[1/3] AI Leadership Summit 2025 - AI Leadership (AI column)
I looked with interest. video The first day of the AI Summit, which took place recently in New York. On this day, there were reviews of investors from AI funds that invest in AI, as well as founders of startups or representatives of large companies that are now actively experimenting with Gen AI technologies. Separately, it is very uncomfortable when all the performances are arranged in one video, which I watched a week in the morning:) Let me share with you a brief summary of the presentations that took place at this conference.
Lux Capital: Trends in 2025 (00:17:20) Grace Isord of Lux Capital presented an overview of the current state of AI. She noted the exponential growth of AI technologies in recent years. 2,5 She mentioned Stargate, OpenAI and DeepSeek, and then went on to discuss the “moment of AI agents” and noted that autonomous agents do not work perfectly yet due to various types of errors and showed this by the example of an agent that helped her with ordering tickets for confu
- Decision error - wrong facts (confused)
- Implementation error - wrong access/integration (We need to overcome Captcha.)
- Heuristic error - wrong criteria (Not counting traffic on the way to the airport.)
- Taste error - wrong personal preferences (The speaker is not ready to fly Boeing 737) She described five strategies for improving agents:
- Data curation (data management)
- Important of evals (importance)
- Scaffolding systems (Creation of reliable systems)
- UX is the moat that matters (UX - the primary meaning)
- Mutlimodality (multimodality.)
AI-execs: Building an AI Strategy That Fails (00:35:11) An ironic report from two tops on how to create an AI strategy that fails. The story is in the form of a book called Harmful Advice by Gregory Oster, which I loved to read as a child. Speakers described common mistakes: creating disparate teams, unclear definition of strategy, using incomprehensible jargon, misallocating tasks, focusing on tools instead of processes, and ignoring data. In fact, acting from the opposite can avoid many problems when developing an AI strategy.
Privacera: Building with AI Safety and Security (00:52:05) Don Bosco Duri, co-founder and CTO Privacera, spoke about the safety and reliability of AI agents. Agents, tasks, and tools typically work in a single OS process, which creates security issues. Dury proposed a three-tiered approach: risk assessment before implementation, ensuring proper authentication and authorization, and real-time surveillance of agents.
Augment Code: Self-Coding Agents (01:10:07) Colin Flary of Augment Code demonstrated an AI agent that helps write your own code. More 90percentage 20,000 Their agent's code lines were written by the agent himself under human supervision. The agent was able to add third-party integrations, write tests, and even optimize its own performance. Flary emphasized the importance of a context engine, reasoning capabilities, and a secure code execution environment.
Neo4j: Knowledge Graphs (02:07:32) Steph Chin of Neo4j and Jonathan Low of Pfizer discussed using graph databases to improve RAG systems. Lowe spoke about a project at Pfizer where graph databases helped accelerate the transfer of technology from laboratory research to industrial drug manufacturing. They explained how graph databases provide better context for LLMs and reduce hallucinations. In fact, the story is just like a promo combination of using embeddings for documents + additional links between the medical records themselves, which can be stored as links in a graph.
Continue the review of the AI Summit conference in the following posts.
#AI #Software #Product #Management #Leadership