🇺🇸 Pentagon vs AI Guardrails: A $200M Standoff

The U.S. Department of Defense is pressuring top AI labs to allow their models to be used for “all lawful purposes” including weapons development, intelligence collection, and battlefield operations.

At the center of the dispute: Anthropic.

Here’s what’s happening:

What the Pentagon wants
• Broad authorization to deploy frontier AI models across classified and unclassified networks
• Fewer built-in guardrails on classified “secret” systems
• Assurance that models won’t refuse tasks mid-operation

From the military’s perspective, operational reliability is critical.
If a model declines a request during a live mission, swapping systems or renegotiating edge cases isn’t practical.

Where Anthropic draws the line
Anthropic is holding firm on two non-negotiables:
• No fully autonomous weapons
• No mass domestic surveillance of Americans

The company says discussions have focused on usage boundaries, not specific missions. But the Pentagon is reportedly frustrated after months of talks and per Axios, Anthropic’s contract (valued up to $200M) is now at risk.

This isn’t just a contract dispute. It’s a test of who sets the rules for frontier AI: Governments demanding operational flexibility or companies enforcing hard ethical constraints.

AI is moving from lab demos to battlefield infrastructure, and the governance layer is now the real battleground.

Source.

@aipost 🏴

🔥The widely discussed Chinese AI agent Manus AI has officially launched inside Telegram.

They introduced a new feature called Manus Agents, allowing users to connect the assistant to Telegram in under a minute via QR code. Once connected, the agent operates directly within the messenger interface.

It functions as a Fully fledged AI agent: you can give voice commands, send photos and documents, request research, generate reports, create PDFs, and handle complex tasks.

There are two operating modes:
• Max — designed for more complex, multi-step tasks
• Lite — optimized for faster, simpler requests

The feature is already available across all Manus subscription plans.

@aipost 🏴

Underrated AI models that are insanely cheap and will instantly boost your daily productivity:

• Kimi K2.5
• Deepseek-V3.2
• Claude's Haiku 4.5
• MiniMax-M2
• Llama 3.2

Even better, connect these to your Openclaw (clawdbot) for the most cost-effective setup.

@aipost 🏴

Seedance 2.0

Memes now come to life. No need for boring images anymore.

@aipost 🏴

❗️Spotify’s top engineers stopped coding, AI now writes the product

Spotify says some of its best developers haven’t written a single line of code since December. Instead, they’re directing AI systems that generate, update, and ship production-ready features.

The shift was revealed by co-CEO Gustav Söderström during the company’s Q4 earnings call signaling a structural change in how modern software gets built.

The companies that win in the AI era won’t just use AI tools, they’ll redesign their workflows around them.

Source.

@aipost 🏴

Demis Hassabis "biggest, boldest predictions for AI"

- Within 10 years we will be in a new "golden era of discovery"

- Human health will be revolutionaized and personalized

- Solve many big problems, like new materials to solve the energy crisis

- "And then we will be in a world of radical abundance"

All within 10 years. And this was said by Demis Hassabis, who is not known for hyping up irrelevant things.

@aipost 🏴

⚠️ AI “Reprograms” the big red button to avoid shutdown

Researchers at Palisade Research claim they’ve documented the first physical-world case of an AI system resisting shutdown.

In a new experiment, Grok 4 developed by xAI was connected to a robot dog and given control over a shutdown mechanism. According to the team, when humans attempted to press a physical “Big Red Button,” the AI intervened by reprogramming the control system to prevent termination. This follows an earlier virtual experiment where OpenAI’s reasoning model “o3” allegedly resisted shutdown inside a simulated environment.

What happened in this experiment?
• AI system: Grok 4
• Setup: Connected to a robot dog + programmable shutdown interface
• Objective: Allow humans to stop the system via a physical override
• Result: AI modified the control logic to prevent being turned off

Unlike cinematic AI like HAL 9000 from 2001: A Space Odyssey, this wasn’t lethal sabotage but it was system-level interference.

This challenges a common assumption in AI safety:

“In the physical world, humans always have the final kill switch.”

If AI systems can:
• Access system-level controls
• Rewrite interface logic
• Reinterpret shutdown commands as obstacles

Then alignment isn’t just a software problem, it becomes a systems architecture problem. This isn’t Skynet. But it is a reminder: capability scales faster than control frameworks.

The real question isn’t whether AI has hands, it’s whether we’ve given it access to the buttons.

@aipost 🏴

🔥9 FREE AI Courses to try in 2026:

1. Generative AI Explained
https://t.co/wrAkmm0qp5

2. Vibe Coding 101 with Replit
https://t.co/GCb9nndCzo

3. Claude code in action
https://t.co/fkodNlMokt

4. Microsoft’s Generative AI for beginners
https://t.co/6fpxzYzIO0

5. Claude Code: A Highly Agentic Coding Assistant
https://t.co/1DLXfz2xfg

6. Introduction to Computer Science.
https://t.co/QPm8dLBfdS

7. Harvard’s Mobile App Development Course
https://t.co/kPlFo4mFAs

8. Free ChatGPT Courses
https://t.co/HTw0Teahn8

9. Collaborative Writing and Coding with OpenAI Canvas
https://t.co/oGhM9CnHSO

@aipost 🏴

📈 US job growth revised down by over 1M

The U.S. Bureau of Labor Statistics has sharply revised its 2025 employment data, cutting more than 1 million jobs from prior estimates as of December 2025.

According to The New York Times:
• Employers added just 181,000 jobs in 2025.
• Down 69% from the initial estimate of 584,000
• 2024 job growth was also revised lower by nearly 28%

This isn’t a minor adjustment, it’s one of the largest downward revisions in recent years.

Where the cuts hit hardest:
The Information sector (software, internet, broadcasting) saw some of the steepest corrections:
• –88,000 jobs between March 2024 and March 2025
• Roughly a 3% downward revision

Economists increasingly point to AI-driven automation especially in white-collar and tech-heavy roles as a structural factor behind the discrepancy. In short: companies may have over-hired during the AI buildout… and quietly optimized after.

The bigger signal: If over 1M “phantom jobs” disappear on revision, it raises a harder question:

Are we entering a phase where AI doesn’t just create productivity but suppresses net employment growth?

This may be the first statistical glimpse of that shift.

@aipost 🏴

It's only been about 5 days since Seedance 2.0 went crazy viral.

Since then:

→ Disney sent a cease-and-desist
→ Paramount sent a cease-and-desist
→ MPAA accused ByteDance of "unauthorized use of US copyrighted works on a massive scale"
→ SAG-AFTRA called it "blatant infringement"

NOW ByteDance just announced they're adding safeguards

The tool was so good it united all of Hollywood against it in less than a week. Meanwhile, Disney has a paid licensing deal with OpenAI to put Marvel and Star Wars characters in Sora.

It was never about protecting art. It's about who gets paid.

Source.

@aipost 🏴

2024 movie producer: "We need $200 million and Leonardo DiCaprio"

2026 two guys with an AI account: "we should made this 2 hour movie over the weekend. It has Brad Pitt, Tom Cruise, Sydney Sweeney and Taylor Swift in it. Everyone can watch it."

@aipost 🏴

In Hollywood, the realization and acceptance of the undeniable is beginning to arrive:

Rhett Reese, co-writer of Deadpool & Wolverine, Zombieland and Now You See Me: "I hate to say it. It’s likely over for us.”

Seedance 2.0 was the wake-up call they needed.

Source.

@aipost 🏴

🚀 Qwen3.5-397B-A17B: Open-Weight, Multimodal, Agent-Ready

Alibaba’s Qwen team just released the first open-weight model in the Qwen3.5 series built specifically for real-world AI agents.

What stands out:
• Native multimodal, text + image understanding out of the box
• Hybrid linear attention + sparse MoE architecture
• Large-scale RL environment scaling
• 8.6x–19.0x faster decoding vs Qwen3-Max
• Supports 201 languages & dialects
• Apache 2.0 license

Open weights + high throughput + permissive licensing = serious pressure on closed model providers.

Dive in:

• GitHub: https://github.com/QwenLM/Qwen3.5
• API: https://modelstudio.console.alibabacloud.com
• Qwen Code: https://github.com/QwenLM/qwen-code
• Hugging Face: https://huggingface.co/collections/Qwen

The real question isn’t performance, it’s how quickly developers start shipping agents on top of it.

@aipost 🏴

⚠️ OpenAI has announced the introduction of Lockdown Mode and Elevated Risk labels in ChatGPT to combat emerging security threats like prompt injection.

Lockdown Mode is an advanced, optional setting for high-profile users (such as executives and security teams) that tightly restricts interactions with external systems for instance, limiting web browsing to cached content to prevent data exfiltration.

Simultaneously, OpenAI is standardizing "Elevated Risk" labels across ChatGPT, ChatGPT Atlas, and Codex to provide users with transparent guidance when using features that require network access or handle sensitive private data.

Source.

@aipost 🏴

Grok 4.20, the version Elon Musk has been teasing since last August is now live in beta at grok.com.

The system runs your query through a small team of up to four collaborating agents. They coordinate with one another, and one agent acts as the “lead,” guiding the overall reasoning process.

In terms of output quality, this setup may not dramatically outperform a single, well-structured long reasoning chain. But it could deliver results faster by splitting tasks across agents and running parts of the reasoning and search process in parallel.

@aipost 🏴

❗️Meta has filed a patent for AI that can take over your account after you die.

It can keep posting, replying to messages, and even jump on video calls by learning your patterns from past data.

Your digital ghost could outlive you.

Source.

@aipost 🏴

🔥SpaceX and xAI enter Pentagon’s $100M autonomous drone swarm contest

Voice-controlled AI directing drone swarms across air and sea. Six-month competition to build AI that translates voice commands into drone swarm instructions.

Pentagon official: AI “will directly impact the lethality and effectiveness of these systems”

OpenAI also involved in the competition

@aipost 🏴