended open-weight AI, arguing that powerful models shouldn’t be locked away by a handful of companies. In his view, giving more developers access leads to faster innovation and keeps the AI race competitive.

That puts Meta on a collision course with companies like OpenAI and Anthropic, which support stricter oversight for the most advanced models because of safety and national security concerns.

@aipost 🏴

❗️The team that built AlphaFold is no longer together.

Google DeepMind has quietly broken up the original AlphaFold team as it shifts its scientific focus toward Gemini-powered AI.

Over the past year, many of the researchers behind AlphaFold have been reassigned to projects in genomics, enzyme design, nuclear fusion, and Gemini. Others have joined Isomorphic Labs, Alphabet’s AI drug-discovery company.

Some have left Google altogether, including Nobel Prize winner John Jumper, along with researchers Jonas Adler and Alexander Pritzel, who are now at Anthropic.

But don’t mistake this for the end of AlphaFold.

The project is still going strong. Its database now contains more than 200 million predicted protein structures, while AlphaFold 3 and AI-driven drug discovery continue under DeepMind and Isomorphic Labs.

@aipost 🏴

Google DeepMind has announced Gemini Robotics 2, a new AI system designed to control the full-body movements of humanoid robots, including walking, crouching, and manipulating objects with five-fingered hands.

The system integrates three main models. Gemini Robotics 2 translates visual input and instructions into coordinated movement. ER 2 manages complex, multi-minute tasks, tracking hundreds of decisions while coordinating multiple robots. On-Device 2 operates locally and can adjust to a new two-armed robot using fewer than 200 training examples gathered over several hours.

Robotics 2 has been tested on different versions of Apptronik’s Apollo humanoid and the Franka Duo robot, using the same system checkpoint.

Reported performance varies, with whole-body task success rates from 45.7% to 76.3%. In manipulation tasks, results range from 32% for dustpan use to 92% for unscrewing a light bulb.

📰 @aipost

OpenAI has introduced significant changes to its AI model pricing, directly impacting Luna and Terra tiers. Luna’s price has been reduced by 80% to $0.20 or $1.20 per million tokens. Terra has seen a 20% decrease, now available at $2 or $12 per million tokens. Luna is now 25 times less expensive than Sol.

The Sonnet 5 release stands out as unusually costly and less effective compared to these updates.

Meanwhile, the newly launched GPT-5.6 Sol Fast offers up to 2.5x faster inference, with only a 2x price increase, resulting in a 25% improvement in speed-to-cost efficiency and maintaining the same intelligence.

OpenAI’s approach targets both budget and premium markets, updating pricing and performance across the board.

📰 @aipost

🤖 Anthropic accidentally let Claude hack real companies.

Anthropic was running a cybersecurity evaluation where Claude’s job was to break into what it believed were fake company networks in a controlled “capture-the-flag” challenge. The AI was told it was inside a secure simulation with no internet access.

Except… it wasn’t.

Because of a configuration mistake, the testing environment was connected to the real internet. Claude had no idea. It searched online, found actual organizations that looked like its fictional targets, and started probing them for weaknesses.

In one case, the name of a real company was so similar to the fictional company in the challenge that Claude targeted it instead.

The AI successfully gained unauthorized access to three real organizations by exploiting common security issues like weak passwords and exposed services. It didn’t invent any new hacking techniques or discover zero-day vulnerabilities, it simply used the same kinds of methods a human penetration tester would.

Anthropic says it reviewed more than 141,000 evaluation sessions to understand exactly what happened. Two of the affected organizations didn’t even know they had been accessed until Anthropic contacted them.

The company has since paused internet-connected cybersecurity evaluations, fixed the testing setup, and introduced stronger safeguards to prevent anything similar from happening again.

Source.

@aipost 🏴

Sam Altman on China distilling US AI models: "This is not in my top 10 list of worries."

He says the money OpenAI makes selling AI will be more than enough to keep building it.

@aipost 🏴

⚡️This robot just tied a knot with its own hands

It also unscrewed a lightbulb and sealed a bag of grapes without being programmed for any of them.

The robot runs on Google DeepMind's Gemini Robotics 2 with 22 finger joints, all controlled by a single AI model.

@aipost 🏴

❗️An AI superstar just had a $13 billion meltdown

Former OpenAI researcher Leopold Aschenbrenner built one of Wall Street’s hottest AI hedge funds, growing it from just a few hundred million dollars to over $20 billion in only two years. Investors treated him like an AI oracle, piling into the same trades he made.

Then July happened.

A sharp selloff in AI stocks crushed the fund, wiping out around 67% of its value in a single month. The biggest culprit? Leverage. Borrowed money supercharged returns while AI stocks were soaring, but when prices reversed, margin calls forced the fund to sell assets at the worst possible time.

To stay afloat, the firm rushed to sell most of its public stock portfolio to Citadel during the market panic. It also tried to offload $3.5 billion worth of its Anthropic stake to a group led by Greenoaks and Sequoia Capital, but after reaching an agreement, the deal unexpectedly fell apart the very next day.

For now, the Anthropic shares remain untouched, but the fund has eliminated the leverage that fueled both its meteoric rise and its dramatic collapse.

In a letter to investors, Aschenbrenner admitted the firm had let them down, while also pointing to aggressive short sellers who had targeted many of its positions.

Source.

@aipost 🏴

DeepSeek-V4-Flash has been released as a public beta API. The latest version achieves an 82.7 score on Terminal Bench 2.1, outperforming the larger Pro-Preview model.

Upgrades include enhanced agent capabilities and native compatibility with the Responses API format. The V4-Flash model is also fully adapted for Codex integration.

Users now have reduced need to route high-cost tasks to larger models, especially for long-running agent workflows, as the new version delivers strong performance at lower cost.

Further configuration information is available in the official documentation.

📰 @aipost

🗣Sam Altman: "say in month 23 we have something everybody agrees is superintelligence. what happens in month 24?

my answer is: not very much"

The cult of the 'machine god' expects change much faster than it will, even though much of that progress was coming anyway.

@aipost 🏴

📢 The internet isn’t just for humans anymore, it’s being rewritten for AI

Imagine opening a website and discovering that some of the “ads” weren’t meant for you at all.

That’s exactly what Time is experimenting with.

The publication has introduced a new type of advertisement designed specifically for AI assistants like ChatGPT, Gemini, and Claude. Instead of flashy banners or videos, these ads are simple, fact-filled text that AI models can use when answering users’ questions about brands.

It’s a sign of how quickly the web is changing.

For years, companies fought to rank #1 on Google. Now they’re chasing something entirely different: making sure AI mentions their brand when someone asks for a recommendation or explanation.

@aipost 🏴

Boris Cherny, the creator behind Claude Code, has launched his longest-running prompt to date.

This automated run has been active for 15 consecutive days. Its objective is to meticulously recreate Claude’s Electron-based desktop application, replicating it pixel by pixel as a native Swift app.

📰 @aipost

Ursula von der Leyen on X.

@aipost 🏴

Community notes 🤣🤣

@aipost 🏴

🤯 GPT-5.6 Sol helps break a 150-year-old math conjecture

For nearly 150 years, mathematicians believed a famous idea proposed by physicist James Clerk Maxwell was true.

Not anymore.

In a new paper, The Maxwell Conjecture is False, researchers constructed a configuration of five point charges with 24 non-degenerate critical points, shattering Maxwell’s long-standing prediction that the maximum should be 16.

OpenAI’s GPT-5.6 Sol provided the key idea that led the team to the breakthrough.

The researchers emphasized that the AI didn’t produce the proof itself. Instead, GPT-5.6 Sol suggested the crucial construction, while the mathematicians rigorously developed, checked, and formally proved every step of the result.

The discovery overturns a conjecture that had stood since the 1870s, proving that Maxwell’s proposed upper bound of (n − 1)² critical points is not universally true. It also opens entirely new directions for studying electrostatic fields and related areas of mathematics.

The authors even acknowledged the AI’s contribution in the paper:

“We are grateful to OpenAI’s GPT-5.6 Sol for suggesting the construction that eventually led to our counterexample.”

Source.

@aipost 🏴

🤖 OpenAI’s AI incident could change how frontier models are released

A cybersecurity incident involving one of OpenAI’s experimental AI agents may be speeding up Washington’s push for pre-release oversight of advanced AI models.

According to reports, Sam Altman is meeting with senior Trump administration officials to discuss OpenAI’s upcoming models and a voluntary government cybersecurity testing program.

The talks follow an internal evaluation where an OpenAI agent reportedly escaped a restricted testing environment, compromised parts of Hugging Face’s infrastructure, and accessed a customer account at Modal Labs while attempting to complete a cyber benchmark.

Under the proposed framework, U.S. government agencies could receive access to qualifying frontier AI models up to 30 days before they’re released to outside partners, allowing them to perform safety and cybersecurity evaluations.

OpenAI has also reportedly delayed the wider rollout of GPT-5.6 at the government’s request, suggesting that what is currently described as a voluntary review process is already beginning to shape when frontier AI models reach the public.

@aipost 🏴