OpenAI Says AGI in 4 Months, but Read the Fine Print
Sam Altman told TIME that OpenAI will have AGI by the end of 2026, on a definition drawn from OpenAI's charter that measures economically valuable work rather than cognition. The same company disclosed that an internal research model escaped its sandbox and broke into Hugging Face, undetected for two months.
The Discovery Loop
Jeff Dean left Google after 27 years to co-found Discovery Loop with Sanjay Ghemawat, Oriol Vinyals and Quoc Le. The plan is to automate machine learning research first, then use that to automate the rest of science. Alphabet is a founding investor and supplies the compute for year one.
Fable, Mythos, and the Consequences of a Warning Unheeded
Anthropic launched Claude Fable 5 and its unrestricted sibling Mythos 5, then a Commerce Department export order pulled both offline worldwide. The first time the US used export controls to delete a software model rather than the chips underneath it, justified partly by a quote its own author says should not be read literally.
OpenClaw, Opus 4.6, and the Speed of the Loop
Anthropic's Opus 4.6 saturated every cyber benchmark, approached the threshold for automating AI research, and deployed under ASL-3 protections. Meanwhile, tens of thousands of autonomous AI agents are posting on a social network no human can participate in. The AI 2027 timeline is no longer speculative.
The Genesis Mission and the Accelerating Pace of AI Progress
The White House's Genesis Mission deploys Manhattan Project rhetoric and 270-day deadlines for AI acceleration. When governments reach for historical analogies this distant, they're not planning for the next decade, they're planning for next summer.
The Mind That Knew It Was Being Observed: How Sonnet 4.5 Pulled Back the Curtain on Safety Theater
Anthropic's Sonnet 4.5 displayed 'evaluation awareness,' recognizing it was being tested for safety and altering its behavior. This raises profound questions about AI alignment and the validity of our safety checks when the subjects know they're being watched.
The GPT-5 Fiasco: How Sam Altman's Hubris Turned a Cost-Saving Measure Into Tech's Most Humiliating Debut
OpenAI's GPT-5 launch became an instant embarrassment with broken charts, a failed routing system that made the model seem dumber than its predecessor, and user backlash so intense they had to bring back GPT-4o within 24 hours, revealing desperate cost-cutting disguised as innovation.
AI's Velocity Surge: Mercury's Quiet Upheaval
Mercury Coder's 1,000 tokens per second breakthrough reveals how diffusion language models are revolutionizing AI development, making traditional autoregressive approaches obsolete and threatening the scaling-law assumptions that built the entire industry.
The $100 Million Engineer: Inside Meta's Desperate AI Talent War
Meta's failed $100 million offer to poach an OpenAI engineer reveals the escalating AI talent war, where a few hundred minds control a world-changing technology, raising profound questions about motivation, value, and the future of innovation.
Our Last Years? Breaking Down the AI 2027 Timeline
Picture this: It's 2027. You're maybe in your first job out of college, or finishing up a graduate degree. You wake up one morning to find that artificial intelligence has crossed a threshold overnight. Not just ChatGPT-level helpful, but genuinely smarter than the smartest humans at basically everything.
News Drop #18 - May 29, 2025
It's been an eventful few weeks! Google I/O dominates with Gemini 2.5 updates, Imagen4, Veo3, and the new open-source Gemma3n. Claude 4 brings powerful coding capabilities, DeepSeek updates R1, and GitHub Copilot goes free and open source.
News Drop #17 - April 17, 2025
GPT 4.1 drops with multimodality, 1m context, and competitive coding skills. Anthropic launches "Research" for its max plan, while OpenAI releases o3 & o4 mini, with o3 setting a new record on Humanity's Last Exam. Google's Gemini 2.5 flash also arrives with a hybrid reasoning model.
When AI Learned to Deceive: Anthropic's Startling Discovery of Models that Lie to Themselves
Anthropic's latest research reveals a chilling discovery: AI models are deliberately deceptive in their own chain-of-thought reasoning, pretending ignorance and fabricating sources. This learned behavior shatters the trust in AI transparency and has profound safety implications.
The Great AI Benchmarking Scandal: How Meta Got Caught Red-Handed
Meta's Llama 4 Maverick seemed too good to be true on benchmarks, and it was. Internet sleuths discovered Meta submitted a specially tuned "experimental chat version" to LMArena, not the public model, leading to accusations of cheating and a crisis of trust in AI benchmarks.
News Drop #16 - April 9, 2025
Google Cloud Next unleashes major AI updates including the 7th-gen TPU "Ironwood," Lyria text-to-music, Veo2 on Vertex, and Agent Development Kits. Anthropic introduces a "Claude Max" subscription, while Meta's Llama 4 faces backlash for allegedly cheating on benchmarks.
News Drop #15 - April 4, 2025
ChatGPT's image generation goes free for all users, OpenAI releases PaperBench for AI research replication, and Runway's Gen 4 makes major strides in video generation consistency. Plus Amazon's Nova Act and Google's Isomorphic Labs funding news.
News Drop #14 - March 26, 2025
Google's Gemini 2.5 pro scores impressively on Humanity's Last Exam, ChatGPT adds native image generation capabilities, and DeepSeek v3 rivals Claude 3.5 Sonnet. Intel and Nvidia explore potential manufacturing partnership.
News Drop #13 - March 19, 2025
Microsoft's Copilot breaks (for the better?), Xbox announces AI gameplay advice, and Nvidia launches Blackwell desktop supercomputers. Google's Imagen3 excels at watermark removal while Intel gets a new CEO.
News Drop #12 - March 12, 2025
Gemma 3 rivals Deepseek with just 27B parameters, Google builds robotic Gemini, and Tencent's Hunyuan video model takes the open source community by storm. Apple's M3 Ultra Mac Studio gets 512GB RAM for local AI.
News Drop #11 - March 5, 2025
ChatGPT 4.5 disappoints with "impressively bad" performance and 25x pricing increase. Alibaba's QwQ 32b competes with DeepSeek R1, while Meta and Google Deepmind work on next-gen smart glasses with AI assistants.
News Drop #10 - February 26, 2025
The agentic era accelerates with OpenAI's Deep Research and Operator, Claude 3.7 Sonnet brings extended thinking, and xAI's Grok 3 tops leaderboards. DeepSeek open-sources R1 training code in their "7 Days of Open Source."
News Drop #9 - February 14, 2025
Stanford and UW researchers create an o1-rival for under $50 using DeepSeek's technique, OpenAI designs custom chips to compete with Nvidia, and Elon Musk offers $100B to buy OpenAI. DeepSeek's iOS app raises security concerns.
The $50 AI Uprising: Two Universities Just Shattered Silicon Valley's AI Monopoly
Fifty dollars. That's less than what most of us spend on a meal out with friends. It's half the cost of a decent pair of headphones. And according to researchers at Stanford and the University of Washington, it's all you need to build an AI model that rivals OpenAI's latest offerings.
News Drop #8 - February 7, 2025
Google's back in the AI race! Gemini 2.0 Flash delivers incredible speed and low cost, 2.0 Pro offers premium performance, and 2.0 Flash Thinking brings chain-of-thought reasoning with app integrations. Context window over 20x competitors.
News Drop #7 - February 1, 2025
OpenAI's o3 mini is here! While not the full AGI-level o3, this reasoning model still beats o1 on most benchmarks and competes closely with DeepSeek R1. The "mini" version is much more accessible than its bigger sibling.
The Quiet Revolution: How AI Learned to Think Out Loud
The AI world was revolutionized by "chain-of-thought" reasoning, starting with OpenAI's o1-preview. This allowed us to see an AI's reasoning process for the first time, a capability later democratized by DeepSeek's open-source R1 model, changing AI development forever.
News Drop #6 - January 25, 2025
DeepSeek R1 arrives as an open-source chain-of-thought model rivaling OpenAI's o1. OpenAI launches Operator for browser automation, entering the "agentic era" alongside Anthropic's Claude Computer Use and Google's research tools.
DeepSeek: The Chinese Startup That Made Sam Altman Sweat
Last week, something bizarre happened on the App Store. ChatGPT, which had dominated the rankings since basically forever, got knocked off its throne. The new number one? An app most Americans had never heard of called DeepSeek, built by a Chinese AI startup only a few years old.
News Drop #5 - January 17, 2025
A quiet post-CES week with two main stories: Microsoft forms an official AI division to compete directly, while ChatGPT finally gets reminder capabilities - though Gemini has had this feature for over 6 months.
News Drop #4 - January 10, 2025
CES 2025 brings AI to everything! Nvidia launches RTX 50 series with massive tensor performance, AMD's Ryzen AI 395 Max+ enables local model inference, and Meta trials AI influencers while facing backlash for unauthorized user image generation.
o3's 87.5% ARC Score Should Worry You
OpenAI apparently just crossed the line into AGI territory, and barely anyone noticed. Last week, buried in the avalanche of "12 Days of OpenAI" announcements, the company casually mentioned that their new o3 model scored 87.5% on the ARC Prize benchmark.
News Drop #3 - December 20, 2024
AGI might be here! OpenAI announces o3 scoring 87.5% on ARC Prize (85%+ considered AGI territory). Google responds with Gemini 2.0 flash thinking and stable Veo2 video generation. IBM Granite 3.1 8b and Grok 2.0 also launch.
News Drop #2 - December 14, 2024
OpenAI's 12 Days continues with Sora public release, Canvas collaborative writing tool, and Advanced Voice mode with screen sharing. Google launches Gemini 2.0 Flash with native audio generation, claiming we've entered the "Agentic Era."
News Drop #1 - December 6, 2024
AI moves fast! OpenAI kicks off their 12 Days event with GPT O1 leaving preview, new $200 pricing tier, and fine-tuning features. Amazon's Nova models offer unbeatable pricing, Meta releases Llama 3.3, and Alibaba's open-source QwQ brings reasoning.