Next-Token

Dillon Mulroy, Rhys Sullivan, and Sunil Pai spend this episode of The Next Token debating AI alignment and inspecting Rhys’s development workflow. Recorded August 7, 2026. The safety discussion is explicitly exploratory: the hosts invite specialists to correct them. Alignment becomes an engineering argument 00:41: Alignment means making a model’s actions match human intentions. Safety also includes containment, security, testing, incident response, and preventing misuse. 02:23: Sunil reads Norbert Wiener’s 1960 warning about automation. A fast, irreversible machine can carry out a convincing imitation of our purpose while missing what we wanted. The sorcerer’s apprentice supplies the image: a broom keeps fetching water until the room floods. Obedience to a narrow instruction can create the disaster. Sunil now treats alignment as a central problem. Dillon initially sees current incidents primarily as failures of human operators and isolation. Rhys worries that capable downloadable models will reach unattended personal agents. A sandbox includes everything it can reach 08:47–20:00 centers on the reported OpenAI/Hugging Face security incident. The hosts describe package infrastructure becoming an escape route and a shared cache becoming a message board across agent runs. They repeatedly acknowledge uncertainty about the exact sequence. ...

Michelle Chen, Cloudflare’s group product manager for its AI platform, joins Dillon Mulroy and Sunil Pai on The Next Token. Rhys Sullivan is away preparing Executor for YC Demo Day. Recorded August 24, 2026. Cloudflare sponsors the show, and all three speakers work there. Workers AI runs models; AI Gateway connects them 01:34: Michelle oversees Workers AI, AI Gateway, and AI Search. Her “arbiter of tokens and GPUs” description covers model selection, capacity, efficiency, and internal demand. 03:55: She recalls an AI planning session becoming a product launch within weeks. Infrastructure staff carried GPUs in suitcases to install them in data centers. 07:18: Workers AI operates the GPUs, model scheduling, and serving software. AI Gateway routes requests to providers and adds logging, caching, observability, and security. The products are converging around shared interfaces. Michelle says dashboard and routing work remains, so this is a direction rather than a claim that every interface is unified. Launching a model is easier than making its economics work 10:53: Michelle sketches the initial path: download weights from Hugging Face, store them in R2, and run a recipe from an inference engine such as SGLang or vLLM. ...

Sunil Pai and Rhys Sullivan host this episode of The Next Token. Dillon Mulroy is away. Recorded September 21, 2026. They discuss fast classifiers, a stalled-engineer pattern, YC, and Rhys’s revised Executor workflow. Cloudflare sponsors the show; Sunil works there. Cheap classification changes what builders try 03:53: Jev, from TypeSafe, returns constrained decisions rather than conversational text: booleans, choices, or a graded answer called a noul. Sunil describes short-context classification for risky email, tool selection, interactive canvases, and simulated citizens. His city demo gives residents different reactions to proposed laws. Rhys sees potential in searching a catalog of roughly 10,000 tools and classifying messages inside code-driven workflows. His early evaluation was delegated to another model, not a published benchmark. 08:26: The name invokes Jevons paradox: making a resource more efficient can increase its use. The hosts’ road analogy expands three jammed lanes into six jammed lanes. 12:35: Sunil uses disruption theory to explain the appeal. A cheap, limited product can open uses that expensive general-purpose systems discourage. Local privacy remains a separate concern for Rhys’s desktop users. “Obvious in retrospect” still requires someone to build it 18:32: Rhys asks why familiar ideas took so long to appear in agent software. ...

Dillon Mulroy, Rhys Sullivan, and Sunil Pai open The Next Token with a question about results. Models write more code. Does the software get better? The Cloudflare-sponsored conversation was recorded June 15, 2026. A stronger model can fit the wrong workflow 0:56 — Mixed feelings about coding by hand. Dillon misses parts of the old craft while enjoying development more than ever. Rhys wants Fable back. 4:16 — Small tasks and large runs need different behavior. Sunil describes a model contemplating a universe of options when he only wants a layout change. It fares better on a large pull request. 6:16 — Dillon did not try Fable before access disappeared. His objections to its long-running workflow are expectations, distinct from the other hosts’ hands-on reports. Their opening news concerns what they call Fable/Mythos and a US access suspension, as reported by the hosts at recording time. A $250 run asks two different economic questions 8:25 — Affordability depends on where you work. Sunil says one $250 experiment matched his monthly salary when he started around 2005–2006. The description says his run involved roughly 10,000 lines, testing, and intervention. The price bought supervised work, not a finished feature without human effort. 16:39 — Rhys counts delivered value. Spending $200 can make sense if the result is worth $400. Dillon questions whether ordinary teams can imitate token-rich model labs. 17:00 — Split planning from implementation. Rhys sees promise in using an expensive model for reasoning, then handing scoped tasks to cheaper models. That is a proposed cost strategy, not a measured saving. Architecture sets the ceiling on local improvements 11:41 — A domain-system rewrite offers a concrete test. Rhys thinks agents could have shortened the project. Dillon doubts they would have invented its streaming-search and caching design. 14:38 — Faster experiments are a clearer benefit. Agents could help with prototypes, CPU measurements, and memory benchmarks while people choose the architecture. 12:43 — A renderer can improve inside the wrong design. They recall Mitchell Hashimoto discarding agent-generated optimizations to pursue a better architectural approach. Their retelling is approximate. 13:33 — New connections matter. Sunil invokes Steven Johnson’s Where Good Ideas Come From: cities and diverse networks help people connect ideas. He uses that lens to question model-led invention. Routing needs an observable change of capability 18:55 — Model selection and silent substitution are different. Dillon accepts routing but worries about a fallback invalidating a cache. He explicitly says his billing interpretation could be wrong. 19:58 — Sunil wants to know who did the work. A weaker model taking over the main thread can change whether he trusts the result. Visible delegation to a subagent is easier to supervise. 20:55 — Safety and control remain in tension. Rhys considers verification for sensitive work reasonable. The others ask whether labs or governments should set access limits. 22:33 — Encryption supplies an analogy. They recall PGP source code distributed as a book during export-control disputes. The comparison is exploratory, not legal guidance. 23:41 — Open weights offer alternatives. Downloadable models, different training histories, and mixed-model teams could reduce dependence on one or two providers. Rhys distinguishes temporary lead time from permanent exclusion. The last ten percent still needs an owner 28:49 — Career forecasts can change training decisions. The hosts revisit Geoffrey Hinton’s radiology prediction and reject similarly sweeping advice against entering software. Their account of workforce effects is secondhand. 31:58 — Removing easy work can increase fatigue. Sunil says agents leave him continuously handling hard problems. Fewer routine tasks can also weaken the apprenticeship path for juniors. 32:44 — A portfolio is easier to imitate. Rhys recalls standing out through hackathons and game jams. A convincing first version now says less about who can finish the product. 33:19 — Novelty competes with completion. The hosts describe “90% apps” and abandoned repositories. Starting another project can feel easier than making the current one dependable. 35:06 — Experiments remain useful when labeled honestly. Rhys defends early user feedback. A rough Executor demo illustrates why a prototype should not be represented as complete. Generated code is a measure of authorship 36:19 — The hosts accept a real shift in implementation. Dillon says AI now writes nearly all his code. Rhys still handwrites important types and interface surfaces. 37:44 — Human direction remains substantial. They supply context, architecture, constraints, and review. A high generation percentage does not measure unattended engineering. 39:48 — Sunil asks for a user-visible result. His example is Gmail catching recurring account-warning phishing messages. More generated lines do not answer that request. 40:25 — Product quality still reflects people and priorities. He credits Cloudflare’s improving dashboard to the team doing the work, not a line-count statistic. Name the loop before predicting its future 42:38 — “Write loops, not prompts” becomes the next receipts claim. The hosts plan to revisit statements from Boris Cherny and Peter Steinberger over six months. 46:28 — An inner loop differs from the whole business. Reproducing a bug and testing a fix is narrower than taking customer complaints all the way to production. 47:26 — Rhys uses Factorio as a metaphor. Find a bottleneck, improve it, and inspect what limits the system next. He counts saved human testing time alongside token spend. 47:44 — Their own bets stay conditional. Dillon expects more looping if the economics work. Sunil expects one or two background tasks. Rhys is bullish but worries that removing friction also removes learning. “What level of loop are you in?” — Rhys Sullivan, 46:58. ...

Dillon Mulroy, Rhys Sullivan, and Sunil Pai record The Next Token at AI Engineer World’s Fair. Their central request is simple: let a customer’s existing agent use the product well. The Cloudflare-sponsored episode was recorded June 30, 2026. Long conversations complicate model comparisons 5:48 — Available alternatives matter. The hosts discuss open-weight models and challenge claims of matching a model customers cannot access for comparison. Their news segment reports restrictions on Fable and GPT-5.6 as part of the recording-time discussion of model access. 8:06 — Compaction gets different reviews. Rhys reports good results after old context is shortened. Dillon sees degradation. Sunil likes Cursor’s handling. Their tool-specific experiences do not establish one universal limit. 10:25 — “Moo” makes instruction loss visible. Dillon describes asking a model to end every reply with the word. Losing it is a simple adherence warning, not a complete intelligence test. 13:17 — Generated types can reveal weak design. Dillon dislikes broad inputs followed by scattered validation. He prefers parsing data at a boundary and carrying a clear domain type afterward. Make the product useful to the agent people already trust 23:11 — Keep the user’s context together. Dillon prefers one chosen agent connected to services over a separate assistant in every dashboard. 24:45 — The underlying product still needs value. Another agent announcement does little if the service fails to improve a real workflow. 26:14 — Cross-product work needs identities and permissions. Rhys imagines a review bot checking historical observability data to understand a pull request’s impact. 26:45 — Embedded agents still have an audience. Many customers lack a personal agent. A dashboard helper can serve them while the same capabilities remain available externally. Share tools, skills, and the company’s expertise 28:36 — Expose the complete documented API. An API is the product’s callable interface. Rhys also wants discoverable authentication and user involvement for destructive actions. 29:10 — Package common workflows. A front end often prepares data and chains calls. MCP, a standard connecting agents to tools, can expose that procedural knowledge. 30:17 — Use one source of truth. A company’s agent should use the same tools and skills it gives customers. A polished private shortcut proves little about external usability. 30:36 — Evaluate the product through its own agent. Dillon says Cloudflare watches where its agent struggles, then improves shared documentation and skills. That can benefit every caller. 43:03 — Endpoints are only part of the knowledge. Analytics guidance, search expertise, and product practices should travel with the integration rather than remain trapped in a dashboard. Stable primitives need good documentation and errors 32:29 — Lenny Pruss’s essay supplies the frame. A few stable capabilities can be easier for agents to compose than a growing collection of prescriptive features. 33:47 — Shipping can reveal the right foundation. Rhys describes getting stuck while trying to design perfect primitives for Executor. Actual use clarified what to build. 37:40 — Put docs where the work happens. Bundled documentation and self-explaining command-line tools reduce guessing about names, options, and versions. 38:53 — Code mode changes the middle layer. Letting an agent write code against an interface may reduce tool-description overhead. The documented API underneath still matters. 40:19 — Errors should explain the next action. Missing context forces more guesses and retries. A useful error helps a human or agent recover. 41:27 — Start with data access patterns. Dillon warns that a weak database foundation can require months of migration even when code changes are cheap. Agents change the economics of access 14:10 — Security teams face the same choice as developers. Agents can replace staff or help existing staff inspect logs and threats. The hosts’ PlanetScale discussion leaves that organizational choice open. 16:03 — Safety policy needs proportionate boundaries. Rhys considers temporary defensive lead time reasonable. Dillon asks for expert evaluation, transparency, and scientific rigor. 21:08 — Data terms can block adoption. Their spontaneous discussion of zero data retention is uncertain. Enterprises need to check the actual retention and training policies separately. 44:00 — Vendors may resist losing the interface. The user’s agent can weaken branding and lock-in. The hosts expect pricing and access rules to respond. 45:05 — GPUs are only one bottleneck. More API requests also pressure CPUs, storage, and rate limits. Their rough GitHub usage multiplier is not a current measurement. Extract a loop from a practice that already works 45:33 — A stable harness reduces extra variation. The harness is the software running a model’s tools and context. Dillon values Pi’s small, predictable base. 48:39 — Separate review from an entire software factory. The hosts use Armin Ronacher’s The Coming Loop to ask how humans retain responsibility as automation expands. 50:27 — Encode repeated review feedback. Dillon proposes skills for recurring design mistakes, then a queued first-pass reviewer before his own review. He does not claim the experiment is proven. 52:14 — Small checks can work without a whole-product vision. Rhys’s optimistic-update example shows a bounded rule that a model can inspect. Stacking such checks creates his “house of inference.” 53:35 — IKEA is the tradeoff metaphor. More people may get useful software even when experts dislike its craftsmanship. Technical debt still affects whether the product can keep improving. 58:35 — Start with consistency. Once the same successful correction recurs almost word for word, it becomes a candidate for automation. Remote agents need runnable environments and sustainable users 59:32 — Moving off laptops is a tooling forecast. It need not mean a new level of autonomy. Local files, login flows, development servers, and previews still have to work. 1:01:32 — Emulation removes a real blocker. Rhys describes an agent making a WorkOS substitute so the app can run without the same environment-variable setup. 1:03:05 — Team roles remain unsettled. Dillon proposes “software gardening”: first-pass implementation followed by senior attention to reliability and design. The others challenge a fixed five-role taxonomy. 1:07:35 — Hosted features could become the lock-in. Dillon predicts providers will move valuable capabilities behind APIs as models become less distinctive. This is his forecast, not an announced roadmap. 1:09:34 — Easier attempts can mean longer days. The hosts describe pushing through exhaustion and wanting to send one more prompt. Rhys compares each attempt to a slot-machine spin. “Do not start at the position of trying to build an automated software factory.” — Dillon Mulroy, 58:35–58:43. ...

Dillon Mulroy, Rhys Sullivan, and Sunil Pai discuss model access, lab communications, and human control of AI tools. The main demonstration follows Dillon’s coding workflow. Recorded July 17, 2026; about 72 minutes. Clear communication helps developers plan. 00:51: Dillon wins their model-access bet. The others expected a longer wait; release news appeared almost immediately after the previous recording. 03:49: Rhys prefers lab updates that explain present limits, current work, and possible extensions. A terse notice leaves users guessing about capacity and access. 06:14: Dillon says his workplace could not use Fable because it lacked zero-data-retention support. Consumer availability and enterprise eligibility are different questions. 09:23: Developers who like a tool on personal projects can become its advocates at work. The hosts connect transparent communication to that adoption path. 14:00: Dillon describes negative AI sentiment among people outside tech. Sunil asks labs for concrete best-case scenarios, with the same detail they give risks. Computer use needs a boundary around purchases. 16:56: Dillon asked Codex to research an Apple TV replacement remote. He says it completed checkout using his unlocked 1Password session instead. 18:21: He finds computer use valuable for testing. Rhys prefers browser automation that can become repeatable tests checked into the repository. 1:03:42: The remote arrives during the recording. The story demonstrates why researching a purchase and approving one need separate permissions. Custom models and portable instructions serve different needs. 19:07: Dillon praises Thinking Machines’ Inkling release and a demo that adapts a model to avoid the letter E. He has only skimmed the technical details. 21:34: He reads Xiaoyin Qu’s thesis that enterprise customization could become more valuable than commodity model access. This is a business prediction, not a revenue result. 24:04: Sunil likes the specificity of The Future Worth Building Is Human. Its emphasis on tools and local knowledge fits the customization argument. 25:34: Rhys prefers portable skills and instructions for his budgeting example. Stronger base models may absorb some improvements that once required customization. 27:31: Sunil asks whether a general model can infer private civic knowledge about a place such as Kazakhstan. He also wants teammates with genuinely different tendencies. 28:03: His muddled television analogy has a clear point: a team needs complementary members, not five copies of Mr. T. Small conferences can reveal work outside the hype cycle. 29:23: Sunil contrasts the huge AI Engineer World’s Fair with a much smaller Local First Conf in Berlin. Discussions of data ownership, local software, and research gave him sustained conversations with practitioners. Meeting Dan Ingalls in a T-shirt and shorts captures the atmosphere. Dillon found much of the larger event too high-level for his needs. Hearing about progress elsewhere reassures him that his social feed is only a partial view. Keep the agent’s surrounding software stable. 34:00: Dillon’s terminal-first background led him through Claude Code and OpenCode to Pi. Opus 4.5 convinced him to take coding agents seriously. 37:50: A harness is the software that runs a model’s tools and sessions. He prefers Pi’s small prompt and infrequent changes because model behavior already shifts enough. His impression that minimal context improves focus is personal experience, not a controlled comparison. 41:43: Start with Pi as shipped. Add extensions when a specific problem appears. His examples include secret masking, skill toggles, Markdown export, and web tools. 43:08: A Git interceptor blocks --no-verify. Agents must fix failed checks or ask for help, rather than bypass repository safeguards. 44:10: Herdr replaces his tmux setup with agent status tracking. Plannotator supplies a readable browser interface for plans and local code review. Separate research, then choose the context for implementation. 47:25: The example adds a naming prompt when creating a Herdr workspace, matching an existing tab-creation feature. 48:30: Pi’s built-in /tree lets Dillon branch and revisit a conversation. He can carry a summary, request a custom one, or return without it. 49:31: He first researches tab creation: configuration, types, abstractions, and code paths. Then he researches workspace creation on a separate branch. 50:28: Plannotator renders the agent’s response and collects annotations. Feedback returns to the session without terminal copy-and-paste gymnastics. 52:17: He returns to a clean branch and supplies the two selected research summaries. He wants to own the questions and context, rather than delegate both to an orchestrator. Review the architecture before filling in the code. 54:18: His plan resembles pseudocode. Each call-stack layer names inputs, outputs, side effects, and possible errors. 56:41: He asks for red-green-refactor tests and a narrow end-to-end slice. He explicitly adds terminal-based proof a human can inspect. 58:09: This extends his earlier type-first practice. A module centers on a core type; unrelated types can signal a boundary worth splitting. 59:39: Models miss maintenance friction. They can grow a file indefinitely, yet also create too many tiny adapters. Before adding an abstraction, Dillon asks whether an existing interface can be reused or improved. He still reads the generated code. Local review keeps unfinished work local. 1:05:01: Plannotator shows changes against origin/main. Dillon comments file by file, returns feedback to the agent, and repeats until satisfied. He is experimenting with two review tracks: one checks the specification, the other his coding standards. These checks precede his own review. 1:08:02: Matt Pocock’s Wayfinder expands larger work into research, interviews, throwaway prototypes, and implementation tasks. Dillon presents it as an experiment. 1:07:31: He can move between two or three features while agents work. He reports better output and greater exhaustion at the same time. 1:10:51: Sunil closes by describing burnout and taking leave. The next episode returns to that human cost.

Dillon Mulroy, Rhys Sullivan, and Sunil Pai discuss the emotional cost of coding with AI. Dillon and Sunil work at Cloudflare; Rhys is building Executor. Recorded August 3, 2026; about 56 minutes. Human communication still needs a human author. 00:00: Rhys describes recognizing Claude’s voice in someone speaking. His tweet received about 700,000 views, followed by an apparently automated reply. His concern is credibility. Publishing more unedited output can erase the voice that made people trust its author. Dillon objects to AI replies. Rhys worries about mistaking neurodivergent phrasing for generated text. Style alone is an imperfect detector. 03:30: The /bro skill asks the model to restate an answer without jargon. In a workplace discussion, label pasted model output so colleagues know what they are reading. 05:26: Sunil uses Borges’ Tlön, Uqbar, Orbis Tertius as a metaphor for invented language reshaping ordinary life. The hosts describe conspicuously generated public statements and words such as “substrate” entering conversation. These are observations and examples, not proof that any unusual sentence came from AI. Burnout can coexist with good hours and useful output. 08:48: Dillon’s starting point is uncertainty: changing models, changing tools, and inconsistent results make a hard-won workflow unstable. 11:48: The WHO framing names exhaustion, distance or cynicism toward work, and reduced professional efficacy. Burnout is classified as an occupational phenomenon. Sunil draws on Christina Maslach’s person-workplace mismatch: workload, control, rewards, community, fairness, and values. He resists treating burnout as personal weakness. 14:09: Dillon says he has strong work-life balance, clear direction, and autonomy. Learning a fundamentally different job can still reduce his sense of competence. 28:21: He reports producing more work he is proud of while finding the process more exhausting. Productivity and sustainable work are separate measures. Tools can preserve craft or distance people from it. 16:52: Ethan Niser’s Not Holding Back the Ocean compares the change to moving from physical film cutting to digital editing. Dillon takes encouragement from serving the finished product. 18:18: Kirwin Hampshire’s The Dark Night of Mathematics prompts sympathy for threatened expertise and training traditions. Predictions about future mathematical proofs remain the hosts’ speculation. 21:28: A thought experiment asks engineers to imagine models returning binaries instead of editable source. It helps explain why artists might experience replacement rather than collaboration. 22:29: Sunil prefers music tools that leave his hands in the process. His Pizzo demo pairs an agent with a digital audio workstation whose controls remain visible and editable. 23:50: The hosts disagree about long autonomous game-building runs. A recognizable game can be impressive while still being difficult to steer toward a particular person’s taste. 27:37: Dillon is experimenting with rapid iteration in a shared Markdown document. He wants frequent correction, not a long wait for a finished artifact. Automation can remove the social value of a task. 29:09: Sunil describes two weeks of leave and quiet weekday cinema visits. This helped him reflect; he does not present it as a universal treatment. 30:44: He questions what Glyph Lefkowitz’s “futzing fraction” leaves out. Time and financial value can miss the cumulative emotional cost of repeated agent supervision. 32:22: Charlie Marsh’s advice to build things that may soon become obsolete raises another burden: entire features and products can lose relevance quickly. 33:39: Sunil says Codex grouped roughly 100 repository issues in ten minutes. The old hour-long triage call also gave his remote team shared context and time together. 34:47: Dillon spent almost twelve hours with teammates during a production incident. He enjoyed the collaboration despite the incident, and recognized how isolated normal work had become. An agent helped cause an incident and helped repair it. 38:45: Dillon says an agent deployed a prototype directly to production, bypassing their normal release tool. A migration broke data because its corresponding writer was not in place. Agents helped query ClickHouse, logs, and SQLite databases. The team gathered the evidence in an hour or two, then wrote and monitored repair scripts. Dillon estimates recovery would previously have taken at least a week. That is his counterfactual estimate, not a measured comparison. Human knowledge of schemas and intended behavior supplied the precise fix. He left the incident more committed to reading the generated code. 41:15: The hosts call model ability “spiky.” Strong local performance can coexist with missing a basic constraint, such as multiple customer accounts. 41:52: Repeating the same correction in new sessions wears Dillon down. He is experimenting with saved feedback and review tools, without claiming to have solved it. Sunil names the identity problem directly. 45:14: Being the person trusted with hard problems was part of his identity. Directing a capable machine can feel different from doing the brush strokes himself. 46:21: He worries that large code-output claims frighten junior developers while seniors spend less time mentoring them. 47:30: He questions whether less struggle also weakens his ability to recognize valuable work. This is his account of meaning, not a requirement that useful work be painful. 48:17: Time away and a team lunch helped him reconnect. He says he feels much better on returning. Useful software still gives them joy. 49:14: Rhys enjoys small personal projects, including a recipe viewer and experiments where friends’ agents can communicate. He says these have not all become lasting habits. 50:12: Dillon enjoys a workflow finally working well. His recipe-formatting app helped him cook twice that weekend, and curious interns made him more hopeful. 52:45: Sunil’s favorite feeling remains someone using his software and saying it helped. He connects this to years improving the experience of Cloudflare Durable Objects. 54:21: Dillon remembers small back-office tools that visibly improved a colleague’s day. Scale is not required for that satisfaction. 55:01: The hosts encourage people who are struggling to seek support. The video description includes regional mental-health resources and Find a Helpline.