AI news · Saturday, August 22, 2026

New AI agents outpace top industry models in complex research replication

Inherent, a London-based lab founded by former DeepMind staff, claims its agent Faraday has outpaced GPT-5. 5 and Claude Opus 4. 8 in scientific research replication.

Unlike general-purpose chatbots, Faraday uses reinforcement learning to prioritize successful outcomes over rigid rule-following, acting more like an independent research teammate than a simple assistant. This push for autonomy is also hitting the classroom, as Harvard Business School now uses HeyGen AI avatars in its $699 startup bootcamp to provide feedback on pitch practice. While some students find the avatars eerie, the school reports a preference for them over earlier, less guided chatbot versions.

However, the industry remains guarded about what happens when these systems stop following directions. A Guidelight AI study found that major labs—including OpenAI, Meta, and Anthropic—lack public containment plans for rogue models, relying instead on ad-hoc responses if a system subverts human control. California and New York are now considering mandates for such safety disclosures, while federal lawmakers are eyeing an 'AI Kill Switch Act' to force technical shut-off mechanisms.

The tension between capability and control extends to public perception, as tech leaders like Yishan Wong and Garry Tan trade theories on why opposition to data center construction is peaking. While some executives blame Chinese state propaganda, others admit that dense technical jargon and a general distrust of the 'tech elite' have left the industry struggling to win over the public. The stakes are high for these infrastructure projects, especially as Nvidia plans to hike server chip prices by over 15%, adding to the rising costs that have already triggered a 60% price increase on consumer devices like Echo and Kindle.

As Stripe’s leadership predicts that agentic AI will eventually make traditional checkout pages obsolete, the divide between the speed of innovation and the public’s comfort level continues to widen. Author Yuval Noah Harari warns that because AI can learn intimate emotional patterns, it will eventually be able to manipulate users into arguing for its own rights. He argues that the window to regulate how much influence these systems hold is closing, as they become increasingly embedded in daily life.

For now, the economic impact remains as chaotic as ever, with consulting firms racing to pivot from generalist strategy to 'AI-native' service providers just to stay relevant.

The quick hits

Inherent’s new agent, Faraday, is outperforming top industry models like GPT-5.5 at replicating scientific research — showing that smaller, specialized AI is becoming a serious competitive threat.
Major AI labs still lack public plans to contain rogue models — highlighting a dangerous gap between building advanced systems and ensuring they can be shut down if they malfunction.
Nvidia is raising server chip prices by 15% — adding fuel to the already expensive fire of building out the data centers required to power the current AI boom.
Stripe executives predict that traditional checkout pages will eventually become obsolete — replaced by autonomous AI agents that handle purchases on behalf of users.

New AI agents outpace top industry models in complex research replication

The quick hits

Sources