VibeDigest
Podcast libraryPricingFAQ

© 2026 VibeDigest. All rights reserved.

Privacy PolicyTerms of Service

Podcasts, already organized

Topics

All topicsAI AgentsAI CodingProductStartupsResearch

274 digests ready

AI Engineer

From VLM/VLA's to Embodied Agents — Armen Aghajanyan, Perceptron AI

Perceptron proposes an embodied foundation model that unifies perception, embodied reasoning, and control while allocating computation selectively across highly imbalanced multimodal data. The speaker claims this approach improves efficiency, robustness, and data scaling, including a reported 10-fold substitution between video pretraining and costly teleoperation data, although deployment readiness remains unresolved.

10 key points

View digestOriginal episode
a16z

Fei-Fei Li is Solving the Hardest Problem in Robotics | World Labs with a16z

10 key pointsOriginal episode
Latent Space

🔬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)

9 key pointsOriginal episode
AI Engineer

From Scratch to SOTA: Training a 3B State-Space Vision Model — Krishna Prasad Srinivasan, Sarvam

Sarvam argues that a sovereign, 3-billion-parameter state-space vision-language model can outperform far larger systems in Indian document intelligence by combining block-level OCR, multilingual pretraining, verifiable reinforcement learning, and a specialized data-and-evaluation engine. Its broader objective is to convert India’s largely un-digitized multilingual knowledge into reliable, locally controlled data.

9 key points

View digestOriginal episode
The Pragmatic Engineer

Design Engineering with Maggie Appleton

AI makes prototyping and implementation dramatically faster, but it does not replace the designer’s responsibility to frame the right problem, understand the medium, and exercise contextual judgment. The next major challenge is collaborative tooling that helps humans and agents make, explain, and preserve decisions together.

10 key points

View digestOriginal episode
a16z

Replit CEO Amjad Masad on What Young People Should Learn in the AI Era

Education should give young people the trust, freedom, resources, and ethical grounding to pursue deep curiosity rather than forcing them through standardized credentials or premature careers. In an AI-transformed world, the strongest model combines project-based exploration, human development, and respect for existing knowledge with permission to challenge it.

9 key points

View digestOriginal episode

More organized episodes

AI predictions: Job markets, Codex beats Claude, and the death of org charts | Dan Shipper
Lenny's Podcast

AI predictions: Job markets, Codex beats Claude, and the death of org charts | Dan Shipper

Dan Shipper predicts that AI will reorganize work without eliminating the need for humans: company-wide agents and agent-centered desktop environments will become common, while people shift toward directing, reviewing, and improving automated work. The biggest beneficiaries will be AI-fluent product managers, full-stack designers, engineers who build reliable agent systems, and anyone who continuously experiments with new models.

Building the Document Context Layer for AI Agents — Jerry Liu, LlamaIndex
AI Engineer

Building the Document Context Layer for AI Agents — Jerry Liu, LlamaIndex

Modern RAG is no longer a fixed retrieval pipeline; it combines an increasingly capable agent with a document context layer that parses, stores, searches, and operationalizes unstructured information. Effective document infrastructure must balance accuracy, cost, and latency through specialized and hybrid parsing strategies.

🔬 "The Most Innovative Diffusion Research Is Happening in Drug Discovery, Not Image Generation"
Latent Space

🔬 "The Most Innovative Diffusion Research Is Happening in Drug Discovery, Not Image Generation"

Genesis Molecular AI argues that useful AI drug discovery depends on sub-angstrom structural accuracy, physics-guided diffusion, synthetic data, and rapid laboratory feedback—not structure prediction alone. Its long-term goal is to turn validated molecular models into agentic systems that let scientists direct many parallel drug-design campaigns.

I Gave an AI a Body — Cyrus Clarke, MIT Media Lab
AI Engineer

I Gave an AI a Body — Cyrus Clarke, MIT Media Lab

Cyrus Clarke’s experiments suggest that embodied AI can become more expressive and conversational without adopting human or animal form. His shape-display agent uses a learned gesture vocabulary to reduce latency and create a sensory, emotionally legible relationship with people.

The arrival of AGI | Shane Legg (co-founder of DeepMind)
Google DeepMind

The arrival of AGI | Shane Legg (co-founder of DeepMind)

Shane Legg expects minimal AGI within a few years and full AGI within a decade, followed eventually by superintelligence. The central challenge is not only making increasingly capable systems safe, but redesigning economic and social institutions so their benefits are broadly shared.

🔬 Google's AI Scientist Started as an Attempt to Automate Kaggle — John Platt, Google Fellow
Latent Space

🔬 Google's AI Scientist Started as an Attempt to Automate Kaggle — John Platt, Google Fellow

ERA turns scientific questions into scorable coding tasks and uses Gemini-guided search to improve candidate solutions. Its power increases rather than removes the need for human creativity, domain knowledge, hidden holdouts, and rigorous validation.

Why Education Has to Change
a16z

Why Education Has to Change

Horowitz and Gagan Academy is designed as an AI-era alternative to conventional higher education for young builders, combining project-based learning, peer community, and direct exposure to Silicon Valley companies. Its core wager is that students will learn faster by building consequential things than by preparing for tests within a system designed for the industrial age.

Harness Engineering: Building the Production Cage for Powerful Domain Agents — Mike Chambers, AWS
AI Engineer

Harness Engineering: Building the Production Cage for Powerful Domain Agents — Mike Chambers, AWS

Production-grade agents require more than a model, prompt, and tools: their harness must independently manage runtime, memory, identity, scaling, observability, evaluation, and deployment. AWS Bedrock AgentCore is presented as a composable way to supply that infrastructure without forcing developers to build every component themselves.

Why Rust is different, with Alice Ryhl
The Pragmatic Engineer

Why Rust is different, with Alice Ryhl

Rust combines low-level performance with compile-time protections against memory errors, null misuse, ignored failures, and incomplete refactors. Its ownership model creates a learning curve, but its tooling, compatibility strategy, and consensus-driven governance make reliability a system-wide property rather than a single feature.

David Reich – Bronze Age shock, the Neanderthal puzzle, & the sudden spread of farming
Dwarkesh Patel

David Reich – Bronze Age shock, the Neanderthal puzzle, & the sudden spread of farming

A new ancient-DNA method suggests that directional natural selection was widespread in Europe and the Middle East over the last 18,000 years, with especially strong immune and metabolic adaptation during the Bronze Age. Most observed allele-frequency change still came from migration and genetic drift, and many provocative trait-level findings require cautious interpretation.

Why I couldn't build Jev at OpenAI — Diogo Almeida, TypeSafe Co-founder & CEO
Latent Space

Why I couldn't build Jev at OpenAI — Diogo Almeida, TypeSafe Co-founder & CEO

Typesafe’s thesis is that AI will create more economic value when optimized as reliable, programmable infrastructure for software rather than as a human-facing chatbot. Its first model, Jev, prioritizes intelligence per dollar, calibrated decisions, structured inputs, and composability, while treating reliability—not benchmark performance—as the central constraint on automation.

The Dark Arts of Skill Engineering — Paul Bakaus, Renaissance Geek (Impeccable)
AI Engineer

The Dark Arts of Skill Engineering — Paul Bakaus, Renaissance Geek (Impeccable)

Reliable skills are harness extensions, not packaged prompts. They combine adversarial agents, deterministic checks, memory, scripts, hooks, interactive tools, cross-harness compilation, and model-specific enforcement.

Waymo Co-CEO Dmitri Dolgov: The Demo Is Only 1% Of The Work
Y Combinator

Waymo Co-CEO Dmitri Dolgov: The Demo Is Only 1% Of The Work

Building physical AI requires far more than a compelling model or demo: it demands extreme reliability, multimodal sensing, closed-loop simulation, system-wide evaluation, and evidence earned through real-world operation. The central lesson is to move fast while shipping safely.

The Story Behind Cerebras’ $63 Billion IPO with Founder and CEO Andrew Feldman
No Priors

The Story Behind Cerebras’ $63 Billion IPO with Founder and CEO Andrew Feldman

Cerebras CEO Andrew Feldman argues that radically faster AI inference, enabled by a contrarian wafer-scale architecture, is moving from a technical curiosity to critical infrastructure. He expects speed to produce entirely new business models and productivity gains, not merely improve today’s AI tools.

How a Writer Uses AI Without Losing His Voice
Every

How a Writer Uses AI Without Losing His Voice

AI makes bespoke software extraordinarily cheap to build, but durable value now depends on maintenance, purpose, and distinctly human judgment. Craig protects his writing by separating networked experimentation from offline mornings devoted to deep attention.

“Every small business should run itself” | Lassie with a16z
a16z

“Every small business should run itself” | Lassie with a16z

Lassie is building autonomous agents that perform administrative labor for dental practices rather than merely giving staff another software tool. Its opportunity rests on severe labor shortages, hard-won workflow data, near-frictionless onboarding, and expansion from one deeply understood vertical into other small businesses.

The most rational take on AI you’ll hear this year
Lenny's Podcast

The most rational take on AI you’ll hear this year

AI is a platform shift on the scale of the internet or mobile, but its eventual products, winners, labor effects, and value distribution remain radically uncertain. The practical response is to learn the technology deeply while distinguishing automatable tasks from the broader jobs, institutions, and human judgments surrounding them.

Why the People Who Built Hip-Hop Ended Up With Nothing
a16z

Why the People Who Built Hip-Hop Ended Up With Nothing

The Paid in Full Foundation combines financial support, public recognition, and sustained personal care for hip-hop pioneers whose cultural influence was never matched by compensation. Its broader mission is to turn hip-hop’s professed values of solidarity and respect into visible action.

90 minutes of unfiltered product advice from Snap and Discord’s product chief | Peter Sellis
Lenny's Podcast

90 minutes of unfiltered product advice from Snap and Discord’s product chief | Peter Sellis

Peter Sellis argues that exceptional product organizations replace excessive coordination with ideological clarity, explicit decision ownership, and unusually high trust. Their advantage comes from strengthening the core product, exercising disciplined taste, modeling systems mathematically, and protecting long-term value from short-term pressure.

The Future of AI Infra: from Kubernetes to Agent Sandboxes — Akshat Bubna, Modal CTO
Latent Space

The Future of AI Infra: from Kubernetes to Agent Sandboxes — Akshat Bubna, Modal CTO

Modal is evolving from a developer-friendly serverless runtime into an agent-oriented cloud for elastic AI inference, training, and sandbox execution. Its advantage is a software layer that hides fragmented compute capacity while preserving code-level control, observability, and hard security boundaries.

Large clusters for small models — Daniel Svonava, Superlinked
AI Engineer

Large clusters for small models — Daniel Svonava, Superlinked

Task-specific small open-source models can approach frontier quality while sharply reducing cost and latency, but serving a diverse model fleet requires a different infrastructure design. The proposed solution replaces top-down routing with worker-pulled centralized queues, runtime abstraction, model packing, and pre-tuned configurations.

The future of intelligence | Demis Hassabis (Co-founder and CEO of DeepMind)
Google DeepMind

The future of intelligence | Demis Hassabis (Co-founder and CEO of DeepMind)

Demis Hassabis argues that AGI will require both continued scaling and major research breakthroughs, especially in reasoning, continual learning, uncertainty estimation, and world models. He expects increasingly autonomous agents to create enormous scientific and economic benefits while sharply increasing safety, governance, and social-transition risks.

Two Bugs That Hid in Plain Sight: A vLLM Debugging Detective Story — Asaf Gardin & Yuval Belfer
AI Engineer

Two Bugs That Hid in Plain Sight: A vLLM Debugging Detective Story — Asaf Gardin & Yuval Belfer

Two silent failures in Jamba’s stateful inference path produced confident but corrupted outputs: one scheduled decode before prefill, while another overflowed a 32-bit cache index. Both were isolated by deterministic reproduction, token-level log-probability comparison, and tracing request identity into low-level kernels.

Building more effective AI agents
Anthropic

Building more effective AI agents

Effective agents emerge from long-horizon training, iterative tool use, and the ability to verify their own work. Multi-agent systems can improve speed, context efficiency, and answer quality, but only when delegation is clear and coordination overhead remains controlled.

Weight Folding, CUDA Streams, and the Bug That Made My Model Speak Backwards — Filip Makraduli
AI Engineer

Weight Folding, CUDA Streams, and the Bug That Made My Model Speak Backwards — Filip Makraduli

FlashNorm restructures RMSNorm with weight folding, deferred normalization, and—in compatible architectures—normalization cancellation, reducing GPU launch, synchronization, and memory overhead. Its implementation also demonstrates why asynchronous CUDA streams require explicit synchronization, even when unit tests and perplexity checks appear correct.

The spelled-out intro to neural networks and backpropagation: building micrograd
Andrej Karpathy

The spelled-out intro to neural networks and backpropagation: building micrograd

Micrograd demonstrates that neural-network training reduces to building a computation graph, applying the chain rule backward through it, and repeatedly updating parameters against the gradient. Production frameworks such as PyTorch use the same mathematics, packaged into efficient tensor operations.

The Frontier AI Inference Cloud for Agents — Byung-Gon (Gon) Chun, FriendliAI
AI Engineer

The Frontier AI Inference Cloud for Agents — Byung-Gon (Gon) Chun, FriendliAI

Friendly AI argues that production agents require inference systems optimized for end-to-end task completion rather than individual model calls. Its proposed stack combines open-weight models with prefix caching, hierarchical KV-cache management, cache-aware routing, and agent-aware scheduling to improve speed, reliability, and cost.

Mousepower: agents that can’t be measured, can’t be managed. — Maximillian Piras, Yutori
AI Engineer

Mousepower: agents that can’t be measured, can’t be managed. — Maximillian Piras, Yutori

Agents have a measurement problem: token consumption only becomes meaningful when tied to verifiable outcomes and business objectives. The best agent tasks involve meaningful execution uncertainty but relatively clear, inexpensive acceptance criteria.

Building OpenCode with Dax Raad
The Pragmatic Engineer

Building OpenCode with Dax Raad

AI coding agents make implementation easier, but they do not resolve the harder problems of product judgment, motivation, architecture, or quality. Without restraint and strong feedback loops, faster code generation can produce more features, more technical debt, and only the illusion of greater productivity.

What rebuilding AlphaGo teaches us about self-play, RL, and future of LLMs - Eric Jang
Dwarkesh Patel

What rebuilding AlphaGo teaches us about self-play, RL, and future of LLMs - Eric Jang

AlphaGo makes an otherwise intractable search practical by using policy and value networks to reduce the breadth and depth of Monte Carlo tree search, then distilling search results back into the network. Eric Jang argues that this stable loop of search, soft-label supervision, and self-play offers lessons for efficient AI research, while also exposing why open-ended LLM reasoning is harder to search and automate.

Why Would AI Companies Want to Slow Down?
a16z

Why Would AI Companies Want to Slow Down?

Ali Ghodsi argues that present-day AI poses serious, engineerable cybersecurity risks but offers little evidence of imminent superintelligence. Enterprises can capture far more value from existing models by supplying organizational context, automating security, and routing each task to an appropriately priced model.

A Worm With 302 Neurons Inspired Their Architecture — Ramin Hasani, Liquid AI
Latent Space

A Worm With 302 Neurons Inspired Their Architecture — Ramin Hasani, Liquid AI

Liquid AI builds hardware-aware hybrid foundation models by searching across attention, convolution, recurrence, and state-space operators rather than betting on one architecture. Its strategy is to deliver efficient, production-grade intelligence on devices while expanding toward automated model customization, continual learning, multimodality, and robotics.

The Case For Data Centers In Space
Y Combinator

The Case For Data Centers In Space

StarCloud is betting that falling launch costs and tightening terrestrial energy constraints will make orbital data centers economically viable. Its strategy is to validate the hardest technical risks through progressively larger spacecraft while building an unusually strong engineering team.

Total Recall: Agent Memory and Harness Engineering — Ignacio Martinez, Oracle
AI Engineer

Total Recall: Agent Memory and Harness Engineering — Ignacio Martinez, Oracle

An AI agent is an interchangeable language model wrapped in a seven-layer harness that supplies memory, semantics, retrieval, tools, context management, and a resilient execution loop. The workshop builds this harness incrementally, using hybrid file/database storage and reusable skills to make nondeterministic models more reliable and continuously adaptable.

Building an AI Guardian for Enterprise with Onyx Security CEO Maxim Bar Kogan
No Priors

Building an AI Guardian for Enterprise with Onyx Security CEO Maxim Bar Kogan

Autonomous agents are spreading faster than enterprises can safely constrain them, making intent-aware oversight a foundational security requirement. Onyx proposes independent, specialized models that escalate only risky actions to more capable reviewers, balancing control with cost, latency, and productivity.

Why Diffusion Will Win AI Inference with Inception Co-Founder and CEO Stefano Ermon
No Priors

Why Diffusion Will Win AI Inference with Inception Co-Founder and CEO Stefano Ermon

Inception is betting that diffusion language models can challenge autoregressive LLMs by generating tokens in parallel, improving latency and hardware utilization without sacrificing model quality. Its immediate advantage is speed, while controllability, data efficiency, and better inference-time scaling remain promising but unproven at frontier scale.

Kubernetes and retiring at the top with Kelsey Hightower
The Pragmatic Engineer

Kubernetes and retiring at the top with Kelsey Hightower

Kelsey Hightower’s career shows that durable success comes from compounding fundamentals, public contribution, measurable impact, and human judgment—not from chasing titles or tools. His view of AI follows the same principle: use it strategically to solve human problems, but do not confuse code generation with engineering.

Realtime Voice Agents with Frontier Intelligence — Bohan Li, EliseAI
AI Engineer

Realtime Voice Agents with Frontier Intelligence — Bohan Li, EliseAI

The architecture combines speculative transcription, background tool execution, eager LLM generation, and cached speech prefixes to deliver fast, intelligent voice interactions. Its central strategy is to overlap work and cancel obsolete computation rather than weaken the underlying models.

Chip design from the bottom up – Reiner Pope
Dwarkesh Patel

Chip design from the bottom up – Reiner Pope

AI-chip design is largely an effort to maximize useful arithmetic relative to the area, bandwidth, and energy spent moving and synchronizing data. Low-precision arithmetic, systolic arrays, explicit memory management, and carefully sized pipelines all exploit this principle with different flexibility trade-offs.

"My name is... my name is...": A Linguistic Map for Voice Agents — Midam Kim, ServiceNow
AI Engineer

"My name is... my name is...": A Linguistic Map for Voice Agents — Midam Kim, ServiceNow

Voice AI should be designed as a real-time joint activity, not a sequence of speech-processing components. Success depends on coordinating recognition, language, turn-taking, context, and adaptation around the user's evolving mental model.

Garry Tan: Own Your Intelligence
Y Combinator

Garry Tan: Own Your Intelligence

The speaker argues that personal AGI is not a future universal event but an owned system combining rented models, personal context, and reusable skills. Building such a system can radically expand individual agency, but its memory, procedures, and infrastructure must remain under the individual's control.

How the Internet Turned Biohacking Into a Cultural Obsession
a16z

How the Internet Turned Biohacking Into a Cultural Obsession

Online wellness culture is driven as much by commentary, aspiration, and marketing as by actual adoption. Peptides, wearables, and biohacking reflect a broader shift toward treating the body as the next platform for technological optimization, even as people increasingly resist excessive quantification.

Act, Confirm, or Stop? Smarter behavior for AI assistants, wearables & robots — Amit Desai, Roku
AI Engineer

Act, Confirm, or Stop? Smarter behavior for AI assistants, wearables & robots — Amit Desai, Roku

Voice AI can improve user satisfaction without becoming more accurate by choosing safer behaviors under uncertainty. The proposed method minimizes weighted user recovery costs—called “OUCH”—across actions such as executing, rejecting, or confirming a request.

We Need An Ecosystem in AI, And Every Company Can Win A Place In It
No Priors

We Need An Ecosystem in AI, And Every Company Can Win A Place In It

Satya Nadella argues that durable AI advantage will come from combining multiple models with proprietary context, tools, traces, and private evaluations—not from dependence on one frontier model. AI will transform work and software economics, but the industry will retain public permission only by delivering tangible, broadly shared benefits.

We let an AI agent execute Bash and lived to talk about it — Sarah Sanders, PostHog
AI Engineer

We let an AI agent execute Bash and lived to talk about it — Sarah Sanders, PostHog

Command-running agents resemble a malware toolkit unless every capability is tightly constrained. PostHog’s approach keeps detection and enforcement deterministic, scans its own context supply chain twice, and uses LLM judgment only to reduce noise after blocking decisions are made.

Granola's Chris Pedregal on Building a $1.5B AI Company
Every

Granola's Chris Pedregal on Building a $1.5B AI Company

Granola’s ambition extends beyond meeting notes: it wants to become the best system for capturing and interpreting meeting context while making that context available to users’ preferred agents. The broader opportunity is a new computing model in which humans and agents share interfaces, delegate work asynchronously, and rely on specialized products to turn raw context into dependable action.

No Memory, No Harness: Why the Database Is the Last Line of Defense — Kay Malcolm, Oracle
AI Engineer

No Memory, No Harness: Why the Database Is the Last Line of Defense — Kay Malcolm, Oracle

AI-assisted coding can accelerate individuals without improving team throughput because Git preserves code, not the agent context and human intent behind it. The speaker argues that shared, selectively retrieved memory is the essential collaboration layer in an enterprise agent harness and promotes Oracle AI Database as a unified store for that memory.

The Nuclear Renaissance - Radiant |  a16z American Dynamism
a16z

The Nuclear Renaissance - Radiant |  a16z American Dynamism

Radiant argues that factory-built, transportable microreactors can restore U.S. nuclear innovation by delivering reliable power with far less customer and construction friction. Its near-term proof point is a planned full-power prototype test at Idaho National Laboratory in 2026.

Load more