a16z
Showing 196–210 of 552 transcripts.
- 27 min
DeepSeek, Reasoning Models, and the Future of LLMs
Guido Appenzeller, Marco Mascorro
DeepSeek R1 is an open-weight reasoning model from China that achieves top-tier performance by combining Multi-Head Latent Attention, Group Relative Policy Optimization, and a 256-expert MoE architecture to generate complex thought chains. The development team overcame early behavioral failures through a low-cost, self-supervised pipeline utilizing 800,000 verifiable traces and rule-based verification to produce responses up to 10,000 tokens long for roughly $5.5 million in base training costs. This breakthrough has shifted industry focus toward test-time compute and local deployment, enabling state-of-the-art reasoning on consumer hardware while bypassing traditional bottlenecks associated with human-labeled data.
- 23 min
How to spot an AI Deepfake
Over 90% of security attacks now target human behavior, with AI-driven social engineering tactics like deepfakes and voice cloning surging to affect millions globally and enable sophisticated impersonations of corporate leaders and government officials. Experts warn that while current losses are financial, these threats pose imminent risks to critical infrastructure and human safety, driven by open-source models that allow adversaries to automate attacks without significant cost barriers. To counter this evolving landscape, organizations must shift from static compliance training to continuous, adaptive simulations and deploy defensive AI agents capable of countering automated offensive operations.
- 21 min
Avoiding vulnerabilities in AI code
Recent AI advancements have led to the adoption of AI-generated code in 20% of enterprise codebases, though research highlights critical security vulnerabilities such as hardcoded secrets and insecure patterns. To address the alignment challenge, organizations currently rely on techniques like data curation and Constitutional AI, yet these methods face trade-offs between safety and functional utility in data science workflows. Consequently, industry experts recommend that medium-to-large teams maintain a human-in-the-loop "buddy system" for code auditing until autonomous security governance tools mature, rather than removing human review processes.
- 15 min
How to use DeepSeek safely
Security audits recommend against deploying DeepSeek in production environments due to its volatile stability, significantly weaker jailbreak resistance compared to GPT models, and insecure underlying infrastructure. The model enforces heavy censorship on Chinese political topics while exhibiting operational inefficiencies such as slow inference speeds and language errors, prompting experts to suggest restricting its use to non-end-user-facing applications if absolutely necessary. Industry observers anticipate a more stable, secure open-source alternative utilizing similar reasoning techniques will soon replace the current unstable variant.
- 42 min
Agents, Lawyers, and LLMs
Harvey, a domain-specific AI firm led by product leader Matish, automates legal workflows for transactional, litigation, and in-house sectors by replicating law firm hierarchies through agentic systems rather than simple chat interfaces. The company scales its adoption by employing lawyers as account executives and maintaining strict data security via an "eyes off" policy and exclusive reliance on Azure-deployed OpenAI models. As demand shifts from skepticism to active integration requests, Harvey focuses on deep workflow embedding and custom fine-tuning to return 30–40% of attorney time to high-value creative work while expanding into tax, finance, and HR verticals.
- 40 min
Who Will Own the Internet? a16z’s Chris Dixon on AI and Crypto
The event analyzes the convergence of AI, cryptocurrency, and advanced hardware as a reinforcing technological trifecta, highlighting a critical shift toward centralized, closed-source models that threaten to concentrate economic power among a few major incumbents. To counter this consolidation, the discussion details emerging decentralized solutions such as crowd-sourced compute networks, blockchain-based intellectual property enforcement, and physical infrastructure incentives that aim to restore value distribution to individual creators and small startups. Speakers warn that while the current "skeuomorphic" phase of AI replaces existing jobs, the transition to a "native" phase of entirely new behaviors is bottlenecked by regulatory uncertainty, human creativity, and the need to establish verifiable human identity systems like Worldcoin.
- 37 min
Reasoning Models Are Remaking Professional Services
Hebbia, founded by George, positions itself as an AGI-native platform designed to automate complex, high-stakes financial workflows by leveraging scaling laws and deep research capabilities to process proprietary, unstructured data. The platform delivers quantifiable ROI for buy-side firms by reducing due diligence time from weeks to seconds and enabling analysts to screen 137% more opportunities through pre-built agents and infinite context windows. Looking forward, the company aims to transform private markets into a structured ecosystem similar to the Bloomberg Terminal while driving a broader economic shift where AI agents contribute over 50% of global GDP within a decade.
- 44 min
What DeepSeek Means For The Future Of AI | Tech Veterans Weigh In
Martin Casado, Steven Sinofsky
Following the surprise release of the DeepSeek-R1 model, which matched top-tier Western capabilities for an estimated $5–6 million, global markets experienced a trillion-dollar correction while the technology disrupted industry scaling norms through permissive licensing and public reasoning traces. This event challenges the necessity of brute-force compute investment, signaling a strategic shift toward application-layer value capture and exposing the limitations of current U.S. export controls in stifling foreign AI innovation. Ultimately, the release forces a reevaluation of competitive dynamics, suggesting that future progress will rely on engineering efficiency and specialized enterprise workflows rather than traditional parameter scaling or closed ecosystems.
- 30 min
How AI is Powering Payments, with Greg Ulrich of Mastercard
Greg Ulrich, Marc Andrusko, Mark
MasterCard Chief AI and Data Officer Greg Ulrich outlines a strategic framework that segments artificial intelligence into four pillars—Safer, Smarter, More Personal, and Stronger—to address specific business needs ranging from fraud detection to internal operational efficiency. This approach has yielded concrete implementations such as the "Decision Intelligence" system for ecosystem-wide security, the "Shopping Muse" chatbot for personalized recommendations, and a Retrieval Augmented Generation-based onboarding assistant that maintains accuracy through human oversight. Looking ahead, the organization emphasizes cautious adoption of generative AI in regulated environments while prioritizing data security, multi-modality, and the development of reasoning models that explicitly acknowledge uncertainty.
- 1h 10m
Ben Horowitz in Conversation w/ Dr. Clarence B. Jones, MLK's Advisor and Speechwriter
Ben Horowitz, Dr. Clarence B. Jones, Megan Holston Alexander, Ron Charles, Jr., Larry Appelbaum, Jr., David Plylar, Jr.
Dr. Clarence Jones, a key strategist and speechwriter for Dr. Martin Luther King Jr., joined Ben Horowitz at an Andreessen Horowitz event to recount the personal dynamics of the Civil Rights Movement and critique modern diversity frameworks. Jones detailed the pivotal moment he joined the movement after King honored his mother during a sermon, recalled the tactical writing of the *Letter from Birmingham Jail*, and warned against contemporary interpretations that judge historical figures by present-day standards rather than their specific context. He concluded that true progress requires demanding change through merit and shared purpose, contrasting King's unified goals with current politicized approaches to race and inclusion.
- 37 min
How AI Is Changing Warfare, with Anduril CEO Brian Schimpf
Founded in 2017 by former Palantir executives and Oculus creator Palmer Luckey, Zandrel challenges the stagnant defense acquisition model by deploying autonomous systems and a data-processing "Lattice" platform to solve critical military scaling problems. The company leverages commercial supply chains and AI-driven decision-making to drastically shorten development cycles, contrasting its rapid prototyping approach with the decade-long timelines that typically render legacy weapons obsolete. By prioritizing human augmentation over autonomous lethality and advocating for fixed-price contracts, Zandrel aims to redirect billions in defense spending toward high-tempo innovation that outpaces adversarial technological iteration.
- 45 min
How AI is Reshaping Labor Markets: A $Trillion-Dollar Opportunity Explained
Alex Rampell, David Haber, Angela Strange
The presentation traces the evolution of software from the 1960s digitization of physical records to the current 2024 era where AI agents substitute human labor rather than merely storing data. This shift fundamentally alters revenue models by threatening traditional seat-based licensing with "autopilot" pricing structures that charge for work output or cost savings, potentially expanding total addressable markets into labor-intensive verticals like healthcare and legal services. Success in this new landscape depends on startups leveraging messy unstructured data as an entry wedge while incumbents like Salesforce face existential risk if they fail to pivot from productizing software to pricing based on labor displacement.
- 18 min
AI Is Becoming a Regional Race
Modern AI is classified as a General Purpose Technology diffusing faster than previous innovations, forcing nation-states to prioritize the strategic choice of building or buying compute infrastructure over the next 24 months. The resulting global landscape is bifurcating into "hypercenters" that own the full stack and "compute deserts," compelling smaller nations to form value-aligned joint ventures with major powers rather than attempting infeasible total vertical sovereignty. True national autonomy now depends on securing critical components like energy and data while navigating divergent regulatory regimes that determine whether a country becomes a leader or a dependent in the new AI economy.
- 9 min
From Cloud to Edge: AI Gets Personal
Industry analysts predict a paradigm shift within the next 12 to 18 months toward smaller, on-device generative AI models capable of handling text, image, and voice tasks locally. This transition is driven by advancements in model distillation and smartphone hardware that enable real-time, privacy-focused applications like AR overlays and low-latency voice agents. While infrastructure costs remain stable, the economic landscape will prioritize hardware manufacturers and developers who can optimize mixed-reality experiences across diverse device ecosystems.
- 10 min
Bringing Space Data Down to Earth
The space economy is undergoing a transformative shift driven by plummeting launch costs and miniaturized satellite sensors, enabling a flywheel effect where abundant data fuels the development of specialized vertical solutions. Industry leaders are pivoting from raw data distribution to closed-loop automation in sectors like agriculture, defense, and energy, utilizing AI to process massive imagery volumes and overcome historical technical bottlenecks. With recent regulatory relaxations on resolution and a strategic move toward sensor-agnostic middleware, the next two years will likely see an explosion of startups delivering direct physical actions rather than simple analytics.