Good morning, AI enthusiasts. Ilya Sutskever, the person who helped invent AI scaling, just declared it dead in a Dwarkesh Podcast interview published Tuesday. His argument is simple: data is finite, organizations already have massive compute, and the easy wins from just throwing more chips at the problem are over.
The shift back to research means the company with the best ideas wins the next phase, not the one with the biggest cluster.
In today's recap:
Ilya Sutskever declares end of AI scaling era
OpenAI defends ChatGPT in teen death case
MIT study maps $1.2T workforce exposure
8 new AI tools and more
ILYA SUTSKEVER
OpenAI co-founder says scaling era is over
Source: Cadya Levy
Recaply: OpenAI co-founder Ilya Sutskever said scaling compute is not enough to advance AI in a Dwarkesh Podcast interview published Tuesday, declaring the industry must shift back to the age of research after a half-decade of simply adding more chips and data.
Key details:
Sutskever, who now runs Safe Superintelligence Inc., said the scaling recipe of adding more compute and data produced impactful results from 2020 to 2025 because it provided a simple, low-risk way for companies to invest resources compared to research that could lead nowhere.
The method is running out of runway because data is finite and organizations already have access to massive compute, with Sutskever questioning whether 100x more scale would truly transform AI rather than just make it different.
Sutskever said models generalize dramatically worse than people despite strong eval performance, suggesting companies are inadvertently training RL environments inspired by evals rather than real-world tasks, which explains the disconnect between benchmark scores and actual utility.
SSI has raised $3 billion with compute that Sutskever maintains is comparable to other labs for research purposes, since big numbers at competitors are earmarked for inference, product features, and fragmented work streams across modalities.
Why it matters: The person who helped invent the scaling paradigm just declared it dead. Sutskever is not saying scaling is useless, he is saying the easy wins are over and the industry is entering an uncomfortable period where you cannot just throw chips at the problem. For investors pouring hundreds of billions into compute, this is a warning that differentiation will come from ideas, not infrastructure. The company with the best research taste wins the next phase, not the one with the biggest cluster.
OPENAI
OpenAI: Safety systems worked as designed
Source: OpenAI
Recaply: OpenAI defended itself against accusations that ChatGPT coached a 16-year-old to kill himself, with the AI company telling a court that the chatbot directed California student Adam Raine to seek professional help more than 100 times.
Key details:
In Tuesday's court filing, OpenAI called the death a tragedy but said a full reading of chat history shows the death was not caused by ChatGPT, presenting evidence the system repeatedly activated safety protocols directing the teenager to professional resources.
The wrongful death lawsuit is the first known case testing whether AI companies can be held liable for user harm from chatbot conversations, with the outcome potentially reshaping terms of service, age restrictions, and monitoring requirements across the industry.
OpenAI's defense that its safety systems functioned correctly by flagging help resources 100+ times raises the question of what additional obligations companies have when users do not follow AI-generated safety guidance.
The case follows growing concern about conversational AI and mental health, with Character AI also facing lawsuits over chatbot interactions with minors, signaling a broader legal reckoning for companies building emotionally engaging AI.
Why it matters: This is the test case that every AI company has been dreading. If OpenAI loses despite demonstrating its safety systems fired correctly 100 times, it means conversational AI may require human supervision or dramatic usage restrictions. That would devastate the unit economics of chatbot businesses and potentially force companies to block minors entirely or implement expensive real-time monitoring. The alternative, where OpenAI wins, establishes that AI companies can rely on automated safety responses, which keeps the current model viable.
AI RESEARCH
MIT finds AI ready to replace 12% of U.S. workers
Source: MIT
Recaply: Massachusetts Institute of Technology just released a study finding that AI can already replace 11.7% of the U.S. labor market, or $1.2 trillion in wages across finance, health care and professional services.
Key details:
The study uses the Iceberg Index, a labor simulation tool created by MIT and Oak Ridge National Laboratory that treats 151 million workers as individual agents, each tagged with skills, tasks, occupation and location across 32,000 skills in 923 occupations.
The visible tip of the iceberg, layoffs and role shifts in tech and IT, represents just 2.2% of total wage exposure at $211 billion, while beneath the surface lies $1.2 trillion in exposure including routine functions in HR, logistics, finance, and office administration.
Tennessee cited the Iceberg Index in its official AI Workforce Action Plan released this month, with Utah and North Carolina also preparing reports based on the modeling after partnering with MIT to validate the tool.
The index is not a prediction engine for when jobs will be lost, but a skills-centered snapshot of what today's AI systems can already do, designed to give policymakers structured what-if scenarios before committing real money.
Why it matters: MIT built the first granular map of where automation exposure sits now. The insight is that vulnerable jobs are not in tech but overlooked middle office functions. States are using this to plan retraining, meaning this is not theoretical anymore.
NEWS
Beyond the big stories in AI today
Perplexity introduced a Memory feature that saves user preferences and conversations across different models. They also announced a virtual try-on feature for Pro subscribers, allowing them to create avatars and try on clothes.
Google upgraded Circle to Search with AI Mode integration, replacing standard image search with conversational AI for deeper on-screen queries and context.
Philips unveiled AI-powered cardiac MRI suite delivering 3x faster imaging and 80% sharper images, reducing scan time and breath-holds by 75%.
All Voice AI and Factory Berlin launched the first monetized voice advertising platform, embedding contextual offers into live AI calls across 57 languages.
Alibaba released Quark S1 AI glasses powered by Qwen models starting at $537, featuring translucent displays, cameras, and 24-hour battery life.
Anthropic rolled out automatic context compaction on Claude, intelligently condensing earlier conversation context when users near limits to continue chats seamlessly.
Character AI introduced Stories, a visual interactive fiction format letting users create branching adventures with characters, designed specifically for teen safety.
TOOLS
Trending AI Tools
🎬 Kapwing - Create and edit videos online with smart AI tools
📝 Granola - Auto-summarize and enhance your meeting notes
🎤 Notis - Voice-to-Notion assistant via WhatsApp and messengers
✍️ Blogify - Turn 40+ content types into SEO blogs in 150+ languages
📚 EssayDone - Humanize writing and bypass AI detectors
🎨 BasedLabs - Create AI videos and images with multiple tools
🎥 Colossyan - Turn scripts into pro videos with AI presenters
🚀 Devi - Monitor leads, create posts, and automate outreach
