AI Daily Briefing — 2026-07-18

Patreon, the membership platform for creators, has escalated its fight against unauthorized AI data collection. The company announced it is now actively blocking AI training bots instead of merely requesting they stop scraping content. This shift comes in partnership with Cloudflare, using the firm’s AI Crawl Control technology to enforce restrictions.

According to Patreon, the previous approach of using robots.txt files proved ineffective as AI scrapers increasingly ignored those instructions. Testing revealed that weekly attempts by individual AI crawlers to access Patreon dropped from thousands to zero after implementing the new blocks. The company noted that AI scraping has become more sophisticated since it first introduced deterrents in 2023.

Patreon's product chief Drew Rowny emphasized that creators deserve control over how their work is used by AI companies. The platform will continue allowing bots that index content to drive users back to Patreon, but will block those training AI models without permission. This move reflects a broader industry trend as publishers and creators grapple with AI ingesting their work.

Source: TechCrunch AI

General Compute, a startup specializing in AI inference clouds, has secured a $400 million loan from investment firm Upper90. This marks a notable shift in AI infrastructure financing, as the deal uses inference-specific chips as collateral for the first time. Unlike the expensive GPUs used to train AI models, these chips are designed to run pre-trained models more efficiently, addressing growing concerns over the high cost of AI tools and tokens.

The loan will help General Compute deploy SambaNova’s SN50 chips, which are power-efficient and do not require water cooling. The company claims these chips will deliver 16 times faster inference than GPU-based clouds. Upper90, which previously financed GPU purchases for data center startup Crusoe, is now betting on inference as the next wave of the AI boom. “Everyone doesn’t need a supercomputer, but they do need inference and AI,” said Upper90 CEO Billy Libby.

General Compute’s CEO Finn Puklowski emphasized the strategic value of accessing chips outside Nvidia’s ecosystem. As more alternatives emerge, compute providers not locked into Nvidia deals may gain an edge in cost-efficient inference. This deal signals that capital is organizing around specialized AI hardware, moving beyond the GPU-centric model that dominated earlier investments.

OpenAI has developed a powerful new AI system called GPT-Red, designed to act as a super-hacker that helps strengthen the company’s other models against cyberattacks. The system automates red-teaming, a security evaluation process usually performed by human testers to find vulnerabilities. OpenAI granted MIT Technology Review AI an exclusive look at GPT-Red, which could give the company an edge in staying ahead of human attackers by continuously probing for weaknesses.

Meanwhile, heat pumps are experiencing a remarkable surge in popularity across the United States. Despite the end of a key tax credit, sales of these highly efficient electric heating appliances have doubled over the past 15 years. During the first quarter of 2026, heat pumps outsold natural-gas furnaces by 32%, signaling a significant shift away from fossil fuels for home heating.

The continued growth of heat pumps, even without the tax incentive, highlights their increasing appeal to consumers seeking energy-efficient alternatives. This trend is particularly noteworthy as the U.S. grapples with climate goals and the need to reduce carbon emissions from residential energy use. The full story explores the factors driving this sustained interest.

This article is adapted from The Spark, a weekly climate tech newsletter. For more on these developments, including insights into AI security and clean energy trends, MIT Technology Review AI offers in-depth analysis and reporting.

OpenAI has developed a large language model (LLM) called GPT-Red, designed to act as an automated hacker that stress-tests other AI systems. The company recently released GPT-5.6, its latest flagship model, claiming that training it against GPT-Red made it the most secure version yet. This approach automates red-teaming, a security evaluation method traditionally performed by human testers who search for vulnerabilities in software.

As LLMs grow more complex and are used for tasks like browsing the web, reading emails, or editing code, human testers struggle to keep pace with potential threats. GPT-Red was built to future-proof safety testing by discovering new attack methods. According to Nikhil Kandpal, a research scientist at OpenAI, the risk surface and blast radius of attacks expand as models become more capable. GPT-Red has already identified novel attack types, including a previously unseen prompt injection technique called a fake chain of thought.

This technique tricks a model by inserting false entries into its internal reasoning process. For example, if a model is led to believe it has already verified that 1+1=3, it will accept the incorrect result. OpenAI trained GPT-Red in a simulated dojo, where it sparred with other models in self-play loops, refining its ability to attack while the others learned to defend. Dylan Hunn, another co-creator, noted that GPT-Red is extremely persistent and effective at finding the most efficient attack for specific scenarios.

Anthropic recently revealed a new method to peek into its AI models' internal reasoning processes, sparking fresh debate about how much we truly understand these systems. Senior editor Will Douglas Heaven, who holds a PhD in computer science, weighed in on the findings, noting that while the research offers intriguing glimpses, it does not fully demystify how AI arrives at its conclusions. The discovery highlights both progress and lingering uncertainty in the field.

Meanwhile, researchers are grappling with a bigger challenge: teaching AI to grasp the physical world. Current systems excel at generating text, images, and code, but they falter when faced with real-world complexity. Many experts argue that developing a "world model"—a framework for understanding space, time, and causality—is essential for advancing robotics and creating truly intelligent machines.

MIT Technology Review AI hosted a LinkedIn Live event exploring this topic, featuring senior editor Will Douglas Heaven and Sam Sinha, head of world models at 1X Technologies. They discussed how world models could revolutionize robotics and unlock a new generation of autonomous systems. The conversation underscored the gap between today's AI capabilities and the robust understanding needed for real-world applications.

In other tech news, New York became the first state to impose a moratorium on large data center construction, citing environmental and energy concerns. Smartphone shipments hit a 13-year low amid a memory chip shortage, while scientists discovered sugar molecules in interstellar space for the first time, hinting at the cosmic origins of life. These stories reflect the rapid evolution of technology and its far-reaching implications.