METR Releases New Metric "Expenditure Horizon" for Measuring AI Agent Cost Efficiency

METR, an AI evaluation organization, has announced a new metric called the "expenditure horizon" that visualizes in dollar terms the point at which an AI agent becomes more expensive than a human worker. According to The Decoder, the metric is a first-pass attempt to give AI agent cost efficiency a concrete dollar-based benchmark, and the initial results from applying it to NanoGPT's speed-verification tasks were "modest."

**Why it matters**: AI agent evaluations have tended to focus on accuracy and success rates, leaving the question of "how much is being spent" on the back burner. If cost-oriented metrics mature, the inputs for business adoption decisions could change significantly. METR itself acknowledges blind spots, though, so how the results shift with next-generation models remains to be seen.

MIT Tech Review Maps the Path to "Decentralized Superintelligence"

On July 27, MIT Technology Review published an article laying out a route to "decentralized artificial superintelligence (ASI)" built on coordination across multiple AI agents. The piece centers on the vision of Vijoy Pandey, Senior VP/GM at Cisco's "Outshift by Cisco" organization, who argues that "the intelligence is already there. What's missing is the connective tissue that turns four strangers into one team."

The article describes two layers that provide this connective tissue. A lower connectivity layer, the "Internet of Agents," handles discovery, authentication, and message exchange between agents, while the semantic layer above it, the "Internet of Cognition," enables the sharing of intent, context, and reasoning. As implementations, the article cites "AGNTCY," an open-source project under the Linux Foundation, for the connectivity layer, and "Mycelium" as a coordination protocol for the semantic layer.

What's particularly interesting is that Pandey cites internal validation of the effect of adopting shared protocols. Unstructured groups reached a decision in only about a third of 14 scenarios, but inserting a coordination protocol that mandates goal declaration, surfacing of missing information, and upfront conflict resolution improved that figure to 93%. Meanwhile, a study that evaluated seven open-source multi-agent systems reported failure rates ranging from 41% to roughly 87%, and the piece also offers the view that multi-agent coordination is still "an architecture problem, not a prompting problem" (Pandey).

**Why it matters**: The main battleground is shifting from the race to scale single models toward how fleets of agents are coordinated. Even at Japanese companies, moves are emerging to trial multi-agent configurations for workflows that span multiple departments, and the gap between failure and success rates suggests it could narrow depending on how the "coordination layer" is designed.

MIT Tech Review: AI Drug Discovery Runs Into the "Data Loop" Wall

Another July 27 article from MIT Technology Review reports on how "data quality and closedness" has become a bottleneck in AI-driven drug discovery. The piece is built around an interview with Paul Belcher, director of protein research strategy at life sciences company Cytiva.

On the drug discovery front, while AI is improving the efficiency of identifying hit compounds, it is also swelling the workload of validation labs. Traditional screening is a mechanism for gathering large volumes of low-precision yes/no data, so when AI churns out many high-quality candidates, the step of evaluating each one in detail can't keep up. Belcher notes that "by increasing the hits, AI is driving demand for higher-throughput, information-rich technologies for validation and characterization."

The more fundamental challenge is the data itself. Many existing models are trained on public data and are hitting a "data wall." Belcher is sharply critical of publication bias: "Most datasets and papers only deal with positive outcomes. Because failures aren't shared, AI can learn the patterns of success but can't capture the full picture of failure." As the industry joke — "if only there were a journal of negative data" — suggests, failure data remains buried in lab notebooks.

The risk of image manipulation is also rising. Western blot, a protein detection technique, is one of the elements of biomedical papers most frequently targeted for manipulation. Microbiologist Elisabeth Bik's 2016 finding that roughly 4% of biomedical papers contain duplicated or manipulated images predates generative AI. Cytiva offers an "Image Integrity Checker" that detects manipulation using the same hashing technology as blockchain, and it is reportedly aiming to standardize the approach in the publishing industry.

**Why it matters**: Drug discovery is one of the AI application areas where the most money is flowing, and no "drug made by AI alone" success story has yet reached FDA approval. Data quality and lab-integration challenges look likely to be key factors in whether AI investments pay off as expected.

Heated Debate on Hacker News: "AI Companies Destroying Rare Books"

A Hacker News post titled "AI companies are shredding rare books" gathered 66 points and drew 18 comments. The post itself is short, but it has sparked an active community debate over how AI training data is actually sourced, and how the preservation of intellectual property and cultural heritage should be handled.

**Why it matters**: Ethical and legal friction over training data sourcing is an unavoidable issue for both the open-source and closed camps. Set against the recent Hugging Face infringement case, this story looks likely to further intensify demands for "transparency" across the AI industry.

TrueFoundry Releases "Aitori" — Local Monitoring of AI Traffic from Claude Desktop and Other Apps

On July 27, TrueFoundry released "Aitori" on Hacker News's Show HN, an open-source tool for monitoring and controlling outbound AI traffic from employee PCs. It is Apache-2.0 licensed, the forwarding specification is documented, and it works with gateways other than the company's own.

Traditionally, when developers call LLM APIs directly, the traffic can be routed through a gateway. But tools that invoke models via the vendor's backend — such as Claude Desktop, the ChatGPT app, and claude.ai — cannot be pointed at a proxy through endpoint configuration. Aitori is a local proxy that runs on the employee's device: it issues a per-device CA and, using a mechanism similar to mitmproxy, intercepts and forwards traffic destined for the AI apps you select. It is designed not to touch other apps' communications, and once the traffic is visible, it can be freely used for logging, policy enforcement, routing, and more.

**Why it matters**: Communications from AI tools that sit on endpoints tend to be a blind spot in corporate security governance. Local interception approaches like Aitori serve as a useful reference — one practical way to bring AI traffic onto the same footing as Data Loss Prevention and zero trust.

---