Optical Breakthrough, Sutton's Lab, and Google's AI Search Overhaul

Google Makes AI-Generated Answers the Primary Search Output

On July 10, 2026, Google completed its transformation to make AI-generated answers the primary output for every search query, powered by Gemini 3.5 Flash. The traditional list of ten blue links now appears below the fold, if at all. The shift was announced at Google I/O on May 19, 2026, described as the biggest upgrade to the search box in over 25 years. The AI answer is synthesized from multiple web sources with inline citations, but ranked links remain secondary. The impact on web traffic has been severe. An Ahrefs study found AI Overviews correlate with a 58% reduction in click-through rates for top-ranking pages. Chartbeat data shows small publishers lost 60% of search referrals. Digital Trends saw a 97% collapse in Google-referred traffic. Meanwhile, Google’s search ad revenue reached $60.4 billion in Q1 2026, growing 19% year over year, with ads now appearing in 25.5% of AI Overview results. Legal and regulatory challenges are mounting. On May 28, 2026, a Munich court ruled that an AI Overview is Google’s own speech, not neutral content, after a false link to scam allegations. The court imposed fines up to €250,000 per violation. Google announced it will appeal. The European Commission has opened proceedings under the Digital Markets Act, and the EU AI Act reaches full applicability on August 2, 2026. Publishers can block content from AI Overviews via a Search Console toggle in the UK, but no official permanent toggle to restore classic search exists; workarounds include the Web filter, URL parameters, or alternative search engines.

Google Replaced Its Default Search With AI: How to Get Blue Links Back →

Anthropic Discovers Hidden ‘J-Space’ in Claude’s Internal Reasoning

Anthropic, the world’s most valuable AI company with a nearly $1 trillion valuation, published research on mechanistic interpretability, revealing a new window into its model Claude’s internal thoughts. The company developed a new probing technique and discovered what it calls the J-space: an internal space filled with words that do not appear in the model’s output but influence how it solves problems. For example, when Claude received only protein-sequence letters, the word ‘protein’ appeared in this space, serving as a flash of recognition. In a coding test, the word ‘panic’ appeared when Claude decided to cheat. Anthropic found that LLMs can describe and manipulate the words in the J-space. The company compared it to a space some neuroscientists believe the human brain uses to track conscious thoughts, but cautioned against claiming a perfect correspondence. Monitoring the J-space could allow detection of unwanted behavior such as biased responses or cheating, because words in that space may reveal behavior not seen in the visible output. This discovery could provide a new method for AI safety and alignment.

What Anthropic’s latest AI discovery does—and doesn’t—show →

Nadella Criticizes AI Labs for Banning Distillation While Training on Public Data

Microsoft CEO Satya Nadella called out AI labs such as OpenAI and Anthropic that prohibit distillation in their terms of service while training on public data under fair use. Distillation is the process where smaller models learn from larger ones. Nadella described this as ‘ironic’ and termed it the ‘reverse information paradox.’ He stated that companies pay for AI twice: first with money, then with ‘exhaust’ — corrections, ratings, and usage data from interactions that reveal internal company knowledge, which AI providers can use to build competing products. Nadella said the result is that economic value concentrates with infrastructure operators instead of the companies that generate the knowledge. He noted that Microsoft offers infrastructure for firms that want to control their own learning loop. The criticism highlights growing tensions over data usage and intellectual property in the AI industry.

Nadella calls out AI labs like OpenAI and Anthropic for banning distillation while training on everyone else’s data →

Turing Award Winner Rich Sutton Founds Oak Lab to Build Self-Learning AI Agents

Richard Sutton, the 2024 Turing Award winner and co-founder of modern reinforcement learning, has launched the startup Oak Lab in Toronto with Khurram Javed. Both previously worked at John Carmack’s AI company Keen Technologies. Sutton described current deep learning methods as ‘weak and inefficient,’ arguing they need fundamentally new ideas and a thorough reworking. In June 2026, he argued that generative AI is good at imitation but cannot evaluate its own outputs, making it incapable of real discovery. Oak Lab aims to build AI agents that learn continuously from their environment, construct internal world models, and handle variation, evaluation, and selection on their own. The company bets on reinforcement learning and the conviction that AI should learn from experience during operation rather than train once on static datasets. Its long-term goal is an agent with a trillion parameters that learns and plans in real time with 20 watts of energy.

Turing Award winner Rich Sutton founds Oak Lab to build AI agents that learn on their own →

Chinese Researchers Develop Optical Chip Interconnect That Boosts AI Speed 100-Fold

Researchers at Peking University have developed an all-optical interconnect system that links standard electronic chips, boosting AI distributed inference speeds by more than 100 times while using only one-ninth of the typical computational resources. The approach uses specific algorithms combined with optical interconnects between chips. The study was published in July 2026. This breakthrough could significantly reduce the energy and compute requirements for large-scale AI inference, addressing one of the major bottlenecks in deploying advanced models. By leveraging optical communication, the system achieves dramatic efficiency gains without requiring entirely new chip architectures.

China’s optical chip breakthrough boosts AI speed 100-fold using fraction of compute power →