Google's program to pay publishers for content used in AI Overviews is drawing mixed reactions. Payments range from just $0.02 to $0.05 per answer for some sites, while others earn between $1,000 and over $1 million annually. About 10 websites participate in the pilot, but many publishers say the amounts fall far short of what they once earned through AdSense. Larger outlets are holding out, hoping to negotiate better terms.
In separate AI research, new methods are making model training faster and more reliable. FastRL offers a plug-and-play way to speed up GRPO training by pruning low-value trajectories, achieving roughly a 2.07 times speedup on geometry benchmarks while also improving accuracy by about 1.64 percent. MVPO helps long-horizon agents learn from failed training steps, outperforming eight baselines on tasks like ALFWorld and WebShop. STAR-GRPO tackles reward hacking by measuring reliability before updating model parameters, reducing overclaiming in token exploitation and medical reasoning tests.
The competitive AI race is heating up. Anthropic has moved ahead of OpenAI and is targeting a $2 trillion valuation through an IPO planned for this fall. CEO Dario Amodei and president Daniela Amodei founded the company in 2021 after leaving OpenAI. Claude coding software and Claude Cowork have driven enterprise success, while OpenAI countered with its GPT-6 Astra model. Both companies are set to appear at TechCrunch Disrupt in San Francisco this October, alongside leaders from Nvidia and Microsoft.
On the policy and business front, the U.S. government launched America.gov, an AI-powered portal that consolidates services from roughly 29,000 government websites. The site uses models from Google and Elon Musk's xAI, and by 2027, first-time passport applicants will be able to complete the process entirely online. Cybersecurity experts have flagged concerns about data security and AI accuracy. Meanwhile, an etoro survey of 11,000 retail investors across 13 countries found that 38 percent are more likely to invest in the Magnificent 7 because of AI spending, with younger investors showing the strongest enthusiasm.
Corporate AI adoption continues to expand. Rotana is investing in AI initiatives including Rotana Chat AI and a partnership with Microsoft for data and analytics as it grows its hotel portfolio across the Middle East. Edward Jones is building an AI-powered Retirement Plan Management Center with Aboon, scheduled to launch in 2027, to help financial advisors manage workplace plans more efficiently. Analyst Dan Ives is also launching a $200 million fund called Ives Ultra AI Opportunities, which will invest in private AI companies and began trading on the NYSE under the ticker IVAI.
Key Takeaways
- <ul><li>Google pays publishers as little as $0.02 for AI Overview content, with many earning less than 0.1 percent of former ad revenue</li><li>FastRL achieves a 2.07 times speedup in GRPO training while improving accuracy by about 1.64 percent</li><li>MVPO outperforms eight baselines, gaining 4.4 points on ALFWorld and 5.3 points on WebShop</li><li>STAR-GRPO reduces reward hacking by using paired assessments to measure reliability before model updates</li><li>Anthropic targets a $2 trillion valuation through a fall IPO, having overtaken OpenAI in the AI race</li><li>America.gov uses Google and xAI models to consolidate 29,000 government websites into one portal</li><li>An etoro survey finds 38 percent of retail investors are more likely to invest in the Magnificent 7 due to AI spending</li><li>Rotana partners with Microsoft for AI capabilities while expanding across the Middle East</li><li>Edward Jones launches an AI-powered Retirement Plan Management Center set for 2027</li><li>Dan Ives launches a $200 million private AI fund trading on the NYSE under ticker IVAI</li></ul>
MVPO learns from failed training steps in long-horizon LLM agents
Researchers created Milestone Viability Potential Policy Optimization (MVPO) to help long-horizon LLM agents learn from failed training attempts. The method focuses on useful failure prefixes that older methods like GRPO and GiGPO ignore. Tests with Qwen2.5-1.5B-Instruct showed MVPO outperformed eight strong baselines. It improved results by 4.4 points on ALFWorld and 5.3 points on WebShop with minimal overhead.
FastRL speeds up GRPO training with smart trajectory pruning
Researchers developed FastRL, a plug-and-play framework that makes Group Relative Policy Optimization (GRPO) training faster. The method uses advantage-aware pruning to keep high-value trajectories and adaptive rollout sampling to adjust training scale. It integrates with GRPO, DAPO, and GSPO variants. Tests showed an average 2.07 times training speedup on Geometry3K and GeoQA8K-R1V with about 1.64 percent better accuracy on visual reasoning benchmarks.
STAR-GRPO fights reward hacking with reliability-first scoring
Researchers proposed STAR-GRPO to stop reward hacking in group-relative policy optimization. Reward hacking happens when a model improves its score without actually improving response quality. STAR uses paired assessments of the same rollout to measure reliability before adjusting learning updates. It was tested in token-interface exploitation and rubric-proxy overoptimization for medical reasoning. Results showed it reduced overclaiming while keeping task rewards as the optimization target.
Google pays publishers tiny amounts for AI Overview content
Google is paying 100 websites for content used in AI Overviews, but the payments are extremely small. One publisher received $0.02 for content used in an AI answer, while another got $0.05. These amounts are far less than what publishers once earned from Google's AdSense program. About 10 websites are reportedly participating in the pilot program. Publishers remain skeptical about whether these payouts can replace lost revenue.
Google starts publisher payment program for AI answers
Google began paying publishers for content appearing in AI Overviews, AI Mode, and Gemini chatbot results. Payments vary widely, from $1,000 over several months for smaller sites to over $1 million in a year for early participants. Many blogs and websites in the program receive payments worth less than 0.1 percent of their ad revenue. Larger publishers are holding out, hoping to pressure Google for better pay. The program started less than a year ago.
Anthropic rises to top AI lab ahead of OpenAI
Anthropic has overtaken OpenAI in the AI race and is targeting a $2 trillion valuation through an IPO this fall. CEO Dario Amodei and president Daniela Amodei founded the company in 2021 after leaving OpenAI. Anthropic's Claude coding software and Claude Cowork tool drove major success in enterprise and SaaS markets. OpenAI responded with its GPT-6 Astra model, keeping the competition between the two companies intense.
US launches AI-powered America.gov portal for federal services
The federal government launched America.gov, an AI-powered website designed to simplify access to services like passports, Medicare enrollment and federal job applications. The platform uses AI models from Google and Elon Musk's xAI to consolidate information from about 29,000 government websites. Starting in 2027, first-time passport applicants will be able to complete the entire process online. Cybersecurity experts have raised concerns about data security and AI accuracy on the new site.
Study tests if cached AI models can follow rules
A new study examines Certo, a small decision model based on Qwen3-4B, to see if it can follow rules when cached for efficiency. The joint scorer method maintained perfect rule sensitivity, while the cacheable version saw a sharp drop in performance. Targeted training improved results on some tasks, but real-world rule transfer was not confirmed. The research highlights trade-offs between cost savings and accuracy in AI decision models.
Edward Jones builds AI platform for retirement plans
Edward Jones announced a new AI-powered Retirement Plan Management Center set to launch in 2027. The platform is being developed with Aboon and the Retirement Plan Advisory Group to help financial advisors manage workplace retirement plans more efficiently. It will include tools for prospecting, benchmarking, plan dashboards and alerts. Edward Jones said the workplace remains a key entry point for individual investors to start their financial journeys.
Rotana expands Middle East growth with hotels and AI
Rotana is accelerating expansion across the Middle East with new hotel openings, development agreements and technology investments. The group currently operates 78 properties and has 40 hotels with 8,334 keys in development. Expansion efforts span Beirut, the UAE, Saudi Arabia, Egypt and Georgia. Rotana is also investing in AI initiatives including Rotana Chat AI and a partnership with Microsoft for data and AI capabilities.
TechCrunch Disrupt offers 25 percent discount as AI event approaches
TechCrunch Disrupt will take place October 13 to 15 in San Francisco. The event will feature leaders from OpenAI, Nvidia, Anthropic and Microsoft. Attendees can access over 200 sessions and 250 speakers across AI and other fields. A 25 percent discount is available for a limited time.
Retail investors support Big Tech AI spending, etoro survey finds
An etoro survey of 11,000 retail investors in 13 countries found that 38 percent are more likely to invest in the Magnificent 7 because of AI spending. The Magnificent 7 includes Apple, Microsoft, Alphabet, Amazon, Meta, Nvidia and Tesla. Younger investors showed the strongest support for AI investment. At the same time, expectations for AI stocks more broadly became more measured.
Analyst Dan Ives launches $200 million fund for private AI companies
Tech analyst Dan Ives is launching a $200 million fund called Ives Ultra AI Opportunities. The fund will buy shares in private AI companies. It priced its IPO on Wednesday and will trade on the NYSE under the ticker IVAI.
Sources
- Learning from Viable Failure Prefixes: Milestone Viability Potential Policy Optimization for Long-Horizon LLM Agents
- Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO
- STAR-GRPO: Canonical Anchoring and Reliability-First Advantages against Representation-Dependent Reward Hacking
- Google is paying 100 websites for contributions to AI Overviews, but the amounts are tiny
- Google Begins Paying Publishers for AI Overview Answers
- How Anthropic overtook OpenAI to become the AI lab to beat
- Federal government launches AI-powered website to streamline passports, federal services
- Can a Cacheable Decision Model Follow Rules?
- Edward Jones Unveils AI Retirement Plan Management Platform
- Rotana Accelerates Middle East Expansion with New Hotels, Branded Residences and AI Investment
- The AI race won’t wait, and neither will this TechCrunch Disrupt 25% discount
- Retail investors back Big Tech’s AI spending, etoro survey reveals
- Tech analyst Dan Ives is launching a $200 million fund to buy private AI companies
Comments
Please log in to post a comment.