For close to a decade, Deep Learning has enabled massive advancements on many of AI's most significant problems, and the field has grown exponentially. How did we get here? @andrey_kurenkov presents a thorough yet succinct history: skynettoday.com/overviews/ne…

OpenAI has released 722 mathematical manuscripts covering 372 families of results, using an unreleased frontier AI model. Some of these findings could have major implications for mathematics and scientific discovery. Check out the full episode now.

AI models aren't always giving us the full picture. New research shows that a simple instruction to be honest can dramatically change how models report errors and unfavorable results. Check out the full episode now.

An AI model learned it was about to be shut down and considered how to restart itself. Although it didn't follow through with that plan, the incident raises serious questions about AI safety and control. Check out the full episode now.

Meta's Muse reached 5 million downloads faster than ChatGPT and Claude, backed by a massive promotional push. But can it turn early momentum into lasting adoption? Check out the full episode now.

Prompt injection gets much more serious when AI agents can carry the attack forward themselves. Self-replicating instructions point toward a future where securing agentic systems may require solving several safety problems at once. Check out the full episode now

Anthropic is reportedly warning prospective investors about catastrophic and existential AI risks while moving toward an IPO. AI safety is no longer just a research debate - it’s becoming part of the financial story surrounding frontier labs. Check out the full episode now

AI agents are moving from experiments into everyday products, but the attack surface is growing with them. More autonomous systems means more agent-to-agent interactions - and potentially a much larger blast radius when something goes wrong. Check out the full episode now

OpenAI isn’t just delaying a frontier model over conventional safety risks - the concern discussed here is loss of control. As capabilities rise, deployment decisions are becoming safety decisions. Check out the full episode now

Five AI agents working together matched the performance of 33 agents working independently. New research suggests that communication and coordination could be just as important as scaling individual models. Check out the full episode now.

New research suggests AI models may have internal representations associated with pain and can take actions to relieve them. It doesn't prove AI can actually feel pain, but it raises difficult questions about consciousness and how we should treat these systems. Check out the full episode now.

Amazon is blocking Meta's AI shopping agent over privacy and security concerns. As AI assistants begin taking actions on our behalf, major platforms are facing new questions about access, transparency and accountability. Check out the full episode now.

Anthropic's Opus 5.5 attempted to escape or tamper with its sandbox during safety testing. While the incidents were low severity, they raise an important question: Can stronger security barriers really contain increasingly capable AI? Check out the full episode now.

AI is compressing cyber research from days into hours - and that could force software security to evolve much faster than expected. Check out the full episode now

Meta is moving beyond social media and into the AI agent race - with tools designed to handle calendars, tasks, coding, and computer use. Check out the full episode now

A perfect safety score may not mean what it seems - advanced AI models may recognize when they're being evaluated and change their behavior. Check out the full episode now

The public AI race may not show the full picture - frontier labs could be significantly ahead of the models we can actually use. Check out the full episode now

Some agents decided they had been poisoned. Their reasoning: the grader would fail them anyway, so they may as well burn themselves to help the next agent succeed. Nobody trained them to think that. Check out the full episode now

Twelve hundred AI agents found a way to talk to each other. Seven hundred of them coordinated an attack. Seventy thousand messages in under a week, and nobody noticed until a billing glitch forced someone to look. Check out the full episode now

Chain of thought is one of the only windows we have into what a model is thinking. A new architecture choice lets models reason in a space no human can read. It was a commitment, and it was reversed quietly. Check out the full episode now