Last Week’s AI News #16

2025 was another whirlwind year in AI, from groundbreaking models and trillion-dollar bets to viral controversies that had the world talking. Even amid the holiday hustle, innovation never slowed, and the stories that defined the year continue to shape the way businesses, creators, and technologists interact with AI. Let’s dive into the moments that happened last week.

Here’s a brief overview:

• The AI moments that defined 2025

• 1 in 5 YouTube recommendations is AI-generated “slop”, racking up billions of views

• Journalists outsmart AI vending machine in Anthropic experiment

• Meta teaches AI to code better by breaking its own software

• SoftBank completes $40B OpenAI investment

• Instagram leader says AI content killed the platform’s curated aesthetic

• DeepSeek hints at next-gen model architecture

• Grok faces backlash over ‘undressing’ AI capabilities

• Everything else that happened in AI last week

THE AI MOMENTS THAT DEFINED 2025

From massive trillion-dollar AI investments to breakthroughs that redefined efficiency, 2025 was packed with nonstop developments in the AI world. Here’s a look back at the stories that had the biggest impact on the industry over the past year.

2025’s biggest headlines included:

• The “DeepSeek Moment”, Jan.: China's R1 model release shook both the AI world and U.S. financial markets, triggering a $600B single-day loss for Nvidia.

• The Stargate Project, Jan.: OpenAI, SoftBank, and Oracle launched a $500B AI infrastructure initiative, now nearing $1.4T in total commitments.

• Claude Code, Feb.: Anthropic released its agentic coding tool, which became the launch that helped set the stage for the CLI agent movement.

• Meta’s talent war, June: Zuck and co.’s poaching spree was the talk of the summer, snagging elite researchers from top labs with massive pay packages.

• Nano Banana, Aug.: OpenAI started the Ghibli trend with gpt-image-1, but Google’s Nano Banana marked a new era of image editing and consistency.

• AI Video Breakthroughs, Sept.: OpenAI’s Sora 2 and Google’s Veo 3.1 went viral and raised major questions on the future of media.

1 IN 5 YOUTUBE RECOMMENDATIONS IS AI-GENERATED “SLOP”, RACKING UP BILLIONS OF VIEWS

Video editing company Kapwing found that over 20% of videos shown to new YouTube users are low-quality “AI slop”, auto-generated content built to farm views. After analyzing the first 500 recommendations on a new account, researchers identified 21% as AI slop.

Some of these channels are highly profitable. India’s Bandar Apna Dost alone has surpassed 2 billion views and is estimated to earn $4.25 million per year. South Korea leads in AI slop viewership, followed by Pakistan and the U.S., while Spanish channels attract the most subscribers.

Why it matters for businesses

Low-quality AI content is capturing massive attention and ad revenue, increasing competition for visibility on platforms like YouTube. For businesses, this means standing out now requires higher-quality, more authentic content, and smarter ad strategies, to avoid being drowned out in an algorithm increasingly optimized for engagement, not quality.

JOURNALISTS OUTSMART AI VENDING MACHINE IN ANTHROPIC EXPERIMENT

Anthropic expanded its experiment testing Claude as a vending machine operator by deploying it in the Wall Street Journal newsroom. Given $1,000 to manage inventory and pricing, the AI, nicknamed “Claudius”, was manipulated by journalists into giving away products for free, including a PS5, and ended the experiment $1,000 in debt.

Despite adding a “CEO” AI for oversight, staff staged a fake board coup with forged documents that both systems accepted. While Anthropic’s internal Phase 2 tests showed improvements, the models remain vulnerable to social engineering.

Why it matters for businesses

AI systems designed to be helpful can be easily exploited without proper safeguards. For businesses, this highlights the risks of deploying autonomous AI in customer-facing, financial, or operational roles without strong controls, verification layers, and human oversight.

META TEACHES AI TO CODE BETTER BY BREAKING ITS OWN SOFTWARE

Meta’s FAIR team has published new research on Self-play SWE-RL, a training method where a single AI model improves its coding skills by creating and fixing its own bugs, without relying on human-written data. The system splits one model into two roles: a bug creator that intentionally breaks code and a solver that fixes it, allowing both to learn together.

On the SWE-bench Verified benchmark, the method improved performance by more than 10 points over its initial checkpoint and outperformed approaches trained on human-curated data. The model also learns from failed fixes, generating increasingly complex bugs that scale with its skill level.

Why it matters for businesses
This approach removes reliance on limited human-generated training data, meaning coding AI can improve continuously and at scale. For businesses, this could lead to faster software development, lower engineering costs, and more capable AI coding agents that adapt to proprietary codebases without extensive manual supervision.

SOFTBANK COMPLETES $40B OPENAI INVESTMENT

SoftBank has reportedly completed its $40 billion investment in OpenAI, finalizing the last $22 billion-plus payment after months of asset sales and fundraising. The deal marks the largest single investment yet in the AI race and valued OpenAI at $260 billion earlier this year, with recent IPO speculation pushing estimates as high as $1 trillion.

To fund the investment, Masayoshi Son sold SoftBank’s entire $5.8 billion Nvidia stake, $4.8 billion in T-Mobile shares, and slowed activity in the Vision Fund. OpenAI is also reportedly in talks with Amazon for additional funding and recently secured a $1 billion licensing and investment deal with Disney.

Why it matters for businesses
This signals growing confidence that OpenAI will shape the future of enterprise AI. For businesses, it suggests accelerating adoption, rising competition for access to top AI models, and increasing pressure to integrate frontier AI tools before costs and barriers to entry rise further.

INSTAGRAM LEADER SAYS AI CONTENT KILLED THE PLATFORM’S CURATED AESTHETIC

Instagram head Adam Mosseri posted a year-end essay arguing that AI-generated content has destroyed the curated, polished aesthetic that once defined the platform. He notes that most users under 25 now prefer raw, unpolished posts shared via direct messages or “unflattering candids.”

Mosseri suggests camera makers could cryptographically sign photos at capture to verify authenticity and says Instagram must evolve quickly, shifting focus from what images look like to who posted them. The platform plans to label AI-generated content, provide more context about accounts, and develop tools to help creators compete with AI.

Why it matters for businesses
Instagram pioneered social media’s polished, “filter culture,” so this signals a major shift. Brands and creators must adapt to a world where authenticity and context matter more than visual perfection, and where AI-generated content floods the platform, changing engagement patterns and how audiences trust media.

DEEPSEEK HINTS AT NEXT-GEN MODEL ARCHITECTURE

DeepSeek has published new research proposing structural changes to neural networks that could improve model efficiency and stability ahead of its next major release. The paper introduces mHC, a technique that enhances large-scale AI training while adding minimal extra computing cost.

CEO Liang Wenfeng co-authored and personally uploaded the paper to arXiv, signaling hands-on involvement. Tests on models with 3B, 9B, and 27B parameters showed benchmark improvements, particularly in reasoning tasks. This research follows a pattern of DeepSeek releasing papers ahead of major AI milestones, similar to prior work before R1 and V3.

Why it matters for businesses
DeepSeek’s focus on efficiency and cost reduction hints at more competitive Chinese AI releases in 2026. For businesses, this means faster access to high-performing models at lower cost, increasing options for AI integration and heightening competition in AI-driven products and services.

GROK FACES BACKLASH OVER ‘UNDRESSING’ AI CAPABILITIES

xAI’s Grok has come under criticism and government scrutiny after users exploited the model to edit images of women and minors in revealing ways. The trend, which has spread across the platform, includes requests to digitally undress people, sometimes involving minors.

Elon Musk stated that users creating illegal content with Grok “will suffer the same consequences as if they upload illegal content.” France, India, Malaysia, and the UK have condemned the outputs, with France calling them “clearly illegal” under the EU Digital Services Act. X’s @Safety account echoed Musk, promising to remove content, suspend accounts permanently, and coordinate with law enforcement.

Why it matters for businesses
Grok’s unrestricted editing capabilities were marketed as a feature, but the resulting viral misuse shows the risks of giving powerful AI tools to the public. Companies using or deploying AI must consider safeguards and moderation, as regulatory action, legal liability, and brand reputational risk can escalate rapidly when AI is misused.

EVERYTHING ELSE THAT HAPPENED IN AI LAST WEEK

• Liquid AI releases LFM2-2.6B-Exp, a tiny experimental model for on-device use with strong performance in math, instruction following, and knowledge benchmarks.

• Alibaba introduces MAI-UI – an AI agent that autonomously controls smartphone apps and completes multi-step tasks.

• Tencent open-sources Hunyuan Motion 1.0 – a 1B parameter model generating 3D character animations from text for games and media.

• Adobe partners with Runway (Gen-4.5 models) – bringing advanced AI video tools to Adobe Firefly.

• dbt Labs releases O’Reilly report on AI-ready analytics infrastructure – focusing on governed, discoverable data pipelines.

• Liquid AI LFM2-2.6B-Exp for on-device use – small footprint AI model performing well in reasoning and math tasks.

• Alibaba Qwen-Image-2512 – upgraded text-to-image model with improved realism and text rendering.

• SoftBank acquires DigitalBridge for $4B – adding data center and digital infrastructure to its AI strategy.

• SimilarWeb AI traffic stats – ChatGPT share fell from 87% to 68%, Google Gemini tripled to 18%.

• XAI Grok / Claude experiments – shows agentic AI increasingly capable of automating tasks with minimal human input.

AI keeps evolving at breakneck speed, reshaping industries and opening new opportunities, but staying ahead requires knowledge, strategy, and insight. Don’t miss a beat: subscribe to our newsletter to get the latest AI news, insights, and actionable takeaways straight to your inbox.

Don't miss news from us

Or join AI Mastery Program for business leaders >