Thursday, October 15, 2026
Today's briefing explores the rapid evolution of AI infrastructure and the rise of agentic models. We also cover significant shifts in how tech companies are navigating international policy and security partnerships.
Must know
AI and compute
Recent analyses show that compute usage for leading AI training runs has been growing with a 3.4-month doubling period since 2012. Additionally, new developments in private AI compute and chip design automation are shifting how infrastructure is being approached.
The exponential demand for compute is far outpacing historical trends like Moore's Law, signaling a rapid escalation in infrastructure requirements.
Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection
Google DeepMind and the Singapore government have partnered to use lightweight Gemma models for on-device scam detection that maintains data privacy.
On-device, private AI models offer a solution to high-stakes security threats without requiring sensitive data to be sent to the cloud.
AI labs are buying data from dead startups to train their models
Founders of shuttered startups are increasingly finding value in selling their companies' codebases and datasets to AI labs for model training purposes.
This secondary market for data provides a new revenue stream for failing ventures while satisfying the AI industry's insatiable need for training data.
Research
How Third Moment Research runs 10,000 agents for 36+ hours
Third Moment Research is utilizing advanced models like GPT-6 Astra and Codex to deploy and manage 10,000 long-running AI agents in the cybersecurity sector.
Demonstrates the practical application of high-capacity agent swarms in persistent, complex technical environments.
The Download: open model trade-offs and geothermal power’s promise
An industry report by Anthropic highlights significant cybersecurity risks posed by the release of powerful open models, fueling the ongoing debate over the balance between openness and safety.
The tension between the benefits of open-source models and the potential for misuse remains a central theme in AI safety discourse.
Open models have real benefits. Do they outweigh the risks?
Anthropic's latest report identifies specific cyberattack risks associated with the GLM-5.3 open model, highlighting ongoing industry concerns about model safety.
Focusing on specific model vulnerabilities provides a concrete look at why regulators and developers are debating open-source guardrails.
Tools and products
Amp puts AI agents to work beyond code with OpenAI models
Amp is utilizing OpenAI's newest models, including GPT-6 Astra, to expand the capabilities of AI agents beyond simple code generation.
This highlights the shift from conversational AI to autonomous, multi-functional agents capable of complex tasks.
The EU’s next chapter in AI is agentic
OpenAI has launched an EU Economic Blueprint 2.0 to support AI adoption across Europe, alongside new data residency options and partnerships with the UK government.
Strategic initiatives to address EU-specific regulatory and data concerns are critical for global AI expansion.
Higgsfield takes AI video from 1 minute to over 10
Higgsfield is leveraging new OpenAI models to enable the creation of video content exceeding 10 minutes in length and to facilitate voice-based filmmaking.
Advances in video model duration and precision are lowering the barrier for AI-assisted professional content creation.
Policy and regulation
Our approach to government and national security partnerships
OpenAI has published its framework for government and national security partnerships, emphasizing safety, accountability, and public-private collaboration. This comes amid broader legislative interest in coordinated government approaches to AI.
As AI becomes central to national strategy, tech companies are increasingly defining formal policies to balance innovation with government oversight.