We turn high signal in-person events for the top AI engineers, founders, leaders, and researchers in the world into the best free learning opportunities for millions around the world here on YouTube. Your subscribes, likes, comments, speaking, attendance, or sponsorships goes a long way toward making our biz model sustainable indefinitely. We strongly believe this industry deserves a better class of community and that we know how to do this well; we just need your support.
AI agents utilized for SRE tasks like incident response require broad system access, creating significant security risks if they are manipulated or make errors. Because agents cannot be trusted to police themselves, security must be enforced at the network level rather than within the agent software. Claw Patrol, an op...
Context engineering for AI agents centers on balancing performance, cost, and memory recall within finite context windows. Contrary to common assumptions, aggressive compaction techniques like summarization often degrade performance and increase costs by invalidating prompt caching, which can reduce token expenses by u...
AI performance relies increasingly on robust infrastructure that connects models to real-time, external data rather than just static training sets. Oxylabs demonstrates this shift by building scalable pipelines for video, metadata, and search results that feed directly into AI workflows. Success in this domain requires...
Web data is evolving from a static information source into dynamic context for AI agents, necessitating a shift in how developers approach data retrieval. While AI search engines provide immediate, ad-hoc access, they struggle with high-frequency, persistent knowledge work due to compounding token costs and rapid data ...
Computer use agent (CUA) evaluation currently suffers from deterministic benchmarks that are easily gamed by "replay agents"—scripts that blindly replicate successful trajectories. These static environments render metrics like PASSAT-K misleading, as they fail to account for true model capability. To address this, the ...
The web will not be "API-fied" for AI agents; instead, computer use models that leverage vision will become the primary drivers of web interaction. Because the vast majority of the web is built for human consumption—often relying on complex, non-standardized rendering rather than clean data endpoints—expecting a transi...
Consumer AI memory systems have evolved from simple thread-based context to sophisticated "running profiles" that synthesize user data over time. ChatGPT and Claude have converged on this architecture, though they employ different trade-offs regarding token limits, update frequency, and user visibility. Memory manageme...
Continual learning in enterprise environments relies on a strategic distillation spectrum that balances offline and online data traces with corresponding hinting mechanisms. By mapping these variables into a four-quadrant framework, organizations can improve agent performance without requiring golden answers. Offline p...
Personal knowledge management relies on transforming disorganized, raw thoughts into structured, interconnected systems using LLMs. Voice dictation serves as the primary input method, capturing high-velocity thoughts that are later processed into Markdown files. Automated agents enrich these notes by adding tags, metad...
The professionalization of computer science has created an "unreasonably narrow path" for researchers, limiting who can contribute to AI breakthroughs. Historically, high barriers to entry—driven by elite academic credentials and massive compute requirements—have concentrated power within a few industry labs. To democr...
Intelligence and expertise represent distinct cognitive domains in AI development, with current frontier models excelling at the former while struggling with the latter. While raw intelligence enables reasoning through unfamiliar problems, expertise involves accumulated, situated competence that allows for pattern reco...
Scaling compute on context addresses the fundamental limitation of current AI models: the inability to acquire deep, personalized knowledge from private data after pre-training. While traditional scaling focuses on public datasets like Wikipedia or GitHub, this approach seeks to enable models to learn from specific, un...
Long-horizon AI agents frequently suffer from context bloat, leading to contradictory outputs and task drift. A robust memory harness, structured as a write-manage-read loop, addresses this by integrating specific recall and archival mechanisms. Experiments using local models like Qwen 27B and DeepSeek V4 Flash reveal ...
Continual learning represents the next frontier for AI, shifting from static benchmark training to systems that improve through real-world interaction. Current reinforcement learning methods, such as GRPO, suffer from high infrastructure demands, off-policy task distributions, and sequence-level reward limitations. On-...
Current language model evaluation relies on independent, stateless tasks that ignore a model's ability to learn over time. This paradigm fails to measure "continual learning," defined as stable, sample-efficient online adaptation. A robust evaluation framework requires three fundamental criteria: headroom for online ad...
Follow this podcast in Podwise
Sign in to get AI summaries, transcripts and mind maps for any episode, including new ones.
