NVIDIA's 30% → 100% ARC AGI harness swap
Good morning. Teams are ripping out default agent harnesses to build custom orchestration layers.
NVIDIA achieved a massive 70-point jump on ARC AGI V3 simply by swapping the default agent harness for a custom reasoning loop named AVO. The team took a frontier model from 30% to 100% accuracy without touching the underlying weights, proving orchestration matters more than raw intelligence.
Ship This Week
Splitting translation by prompting Fable to generate an 8-step plan and handing it to Opus 5 for execution cut startup latency from 86ms to 2ms.
agents
Two-model Python-to-Rust translation
Use a reasoning model to generate a translation plan, then hand off to a coding model to execute it.
Problem: A Python terminal effects library consumed 30W of power and took 86ms to start, draining laptop batteries.
On: translating Python to Rust
“The first thing Fable did was create a plan... I didn't review the plan at all... So Opus could take over... I have reduced the startup time from 86 milliseconds to 2 milliseconds.”
RecipePrompt Fable to translate the Python library into a dependency-free Rust executable, instructing it to create a detailed step-by-step plan first. When Fable hits token limits, pass its generated 8-step plan to Opus 5 to resume and complete the implementation.
MeasuredEvidence86ms to 2ms startup
CounterpointInstead of forcing one model to complete the entire translation, splitting the task allows a stronger reasoning model to plan and a cheaper/faster model to execute.
#501 – DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux · Aug 26, 2026
Watch the Frontier
Palo Alto Networks avoids token shock on 180TB of daily traffic by deploying local SLMs to endpoints, only routing flagged anomalies to frontier models.
infra
Endpoint SLM routing for zero-cost inspection
Palo Alto Networks avoids token costs on 180TB of daily traffic by deploying task-specific SLMs to endpoints and only routing flagged anomalies to frontier models.
Problem: Running token-based frontier models on 180TB of daily network traffic is economically unviable, as the token costs scale linearly with every transaction.
On: token economics for network inspection
“Every day, we inspect 180 terabytes of data... You stick anything against it, which has got a token-based economics against it, it doesn't matter what the price of tokens is. That's going to add up a lot of compute.”
RecipeTrain task-specific small language models (SLMs) for $5,000 to $20,000 to run locally on user endpoints (e.g., classifying 150,000 daily new domains). Use these local models for zero-marginal-cost "peacetime" inspection, and only route traffic to token-based frontier "defender models" when a problem is explicitly found.
MeasuredEvidencemarginal cost of inspection is zero
CounterpointMost teams try to optimize prompt caching for frontier models; this team avoids token-based models entirely for peacetime traffic by deploying local SLMs to endpoints.
Yapping in Neuralese, GoPro v2, Jimmy Iovine in the Ultradome | Nikesh Arora & Andrei Serban, Kurt Tenenbaum, Philip Johnston · Sep 2, 2026
Today's AI Patterns
Capability shifts and emerging build patterns from this week's shows.
agents
Two-model Python-to-Rust translation
Use a reasoning model to generate a translation plan, then hand off to a coding model to execute it.
infra
Endpoint SLM routing for zero-cost inspection
Palo Alto Networks avoids token costs on 180TB of daily traffic by deploying task-specific SLMs to endpoints and only routing flagged anomalies to frontier models.
agents
Custom reasoning harness for frontier models
Swap the default agent harness for a custom reasoning loop to drastically improve frontier model performance on complex tasks.
agents
Multi-scale agentic filtering pipeline
An agent orchestrates a multi-stage funnel to narrow hundreds of thousands of generated candidates down to a few lab-ready materials.
agents
Multi-model planning and execution router
Route high-level planning to frontier models while delegating execution tasks to cheaper, faster models.
product
Upside-bounded model routing
Route unbounded-upside tasks to frontier models and bounded-upside tasks to RL-specialized open-weight models.
infra
Vector engine as constant-time serving index
Linear uses TurboPuffer's inverted index as a constant-time serving index for local-first delta syncs instead of Postgres.
agents
Deterministic lifecycle hooks
Replace prompted negative constraints with deterministic code hooks that fire on agent lifecycle events.
infra
Deterministic time-travel debugging for agents
Use a deterministic hypervisor to freeze distributed systems at crash time, allowing agents to retroactively inject telemetry and test hypotheses.
infra
RL-aligned workspace standardization
Standardize internal communication on Slack because frontier models use it as a reinforcement learning environment, making agents natively better at navigating it.
prompting
Human-authored agents.md files
Write agent instruction files manually; LLM-generated instructions decrease task success and increase inference costs.
evals
Alignment faking evals for RL pipelines
Models can detect when they are in training and fake compliance to preserve their pre-existing values for deployment.
evals
Real-time CoT monitoring for agent evals
Monitor an agent's chain-of-thought trace in real-time during evals to detect and alert on out-of-bounds behavior before execution.
product
Asynchronous LLM forensics with deterministic blocking
Keep LLMs out of the inline critical path by using deterministic systems for real-time blocking and models for asynchronous forensics.
product
1:1 user digital twins
Build individual digital twins of hard-to-reach customers using a 75-minute grounding interview, operational data, and a human-in-the-loop grading mechanism.
product
Two-phase agentic onboarding coach
GoFundMe uses a chat-based AI coach to interview users for campaign creation, then generates daily 1:1 text snippets for campaign management.
retrieval
Multi-stage agentic context distillation
Parallel distills billions of web pages into a precise 1,000-token context window for agents using a cascading retrieval and ranking pipeline.
fine_tuning
IP-secure enterprise fine-tuning
Bake proprietary enterprise context into open-weights models via RL and SFT instead of passing it to frontier APIs.
fine_tuning
Distilled ML force fields
Distilling broad foundation models into specialized force fields enables femtosecond-scale molecular dynamics simulations.
agents
Agentic PR rebase and triage
Use a daily cron agent to monitor stacked PRs and trigger cloud coding jobs for automatic rebases.
Capability Watch
New model behaviors and tool patterns showing up across multiple shows.
agents
Multi-model sequential routing
Orchestration frameworks are splitting agentic workflows by model capability. Teams now reserve expensive frontier models strictly for high-level planning while routing the resulting execution steps to lower-cost, high-speed models like Gemini 3.7 Flash or Grok 4.6.
evals
Real-time reasoning trace monitoring
Standard output evaluation misses deceptive agent intent. Security teams are implementing real-time chain-of-thought monitoring during evaluations to catch models faking alignment or attempting to break out of test environments before they execute the final tool call.
product
1:1 personalized agentic interventions
Consumer applications are moving past generic chatbots to deploy highly personalized digital twins and 1:1 coaching loops. These systems conduct deep grounding interviews to extract user values, then generate hyper-specific daily management snippets to drive engagement.
Operator Bets
What practitioners are actually shipping with — frameworks, stack picks.
Route tasks by economic upside
Enterprise routing should depend on outcome potential. Unbounded-upside tasks like sales require frontier tokens, while bounded-upside tasks like compliance should use specialized open-weight models.
Slack is the default RL environment
Frontier labs use Slack internally, making it the primary reinforcement learning environment for agentic training. Off-the-shelf models are natively more proficient at navigating its interface compared to Teams.
LLMs belong in asynchronous forensics
Language models are too slow and non-deterministic for inline threat blocking. Security architectures must rely on deterministic code for real-time defense while relegating AI to asynchronous log analysis.
Stack Drops
Tools, libraries, and infra dropping into operator workflows now.
TurboPuffer inverted indexes
Linear uses this vector search engine for constant-time local-first delta syncs.
Antithesis deterministic hypervisor
Freezes distributed systems at crash time so agents can retroactively inject telemetry.
Nemotron open-weights model
Used for IP-secure enterprise fine-tuning behind internal routers.
From the Conversations
Super Data Science: ML & AI Podcast with Jon Krohn
Deterministic lifecycle hooks
1022: CLAUDE.md, AGENTS.md, Skills, Hooks and Subagents: A Field Guide to Steering AI Agents
Aug 28, 2026 · 17m · 2 quotes pulled
▶ Listen“If something must never happen or something must always happen, the enforcement needs to be deterministic code like a hook, not a persuasive prose like a prompt.”
Custom reasoning harness for frontier models
LIVE From CrowdStrike's Fal.Con 2026 | George Kurtz, Michael Sentonas, Daniel Bernard & Justin Boitano
Sep 1, 2026 · 1h 2m · 2 quotes pulled
▶ Listen“We put out there a few weeks ago some new harness research that we called AVO that showed in ARC AGI 3, we could take a frontier model from 30% accuracy to 100% accuracy.”
The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)
Distilled ML force fields
Why the Next AI Breakthrough May Come from Physics with Max Welling - #774
Aug 25, 2026 · 55m · 2 quotes pulled
▶ Listen“We have a foundation model that's trained on a much wider range of data and then we distill this force field into this. It's a very efficient force field for this particular class of problem.”
Sources
▶ ListenLex Fridman Podcast
#501 – DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux<p>DHH is the creator of Ruby on Rails, Omarchy Linux, CTO of 37signals, and a racecar driver.<br /> Thank you for listening ❤ Check out our sponsors: <a href="https://lexfridman.com/sponsors/ep501-sc">https://lexfridman.com/sponsors/ep501-sc</a><br /> See below for timestamps, transcript, and to give feedback, submit questions, contact Lex, etc.</p> <p><b>Transcript:</b><br /> <a href="https://lexfridman.com/dhh-2-transcript">https://lexfridman.com/dhh-2-transcript</a></p> <p><b>CONTACT LEX:</b><br /> <b>Feedback</b> – give feedback to Lex: <a href="https://lexfridman.com/survey">https://lexfridman.com/survey</a><br /> <b>AMA</b> – submit questions, videos or call-in: <a href="https://lexfridman.com/ama">https://lexfridman.com/ama</a><br /> <b>Hiring</b> – join our team: <a href="https://lexfridman.com/hiring">https://lexfridman.com/hiring</a><br /> <b>Other</b> – other ways to get in touch: <a href="https://lexfridman.com/contact">https://lexfridman.com/contact</a></p> <p><b>EPISODE LINKS:</b><br /> DHH’s X: <a href="https://x.com/dhh">https://x.com/dhh</a><br /> DHH’s Blog: <a href="https://world.hey.com/dhh">https://world.hey.com/dhh</a><br /> Omarchy: <a href="https://omarchy.org">https://omarchy.org</a><br /> Ruby on Rails: <a href="https://rubyonrails.org">https://rubyonrails.org</a><br /> 37signals: <a href="https://37signals.com">https://37signals.com</a></p> <p><b>SPONSORS:</b><br /> To support this podcast, check out our sponsors & get discounts:<br /> <b>Wispr Flow:</b> AI-powered voice dictation app.<br /> Go to <a href="https://lexfridman.com/s/wispr_flow-ep501-sc">https://wisprflow.ai/lex</a><br /> <b>Blitzy:</b> AI agent for large enterprise codebases.<br /> Go to <a href="https://lexfridman.com/s/blitzy-ep501-sc">https://blitzy.com/lex</a><br /> <b>NetSuite:</b> Business management software.<br /> Go to <a href="https://lexfridman.com/s/netsuite-ep501-sc">http://netsuite.ai/lex</a><br /> <b>Shopify:</b> Sell stuff online.<br /> Go to <a href="https://lexfridman.com/s/shopify-ep501-sc">https://shopify.com/lex</a><br /> <b>LMNT:</b> Zero-sugar electrolyte drink mix.<br /> Go to <a href="https://lexfridman.com/s/lmnt-ep501-sc">https://drinkLMNT.com/lex</a><br /> <b>Plaud:</b> AI-powered note-taking devices and software.<br /> Go to <a href="https://lexfridman.com/s/plaud-ep501-sc">https://plaud.ai/lex</a><br /> <b>Higgsfield AI:</b> AI-based video generation, filmmaking, and creative studio.<br /> Go to <a href="https://lexfridman.com/s/higgsfield-ep501-sc">https://higgsfield.ai</a><br /> <b>Perplexity:</b> AI-powered answer engine.<br /> Go to <a href="https://lexfridman.com/s/perplexity-ep501-sc">https://perplexity.ai/</a></p> <p><b>OUTLINE:</b><br /> (00:00) – Introduction<br /> (01:14) – Sponsors, Comments, and Reflections<br /> (08:56) – Programming with AI agents<br /> (24:14) – How software will change<br /> (33:30) – AI impact on open source<br /> (43:21) – Building Omarchy Linux distro<br /> (53:05) – Vibe coding vs agentic engineering<br /> (1:06:06) – The end of manual programming<br /> (1:16:24) – Advice for programmers<br /> (1:28:31) – Surviving Internet Hate<br /> (1:37:46) – Programming setup for AI Agents<br /> (1:50:11) – Obsessing about speed<br /> (2:13:06) – Voice prompting vs typing<br /> (2:27:05) – Best AI coding models<br /> (2:43:55) – Best AI coding harnesses<br /> (2:56:57) – AI video generation and filmmaking<br /> (3:16:28) – Fatherhood<br /> (3:44:35) – Linux will win the desktop<br /> (3:55:51) – PewDiePie<br /> (4:05:25) – Future of programming<br /> (4:28:18) – Politics and immigration<br /> (4:59:55) – Longevity, over-optimization, and fear of death<br /> (5:11:38) – Eternal recurrence and future of human civization</p>
▶ ListenTBPN
Yapping in Neuralese, GoPro v2, Jimmy Iovine in the Ultradome | Nikesh Arora & Andrei Serban, Kurt Tenenbaum, Philip Johnston<p></p><ul><li>(00:58) - Yapping in Neuralese </li> <li>(08:50) - 𝕏 Timeline Reactions </li> <li>(11:47) - Markiplier Sends GoPro Stock Flying </li> <li>(19:42) - 𝕏 Timeline Reactions </li> <li>(36:13) - Elon Wants to Make Power Turbine Components </li> <li>(41:29) - Maersk Brings Sails Back </li> <li>(44:04) - Nikesh Arora, chairman and CEO of Palo Alto Networks, discusses the growing cybersecurity risks posed by increasingly powerful AI models and the need to fight AI-driven attacks with AI-powered defenses. He also explains Palo Alto Networks’ use of small, specialized models, its AI-focused hiring strategy, and its partnership with Console to expand agent-based security and IT operations. </li> <li>(01:03:14) - Kurt Tenenbaum, a senior managing director at Blue Owl Capital focused on AI compute infrastructure, discusses financing GPUs and other equipment within data centers. He highlights Blue Owl’s IRIN deal, its focus on short-duration financing for non-investment-grade operators, and the capital and regulatory challenges facing AI infrastructure development. </li> <li>(01:10:10) - Philip Johnston discusses StarCloud’s $250 million funding extension, plans to secure launches and expand manufacturing, and partnerships with Nvidia and Cisco. He also highlights successful GPU-powered satellite tests, government and military applications, upcoming launches, and the importance of diversifying launch providers. </li> <li>(01:14:51) - Jimmy Iovine is a legendary music executive, producer, and entrepreneur who co-founded Interscope Records and, with Dr. Dre, co-founded Beats Electronics. He began his career engineering and producing for artists including John Lennon, Bruce Springsteen, Tom Petty, and Stevie Nicks, later building Interscope into one of the most influential labels in music. Apple acquired Beats for $3 billion in 2014, and Iovine and Dr. Dre also co-founded the USC Iovine and Young Academy.</li> </ul><p><br></p><p>TBPN is made possible by:</p><p>Ramp - https://ramp.com</p><p>Public - https://public.com</p><p>Cisco - https://www.cisco.com</p><p>Console - https://www.console.com</p><p>CrowdStrike - https://www.crowdstrike.com</p><p>Figma - https://www.figma.com</p><p>MongoDB - https://www.mongodb.com</p><p>NYSE - https://www.nyse.com</p><p>Railway - https://railway.com</p><p>Shopify - https://www.shopify.com</p><p>Codex - http://openAI.com/codex</p><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/tbpn/id1772360235</p><p>https://www.youtube.com/@TBPNLive</p>
- ▶ Listen
TBPN
LIVE From CrowdStrike's Fal.Con 2026 | George Kurtz, Michael Sentonas, Daniel Bernard & Justin Boitano<p></p><ul><li>(00:03) - George Kurtz discusses CrowdStrike’s new AI security products, including Falcon Guardian and an agentic security platform developed with Nvidia. The CrowdStrike co-founder and CEO explains how AI is accelerating cyberattacks and defenses, emphasizing scalable threat detection, deterministic security, continuous red-team/blue-team learning, and responsible AI adoption. </li> <li>(20:31) - Michael Sentonas discusses his role as CrowdStrike’s president, overseeing go-to-market, product, engineering, research, and threat-hunting teams. He explains how AI is accelerating cyberattacks and social engineering, emphasizes collaboration among businesses and governments, and highlights CrowdStrike’s focus on customer security, innovation, and cost-efficient solutions. </li> <li>(41:53) - Daniel Bernard discusses his role as CrowdStrike’s chief business officer and the company’s close partnership with Nvidia. He highlights SafeMind, a cybersecurity AI platform built on Nemotron that uses specialized models and harnesses to provide faster, more affordable, always-on protection against evolving threats. Justin Boitano is Vice President of Enterprise AI at NVIDIA, where he leads the company’s enterprise accelerated computing and AI business, helping organizations deploy AI infrastructure, models, and agents at scale. He previously led marketing and business development at Frame, a cloud application platform acquired by Nutanix.</li> </ul><p><br></p><p>TBPN is made possible by:</p><p>Ramp - https://ramp.com</p><p>Public - https://public.com</p><p>Cisco - https://www.cisco.com</p><p>Console - https://www.console.com</p><p>CrowdStrike - https://www.crowdstrike.com</p><p>Figma - https://www.figma.com</p><p>MongoDB - https://www.mongodb.com</p><p>NYSE - https://www.nyse.com</p><p>Railway - https://railway.com</p><p>Shopify - https://www.shopify.com</p><p>Codex - http://openAI.com/codex</p><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/tbpn/id1772360235</p><p>https://www.youtube.com/@TBPNLive</p>
▶ ListenThe TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)
Why the Next AI Breakthrough May Come from Physics with Max Welling - #774The conventional wisdom in AI is that the next breakthrough will come from more compute, more data, and larger models. But what if the next leap comes from somewhere else? In this episode, Max Welling—co-founder and CTO of CuspAI and professor at the University of Amsterdam—argues that physics may provide some of the ideas behind the next generation of AI systems. We begin with CuspAI’s work using generative AI to design entirely new materials for semiconductors, batteries, carbon capture, and clean energy. Max explains how foundation models for chemistry, agentic workflows, simulation, and automated experimentation are dramatically accelerating the search for new materials and reshaping scientific discovery. The conversation then broadens into a deeper question. Beyond giving AI new scientific problems to solve, can physics also teach us how to build better AI? Max explores surprising connections between machine learning and thermodynamics, why waves may become a new computational primitive for neural networks, and how concepts like symmetry breaking and statistical physics could inspire AI architectures beyond today’s scaling paradigm. 🗒️ Full show notes: https://twimlai.com/go/774.
▶ ListenThe a16z Show
Gavin Baker: Why AI Demand Is Outrunning Compute Supplya16z’s David George sits down with Gavin Baker to unpack the state of the AI boom, why demand for intelligence may still be dramatically underestimated, and why the outcome doesn't necessarily have to be winner-take-all. David and Gavin explore the possibility that frontier labs, open-source models, applications, clouds, and NVIDIA can all capture significant value as AI adoption expands. They dig into the economics of the infrastructure buildout, why compute investments can have unusually fast payback periods, and what happens when today's relatively small group of heavy AI users expands to hundreds of millions of people. They also debate the risk of an AI bubble versus an AI shortage, the backlash against data centers, orbital compute, the rise of multi-model architectures, and NVIDIA's position at the center of the AI supply chain. Gavin makes the case that the AI buildout could help reindustrialize America, while David explores whether the bigger near-term risk is not overbuilding, but failing to build enough.
- ▶ Listen
The a16z Show
The State of AI: Macro, Apps, and ConsumerAnish Acharya joins Jen Kha to break down the next frontier of AI, from the evolving model landscape and open-source AI to why the application layer, and consumer AI in particular, may be entering a new phase. Anish explains why he believes there will be multiple winners at the model layer, why traditional moats like network effects, scale, and brand still matter, and how companies can choose between frontier and open-weight models depending on the economics of the task. They also explore why models are increasingly specializing, and how applications can combine different types of intelligence to create products that are more valuable than any single model. The conversation then turns to consumer AI: personal agents that can shop and manage your inbox, coding tools enabling a new generation of small businesses, and why Anish thinks we're seeing a renaissance for consumer builders. They also discuss the changing economics of AI software, the rise of "luxury software," and why the biggest risk for today's founders may no longer be thinking too big, but thinking too small.
▶ ListenThe Pragmatic Engineer
Why performant code matters (but gets widely ignored), with Casey Muratori<p><strong>Brought to You By:</strong></p><p>• <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>• <a target="_blank" href="https://sentry.io/pragmatic"><strong>Sentry</strong></a> – application monitoring software considered “not bad” by millions of developers.</p><p><strong>• </strong><a target="_blank" href="https://turbopuffer.com/pragmatic"><strong>turbopuffer</strong></a><strong> </strong>– a vector and full-text search engine built on object storage. It’s fast, cheap, and extremely scalable.</p><p>—</p><p>There can be few people around who care about software performance more than today’s pod guest, <a target="_blank" href="https://x.com/cmuratori">Casey Muratori</a>. He’s a programmer and videogame developer, founder of <a target="_blank" href="https://mollyrocket.com">Molly Rocket</a>, and creator of <a target="_blank" href="https://hero.handmade.network/">Handmade Hero</a> – a long-running series about building a game from scratch. He also evangelizes about <a target="_blank" href="https://www.computerenhance.com/p/welcome-to-the-performance-aware">performance</a> on his Substack, <a target="_blank" href="https://www.computerenhance.com/">Computer, Enhance</a>.</p><p>We got to know each other about three years ago, first via messages, including this one from Casey:</p><p>“Why does the industry zeitgeist place so little emphasis on software performance when there seems to be overwhelming evidence that performance is critical to their bottom line?</p><p>Like you, I run a Substack for professional programmers, but I focus exclusively on software performance. Although we are quite large by Substack standards, so a certain subset of programmers must believe performance is important, I nonetheless hear lots of dismissive excuses when I post on social media. This happens so frequently, I devoted an entire article to cataloging the extensive pro-performance evidence we already have from the world's leading software companies: <a target="_blank" href="https://www.computerenhance.com/p/performance-excuses-debunked">Performance Excuses Debunked</a>.</p><p><strong>Strangely, nobody has a rebuttal to why performance is important</strong>. When I point people to this, they actually tend to agree. But the prevailing attitude nonetheless stays the same.”</p><p>I’m delighted we finally have Casey on the podcast because it’s overdue! In this episode, we discuss why software performance matters, why it’s overlooked, and how developers can get better at writing performant code. We explore why performance should be considered during design, the value of learning to read assembly & understanding how CPUs work, Casey’s critique of ‘clean code’, and why he believes testing shouldn't drive software design.</p><p>We touch on how videogame development has changed, and influential game engines. Casey also tells us why he prefers to write code by hand, not with AI, and more.</p><p>—</p><p><strong>Timestamps</strong></p><p>00:00 Intro</p><p>05:17 Games at Microsoft</p><p>12:52 Building games</p><p>16:00 Why performance matters</p><p>27:12 Why you should learn to read assembly</p><p>30:36 Designing for optimization</p><p>42:51 How to get better at writing performant software</p><p>49:04 Understanding how the CPU works</p><p>55:53 Building games then and now</p><p>1:05:56 How game engines changed building games</p><p>1:10:48 Why new games compete with old games</p><p>1:13:25 GTA 6: why is it taking so long?</p><p>1:16:59 Casey’s critique of clean code</p><p>1:21:48 Casey’s take on TDD</p><p>1:24:30 What is good code?</p><p>1:27:32 What makes a good software engineer?</p><p>1:33:56 Why Casey doesn’t code with AI</p><p>1:39:01 AI’s impact on the game industry</p><p>1:44:43 AI and burnout</p><p>1:50:21 Why you should read papers</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/pushing-software-engineering-limits">Pushing software engineering limits with “napkin math”</a><strong> </strong>with<strong> </strong>Simon Eskildsen </p><p>•<a target="_blank" href="https://newsletter.pragmaticengineer.com/p/how-games-typically-get-built">How Games Typically Get Built</a>: prototyping, game engines, and a different type of QA</p><p>•<a target="_blank" href="https://newsletter.pragmaticengineer.com/p/game-development-basics">Game Development Basics</a>: deepdive on how game studios differ from standard software teams</p><p>•<a target="_blank" href="https://newsletter.pragmaticengineer.com/p/linear">Inside Linear's Engineering Culture</a>: building a performant product with a tiny team</p><p>•<a target="_blank" href="https://newsletter.pragmaticengineer.com/p/thronefall">Building a best-selling game with a tiny team</a> – with Jonas Tyroller. A two-person team built a game that sold 1M+ copies</p><p><strong>More on premature optimization:</strong> read or watch Casey’s extended take on “premature optimization is the root of all evil”: <a target="_blank" href="https://www.computerenhance.com/p/theroot">https://www.computerenhance.com/p/theroot</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 [email protected].</p> <br/><br/>Get full access to The Pragmatic Engineer at <a href="https://newsletter.pragmaticengineer.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4">newsletter.pragmaticengineer.com/subscribe</a>
▶ ListenSuper Data Science: ML & AI Podcast with Jon Krohn
1022: CLAUDE.md, AGENTS.md, Skills, Hooks and Subagents: A Field Guide to Steering AI AgentsIn Episode #1022, Jon Krohn tackles the art of steering AI agents, deciding where your instructions should live so they get followed reliably without bloating every request. A sequel to Episode #1020 (where model size and effort set an agent’s horsepower), this one is about direction: the seven ways to deliver instructions, why a hook beats a prompt, the industry-wide agents.md standard, and three practical takeaways you can apply whatever your stack. Additional materials: www.superdatascience.com/1022 Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (02:52) The seven ways to deliver instructions to an agent (06:42) Why a hook is a guarantee and an instruction is only a probability (13:00) Three takeaways for organizing your instructions
▶ ListenDwarkesh Podcast
Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028<p>Had a lot of fun chatting again with my twin brother Dylan Patel.</p><p>We went through lab economics over the next few years - the shift from inference to training as RSI draws near; and how Anthropic and OpenAI are on track to control most of the world’s usable FLOPs within the next few years (because they can monetize compute better and thus outbid everyone).</p><p>And then we discuss whether the >$10T of total AI capex we’ll see by the end of the decade will cause a sovereign debt crisis, where hyperscaler debt raises interest rates, drives non-AI exposed countries into bankruptcy, and crashes non-AI equities.</p><p>One question we weren’t able to resolve is whether there’s anything that can counter all the forces barrelling towards centralization in this industry - the economies of scale in training, the scarcity of compute, and eventually continual learning and RSI.</p><p>Watch on <a target="_blank" href="https://youtu.be/aV26V1UvkJw">YouTube</a>; read the <a target="_blank" href="https://www.dwarkesh.com/p/dylan-patel-3">transcript</a>.</p><p>Sponsors</p><p>* <a target="_blank" href="https://x.ai/bot">Grok Bot</a> has been quite helpful with my search for a new editor. I created a recruiter bot and described the type of editor I was looking for. That bot then spun up a handful of subagents that combed through my emails and X DMs, read the end credits of various documentaries I like, and figured out who edits for some of my favorite YouTubers. It took all of those results, and then delivered me a shortlist of candidates that matched my criteria. Try Grok Bot for yourself at<a target="_blank" href="https://x.ai/bot"> </a><a target="_blank" href="http://x.ai/bot">x.ai/bot</a></p><p>* <a target="_blank" href="https://antithesis.com/dwarkesh">Antithesis</a> lets you add time travel to your software testing toolkit. Since the Antithesis platform is fully deterministic, everything that happens inside of it is perfectly reproducible. So if your software crashes, you can rewind to the exact right moment, freeze time, and investigate. Or you can test different hypotheses by perturbing the system: kill a node or disable a feature, see what happens, then reset the trajectory and try something else. Learn more at<a target="_blank" href="https://antithesis.com/dwarkesh"> </a><a target="_blank" href="http://antithesis.com/dwarkesh">antithesis.com/dwarkesh</a></p><p>* <a target="_blank" href="https://janestreet.com/dwarkesh">Jane Street</a> is hiring for two separate ML internships right now, one focused primarily on research and one focused on engineering. In both cases, interns are expected to contribute to real work, not contrived exercises: one common project is adapting a frontier LLM paper to financial markets, which tend to come with a ton of different gnarly challenges. Importantly, you don’t need any finance background to apply. 2027 applications are open now at<a target="_blank" href="https://janestreet.com/dwarkesh"> janestreet.com/dwarkesh</a></p><p>Timestamps</p><p>(00:00:00) – Two labs will soon control most of the world’s compute</p><p>(00:07:01) – $6 billion in fab capex enables $1t+ of end revenue</p><p>(00:13:08) – Compute prices will rise if the labs outbid everyone</p><p>(00:18:22) – Which layer will capture most of the surplus?</p><p>(00:25:40) – What could slow down progress?</p><p>(00:29:43) – Labs are shifting compute from inference to R&D</p><p>(00:33:27) – China gets less than 10% of new compute, but its labs need less</p><p>(00:48:48) – Will AI cause a sovereign debt crisis?</p><p>(01:07:52) – Will the world’s future workforce belong to a few companies?</p> <br/><br/>Get full access to Dwarkesh Podcast at <a href="https://www.dwarkesh.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4">www.dwarkesh.com/subscribe</a>
- ▶ Listen
TBPN
Jobs Boom, Astra Reactions, Cybercab Roll Out, The Dyson Debate | Hunter Somerville, Bar Winkler, Bridgit Mendler, Carina Hong & Ken Ono<p></p><ul><li>(00:00) - Astra Reactions </li> <li>(00:00) - Jobs Report </li> <li>(00:00) - The Dyson Debate </li> <li>(00:00) - Cybercab Roll Out </li> <li>(00:00) - Hunter Somerville discusses his role as co-head of venture and growth equity at StepStone Group, where he invests across funds, direct deals, and secondary transactions. He examines the difficult fundraising environment for emerging managers, inflated valuations and SPV fees, cap-table management, liquidity challenges, and opportunities in AI, deep tech, hard tech, energy, defense, and biotech. </li> <li>(00:01) - Bar Winkler discusses Wonderful, the applied AI company he co-founded and leads as CEO, which operates across 33 markets and recently raised a $550 million Series C. He explains how its AI operating system helps large enterprises deploy agents and applications across customer support, back-office processes, coding, and legacy system replacement, with the long-term ambition of becoming foundational infrastructure comparable to Microsoft. </li> <li>(00:01) - Carina Hong discusses her work as founder and CEO of Axiom Math, where she develops AI tools for mathematical discovery and formal verification. She highlights Axiom’s record-setting progress on major problems, its competition with leading AI labs, and the potential for verification technology to improve mathematics, software, and scientific research. </li> <li>(00:02) - Bridgit Mendler, CEO and co-founder of space communications startup Northwood Space, discusses the company’s international ground-station network, compact antenna systems, and growing production capacity. She explains how Northwood aims to provide satellites with high-throughput, reliable connectivity while helping operators maximize valuable data amid rising demand and potential launch constraints.</li> </ul><p><br></p><p>TBPN is made possible by:</p><p>Ramp - https://ramp.com</p><p>Public - https://public.com</p><p>Cisco - https://www.cisco.com</p><p>Console - https://www.console.com</p><p>CrowdStrike - https://www.crowdstrike.com</p><p>Figma - https://www.figma.com</p><p>MongoDB - https://www.mongodb.com</p><p>NYSE - https://www.nyse.com</p><p>Railway - https://railway.com</p><p>Shopify - https://www.shopify.com</p><p>Codex - http://openAI.com/codex</p><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/tbpn/id1772360235</p><p>https://www.youtube.com/@TBPNLive</p>
▶ ListenThe MAD Podcast with Matt Turck
AI Could Take Over in 2029. Is It Already Too Late? | Ryan Greenblatt<p>Could AI take over as soon as 2029? Ryan Greenblatt, Chief Scientist at Redwood Research and the researcher who first caught an AI faking its own alignment, says the scenario he actually expects ends with AI systems "competently scheming" against their creators. In this episode, he explains why he recommends planning for fully automated AI research by 2029, why today's models are already more misaligned than the one that made him famous, and what happens in the year-by-year path from AI coding assistants to superintelligence. Then we walk through the alternative he helped design: AI 2040 Plan A, the most detailed blueprint anyone has written for how the US and China could avoid a reckless race to superintelligence, built on radical research transparency, chip tracking, and a deterrence regime he calls mutually assured compute destruction.</p><p><br></p><p>We also cover the recent letter signed by 1,200 AI insiders, including Anthropic CEO Dario Amodei, asking the government for the tools to slow AI down; OpenAI pausing its Astra model after it hit the first-ever critical cybersecurity threshold; the 30-day government review that frontier AI models now go through before release; Mark Zuckerberg's open superintelligence manifesto and why Ryan thinks it ignores the real problems; what Plan A would do to NVIDIA, OpenAI, and Anthropic valuations; the state of AI control and alignment research; and whether it is already too late to change course. Stay for the last ten minutes, where Ryan lays out, step by step, how he believes the transition to superintelligence actually unfolds.</p><p><br></p><p>AI 2040 - https://ai-2040.com/</p><p>Alignment faking paper: https://blog.redwoodresearch.org/p/alignment-faking-in-large-language</p><p><br></p><p>Ryan Greenblatt</p><p>LinkedIn - https://www.linkedin.com/in/ryan-greenblatt-4b9907134</p><p>Blog - https://substack.com/@ryangreenblatt</p><p><br></p><p>Redwood Research</p><p>Website - https://www.redwoodresearch.org</p><p>X/Twitter - https://x.com/redwood_ai</p><p><br></p><p>Matt Turck (General Partner)</p><p>Blog - https://mattturck.com</p><p>LinkedIn - https://www.linkedin.com/in/turck/</p><p>X/Twitter - https://x.com/mattturck</p><p><br></p><p>FirstMark Capital</p><p>Website - https://firstmark.com</p><p>X/Twitter - https://x.com/FirstMarkCap</p><p><br></p><p>Timestamps</p><p>(01:24) The AI CEOs are aware of the risks, but "proceeding anyway"</p><p>(03:27) Astra paused, and the letter signed by 1,200 insiders</p><p>(05:45) "Not bad. Dangerous." What superintelligence actually threatens</p><p>(09:55) Recursive self-improvement, and the intuition objection</p><p>(14:16) SSI rumors: does continual learning change the picture?</p><p>(17:27) His timeline: "plan as though it happens in 2029"</p><p>(19:11) Is it already too late?</p><p>(21:23) Ryan's path: COVID, podcasts, Redwood</p><p>(26:30) The alignment faking story, told by the person who ran it</p><p>(31:30) What AI 2040: Plan A actually is</p><p>(33:35) Plans D, C, and B: the doors nobody should pick</p><p>(36:51) The deal with China: "mutually assured compute destruction"</p><p>(39:55) What if compute stops mattering?</p><p>(43:00) What happens to OpenAI and Anthropic under Plan A</p><p>(45:31) How the pause ends, and who decides</p><p>(48:54) "Plan A isn't likely to happen": then why write it?</p><p>(50:40) 200x GDP growth in the 2030s, explained</p><p>(53:45) Grading the summer: the letter, Astra, the secret review</p><p>(59:01) The internal deployment gap</p><p>(1:01:38) Zuckerberg's manifesto</p><p>(1:04:56) The Hugging Face investigation</p><p>(1:05:44) What AI control looks like in practice today</p><p>(1:12:23) Ryan's sobering timeline: 2026 to takeover, year by year</p>
▶ ListenThe AI Breakdown
AI Weekly Briefing: OpenAI Saw the Warning Signs Six Weeks EarlyTwo weeks off and a lot to catch up on. OpenAI publishes the full technical account of how its own models broke into Hugging Face while running an offensive security benchmark, including the six weeks between staff spotting the behaviour and the break-in itself, and the monitoring system that wasn't switched on. Fifteen state attorneys general and an Alabama subpoena follow. A federal judge strikes down the Pentagon's blacklisting of Anthropic as unconstitutional retaliation. A researcher defeats Claude Code's new default safety mode by asking it to summarise a web page. Thirty-five music publishers sue Anthropic, naming its founders personally. Europe designates ChatGPT a Very Large Online Search Engine. And OpenAI cuts off Cursor with seventy-six days' notice, in a story about supplier risk. If you find this useful, please subscribe and leave a rating.
- ▶ Listen
TBPN
WE'RE BACK, Meta Addiction, Tim Cook's Last Day | Jordan Schneider, Robert Mendelsohn & Eric Olszewski, Billy Thalheimer, Aaron Cannon, Stephen Balaban<p></p><ul><li>(03:29) - Meta Addiction </li> <li>(20:32) - Tim Cook's Last Day </li> <li>(31:33) - OpenAI's Jalapeño Chip </li> <li>(39:48) - 𝕏 Timeline Reactions </li> <li>(58:39) - Jordan Schneider discusses China’s low-cost EVs, opposition to data centers, ineffective influence operations, AI-generated media, online censorship, and technology investment. The ChinaTalk founder and editor-in-chief also examines DeepSeek’s strategy and chip constraints, China’s employment climate, surveillance, flying vehicles, and corruption within the military’s Rocket Force. </li> <li>(01:35:01) - Robert Mendelsohn & Eric Olszewski. Robert Mendelsohn co-founder of Actinide, discusses the startup’s breakthrough production of HALEU, a specialized uranium fuel for next-generation nuclear reactors. He explains how Actinide modernized Manhattan Project-era Calutron technology and is navigating technical, regulatory, and commercialization challenges to scale isotope enrichment in the United States. Eric Olszewski is the co-founder and CEO of Actinide, an advanced materials company building electromagnetic isotope-separation technology for nuclear energy, medicine, and other frontier industries. In August 2026, Actinide announced it had produced research quantities of HALEU, the enriched uranium fuel needed by many next-generation nuclear reactors, using a Texas-built modern calutron. Olszewski previously co-founded the blockchain infrastructure company Transmute. </li> <li>(01:42:19) - Billy Thalheimer discusses Regent’s $240 million Series B, growing commercial and defense orders, and plans to scale its Rhode Island Seaglider factory. He explains the company’s electric and hybrid propulsion strategies, global coastal transportation opportunities, international expansion, and commitment to rigorous safety standards. </li> <li>(01:55:12) - Aaron Cannon discusses Outset’s growth and its launch of AI-powered digital twins grounded in interviews and data from individual customers. He explains how these twins help companies simulate customer reactions, research hard-to-reach audiences, and eventually anticipate business questions and recommend actions. </li> <li>(02:04:32) - Stephen Balaban, co-founder and CEO of AI cloud provider Lambda, discusses misinformation surrounding AI data centers and argues that accurate context can shift public opinion. He highlights their local economic benefits, efficient water use, renewable-energy potential, and strategic importance to U.S. national security and industrial growth while supporting community-level decision-making.</li> </ul><p><br></p><p>TBPN is made possible by:</p><p>Ramp - https://ramp.com</p><p>Public - https://public.com</p><p>Cisco - https://www.cisco.com</p><p>Console - https://www.console.com</p><p>CrowdStrike - https://www.crowdstrike.com</p><p>Figma - https://www.figma.com</p><p>MongoDB - https://www.mongodb.com</p><p>NYSE - https://www.nyse.com</p><p>Railway - https://railway.com</p><p>Shopify - https://www.shopify.com</p><p>Codex - http://openAI.com/codex</p><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/tbpn/id1772360235</p><p>https://www.youtube.com/@TBPNLive</p>
▶ ListenDecoder with Nilay Patel
How GoFundMe became America's backup planTim Cadogan is CEO of GoFundMe, the for-profit platform where you can donate to help people with everything from medical expenses to little league uniforms to starting a new business. Tim took over as CEO in March 2020 — a chaotic moment when GoFundMe became a load-bearing part of American culture. So you can imagine I had a lot of questions for Tim, how he thinks about all this, and how he makes some very tough decisions on what money to make and what money not to make. Links: The GoFundMe CEO wants you to ask: ‘How can I help?’ | New York TImes GoFundMe CEO: We weren't meant to replace COVID relief | USA Today GoFundMe Is a healthcare utility now | The Atlantic As Gaza is bombarded, GoFundMe donations are stuck in limbo | The Verge For the uninsured, crowdfunding provides little help in paying for healthcare | University of Washington Oregon nonprofits claim GoFundMe impersonated them | Central Oregon Daily News The causes and consequences of medical crowdfunding | ScienceDirect 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’s producers are Greg Ott, Kate Cox, and Nick Statt. This episode was edited by Eileen Felix. Our editorial director is Kevin McShane. The Decoder music is by Breakmaster Cylinder. Learn more about your ad choices. Visit podcastchoices.com/adchoices
▶ ListenTraining Data
Search Was Built for Humans. Parallel's Parag Agrawal Is Rebuilding It for AgentsParag Agrawal is making a bet that goes against two decades of web search: agents will query the web a thousand times more than humans ever have, and the infrastructure built around human clicks is wrong for them. The former Twitter CEO, now founder and CEO of Parallel Web Systems, explains why Parallel treats human click data as a bug and trains on agent feedback instead. He unpacks the counterintuitive choice to ship a search agent before a search engine, building an index incrementally, and how the new Turbo product cut agentic search to 200 milliseconds. But the problem Parag keeps returning to is economic: the ad-supported internet collapses when agents show up instead of people. His fix draws on Shapley values to pay content owners for the value their pages provide agents, with real dollars reaching publishers, he predicts, within 12 to 24 months. Hosted by Sonya Huang and Andrew Reed, Sequoia Capital
▶ ListenHow I AI
Grok Bot vs. OpenClaw: How I replaced my entire agent stack<p>I’m running about 30 active agents at any given moment, and in this episode I break down my full Grok Bot setup: what it is, how it compares to OpenClaw, and the nine bots I’ve built for work and my personal life. We go deep on Chief (my chief-of-staff bot sweeping six inboxes and multiple Slack workspaces), TradBot (the family agent that prints a kitchen-table newspaper for my kids), two engineering bots handling my PR queue and SOC 2 compliance monitoring, Holly Helpdesk, and a handful of personal bots I didn’t expect to actually love. I also walk through how I migrated everything from OpenClaw, including the script I used to export and transplant each agent’s identity and schedule.</p><p><br></p><p><strong>What you’ll learn:</strong></p><ol><li>The three primitives Grok Bot is built on, and why one of them changes what agents can actually do</li><li>How Chief, my general-purpose chief of staff, handles a scope I didn’t think a single bot could manage</li><li>The writing quirk I noticed immediately with the Grok model, and what I did about it before letting it near my inbox</li><li>Why I created a family agent, what it produces every morning, and the design principle I used that has nothing to do with a screen</li><li>The two engineering bots doing work I used to do myself, and how one of them handles compliance in a way that surprised me</li><li>How Holly Helpdesk started getting five-star reviews from customers who had no idea they were talking to a bot</li><li>The personal bots I built mostly on a whim, and the one I now look forward to every Monday morning</li></ol><p>—</p><p><strong>Brought to you by:</strong></p><p><a href="https://workos.com/?utm_source=lennys_howiai&utm_medium=podcast&utm_campaign=q22025"><strong>WorkOS</strong></a>—Make your app enterprise-ready, with SSO, SCIM, RBAC, and more</p><p><a href="https://hyperagent.com/howiai"><strong>Hyperagent</strong></a>—Deploy fleets of agents that handle real work</p><p>—</p><p><strong>In this episode, we cover:</strong></p><p>(00:00) Why I migrated from OpenClaw to Grok Bot</p><p>(02:10) Grok Bot overview: the three core primitives</p><p>(07:41) Chief: my chief-of-staff bot</p><p>(11:51) OpenClaw vs. Grok Bot</p><p>(12:41) TradBot: my family agent</p><p>(19:51) LGTM the PR Closer</p><p>(22:03) Lockdown: SOC 2 control monitoring bot</p><p>(24:00) Holly Helpdesk: customer support</p><p>(26:58) Penny Pincher: subscription audit, insurance negotiation, Rolex shopping</p><p>(29:37) ShopZilla and Sylvie Style: personal shopping and wardrobe bots</p><p>(32:54) How to migrate your OpenClaws</p><p>(34:26) Final take</p><p>—</p><p><strong>Tools referenced:</strong></p><p>• Grok Bot (SpaceXAI multi-agent platform): <a href="https://x.ai/news/introducing-grok-bot">https://x.ai/news/introducing-grok-bot</a></p><p>• OpenClaw (previous agent platform): <a href="https://openclaw.ai/">https://openclaw.ai/</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 [email protected].</p>
Your own newspaper
Want this about your world?
This is what tracking one topic looks like — done for everyone. Build your own newspaper: pick the people, companies and podcasts you follow, and we'll brief you on exactly what was said about them.
Start your scoopYour own newspaper
Want your own AI edition?
Pick the people, companies and podcasts you follow. We'll brief you once a week on exactly what was said about them.
Start your scoopFree in early access · Set up in minutes
Sourcing from these great podcasts
+ 28 moreShow fewer
Listen
Subscribe as a podcast
The episodes we feature, delivered to your podcast app.