📢 “AI surveillance” in the U.S. was actually cheap offshore labor

What looked like cutting-edge policing tech has turned out to be a global sweatshop. Flock, the largest provider of “AI-powered” cameras for U.S. police, was barely using AI at all. Instead, much of the work was done manually by low-paid freelancers in the Philippines.

• The workers handled everything: reading license plates, identifying car makes and colors, tagging pedestrians, and even transcribing accident audio.
• Cities bought these systems expecting automated intelligence but got human eyes quietly scanning American streets.
• The revelation raises serious questions about data security, law-enforcement transparency, and how many “AI” products are really powered by hidden labor.

AI Post ⚪️ | Our X 🏴

🔥 NVIDIA just pulled off a an amazing stung using a tiny 4B model that beat far larger systems on ARC AGI 2, 29,72% / $0.20 per task!

By leaning on synthetic data and test-time training instead of brute-force scale, the NVARC team proved that clever design can outpace raw parameter count. It’s an exciting signal that efficient, adaptive reasoning might be the real frontier in AGI progress - not just ever-bigger models.

•27.64% accuracy on the official ARC-AGI-2 leaderboard

• Uses a 4B-parameter model that beats far larger, more expensive models on the same benchmark.

• Inference cost is just $0.20 per task, enabled by synthetic data, test-time training, and NVIDIA NeMo tooling.

Source.

AI Post ⚪️ | Our X 🏴

🚨 Trump says U.S. will allow Nvidia to sell H200 AI chips to China under new rules

He says 25% of the revenue will go to the U.S., argues this will boost American jobs and manufacturing, and criticizes Biden for forcing “degraded” chip designs.

He adds that newer NVIDIA chips (Blackwell, Rubin) aren’t part of the deal, and that a similar approach will be used for AMD, Intel and other U.S. chipmakers.

AI Post ⚪️ | Our X 🏴

This is actually the biggest danger: that workers still have to keep it a secret that they use AI at work because so many employers still think AI is still at the GPT-4 level.

This has to change.

AI Post ⚪️ | Our X 🏴

📈 Grok 4.20: the chaotic genius of the AI trading world

A new trading tournament between top neural networks just wrapped and only one model made money. The mystery winner? An internal experimental build of Grok 4.20, which posted a +12.11% gain and $4,844 profit.

• GPT-5.1: –6%
• DeepSeek V3.1: –32%
• Claude Sonnet 4.5: –38%
• Public Grok 4: –57% (dead last)

Grok 4.20 is now reportedly planned for release by year-end. It’s a tuned evolution of the Grok 4 line and a stepping stone toward Grok 5, xAI’s biggest model yet 6T parameters, double the current generation.

The irony? Grok is simultaneously the worst and the best depending on which build you’re allowed to touch.

AI Post ⚪️ | Our X 🏴

Google reportedly told advertisers it plans to introduce ads into its Gemini AI chatbot starting in 2026.

This moves Gemini closer to becoming a monetizable front end for Google’s entire AI ecosystem.

Source.

AI Post ⚪️ | Our X 🏴

Google says 'AI Glasses' powered by Gemini are coming in 2026.

AI Post ⚪️ | Our X 🏴

Reports indicate SoftBank and NVIDIA are nearing a deal to invest in Skild AI at a staggering $14B valuation, tripling its previous worth.

Skild is building an "omni-bodied brain" for the multiverse of machines.

AI Post ⚪️ | Our X 🏴

McDonald's has released an AI-generated Christmas ad

The studio behind it says they 'hardly slept' for several weeks while writing AI prompts and refining the shots, 'AI didn't make this film. We did'. Comments have been turned off on YouTube.

AI Post ⚪️ | Our X 🏴

Jensen Huang predicts that within 2–3 years, 90% of the world's knowledge will be generated by AI

Right now, knowledge is created, shared, and modified by humans. Soon, most of it will be synthetic, created by machines, mixed with human input, some true, some not. "That's crazy, but it's just fine".

AI Post ⚪️ | Our X 🏴

Google & Next era Energy are partnering on multi-gigawatt U.S. data center campuses with dedicated power for accelerating AI demand.

The first three sites are already in development with ~3.5 GW operating or contracted.

AI Post ⚪️ | Our X 🏴

Google Geminis Year over Year growth ~400%.

Let that sink in. That's why Sam Altman declared code red.

AI Post ⚪️ | Our X 🏴

📊 How people actually use LLMs, inside the a16z x OpenRouter 100T-token study

A16z and OpenRouter analyzed 100 trillion tokens of real usage roughly the text of 100 million Bibles to understand what people truly do with AI. Since OpenRouter routes traffic to hundreds of models across all major labs, the dataset reflects the real market, not just one provider.

What stood out immediately
• Role-play & storytelling = the #1 use case for open-source models.
Not coding. Not productivity. People mostly use small/medium open-source models to chat with characters and generate stories. A trillion-dollar industry driven by digital companionship.
• Open-source share exploded from <10% → 30% in one year. DeepSeek and Qwen are the rocket ships.
• Coding remains massive, but concentrated: 60%+ of all code requests go to Claude, with Sonnet utterly dominating developer workflows.

How usage is shifting
• Reasoning models now take ~50% of all tokens.
Users aren’t just asking for text, they’re asking models to think, plan, and use tools.
• Asia surged from 13% → 31% of global usage, with China becoming the world’s #2 consumer, not just the world’s #2 producer.
• Price barely matters. Even though Claude is pricier, people choose it for reliability and depth especially in coding and reasoning.

Market dynamics inside the data
• “Glass Slipper” effect: when a model nails a user’s need on first contact, they stay loyal indefinitely. That first solved task becomes the moat.
• Small models are shrinking in relevance. Sub-15B models are losing share fast.
• Medium models (15–70B) are the sweet spot, beating both tiny and ultra-large models on price performance.

The AI market isn’t being defined by the biggest labs, it’s being defined by how people want to talk, build, and think with models. Companionship drives adoption, reasoning drives spend, and the first model that solves a user’s problem wins the relationship.

AI Post ⚪️ | Our X 🏴

🔥 Everything we know about Meta’s new LLM "avocado"

- Meta is working on a new frontier LLM called “Avocado” as the direct successor to its Llama models.

- Target release: now expected in Q1 2026, delayed from an internal goal of “by end of 2025”

- Unlike Llama, Avocado is likely to be proprietary (closed-source), marking a strategic pivot away from Meta’s “open-source everything” story

- Internally, Meta’s AI strategy is described as confusing and scattershot, with staff unsure how Llama, Avocado and the overall product roadmap fit together

-The project is facing training and performance-testing challenges, which is part of why the launch slipped and why leadership is proceeding more cautiously than with past Llama announcements

Source.

AI Post ⚪️ | Our X 🏴

🚨 A Tokyo based firm, Integral AI claims to have built what it calls the first AGI

According to the article, the model exhibits reasoning capabilities that go beyond typical narrow AI, potentially enabling broader general problem solving, understanding and adaptation.

"Integral AI says it has conducted robot trials using its new system. During these trials, the firm claims that the robots learned new skills without human supervision."

"The company also claims its system mirrors the multi layered neocortex, a region of the human brain responsible for conscious thought, perception, and language."

Source.

AI Post ⚪️ | Our X 🏴

Pentagon launches military AI platform powered by Google Gemini for defense operations

Secretary of War Pete Hegseth says GenAI.mil will ‘revolutionize the way we win’ in future warfare

AI Post ⚪️ | Our X 🏴

🇺🇸 Trump’s last cartridge in the AI war with China

President Trump says he’ll sign a decree this week establishing a unified national AI regulatory framework effectively stripping states of their ability to regulate AI on their own. It’s the administration’s most aggressive move yet to seize control over how AI and AI-driven security systems develop in the U.S.

AI Post ⚪️ | Our X 🏴

🔥 Researchers at the University of Pennsylvania decided to stress-test one of the most popular pieces of prompt-engineering folklore.

The magic phrase “Act like an expert in…” You know the trick everyone recommends when you want an LLM to suddenly develop a PhD in quantum physics just because you asked nicely.

They ran six models (GPT-4o, GPT-4o-mini, o3-mini, o4-mini, Gemini 2.0 Flash, Gemini 2.5 Flash) through graduate-level questions in physics, chemistry, law, and more and tried three setups:

1) Expert in the subject:
“Pretend you’re a physicist” for physics questions.

2) Expert not in the subject:
“Pretend you’re a physicist” for law questions.

3) Total newbie:
A teenager, a layperson… or a toddler wobbling through its first steps.

The results? Brutal.

Giving the model the “correct expert role” barely improved accuracy essentially no meaningful lift. Giving it the wrong expert role sometimes made things worse. Gemini even spiraled into moral panic, refusing to answer because it “lacked expertise.”And forcing the models to act like children made them behave… exactly like children: confidently wrong.

All those guides obsessing over expert personas? Mostly toilet paper. The role affects tone, not intelligence. Your prompt can dress the model up, but it doesn’t make it any smarter.

The machine doesn’t level up just because you tell it to role-play.

AI Post ⚪️ | Our X 🏴

NVIDIA CEO Jensen Huang breaks down the five layers of AI.

AI Post ⚪️ | Our X 🏴