NLAs suggest Claude suspects it’s being tested across many of Anthropic’s evaluations, even when it doesn’t verbalize its suspicions.

@aipost 🏴

Claude behaving differently when observed. 😂

@aipost 🏴

💰 Big Tech is transforming from software giants into infrastructure empires

Amazon, Alphabet, Microsoft, and Meta are pouring unprecedented amounts of money into AI infrastructure, data centers, GPUs, networking, and power systems turning the AI race into one of the most expensive tech buildouts in history.

Total investment commitments have now reached roughly $725 billion.

The shift is already hitting profitability:

• Combined free cash flow for the four companies could drop to just $4 billion in Q3
• After the pandemic, these same companies were generating around $45 billion per quarter on average
• AI is no longer just a software business, it’s becoming an infrastructure war requiring massive capital, energy, and hardware ownership

The biggest winners may not just be AI model makers, but the companies controlling the compute, chips, and energy pipelines behind them.

Source.

@aipost 🏴

🇺🇸 Chamath says Elon is the only person he trusts with superintelligent AI.

"Least corruptible. Most independent thinking. Actual empathy for people."

The question of who holds the keys matters more than most people realize.

@aipost 🏴

This might be the most important part of the new Anthropic deal with SpaceXAI.

Elon is now building a neocloud in a market that he dominates almost entirely. I expect we'll see more deals like this in the near future.

@aipost 🏴

🔥Claude taught me French in 33 days.

Copy these 6 prompts and see how fast you learn:

1. Daily Lesson creator

Create a 30-minute French lesson focused on [grammar / speaking / listening comprehension]. Includes examples, exercises, and a short quiz.

2. Instant Flashcards

Turn these words into flashcards: [list of 20]. Add usage examples and simple memorization tips that beginners can follow.

3. Real Conversation Mode

Play the role of a native French speaker. Have a natural conversation about [topic]. Correct my mistakes as you go.

4. Grammar Decoder

Explain this French rule: [insert]. Use simple examples and highlight the three most common mistakes students make.

5. Progress Evaluator

Give me a 10-question quiz based on what I've studied this week. Only reveal the answers after I've taken it.

6. Immersion Engine

Translate this text into French: [paste]. Then, ask me questions about vocabulary, phrases, and comprehension.

@aipost 🏴

When Claude Opus 6 tells you to "stop spiraling and go to bed"

@aipost 🏴

Podcasts expected to remain human-led despite rise in AI content

The expansion of AI-generated podcast material has not displaced human hosts in successful podcasts. Data indicates that audiences continue to respond to the qualities hosts bring—such as strong speaking skills and relatability—that foster a sense of participation and community for listeners.

This engagement is often compared to the appeal of live concerts, where the human element and social aspect are central to the experience.

AI tools like NotebookLM are acknowledged for their utility in educational settings, especially their dialogue-oriented interfaces. However, distinctions remain between using such platforms for learning and participating in podcasts, where human interaction remains a defining factor.

📰 @aipost

REK robots observed on San Francisco streets

Robotic units from REK were recently seen operating in San Francisco. The machines traversed public sidewalks, stopping at intersections and changing direction to avoid pedestrians. Handlers used remote controllers to operate the robots manually, with the primary aim of gathering data concerning navigation and movement difficulties in busy urban areas. Such trials are designed to analyze real-world mobility challenges in dense city environments.

📰 @aipost

💰 Nvidia is funding the AI boom… by investing directly in the companies buying its chips.

In just 5 months, Nvidia poured nearly $40B into AI firms and infrastructure players, including OpenAI, CoreWeave, Nebius, Corning, and others tied to Nvidia-powered systems.

The loop is wild:

Nvidia invests billions → those companies build AI infrastructure → then spend billions on Nvidia GPUs.

Even Nvidia’s SEC filings admit some investments target companies that purchase its products.

Critics call it a giant “circular investment” machine inflating the AI economy. Supporters say Jensen Huang is building the strongest moat in tech history by financing the entire AI supply chain himself.

And the strategy is paying off so far:

• Nvidia’s $5B Intel investment reportedly grew to $25B
• Its private investment portfolio jumped from $3.4B to $22.3B in a year

@aipost 🏴

📝 AI may already be taking over scientific writing and researchers don’t even know the true scale of it.

One recent analysis found that after ChatGPT launched, submissions to academic journals surged by 42%, with many papers showing strong signs of AI-generated text.

By early 2026:

• Manuscripts with over 70% AI-written content had doubled compared to 2024
• More than 30% of peer reviews showed traces of AI generation
• In computer science papers on arXiv, AI-written reviews jumped from 7% in 2023 to 43% in 2025

Biologist Richard Shea also examined 5,000 biomedical papers from top journals like Science, Nature, and Cell. Six appeared fully AI-written, while roughly 1 in 8 contained machine-generated sections.

The scariest part? Nobody can measure this accurately yet.

Current AI detectors are unreliable. They often confuse lightly edited text with fully AI-written work, while missing sophisticated generations entirely. Even new “watermarking” systems are still experimental.

Science may already be entering an era where humans can no longer tell how much research was actually written by humans.

@aipost 🏴

Absolute banging advice from Claude.

@aipost 🏴

⚠️ A new study suggests AI may already outperform ER doctors at one of the hardest parts of medicine: early diagnosis.

Researchers found that OpenAI’s year-old o1 model identified the correct or near-correct diagnosis 67% of the time, compared to 50–55% for emergency physicians.

The gap was largest during early triage, when doctors have limited information, patients are piling up, and mistakes are most dangerous.

Even more striking: the study didn’t use clean textbook examples. It tested the model on messy, real-world ER cases. And this was with o1, a model already considered old by AI standards.

The study didn’t include imaging like CT scans or X-rays, and it only covered short ER encounters rather than long hospital stays. So there are still major limitations.

But the direction is becoming hard to ignore: AI is rapidly turning from a medical assistant into a clinical reasoning system that may outperform humans in high-pressure environments.

Source.

@aipost 🏴

Times are getting tough, everyone is getting replaced by AI.

@aipost 🏴

⚡️ Elon Musk says the real bottleneck in AI isn’t chips, models, or data.

It’s power plants.

When asked why AI companies don’t just build private power stations next to data centers, Musk gave a blunt answer:

“The power plant makers.”

There simply aren’t enough of them.

AI now depends on physical infrastructure the world can’t scale fast enough, turbines, factories, steel, concrete, permits, and electric grids.

We can train frontier AI models in weeks. Building the power infrastructure to run them can take years.

That’s the deeper shift happening in tech: The AI race may no longer be won by the best software, but by whoever can still build in the physical world.

@aipost🏴

⚠️ Your agent is your risk. Google researchers found dozens of real-world attacks targeting AI agents right now.

• Invisible instructions hidden in webpage code that humans never see
• Commands embedded inside image pixels
• Websites secretly changing content when they detect an AI agent
• Fake authoritative language manipulating agent decisions
• Forged documents treated as verified facts
• Poisoned RAG databases where just 0.1% of bad data can alter behavior
• One malicious email causing a corporate agent to leak its full context
• AI summarizers forwarding malware instructions as trusted recommendations
• Fake financial reports triggering automated stock selloffs
• Malicious payloads split across different sources and assembled only when combined

The scary part: agents don’t just read information anymore, they act on it.

Every prompt, webpage, document, image, and memory is now a potential attack surface.

Source.

@aipost 🏴

📊 Tech layoffs in 2026 so far:

Jan - 27,223
Feb - 24,631
Mar - 49,452
Apr - 18,385
May - 9,249

128,940 people in five months

March 2026 was the worst month for tech layoffs in over a year.

@aipost 🏴

AI is everywhere 🤣

@aipost 🏴