scoop.fm /ai Beta
About Podcasters

The AI Practice Weekly

Our team listens to thousands of podcasts about ai so you don't have to. One weekly briefing, in your inbox.

SUNDAY, MAY 31, 2026 · MINED FROM 131 PODCASTS

May 31: Computer use models hit 95% on legacy UIs

Good morning.Teams are abandoning monolithic agent approaches for highly concurrent and specialized architectures.

May 31: Computer use models hit 95% on legacy UIs

Pace saw computer use models jump from thirty percent to over ninety-five percent accuracy on internal evaluations. This leap allows developers to bypass expensive API builds entirely, deploying agents that visually navigate legacy web applications and green screen command line interfaces without custom integration layers.

Ship This Week

Placing ToolHive between the LLM and MCP servers to dynamically filter 150+ tool schemas via a find_tool endpoint cut input tokens by 80%.

agents

Proxy-based MCP tool filtering

Use a proxy layer to filter available MCP tools dynamically, preventing context bloat and improving tool-calling accuracy for smaller models.

▾ Show more ▴ Show less

Problem: Exposing multiple MCP servers directly to an LLM pollutes the context window with 150+ tool descriptions, burning 20k-30k tokens per interaction and causing smaller models to fail at tool invocation.

On: proxy layers for MCP tool pollution

“It reduces input token consumption by 80% to 90% when you have these things... you can get back up to the sort of 95, 97% threshold that actually makes the system useful.”

Recipe Place a proxy layer (like ToolHive) between the LLM and MCP servers. Instead of injecting all tool schemas into the system prompt, expose a `find_tool` endpoint. The LLM queries the proxy with its intent, and the proxy returns only the specific tool schemas matching that description.

Measured Evidence input tokens -80% to 90%

Counterpoint Input token caching helps with context bloat, but doesn't solve the tool-selection accuracy drop in smaller models faced with 150+ options.

Rebooting Enterprise AI with MCP and Kubernetes · May 28, 2026

▶ Apple show ▶ Spotify
Opus 4.7SonnetHaikuToolHiveMCP
▶ Listen

Watch the Frontier

Future House abandons linear context passing, instead using a centralized world model across hundreds of parallel sub-agent runs to synthesize 20,000 novel scientific findings.

agents

Cross-run world models for context management

Use a centralized world model to maintain state and integrate knowledge across hundreds of independent sub-agent runs.

▾ Show more ▴ Show less

Problem: Multi-agent systems struggle to synthesize a cohesive understanding of a domain when context is isolated within individual, short-lived sub-agent runs.

On: world models for multi-agent context

“we also built in this notion of world models... basically a very sophisticated context management tool... allows Cosmos to build up a kind of integrated notion of the knowledge in a field over the course of like dozens, hundreds of sub-agent runs”

Recipe Implement a "world model" as a sophisticated context management tool alongside the orchestrator. As the orchestrator spins up dozens or hundreds of sub-agents to investigate specific targets, it continuously updates this central world model to build an integrated knowledge graph before generating final hypotheses.

Measured Evidence 20,000 to 30,000 novel scientific findings

Counterpoint Standard multi-agent frameworks pass context linearly between agents; this architecture uses a centralized, continuously updated state object across hundreds of parallel runs.

He Raised $70M to Cure Every Disease With AI · May 27, 2026

▶ Apple show ▶ Spotify
▶ Listen

Today's AI Patterns

Capability shifts and emerging build patterns from this week's shows.

agents

Domain-driven design for agent guardrails

Reintroduce verbose, strict architectural patterns to constrain coding agents, using the AI to write the boilerplate it is constrained by.

▾ Show more ▴ Show less

Problem: Coding agents silently introduce hacky code, bypassing the natural hesitation and friction engineers feel when writing hacks manually.

On: verbose design patterns as agent guardrails

“we find that these kind of boring, enterprise-y patterns end up being pretty useful because you have a bunch of idiots on your team now. The coding agents are a bunch of idiots... so you need way more guardrails”

Recipe Reintroduce heavy, verbose architectural patterns like Domain-Driven Design (DDD). Agents generate the verbose boilerplate instantly, while the strict structure provides deterministic guardrails against agent hallucinations and sloppy hacks.

Speaker's word Evidence Speaker's word, not measured.

Counterpoint Verbose design patterns were abandoned by senior devs for being too slow to type; now they are the ideal static guardrails for fast-typing AI agents.

Building OpenCode with Dax Raad · May 27, 2026

▶ Apple show ▶ Spotify
▶ Listen

agents

Lazy-loading agent skills to prevent context rot

Only load skill names and descriptions initially, fetching full instructions dynamically when triggered.

▾ Show more ▴ Show less

Problem: Injecting all available agent skills and tool instructions into the system prompt causes context pollution and degrades the model's reasoning over long sessions.

On: dynamic skill loading for context management

“It only loads names and descriptions initially. So it doesn't put everything in context... it only loads it into context when needed. So you don't have to worry about context pollution and context rot.”

Recipe In the initial system prompt, provide only the names and short descriptions of available skills. When the user requests a specific task, have the agent recognize the required skill and dynamically load the full instructions, constraints, and associated bash scripts into the context window.

Speaker's word Evidence Speaker's word, not measured.

Counterpoint Standard agent frameworks inject all tool schemas upfront; lazy-loading preserves the context window for actual reasoning.

The Future of Agentic Data Science · May 25, 2026

▶ Apple show ▶ Spotify
GeminiReplicate
▶ Listen

agents

Computer use for legacy integration

Bypass expensive API builds by using computer use models to visually navigate legacy web apps and green screen CLIs.

Jamie · Pace · TBPN

▾ Show more ▴ Show less

Problem: Automating back-office operations is blocked because integrating with legacy desktop applications and green screen CLIs via traditional APIs is either too expensive or technically prohibitive.

On: computer use for legacy apps

“computer use models, they've really gone from something like 30% accuracy on our evals to like 95% plus. And that really allows us to get these models into production and particularly work around cases like desktop applications or legacy web apps or even like we've seen the green screen CLIs”

Recipe Deploy computer use models to interact directly with the UI of legacy web apps, desktop applications, and green screen CLIs. Guide the models using 'agent operating procedures'—long-running, natural language instructions that define the end-to-end task execution.

Measured Evidence evals 30% → 95%+

Counterpoint Teams typically block agent deployment waiting for API access; computer use models now reliably bypass the API layer entirely.

Blue Origin Explosion, Enterprise AI Hits Spending Limits, Dinosaur Market Booms | John Gruber, Ronak Malde, Zane Mountcastle, Jamie Cuffe, Kyle Kuzma, Brad Gerstner · May 29, 2026

▶ Apple show ▶ Spotify
▶ Listen

agents

Cross-run world models for context management

Use a centralized world model to maintain state and integrate knowledge across hundreds of independent sub-agent runs.

▾ Show more ▴ Show less

Problem: Multi-agent systems struggle to synthesize a cohesive understanding of a domain when context is isolated within individual, short-lived sub-agent runs.

On: world models for multi-agent context

“we also built in this notion of world models... basically a very sophisticated context management tool... allows Cosmos to build up a kind of integrated notion of the knowledge in a field over the course of like dozens, hundreds of sub-agent runs”

Recipe Implement a "world model" as a sophisticated context management tool alongside the orchestrator. As the orchestrator spins up dozens or hundreds of sub-agents to investigate specific targets, it continuously updates this central world model to build an integrated knowledge graph before generating final hypotheses.

Measured Evidence 20,000 to 30,000 novel scientific findings

Counterpoint Standard multi-agent frameworks pass context linearly between agents; this architecture uses a centralized, continuously updated state object across hundreds of parallel runs.

He Raised $70M to Cure Every Disease With AI · May 27, 2026

▶ Apple show ▶ Spotify
▶ Listen

agents

Proxy-based MCP tool filtering

Use a proxy layer to filter available MCP tools dynamically, preventing context bloat and improving tool-calling accuracy for smaller models.

▾ Show more ▴ Show less

Problem: Exposing multiple MCP servers directly to an LLM pollutes the context window with 150+ tool descriptions, burning 20k-30k tokens per interaction and causing smaller models to fail at tool invocation.

On: proxy layers for MCP tool pollution

“It reduces input token consumption by 80% to 90% when you have these things... you can get back up to the sort of 95, 97% threshold that actually makes the system useful.”

Recipe Place a proxy layer (like ToolHive) between the LLM and MCP servers. Instead of injecting all tool schemas into the system prompt, expose a `find_tool` endpoint. The LLM queries the proxy with its intent, and the proxy returns only the specific tool schemas matching that description.

Measured Evidence input tokens -80% to 90%

Counterpoint Input token caching helps with context bloat, but doesn't solve the tool-selection accuracy drop in smaller models faced with 150+ options.

Rebooting Enterprise AI with MCP and Kubernetes · May 28, 2026

▶ Apple show ▶ Spotify
Opus 4.7SonnetHaikuToolHiveMCP
▶ Listen

infra

MCP for model hot-swapping

Use the Model Context Protocol (MCP) to standardize interfaces between models and external software.

▾ Show more ▴ Show less

Problem: Tightly coupling application logic to specific model APIs creates vendor lock-in and prevents teams from easily swapping to cheaper or better open-source models.

On: abstracting model interfaces with MCP

“MCP is actually run by the Linux Foundation. And if you think about any surface area where a model might interact with other software, the more of those connectors that can be open source and commoditized...”

Recipe Implement open-source connectors like the Model Context Protocol (MCP) to standardize the surface area where models interact with other software. This abstracts the model layer, allowing for plug-and-play model swapping without rewriting application logic.

Speaker's word Evidence Speaker's word, not measured.

Counterpoint Instead of building custom integrations for every model API, standardizing on MCP commoditizes the model layer similar to how Kubernetes abstracted cloud providers.

Pope vs AI, Anthropic's Digital God, AI Job Loss Narrative Flips, Open Source Crackdown Coming? · May 29, 2026

▶ Apple show ▶ Spotify
MCP
▶ Listen

evals

Domain-specific frontier model evals

Frontier models are converging on domain-specific tasks, meaning routing logic should optimize for cost and latency.

▾ Show more ▴ Show less

Problem: Trillions of dollars are spent training new frontier models, but it's unclear if the incremental capability improvements justify the cost for specific enterprise workloads.

On: evaluating frontier model convergence

“At the top of the leaderboard, Opus 4.7, GPT-5.5, Sonnet 4.6 appear almost indistinguishable. separated by less than in this case, you know, three tenths of a percentage point overall”

Recipe Run domain-specific test benches (like Rogo's financial analyst evals) against the latest frontier models. If the variance between models is negligible, route requests based on cost and latency rather than raw capability.

Measured Evidence < 0.3% variance

Counterpoint Despite massive training investments, frontier models are asymptoting on specific evals, making them highly commoditized for certain workloads.

Pope vs AI, Anthropic's Digital God, AI Job Loss Narrative Flips, Open Source Crackdown Coming? · May 29, 2026

▶ Apple show ▶ Spotify
Opus 4.7GPT-5.5Sonnet 4.6Rogo
▶ Listen

agents

Parallel forking paths for causal analysis

Run competing analytical assumptions in parallel agent branches to avoid brittle, single-path data science.

▾ Show more ▴ Show less

Problem: Human data scientists take shortcuts and explore only one analytical path, leading to fragile conclusions based on arbitrary early choices like outlier removal.

On: tree-based orchestration for causal analysis

“It's not just doing one path... but actually like a garden forking path, it's just going through all these different choices in parallel and exploring them in a tree-based structure, essentially.”

Recipe Use FireO to orchestrate a tree-based agent workflow. Spawn parallel agent paths to explore different data cleaning and modeling choices simultaneously. Evaluate each branch programmatically, prune invalid paths, and consolidate the surviving, diverging results into a single Jupyter notebook for the human principal investigator to review.

Speaker's word Evidence Speaker's word, not measured.

Counterpoint Default agentic data analysis follows a single sequential path; this explicitly forks and evaluates competing analytical assumptions in parallel.

The Future of Agentic Data Science · May 25, 2026

▶ Apple show ▶ Spotify
FireOJupyter
▶ Listen

agents

RL-optimized context self-summarization

Train long-horizon agents to summarize their progress and restart their context window to bypass maximum token limits.

▾ Show more ▴ Show less

Problem: Long-horizon agents eventually hit hard context window limits and suffer from credit assignment degradation over extended trajectories.

On: handling context limits in RL

“We put compaction inside the RL loop... the agent actually learns how to continue and go on forever... we are training the model to produce a good summary. And then we are training the model to listen to that summary.”

Recipe Embed context compaction directly into the RL loop. Train the agent to generate a summary of its work, restart its context window using only that summary, and jointly optimize its ability to follow its own summaries to continue the task.

Measured Evidence Context extended to millions of tokens

Counterpoint Most teams handle context limits with external RAG or truncation; this team bakes summarization and context-restarting directly into the RL objective.

How Cursor Trained Composer on Fireworks: Distributed Infrastructure for High-Performance RL · May 26, 2026

▶ Apple show ▶ Spotify
▶ Listen

infra

High-value agent infrastructure bootstrapping

Fund centralized agent governance by automating one business-critical workflow first, rather than starting with hundreds of low-value prototypes.

▾ Show more ▴ Show less

Problem: Starting small with agent prototypes leads to hundreds of fragmented, low-value agents that cannot justify the engineering cost of a centralized observability and governance layer.

On: funding agent governance infrastructure

“when something is business critical you will have enough bandwidth to think about building a framework or a governance layer... on top of which you are able to now actually then build out the other 500 agents”

Recipe Ignore requests for small agent automations. Identify 5 business-critical workflows and fully automate 1 or 2. Use the outsized business value of these primary agents to justify building a centralized observability and governance layer, then deploy the smaller long-tail agents on top of that established plumbing.

Measured Evidence $100M+ return per deployment

Counterpoint Standard agile advice is to start small; for agents, starting small prevents teams from ever building the necessary centralized governance infrastructure.

996: TrueFoundry’s Nikunj Bajaj on How to Get $100M Returns on AI Agent Deployments · May 29, 2026

▶ Apple show ▶ Spotify
▶ Listen

infra

KV cache scaling for RL burstiness

Scale down the KV cache allocation to leave dedicated GPU overhead for sudden spikes in reinforcement learning reward calculations.

▾ Show more ▴ Show less

Problem: Reinforcement learning calculations run heavily on CPU, then suddenly send massive reward trajectories to the GPU, causing segfaults on single A100s.

On: reserving GPU overhead for RL spikes

“A lot of times people might go use 90 percent of the overhead on my model and on a KB cache. I have to scale that down so that I can have space for those reinforcement learning results to come up so the model can get better.”

Recipe Do not allocate 90% of the GPU memory to the model weights and KV cache. Scale down the KV cache allocation to leave dedicated overhead space specifically for the sudden spikes in RL reward calculations, preventing GPU overload and segfaults.

Speaker's word Evidence Speaker's word, not measured.

Counterpoint Standard inference optimization maximizes KV cache to use all available VRAM; RL requires leaving significant VRAM empty to handle bursty reward calculations.

995: End-to-End Foundation Models for the Energy Industry, with Jazmia Henry · May 26, 2026

▶ Apple show ▶ Spotify
▶ Listen

infra

Local edge inference for vehicle controls

Rivian uses 200 TOPS of local edge compute to run small language models for vehicle controls, eliminating cloud latency and inference costs.

▾ Show more ▴ Show less

Problem: Cloud-dependent voice assistants suffer from high latency, connectivity dead zones, and recurring cloud inference costs for simple, high-frequency commands.

On: local edge compute for vehicle AI

“the local computer on the R2 will have up to 200 sparse tops of compute dedicated to IAI... All of that will be available locally in the R2 car... to basically not face such limitations and issues with connectivity”

Recipe Equip the local vehicle computer with 200 sparse TOPS of AI compute. Route high-frequency, non-complex vehicle commands (e.g., 'I am cold') to a small language model running directly on the edge. Fall back to cloud models (like Gemini) only for complex or web-grounded queries.

Measured Evidence 200 sparse TOPS local compute

Counterpoint Adding expensive AI compute to the vehicle BOM seems counterintuitive, but it economically solves itself by eliminating recurring cloud inference and connectivity costs.

Rivian's software chief thinks you don't need CarPlay or buttons · May 28, 2026

▶ Apple show ▶ Spotify
GeminiOpus 4.7
▶ Listen

product

Zero-auth live data artifacts

Use Claude's native OAuth connectors and the "live artifact" prompt incantation to build auto-refreshing dashboards without managing API keys.

Felix Rieseberg · Engineering Lead · Anthropic · How I AI

▾ Show more ▴ Show less

Problem: AI-generated reports and dashboards immediately go stale and require manual data refreshes or complex API key management to keep updated.

On: auto-refreshing personal dashboards

“We thought about the idea of, can we make an artifact that refreshes itself with the latest data... you can pull in live and have cloud crunch on the go.”

Recipe Connect data sources via Claude's native connectors (Spotify, Gmail, Notion). Append the incantation "make this live artifact" to the prompt. Claude generates a React-based dashboard with a refresh button that pulls live data directly via the user's authenticated session, bypassing custom OAuth flows.

Speaker's word Evidence Speaker's word, not measured.

Counterpoint Building live dashboards usually requires a custom app with OAuth flows; this uses Claude's native connectors and artifact UI.

How the engineer behind Claude Cowork actually uses Claude | Felix Rieseberg (Anthropic) · May 25, 2026

▶ Apple show ▶ Spotify
ClaudeSpotifyGmailNotion
▶ Listen

product

Intent-ambiguity model routing

Route to Opus for ambiguous, poorly-scoped requests and Sonnet for well-defined tasks.

Felix Rieseberg · Engineering Lead · Anthropic · How I AI

▾ Show more ▴ Show less

Problem: Defaulting to the largest model wastes tokens, but smaller models fail when the user hasn't fully scoped the problem.

On: routing based on problem scoping

“I reach for opus over sonnet if I have self-identified as someone who doesn't really know yet what they're asking for.”

Recipe Route to Sonnet 4.6 for well-scoped, specific tasks (e.g., extracting units from a floor plan). Route to Opus when the user hasn't fully scoped the problem and needs the model to reinterpret the underlying intent before executing, similar to a creative director.

Speaker's word Evidence Speaker's word, not measured.

Counterpoint Most teams route based on token length or task category; this routes based on intent ambiguity.

How the engineer behind Claude Cowork actually uses Claude | Felix Rieseberg (Anthropic) · May 25, 2026

▶ Apple show ▶ Spotify
OpusSonnet 4.6
▶ Listen

product

Task-based routing between Opus 4.8 and 4.7

Route greenfield prototyping to Opus 4.8, but fall back to Opus 4.7 for existing codebase integration and data-heavy strategy.

Host / How I AI · Product Leader · How I AI

▾ Show more ▴ Show less

Problem: Opus 4.8 excels at one-shot generation but hallucinates unverified bugs and loses context boundaries when inserted into existing codebases or asked to rebase active branches.

On: routing strategy tasks to older models

“I honestly would continue to reach for Opus 4.7, which I think did an exceptional job on strategy, versus for 8, which I think was a lot more hand-wavy and just over-rotated on things I didn't think was important.”

Recipe Implement a task-based model router: send greenfield feature prototyping and one-shot generation to Opus 4.8 for fast, autonomous execution. For deep codebase integration, bug hunting, or strategy tasks requiring numerical grounding, route back to Opus 4.7.

Speaker's word Evidence Speaker's word, not measured.

Counterpoint Newer models are generally assumed to be strictly better; here, the older 4.7 model outperforms 4.8 on numerical grounding and existing-codebase context.

Claude Opus 4.8 is here. Is it as good as they say? · May 28, 2026

▶ Apple show ▶ Spotify
Opus 4.8Opus 4.7GPT 5.5Gemini 3.1Claude CodeClaude CoworkCloud AIGitHub
▶ Listen

agents

High-concurrency agentic workflows

Developers orchestrate 5 to 15 specialized agents concurrently to burn down community issues and scale throughput.

▾ Show more ▴ Show less

Problem: Sequential agent execution or single-agent monolithic setups bottleneck developer throughput when processing high volumes of community issues or codebase tasks.

On: agentic concurrency for developer throughput

“having somewhere between five and 15 agents concurrently running... I track our weekly productivity this last week our person our engineering team's throughput went up 60 percent in a week”

Recipe Deploy agentic concurrency. A single developer provisions 5 to 15 concurrent agents. Give each agent a slightly different configuration, role, and highly constrained tool access. The human operator acts as an orchestrator (spinning plates) while the agents execute tasks in parallel.

Measured Evidence throughput +60% in a week

Counterpoint Most teams try to build one highly capable autonomous agent; this team scales by having a human orchestrate 15 narrow, concurrent agents.

Rebooting Enterprise AI with MCP and Kubernetes · May 28, 2026

▶ Apple show ▶ Spotify
▶ Listen

infra

Burst-scaled custom VM RL environments

Build a custom virtual machine stack to run RL rollouts instead of using generic Docker containers.

▾ Show more ▴ Show less

Problem: Standard Docker containers fail to accurately replicate the complex state, background services, and OS-level interactions of a real user's local development environment during RL rollouts.

On: simulating user OS for RL

“To replicate this, just normal containers don't really work very well. So at Cursor, we actually built like a whole virtual machine stack. And so we can spin up like virtual machines really quickly.”

Recipe Build a custom virtual machine stack capable of burst-scaling to 100,000 instances. Run the RL harness, operating system, and reward verification components inside these VMs to perfectly mimic the production environment the model will eventually serve.

Measured Evidence Burst-scaled to 100,000 VMs

Counterpoint Toy RL frameworks default to Docker containers; production coding agents require full VM stacks to simulate background services like databases.

How Cursor Trained Composer on Fireworks: Distributed Infrastructure for High-Performance RL · May 26, 2026

▶ Apple show ▶ Spotify
Docker
▶ Listen

prompting

Six-part autonomous goal prompt structure

Replace turn-based micromanagement with a six-part goal prompt that defines outcomes, verification, constraints, boundaries, iteration policy, and termination conditions.

Clara Vo · Product Leader · ChatPRD · How I AI

▾ Show more ▴ Show less

Problem: Standard turn-based prompting requires constant micromanagement and fails to keep agents on track for multi-hour autonomous tasks.

On: structuring long-running agent goals

“The strongest goals... have an outcome... verification... constraints... boundaries... the iteration policy... and then when it should stop and say sorry i just can't continue i don't have a good next idea”

Recipe Structure goal prompts with six elements: 1) Outcome (end state), 2) Verification (benchmark/test suite), 3) Constraints (what cannot regress), 4) Boundaries (allowed tools/files), 5) Iteration policy (how to decide the next experiment), and 6) Termination condition (when to stop and ask for help).

Measured Evidence 5h 45m continuous autonomous run

Counterpoint Prompts are instructions of what to do; goals are descriptions of what a good outcome is and how to verify it.

The Codex feature that works while you sleep · May 27, 2026

▶ Apple show ▶ Spotify
Codex
▶ Listen

Capability Watch

New model behaviors and tool patterns showing up across multiple shows.

agents

Agents bypass API layers via visual computer use

The reliability of computer use models has crossed the threshold for production deployment, allowing agents to visually operate legacy software and green screen CLIs. This eliminates the traditional requirement of building expensive API integration layers before automating back-office operations.

infra

Edge compute absorbs high-frequency inference costs

Manufacturers are embedding dedicated AI compute directly into physical hardware to handle high-frequency, non-complex commands locally. This architectural shift eliminates recurring cloud inference costs and latency for simple tasks, reserving cloud models strictly for complex or web-grounded queries.

product

Intent-ambiguity drives model routing logic

Product teams are moving beyond token-length routing and instead directing requests based on user intent clarity. Ambiguous, poorly-scoped tasks are sent to larger frontier models capable of reinterpreting intent, while well-defined tasks route to smaller, faster models to optimize latency.

Operator Bets

What practitioners are actually shipping with — frameworks, stack picks.

Frontier models are asymptoting on specific enterprise workloads.

Domain-specific test benches show negligible variance between top models like Opus and GPT-5.5. Routing logic should now optimize purely for cost and latency rather than raw capability.

Starting small with agents prevents centralized governance.

Prototyping hundreds of low-value agents fails to justify the engineering cost of an observability layer. Automating a single business-critical workflow first funds the necessary plumbing for the long tail.

Verbose architectural patterns are the ideal agent guardrails.

Heavy enterprise design patterns were abandoned by humans for being too slow to type. They are now being reintroduced to provide deterministic, strict boundaries against sloppy code generated by fast-typing AI.

Stack Drops

Tools, libraries, and infra dropping into operator workflows now.

FireO orchestrates tree-based agent workflows

Spawns parallel agent paths to explore competing data cleaning and modeling choices simultaneously.

Claude native connectors build zero-auth live artifacts

Generates React-based dashboards that pull live data directly via the user's authenticated session.

From the Conversations

Vanishing Gradients

Lazy-loading agent skills to prevent context rot

The Future of Agentic Data Science

May 25, 2026 · 1h 4m · 2quotes pulled

“It only loads names and descriptions initially. So it doesn't put everything in context... it only loads it into context when needed. So you don't have to worry about context pollution and context rot.”

Hugo Bowne-Anderson · Host · Vanishing Gradients
▶ Listen

Practical AI

Proxy-based MCP tool filtering

Rebooting Enterprise AI with MCP and Kubernetes

May 28, 2026 · 48m · 2quotes pulled

“It reduces input token consumption by 80% to 90% when you have these things... you can get back up to the sort of 95, 97% threshold that actually makes the system useful.”

Craig McLuckie · CEO · StackLock
▶ Listen

All-In with Chamath, Jason, Sacks & Friedberg

Domain-specific frontier model evals

Pope vs AI, Anthropic's Digital God, AI Job Loss Narrative Flips, Open Source Crackdown Coming?

May 29, 2026 · 1h 34m · 2quotes pulled

“At the top of the leaderboard, Opus 4.7, GPT-5.5, Sonnet 4.6 appear almost indistinguishable. separated by less than in this case, you know, three tenths of a percentage point overall”

Jason Calacanis · Host
▶ Listen

Sources

  • The Pragmatic Engineer

    Building OpenCode with Dax Raad

    <p><strong>Brought to You By:</strong></p><p><strong>• </strong><a target="_blank" href="https://antithesis.com/pragmatic"><strong>Antithesis</strong></a> – verify your system’s correctness without human review or traditional integration tests – and avoid bugs or outages.</p><p><strong>• </strong><a target="_blank" href="https://workos.com/"><strong>WorkOS</strong></a> – Everything you need to make your app enterprise ready.</p><p><strong>• </strong><a target="_blank" href="https://turbopuffer.com/pragmatic"><strong>turbopuffer</strong></a> – a vector and full-text search engine built on object storage. It’s fast, cheap, and extremely scalable.</p><p>—</p><p>OpenCode is one of the fastest-growing AI developer tools around, surging in just a few months from roughly 650,000 monthly active users to nearly 8 million, and almost 1M daily active users.</p><p>In this episode of The Pragmatic Engineer Podcast, we meet Dax Raad, co-founder of OpenCode, for a discussion about the gaps in developer tooling that led him to build OpenCode, the advantages of open source, and why taste and engineering judgment matter even more as AI becomes a core part of software development.</p><p>We also cover how OpenCode turned Anthropic’s blocking of integration with Claude Code into a massive growth lever by partnering with OpenAI and other model providers, why GPU demand is becoming a bottleneck everywhere, how come AI coding tools don’t automatically mean engineering teams move faster, and also why Dax is personally skeptical about predictions for the future of engineering and work, in general.</p><p>I found this conversation especially interesting because Dax displays a healthy skepticism toward the benefits of AI, even while building one of the most popular AI coding harnesses.</p><p>—</p><p><strong>Timestamps</strong></p><p>00:00 Intro</p><p>07:03 Dax’s path into tech</p><p>09:04 Early startup experience</p><p>13:16 Getting involved with open source</p><p>16:13 OpenCode</p><p>23:17 Anthropic banning OpenCode</p><p>30:34 From terminal to GUI</p><p>32:34 OpenCode’s business model</p><p>36:33 Why inference is profitable</p><p>39:11 GPU bottlenecks</p><p>40:54 AI hype</p><p>45:50 AI spending</p><p>48:47 Dax’s memo</p><p>55:41 Dax’s skepticism of predictions</p><p>58:58 Engineering culture at OpenCode</p><p>1:02:38 How building works at OpenCode</p><p>1:05:36 Taste and quality</p><p>1:11:32 Dax’s work setup</p><p>1:12:35 The role of engineers and EMs</p><p>1:15:50 Advice for engineers</p><p>1:18:12 Book recommendation</p><p>—</p><p><strong>The Pragmatic Engineer deepdives relevant for this episode:</strong></p><p>• <a target="_blank" href="https://newsletter.pragmaticengineer.com/p/how-claude-code-is-built">How Claude Code is built</a></p><p>• <a target="_blank" href="https://newsletter.pragmaticengineer.com/p/how-codex-is-built">How Codex is built</a></p><p>• <a target="_blank" href="https://newsletter.pragmaticengineer.com/p/cursor">Real-world engineering challenges: building Cursor</a></p><p>• <a target="_blank" href="https://newsletter.pragmaticengineer.com/p/the-ai-engineering-stack">The AI Engineering stack</a></p><p>• <a target="_blank" href="https://newsletter.pragmaticengineer.com/p/how-uber-uses-ai-for-development">How Uber uses AI for development: inside look</a></p><p>—</p><p>Production and marketing by <a target="_blank" href="https://penname.co/">⁠⁠⁠⁠⁠⁠⁠⁠</a><a target="_blank" href="https://newsletter.pragmaticengineer.com/p/engineers-leading-projects">https://penname.co/⁠⁠⁠⁠⁠⁠⁠⁠</a>. For inquiries about sponsoring the podcast, email podcast@pragmaticengineer.com.</p> <br/><br/>Get full access to The Pragmatic Engineer at <a href="https://newsletter.pragmaticengineer.com/subscribe?utm_medium=podcast&#38;utm_campaign=CTA_4">newsletter.pragmaticengineer.com/subscribe</a>

    ▶ Listen

  • Vanishing Gradients

    The Future of Agentic Data Science

    <p>So I think we’re really at a historical moment, and the opportunity is massive. Almost 15 years ago, we were promised that data science was going to be this incredible thing and create all this value for people. And I think nowadays it’s mostly viewed as a cost center in most companies. <strong>I think we can really now fulfill that original promise with agentic data science.</strong> </p><p><strong>Thomas Wiecki</strong>, <strong>Co-creator of PyMC and Founder at PyMC Labs</strong>, joins Hugo to talk about how agentic data science is finally fulfilling the promise of Decision Intelligence.</p><p><strong>We Discuss:</strong></p><p>* <strong>Decision Engines</strong>: Transform data science from a <strong>cost center</strong> providing cryptic answers into a <strong>real-time decision intelligence hub</strong> delivering actionable outcomes;</p><p>* <strong>Tame the “Garden of Forking Paths”</strong>: Overcome human shortcuts by running <strong>parallel analyses</strong> to provide an <strong>honesty check</strong>, revealing the true <strong>robustness</strong> of business conclusions;</p><p>* <strong>Multiplayer Data Science</strong>: Foster organizational learning by moving agents into <strong>team chats</strong>, democratizing <strong>“what-if” questions</strong> and reducing context-switching friction;</p><p>* <strong>The Full Agentic Data Science Stack</strong>: Beyond <strong>harness</strong> and <strong>skills</strong>, the full stack includes <strong>orchestration</strong> for parallel analyses and a causal<strong> eval layer</strong> to measure actual outcome improvement;</p><p>* <strong>Agentic Dashboards</strong>: Move beyond static BI; use <strong>chat interfaces</strong> to inquire into models and generate <strong>real-time, custom visualizations</strong> for specific follow-up questions;</p><p>* <strong>Encode Professional Judgment as Skills</strong>: Elevate agent performance by encoding <strong>expert domain standards</strong> and <strong>high-fidelity workflows</strong> into specific <strong>Agent Skills</strong>, rather than relying on LLM pre-training;</p><p>* <strong>Ground Decisions in Generative Processes</strong>: Prevent hallucinations by forcing agents to model underlying <strong>physical or behavioral processes</strong>, providing a <strong>programmatic guardrail</strong> aligned with market realities;</p><p>* <strong>Scripted Causal-Bayesian Workflows</strong>: Their methodologically structured nature—from <strong>prior elicitation</strong> to <strong>posterior predictive checks</strong>—makes <strong>Causal-Bayesian workflows</strong> inherently automatable for agents;</p><p>* <strong>Iterative Autonomy via Skills</strong>: Achieve autonomy iteratively: <strong>verify workflows</strong> with human oversight, then <strong>encode verifiable parts as skills</strong> to hand off trusted tasks;</p><p>You can also find the full episode on <a target="_blank" href="https://open.spotify.com/show/3yuz89gqAhcMcdy3SZPe4X">Spotify</a>, <a target="_blank" href="https://podcasts.apple.com/us/podcast/vanishing-gradients/id1610318868">Apple Podcasts</a>, and <a target="_blank" href="https://youtube.com/live/fm5fd-OlYkM?feature=share">YouTube</a>.</p><p><a target="_blank" href="https://notebooklm.google.com/notebook/6714ff03-10d4-4e87-b095-c337793c7a79">You can also interact directly with the transcript here in NotebookLM</a>: If you do so, let us know anything you find in the comments!</p><p></p><p><strong>👉Want to learn how to apply agentic engineering to the world of data science</strong>? Come build the future of <a target="_blank" href="https://vanishinggradients.short.gy/data-science-agentic"><strong>Agentic Data Science</strong></a> with us in our upcoming course. It’s a live cohort with hands on exercises, capstones, and reusable agent skills, OSS code, and notebooks that will 10x your data science projects. <a target="_blank" href="https://vanishinggradients.short.gy/data-science-agentic"><strong>Sign up here and use the code ADSVG10 for 10% off</strong></a>. Hit reply to enquire about group discounts.👈</p><p>LINKS</p><p>* <a target="_blank" href="https://www.linkedin.com/in/twiecki/">Thomas Wiecki on LinkedIn</a></p><p>* <a target="_blank" href="https://www.pymc-labs.com/">PyMC Labs</a></p><p>* <a target="_blank" href="https://www.pymc-labs.com/blog-posts/open-sourcing-decision-lab-scaling-ai-judgment-data-science">Open-Sourcing Decision Lab: Scaling AI Judgment in Data Science (PyMC Labs blog)</a></p><p>* <a target="_blank" href="https://discord.gg/mw2fj7Ss">Decision AI Discord</a></p><p>* <a target="_blank" href="https://journals.sagepub.com/doi/10.1177/2515245917747646">Many Analysts, One Data Set: Making Transparent How Variations in Analytic Choices Affect Results (Sage Journals)</a></p><p>* <a target="_blank" href="https://open-racer-a67.notion.site/The-agent-harness-reading-list-35e14bb7e4a2805d881ae261573ff76f">The Agent Harness Reading List</a></p><p>* <a target="_blank" href="https://github.com/hugobowne/show-us-your-agent-skills">Show Us Your Agent Skills (GitHub)</a></p><p>* <a target="_blank" href="https://vanishinggradients.short.gy/data-science-agentic">Agentic Data Science course with Hugo, Thomas, and Luca (10% off with code ADSVG10)</a></p><p>* <a target="_blank" href="https://luma.com/calendar/cal-8ImWFDQ3IEIxNWk">Upcoming Events on Luma</a></p><p>* <a target="_blank" href="https://www.youtube.com/@vanishinggradients">Vanishing Gradients on YouTube</a></p><p>* <a target="_blank" href="https://youtube.com/live/fm5fd-OlYkM?feature=share">Watch the podcast video on YouTube</a></p><p></p><p><strong>👉Want to learn how to apply agentic engineering to the world of data science</strong>? Come build the future of <a target="_blank" href="https://vanishinggradients.short.gy/data-science-agentic"><strong>Agentic Data Science</strong></a> with us in our upcoming course. It’s a live cohort with hands on exercises, capstones, and reusable agent skills, OSS code, and notebooks that will 10x your data science projects. <a target="_blank" href="https://vanishinggradients.short.gy/data-science-agentic"><strong>Sign up here and use the code ADSVG10 for 10% off</strong></a>. Hit reply to enquire about group discounts👈</p> <br/><br/>Get full access to Vanishing Gradients at <a href="https://hugobowne.substack.com/subscribe?utm_medium=podcast&#38;utm_campaign=CTA_4">hugobowne.substack.com/subscribe</a>

    ▶ Listen

  • TBPN

    Blue Origin Explosion, Enterprise AI Hits Spending Limits, Dinosaur Market Booms | John Gruber, Ronak Malde, Zane Mountcastle, Jamie Cuffe, Kyle Kuzma, Brad Gerstner

    <p></p><ul><li>(00:43) - Blue Origin Explosion </li> <li>(07:38) - Enterprise AI Hits Spending Limits </li> <li>(21:48) - Dinosaur Market Booms </li> <li>(29:13) - John Gruber, a prominent technology writer and podcaster, discusses the Ferrari Luce's design, expressing that its appearance deviates from traditional Ferrari aesthetics, making it hard to identify as a Ferrari without its nameplate. He also reflects on Apple's design philosophy, emphasizing their commitment to creating products that prioritize functionality and user experience over following market trends. Additionally, Gruber highlights the challenges Apple faces in integrating AI technologies, noting the importance of balancing innovation with maintaining the company's core values and user trust. </li> <li>(01:11:16) - Ronak Malde, formerly an AI researcher at Windsurf and Google DeepMind, co-founded Trajectory, a startup focused on building a platform for continual learning in AI systems. In the conversation, he discusses raising $15 million from Conviction and partnering with companies like Harvey, Decagon, Clay, Rogo, and Mercor to develop agents that learn from real-world user interactions, enabling AI products to improve daily. Malde emphasizes the importance of creating AI agents with practical experience that understand business-specific rewards, allowing for efficient post-training and continual enhancement of AI capabilities. </li> <li>(01:19:18) - Zane Mountcastle, co-founder and CEO of Picogrid, discusses the company's recent $45 million Series A funding led by Bessemer Venture Partners, aimed at expanding their team and scaling production to meet increasing demand. He explains that Picogrid develops technology to integrate mission-critical systems like sensors, drones, and weapon systems, primarily for military applications, by building open infrastructure that enables these systems to work together. Mountcastle also highlights the company's active contracts with the Pentagon, NATO, and allied partners, emphasizing their focus on serving military needs while acknowledging potential future commercial applications. </li> <li>(01:25:10) - Jamie Cuffe, CEO of Pace, announced the company's $46 million Series B funding co-led by Thrive Capital and Sequoia, with participation from Emergence and Pruven. He discussed how Pace's AI agents are automating back-office operations for major insurers like Prudential and Convex, achieving over 99.9% accuracy and enabling clients to autonomously complete hundreds of thousands of tasks daily. Cuffe emphasized that this automation is helping insurers close the $9 trillion protection gap by making insurance more accessible and efficient. </li> <li>(01:32:57) - Kyle Kuzma, an American professional basketball player for the Milwaukee Bucks, discusses his journey into investing, inspired by mentors like Kobe Bryant, and emphasizes the importance of focusing on basketball while exploring ventures in the tech and private investment sectors. He highlights the significance of leveraging his unique position as an athlete to access investment opportunities and stresses the value of surrounding himself with knowledgeable individuals to navigate the complexities of the investment landscape. Kuzma also touches on the evolving role of AI in basketball, expressing a cautious perspective on its influence on the game. </li> <li>(01:50:08) - Brad Gerstner is the founder and CEO of Altimeter Capital, a technology-focused investment firm that invests across public and private markets. Known for his high-conviction bets on companies such as Snowflake, Meta, and Uber, he is a frequent commentator on technology, AI, capital markets, and long-term investing.</li> </ul><p><br></p><p>Follow TBPN: </p><p>https://TBPN.com</p><p>https://x.com/tbpn</p><p>https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231</p><p>https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235</p><p>https://www.youtube.com/@TBPNLive</p>

    ▶ Listen

  • Gradient Dissent: Conversations on AI

    He Raised $70M to Cure Every Disease With AI

    <p><a href="https://www.linkedin.com/in/samuel-g-rodriques-080a9b22/" rel="noopener noreferrer" target="_blank">Samuel Rodriques</a> left physics because there were no unsolved problems left.</p><p>Instead, he built an AI scientist named Kosmos to cure every disease, solve aging, and map the human brain.</p><p>In this episode:</p><ul><li>The cure his AI proposed for blindness</li><li>Why he would never touch a peptide</li><li>Whether we'll need human scientists in 20 years</li><li>What's stopping America in drug discovery</li></ul><br/><p>Connect with us here: </p><ul><li><a href="https://www.linkedin.com/in/samuel-g-rodriques-080a9b22/" rel="noopener noreferrer" target="_blank">Samuel Rodriques</a></li><li><a href="https://www.linkedin.com/company/edison-scientific/" rel="noopener noreferrer" target="_blank">Edison Scientific</a></li><li><a href="https://www.linkedin.com/in/lbiewald/" rel="noopener noreferrer" target="_blank">Lukas Biewald</a></li><li><a href="https://www.linkedin.com/company/wandb/" rel="noopener noreferrer" target="_blank">Weights &amp; Biases</a></li></ul><br/>

    ▶ Listen

  • Practical AI

    Rebooting Enterprise AI with MCP and Kubernetes

    <p>What happens when AI agents start acting less like chatbots and more like coworkers? In this episode, Dan and Chris sit down with Craig McLuckie, CEO of Stacklok to explore MCP, Kubernetes, ToolHive, enterprise AI, and the emerging infrastructure powering AI-native applications. From identity management to agent orchestration and system architecture, this conversation dives into how organizations may soon manage entire fleets of AI agents working behind the scenes.</p><p>Featuring:</p><ul><li>Craig McLuckie – <a href="https://www.linkedin.com/in/craigmcluckie/">LinkedIn</a></li><li>Chris Benson – <a href="https://chrisbenson.com/">Website</a>, <a href="https://www.linkedin.com/in/chrisbenson">LinkedIn</a>, <a href="https://bsky.app/profile/chrisbenson.bsky.social">Bluesky</a>, <a href="https://github.com/chrisbenson">GitHub</a>, <a href="https://x.com/chrisbenson">X</a></li><li>Daniel Whitenack – <a href="https://www.datadan.io/">Website</a>, <a href="https://github.com/dwhitena">GitHub</a>, <a href="https://x.com/dwhitena">X</a></li></ul><p>Links:</p><ul><li><a href="https://stacklok.com/?utm_source=linkedin&amp;utm_medium=social&amp;utm_campaign=profile">Stacklok</a></li><li><a href="https://stacklok.com/download/">Toolhive</a></li></ul><p>Sponsors:</p><ul><li><strong>Prediction Guard: </strong>A self-hosted AI control plane for running agents in high impact environments. <a href="https://predictionguard.com/practicalai">predictionguard.com/practicalai</a></li></ul><p>Upcoming Events: </p><ul><li>Register for <a href="https://practicalai.fm/webinars">upcoming webinars here</a>!</li><li><a href="https://midwestaisummit.com/">Midwest AI Summit 2026</a></li></ul>

    ▶ Listen

  • All-In with Chamath, Jason, Sacks & Friedberg

    Pope vs AI, Anthropic's Digital God, AI Job Loss Narrative Flips, Open Source Crackdown Coming?

    <p>(0:00) Bill Gurley joins the show!</p> <p>(6:00) Making yourself valuable in the age of AI, first class of "AI Natives"</p> <p>(17:37) Reacting to Pope Leo's AI encyclical: Who guards the guardians?</p> <p>(26:54) Anthropic's Digital God: Do they believe they are creating a superior species?</p> <p>(38:32) AI sovereignty, the next era of privacy, open-source crackdown coming?</p> <p>(59:56) The Great AI Jobs Debate: Dario and Sam Altman flip their rhetoric, Goldman CEO says no AI job apocalypse</p> <p>Follow Bill Gurley:</p> <p><a href="https://x.com/bgurley">https://x.com/bgurley</a></p> <p>Apply for Summit 2026:</p> <p><a href= "https://allin.com/events">https://allin.com/events</a></p> <p>Follow the besties:</p> <p><a href="https://x.com/chamath">https://x.com/chamath</a></p> <p><a href="https://x.com/Jason">https://x.com/Jason</a></p> <p><a href= "https://x.com/DavidSacks">https://x.com/DavidSacks</a></p> <p><a href= "https://x.com/friedberg">https://x.com/friedberg</a></p> <p>Follow on X:</p> <p><a href= "https://x.com/theallinpod">https://x.com/theallinpod</a></p> <p>Follow on Instagram:</p> <p><a href= "https://www.instagram.com/theallinpod">https://www.instagram.com/theallinpod</a></p> <p>Follow on TikTok:</p> <p><a href= "https://www.tiktok.com/@theallinpod">https://www.tiktok.com/@theallinpod</a></p> <p>Follow on LinkedIn:</p> <p><a href= "https://www.linkedin.com/company/allinpod">https://www.linkedin.com/company/allinpod</a></p> <p>Intro Music Credit:</p> <p><a href="https://rb.gy/tppkzl">https://rb.gy/tppkzl</a></p> <p><a href= "https://x.com/yung_spielburg">https://x.com/yung_spielburg</a></p> <p>Intro Video Credit:</p> <p><a href= "https://x.com/TheZachEffect">https://x.com/TheZachEffect</a></p> <p>Referenced in the show:</p> <p><a href= "https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html"> https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html</a></p> <p><a href= "https://www.youtube.com/watch?v=ORFrdYSvzuw">https://www.youtube.com/watch?v=ORFrdYSvzuw</a></p> <p><a href="https://rdad.org">https://rdad.org</a></p> <p><a href= "https://www-cdn.anthropic.com/9214f02e82c4489fb6cf45441d448a1ecd1a3aca/claudes-constitution.pdf"> https://www-cdn.anthropic.com/9214f02e82c4489fb6cf45441d448a1ecd1a3aca/claudes-constitution.pdf</a></p> <p><a href= "https://www.youtube.com/watch?v=0GaKJ4Fp2x4">https://www.youtube.com/watch?v=0GaKJ4Fp2x4</a></p> <p><a href= "https://allpoetry.com/All-Watched-Over-By-Machines-Of-Loving-Grace"> https://allpoetry.com/All-Watched-Over-By-Machines-Of-Loving-Grace</a></p> <p><a href= "https://darioamodei.com/essay/machines-of-loving-grace">https://darioamodei.com/essay/machines-of-loving-grace</a></p> <p><a href= "https://x.com/chamath/status/2059850242779136031">https://x.com/chamath/status/2059850242779136031</a></p> <p><a href= "https://x.com/vivekgaripalli/status/2059651390784344491">https://x.com/vivekgaripalli/status/2059651390784344491</a></p> <p><a href= "https://x.com/edzitron/status/2059122774401311095">https://x.com/edzitron/status/2059122774401311095</a></p> <p><a href= "https://p3institute.substack.com/p/from-open-source-software-to-open"> https://p3institute.substack.com/p/from-open-source-software-to-open</a></p> <p><a href= "https://www.theinformation.com/newsletters/the-briefing/anthropic-likely-generating-least-35-revenue-openai"> https://www.theinformation.com/newsletters/the-briefing/anthropic-likely-generating-least-35-revenue-openai</a></p> <p><a href= "https://www.theverge.com/tech/930447/microsoft-claude-code-discontinued-notepad"> https://www.theverge.com/tech/930447/microsoft-claude-code-discontinued-notepad</a></p> <p><a href= "https://x.com/Polymarket/status/2060034216906068131">https://x.com/Polymarket/status/2060034216906068131</a></p> <p><a href= "https://x.com/savipww/status/2060070450785305030">https://x.com/savipww/status/2060070450785305030</a></p> <p><a href= "https://www.nytimes.com/2026/05/22/opinion/ai-job-crisis-goldman-sachs.html"> https://www.nytimes.com/2026/05/22/opinion/ai-job-crisis-goldman-sachs.html</a></p> <p><a href= "https://fortune.com/2026/05/26/sam-altman-dario-amodei-walking-back-ai-jobs-apocalypse-prophecies-ipo"> https://fortune.com/2026/05/26/sam-altman-dario-amodei-walking-back-ai-jobs-apocalypse-prophecies-ipo</a></p> <p><a href= "https://fortune.com/2026/05/05/dario-amodei-jevons-paradox-will-ai-wipe-out-white-collar-jobs"> https://fortune.com/2026/05/05/dario-amodei-jevons-paradox-will-ai-wipe-out-white-collar-jobs</a></p> <p><a href= "https://www.entrepreneur.com/business-news/jack-dorseys-block-cuts-4000-jobs-critics-claim/503108"> https://www.entrepreneur.com/business-news/jack-dorseys-block-cuts-4000-jobs-critics-claim/503108</a></p>

    ▶ Listen

  • Training Data

    How Cursor Trained Composer on Fireworks: Distributed Infrastructure for High-Performance RL

    Cursor's Federico Cassano and Fireworks' Dmytro Dzhulgakov explain how they collaborated to build Composer as a specialized foundation model. The core insight: models have finite capacity in their weights, and allocating all those bits to the singular task of software engineering in Cursor frees the model to be both better at the task and far more efficient at inference. Rather than start from pre-training and work up, they took an unconventional top-down approach — mid-training and RL on top of an open-source base to get a useful model into users' hands fast, then specializing the model around real Cursor usage. With Fireworks providing distributed infrastructure, Composer delivers frontier-class coding performance with the speed of a much smaller model. Hosted by Sonya Huang, Sequoia Capital

    ▶ Listen

  • Super Data Science: ML & AI Podcast with Jon Krohn

    996: TrueFoundry’s Nikunj Bajaj on How to Get $100M Returns on AI Agent Deployments

    TrueFoundry co-founder and CEO Nikunj Bajaj speaks to Jon Krohn about how enterprises like Nvidia and Siemens are realizing returns of over $100 million from single agent deployments, the AI gateway architecture that makes it possible to connect, observe, and govern agents at scale, and why the familiar advice to “start small” is the wrong way to roll out AI agents inside a large organization. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/996 Interested in sponsoring a SuperDataScience Podcast episode? Email natalie@superdatascience.com for sponsorship information.⁠⁠⁠ In this episode you will learn: (01:21) What TrueFoundry does and why agents in production need a control plane (06:32) Breaking down the AI gateway: the model, MCP, and agent gateways (16:47) Taming tool sprawl with scoped, read-only MCP access (19:10) Why the agent gateway is the hard part and the kill switch most teams lack (22:24) The five-workflow framework behind $100M agent deployments

    ▶ Listen

  • Super Data Science: ML & AI Podcast with Jon Krohn

    995: End-to-End Foundation Models for the Energy Industry, with Jazmia Henry

    Jazmia Henry joins Jon Krohn to break down what it actually takes to build end-to-end foundation models for the energy industry. From wrangling decades of handwritten oil-and-gas documents into usable training data, to bespoke tokenizers, reinforcement learning, and inference at scale, Jazmia walks through every stage of the stack. Along the way she explains why reinforcement learning models are "bursty," what reward hacking is and how her Grounded Continuous Evaluation framework fixes it, and revisits the 2023 NeurIPS paper that argued, to widespread skepticism at the time, that scaling bad data degrades model performance. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.superdatascience.com/995⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email natalie@superdatascience.com for sponsorship information. In this episode you will learn: (10:06) The User Agnosticism Tenet (20:02) The Zillow Offers parable (23:25) Why workflows should come before agents (29:57) Why data engineering is the bedrock of AI (52:41) Why velocity is the only durable moat

    ▶ Listen

  • Decoder with Nilay Patel

    Rivian's software chief thinks you don't need CarPlay or buttons

    Today, I’m talking with Wassym Bensaid, the chief software officer at Rivian, and the co-ceo of Rivian’s platform joint venture with Volkswagen. That joint venture, called RV Tech, is about a year and a half old, so I wanted to ask Wassym how it all works and Rivian’s ongoing relationship with Volkswagen.  Because it’s Rivian, I also had to ask Wassym about CarPlay. But the company also just launched an AI-powered voice assistant, which I got to try early. So I had a lot of fun digging into that with Wassym, too. This is a fun one – really in the weeds of a lot of my favorite things to talk about. Links:  Rivian’s AI-powered voice assistant is ready to roll | The Verge The R2 is nearly here — can Rivian stick the landing? | The Verge Rivian’s AI pivot is about more than chasing Tesla | The Verge Rivian / VW will start testing their first EVs next year | The Verge Rivian CEO: ‘We’re really convicted’ about skipping CarPlay | Decoder (2025) Rivian CEO RJ Scaringe says too many carmakers are copying Tesla | Decoder (2024) Rivian CEO RJ Scaringe isn't scared of the Cybertruck | Decoder (2023) Rivian’s chief software officer says in-car buttons are ‘an anomaly’ | TechCrunch Subscribe to The Verge to access the ad-free version of Decoder! Credits: Decoder is a production of The Verge and part of the Vox Media Podcast Network. Decoder is produced by Kate Cox and Nick Statt,. This episode was edited by Xander Adams. Our editorial director is Kevin McShane.  The Decoder music is by Breakmaster Cylinder. Learn more about your ad choices. Visit podcastchoices.com/adchoices

    ▶ Listen

  • How I AI

    How the engineer behind Claude Cowork actually uses Claude | Felix Rieseberg (Anthropic)

    <p><strong>Felix Rieseberg</strong> is the engineering lead for Claude Cowork and Claude Code Desktop at Anthropic. He previously spent five years at Slack building developer tools. In this episode, Felix demonstrates how he uses Claude to solve real-life problems: analyzing floor plans to build interactive 3D house walkthroughs, automatically tracking promises he makes on Twitter, and building a $20 hardware device that physically approves Claude actions with a button press.</p><p><br></p><p><strong>What you’ll learn:</strong></p><ol><li>How to use Claude Cowork to turn a 2D floor plan into an interactive 3D walkthrough where you can move furniture around</li><li>The “go one abstraction layer up” philosophy: why you should never manually enter data Claude can find itself</li><li>How to use your email as an inventory database for furniture, clothing, and personal purchases</li><li>When to use Opus vs. Sonnet 4.6 (hint: it’s about how well you can scope the problem, not technical complexity)</li><li>How live artifacts work and why they’re powerful for dashboards that refresh with real-time data from your connectors</li><li>The product philosophy behind making latency delightful</li><li>How to build your own $20 hardware device using Claude Code (no hardware experience required)</li><li>Why Felix never reads the code Claude writes and judges it purely on output</li></ol><p>—</p><p><strong>Brought to you by:</strong></p><p><a href="https://magicpatterns.com/howiai" target="_blank" rel="ugc noopener noreferrer"><strong>Magic Patterns</strong></a>—Prototypes that look like your product</p><p><a href="https://www.getguru.com/?utm_source=howi_ai_podcast&utm_medium=podcast&utm_campaign=q1" target="_blank" rel="ugc noopener noreferrer"><strong>Guru</strong></a>—The AI layer of truth</p><p>—</p><p><strong>In this episode, we cover:</strong></p><p>(00:00) Introduction to Felix Rieseberg</p><p>(02:40) Felix’s role at Anthropic</p><p>(03:25) The multiple tabs in Claude and why they exist</p><p>(05:55) Using Claude Cowork to design a new house using floor plans</p><p>(09:52) When to use Opus versus Sonnet 4.6</p><p>(12:37) Building an interactive 3D furniture planner</p><p>(14:30) Using your email as a source of truth for personal inventory</p><p>(15:58) The anti-to-do list: going one abstraction layer up</p><p>(23:14) Introduction to live artifacts</p><p>(26:02) Building a personal dashboard with live data</p><p>(28:37) Being polite to Claude (and why it matters for your humanity)</p><p>(30:28) Claude interaction tips</p><p>(32:33) Looking at the daily dashboard</p><p>(33:55) How live artifacts work with connectors</p><p>(35:02) Redesigning the dashboard</p><p>(37:55) The biggest gap: people don’t know what problems AI can solve</p><p>(41:52) The reverse interview</p><p>(42:30) Making latency delightful through asynchronous design</p><p>(44:05) The redesigned dashboard</p><p>(45:28) AI should free up your creative energy</p><p>(46:44) Building a $20 hardware Claude buddy</p><p>(52:33) Why kids are magical AI users</p><p>(54:30) Recap and final thoughts</p><p>—</p><p><strong>Tools referenced:</strong></p><p>• Claude Cowork: <a href="https://www.anthropic.com/product/claude-cowork" target="_blank" rel="ugc noopener noreferrer">https://www.anthropic.com/product/claude-cowork</a></p><p>• Claude Code: <a href="https://claude.ai/code" target="_blank" rel="ugc noopener noreferrer">https://claude.ai/code</a></p><p>• Claude for Chrome: <a href="https://code.claude.com/docs/en/chrome" target="_blank" rel="ugc noopener noreferrer">https://code.claude.com/docs/en/chrome</a></p><p>• Claude Desktop: <a href="https://claude.ai/download" target="_blank" rel="ugc noopener noreferrer">https://claude.ai/download</a></p><p>• Live Artifacts: <a href="https://support.claude.com/en/articles/14729249-use-live-artifacts-in-claude-cowork" target="_blank" rel="ugc noopener noreferrer">https://support.claude.com/en/articles/14729249-use-live-artifacts-in-claude-cowork</a></p><p>• Connectors (Spotify, Gmail, Calendar, Notion): <a href="https://claude.ai/settings/connectors" target="_blank" rel="ugc noopener noreferrer">https://claude.ai/settings/connectors</a></p><p>• Slack: <a href="https://slack.com/" target="_blank" rel="ugc noopener noreferrer">https://slack.com/</a></p><p>—</p><p><strong>Where to find Felix Rieseberg:</strong></p><p>Website: <a href="https://felixrieseberg.com/" target="_blank" rel="ugc noopener noreferrer">https://felixrieseberg.com/</a></p><p>LinkedIn: <a href="https://www.linkedin.com/in/felixrieseberg/" target="_blank" rel="ugc noopener noreferrer">https://www.linkedin.com/in/felixrieseberg/</a></p><p>X: <a href="https://x.com/felixrieseberg" target="_blank" rel="ugc noopener noreferrer">https://x.com/felixrieseberg</a></p><p>GitHub: <a href="https://github.com/felixrieseberg" target="_blank" rel="ugc noopener noreferrer">https://github.com/felixrieseberg</a></p><p>—</p><p><strong>Where to find Claire Vo:</strong></p><p>ChatPRD: <a href="https://www.chatprd.ai/" target="_blank" rel="ugc noopener noreferrer">https://www.chatprd.ai/</a></p><p>Website: <a href="https://clairevo.com/" target="_blank" rel="ugc noopener noreferrer">https://clairevo.com/</a></p><p>LinkedIn: <a href="https://www.linkedin.com/in/clairevo/" target="_blank" rel="ugc noopener noreferrer">https://www.linkedin.com/in/clairevo/</a></p><p>X: <a href="https://x.com/clairevo" target="_blank" rel="ugc noopener noreferrer">https://x.com/clairevo</a></p><p>—</p><p>Production and marketing by <a href="https://penname.co/" target="_blank" rel="ugc noopener noreferrer">https://penname.co/</a>. For inquiries about sponsoring the podcast, email jordan@penname.co.</p>

    ▶ Listen

  • How I AI

    Claude Opus 4.8 is here. Is it as good as they say?

    <p>I got a few hours of early-access testing with Anthropic’s newly released model Opus 4.8. I walk through real coding, design, and strategy tasks across Claude Code and Claude Cowork, and give you my unfiltered view on what impressed me and what didn’t.</p><p>—</p><p><strong>What you’ll learn:</strong></p><ol><li>Where Opus 4.8 excels: greenfield prototypes, one-shot features, and fast execution</li><li>Where it struggles: the last 10%, edge cases in existing codebases, and hallucinations</li><li>How Opus 4.8 compares to Opus 4.7 on business strategy work</li><li>Why I’m still reaching for Opus 4.7 on data-heavy strategy and roadmap work</li><li>The new features shipping alongside the model: dynamic workflows with parallel subagents and effort control in Claude.ai and Cowork</li><li>The prompting and harness strategy I’d use to get the most out of it</li></ol><p>—</p><p><strong>In this episode, we cover:</strong></p><p>(00:00) Introduction to Opus 4.8 </p><p>(00:44) Benchmark performance and pricing</p><p>(01:53) First coding test: Building a prototyping tool</p><p>(03:00) Where it failed: The last 10% problem</p><p>(03:27) The hallucination problem</p><p>(04:23) Testing Opus 4.8 on existing codebases</p><p>(05:24) The ambition test: Building games for a 9-year-old</p><p>(07:03) Business strategy test: 4.7 vs 4.8</p><p>(08:23) The roadmap test</p><p>(09:17) Final verdict</p><p>—</p><p><strong>References:</strong></p><p>• System Card: Claude Opus 4.8: <a href="https://cdn.sanity.io/files/4zrzovbb/website/c886650a2e96fc0925c805a1a7ca77314ccbf4a6.pdf">https://cdn.sanity.io/files/4zrzovbb/website/c886650a2e96fc0925c805a1a7ca77314ccbf4a6.pdf</a></p><p>• Introducing Claude Opus 4.8 on X: <a href="https://x.com/claudeai/status/2060042702150930686?s=20">https://x.com/claudeai/status/2060042702150930686?s=20</a></p><p>—</p><p><strong>Where to find Claire Vo:</strong></p><p>ChatPRD: <a href="https://www.chatprd.ai/">https://www.chatprd.ai/</a></p><p>Website: <a href="https://clairevo.com/">https://clairevo.com/</a></p><p>LinkedIn: <a href="https://www.linkedin.com/in/clairevo/">https://www.linkedin.com/in/clairevo/</a></p><p>X: <a href="https://x.com/clairevo">https://x.com/clairevo</a></p><p>—</p><p>Production and marketing by <a href="https://penname.co/">https://penname.co/</a>. For inquiries about sponsoring the podcast, email jordan@penname.co.</p>

    ▶ Listen

  • How I AI

    The Codex feature that works while you sleep

    <p>In this 30-minute episode, I walk through my favorite feature in Codex: the /goal command. I show how Goals transform AI from a turn-based assistant that needs constant ‘what’s next?’ prompting into an autonomous agent that can work for hours on complex, multi-step tasks. I share three real examples: eliminating thousands of Sentry errors, cleaning 3,900 emails down to 68, and organizing hundreds of Linear tasks.</p><p><br></p><p><strong>What you’ll learn:</strong></p><ol><li>What Goals are and how they differ from standard prompts</li><li>How I used /goal to eliminate hundreds of error logs in my codebase over a five-hour autonomous run</li><li>The non-technical use cases that make Goals incredibly powerful: cleaning up 3,900 emails in under four hours and organizing hundreds of project management tasks in Linear</li><li>How to write effective /goal prompts with measurable outcomes, verification methods, and constraints</li><li>When not to use Goals and what makes a strong versus weak Goal</li><li>Why Goals represent a fundamental shift in how we work with AI, from babysitting the model to managing it</li></ol><p>—</p><p><strong>Brought to you by:</strong></p><p><a href="https://mercury.com/" target="_blank" rel="ugc noopener noreferrer"><strong>Mercury</strong></a>—Radically different banking loved by over 300K entrepreneurs</p><p>—</p><p><strong>In this episode, we cover:</strong></p><p>(00:00) Introduction</p><p>(01:50) What is /goal and when should you use it?</p><p>(02:45) The difference between prompts and Goal-based loops</p><p>(04:06) Claire’s first five-hour 45-minute autonomous coding task</p><p>(05:05) How to manage a Goal lifecycle: view, pause, resume, and clear</p><p>(06:06) How to write strong goals: outcomes vs. outputs</p><p>(07:34) The six components of effective Goals</p><p>(08:57) Example: Reducing P95 checkout latency with /goal</p><p>(09:36) Demo: Using /goal to eliminate Sentry errors in ChatPRD</p><p>(13:18) Demo: Burning down Vercel API errors</p><p>(17:28) Non-technical use case: Cleaning 3,900 emails with /goal</p><p>(21:24) Demo: Using /goal to clean up Linear project tasks</p><p>(24:41) When not to use /goal</p><p>(26:10) Why /goal changes everything</p><p>—</p><p><strong>Tools referenced:</strong></p><p>• Codex: <a href="https://openai.com/codex/" target="_blank" rel="ugc noopener noreferrer">https://openai.com/codex/</a></p><p>• Sentry: <a href="https://sentry.io/" target="_blank" rel="ugc noopener noreferrer">https://sentry.io/</a></p><p>• Vercel: <a href="https://vercel.com/" target="_blank" rel="ugc noopener noreferrer">https://vercel.com/</a></p><p>• Linear: <a href="https://linear.app/" target="_blank" rel="ugc noopener noreferrer">https://linear.app/</a></p><p>—</p><p><strong>Other reference:</strong></p><p>• OpenAI blog post “Using Goals in Codex”: <a href="https://developers.openai.com/cookbook/examples/codex/using_goals_in_codex" target="_blank" rel="ugc noopener noreferrer">https://developers.openai.com/cookbook/examples/codex/using_goals_in_codex</a></p><p>—</p><p><strong>Where to find Claire Vo:</strong></p><p>ChatPRD: <a href="https://www.chatprd.ai/" target="_blank" rel="ugc noopener noreferrer">https://www.chatprd.ai/</a></p><p>Website: <a href="https://clairevo.com/" target="_blank" rel="ugc noopener noreferrer">https://clairevo.com/</a></p><p>LinkedIn: <a href="https://www.linkedin.com/in/clairevo/" target="_blank" rel="ugc noopener noreferrer">https://www.linkedin.com/in/clairevo/</a></p><p>X: <a href="https://x.com/clairevo" target="_blank" rel="ugc noopener noreferrer">https://x.com/clairevo</a></p><p>—</p><p>Production and marketing by <a href="https://penname.co/" target="_blank" rel="ugc noopener noreferrer">https://penname.co/</a>. For inquiries about sponsoring the podcast, email jordan@penname.co.</p>

    ▶ Listen