#142 Why LLMs Need Their Own Programming Language: From Assembly to AI with Vaibhav Gupta // Co-founder @ BAML

#142 Why LLMs Need Their Own Programming Language: From Assembly to AI with Vaibhav Gupta // Co-founder @ BAML

From HoloLens assembly to AI: Vaibhav Gupta on why LLMs need their own language, and how BAML makes them type-safe and shippable at agent speed. BROUGHT TO YOU BY: Blocks [https://blocks.cloud/alphalist?utm_source=alphalist&utm_medium=podcast&utm_campaign=blocks-podcast-2026] A decade building computer vision and writing assembly at Microsoft (HoloLens), Google, and D.E. Shaw, then a from-scratch bet on a programming language built for LLMs. Vaibhav Gupta joins Tobi to explain why probabilistic compute needs its own tooling, what "shipping at agent speed" actually requires, and why the world's appetite for software is mathematically infinite. Chapters: 00:00:00 — Intro 00:01:00 — From HoloLens to D.E. Shaw: a decade in computer vision and assembly 00:02:00 — Starting from scratch, and the YC pivot ("500K to not build a Slack competitor") 00:04:00 — Falling in love with coding — and bricking a few machines along the way 00:06:00 — His first AI moment: the GPT-3.5 wake-up call 00:09:00 — Code as a means to an end — why 90% of the job is plumbing 00:11:00 — Why he decided to build a language: first principles and BAML 00:13:00 — LLMs as a new compute primitive 00:15:00 — What BAML actually is — embedded, type-safe, callable from any language 00:17:00 — The business model and the "data trench" 00:19:00 — When AI ships code you didn't ask for — and why CI/CD breaks in an agent loop 00:22:00 — Live demo: function versioning and locking the codebase 00:24:00 — Why no one else competes here — Protobuf, Thrift, and Google's playbook 00:29:00 — Getting started: the BAML "hello world" (live coding) 00:32:00 — Everything is a function — type safety that runs in Rust 00:36:00 — Shipping at agent speed = trust plus granular control 00:38:00 — Visualizing code instead of reading it 00:44:00 — Where to start with BAML (docs.boundaryml.com → Agents MD) 00:45:00 — The mathematically infinite appetite for software 00:47:00 — Why we'll have 10x more builders — and why "English isn't a programming language" 00:51:00 — The future of SaaS: PaaS, harnesses, and customer-defined models 00:56:00 — Will design matter more in an agent world? 00:59:00 — The "time travel" decorator: advice to his 2017 self 01:04:00 — Outro Quotes: 00:13:00 — "BAML's a new thing that exists because we have a new compute primitive in the form of LLMs." (verbatim) 00:10:00 — "90% of software engineering — well, in most jobs, 100% of software engineering is plumbing… And AI just takes that 90%, just makes it go away." 00:19:00 — "We can now generate code at machine speed… But we still cannot ship code at machine speed." 00:49:00 — "English can't go down to assembly… And the minute you add that to English, what have you done? You've built a new programming language." 00:45:00 — "I don't think we've yet found a company that hasn't found that I can make more money if I write more code." -- OUR SPONSORS Blocks [https://blocks.cloud/alphalist?utm_source=alphalist&utm_medium=podcast&utm_campaign=blocks-podcast-2026] Save at least 20% on your AWS costs with AI-powered optimization and enterprise discounts. Get your free Cloud Check at blocks.cloud/alphalist.

#141 AI Pat Works Here Now: Why Agents Must Follow Human Rules with Pat Casey // CTO @ ServiceNow

ServiceNow's CTO of 20 years explains why the safest way to deploy AI agents is to treat them like employees, same rules, same approvers, same spending limits, and why AI is reshuffling the deck on who your best engineers are. Intro How do you go from installing software off floppy disks to running engineering for a $13B revenue company — without ever losing the fish? Pat Casey, CTO of ServiceNow and its first engineer after founder Fred Luddy, joins Tobi to talk 20 years of scale, the architecture behind 90,000 databases, and why enterprise AI agents should be treated exactly like slightly untrustworthy employees. Key topics * Pat's nerd path: Atari 400, Wizardry, and building Adobe's first employee tracking system in Microsoft Access * Flipping off Fred Luddy in traffic — and becoming ServiceNow employee-after-one * The stuffed-fish code-ownership system and why productivity dips at ~100 engineers * Inside the architecture: Java metadata engine, hacked Rhino, K8s services, single-tenant clusters * RaptorDB: from MariaDB board seat to buying Swarm64 and forking Postgres * Gen 3 of AI coding: Windsurf vs Claude Code, the 15% average vs the 5x outliers * The five-chessboards theory of AI-native engineering, and Pat's daughter's vibe-coding conversion * "AI Pat": agents in the user table, following human rules * Hybrid pricing: seats for humans, consumption for AI * Is the market wrong about SaaS incumbents? Pat's answer to the Anthropic scare * What Pat would whisper to his younger self (spoiler: it's about Jelly XML — and family) Chapters [00:00:51] Intro: who is Pat Casey [00:01:43] Nerd origin story: Atari 400, tape storage, Wizardry [00:04:12] First job at Aldus/Adobe — accidental ITIL [00:06:12] Flipping off Fred Luddy → joining Glide in 2005 [00:08:28] From 2 job titles to 10,000 engineers — the introvert advantage [00:10:21] The stuffed fish, and the productivity trough at 100 engineers [00:13:44] Architecture: metadata engine, Java monolith, Kubernetes [00:16:41] The single-tenant bet: 90,000 databases, 25B queries/hour [00:21:32] MariaDB, Monty, and building RaptorDB from Swarm64 [00:27:21] Postgres fork, OpenJDK contributions, open-sourcing Raptor? [00:29:35] AI coding gen 1–3: Copilot → 7,000 Windsurf licenses → Claude Code [00:33:35] Five boards of chess: who clicks with AI coding (and who doesn't) [00:36:47] Pat's daughter and the vibe-coding conversion [00:38:21] Enterprise agents: from toolkits to outcomes [00:40:41] "AI Pat": agents that follow human rules [00:42:29] Pricing: seats for humans, consumption for AI [00:46:15] The Anthropic scare, SaaS valuations, and the incumbent advantage [00:53:06] The new bottleneck: not engineering, not product — customers [00:59:00] Pat's advice to CTOs: lean in, don't turf it [01:01:26] Time machine: what Pat would whisper to his 2005 self Quotes [00:41:30] "You should not trust an LLM more than you trust a human being. Modern business processes were designed on the assumption that human beings are a little bit untrustworthy." (fillers removed — verify against audio) [00:34:00] "AI coding, if you get to that next level, is like playing five boards of chess, 'cause you got multiple prompts spinning at the same time." (one "um" removed) [00:16:04] "If it was that easy, none of the world's big monolithic code bases would still exist." (condensed from "It's like, all right, if it was that easy, like, none of…" — verify) [01:01:26] "This is not a time for excessive caution. It's not a time to completely go bonkers and do crazy stuff, but this is a time really to lean into the new technology." (one "uh" removed)

#140 From Stripe's Fifth Engineer to Serving Millions of Developers with Anurag Goel // Founder & CEO @ Render Goel

Anurag Goel was Stripe's fifth engineer before he built Render into a platform millions of developers deploy on. Here's his contrarian read on agents, security, and why "the AI cloud" is the wrong thing to be. Show Notes Anurag Goel joined Stripe as its fifth engineer in 2011 and later ran risk. He left to solve a big problem and landed on the one he'd watched eat Stripe's engineering time: making infrastructure disappear. This conversation is about what Render learned on the way to millions of developers and what changes now that a lot of what gets deployed isn't a website, it's an agent. * Key topics:* From Stripe's fifth engineer to founding Render The "application cloud" vs. "AI cloud" positioning Agents as long-running, stateful applications for a new end user Workflows, sandboxes, and the consolidated AI runtime Executive hiring and reference calls as a growth hack Security: minimizing blast radius, short-lived scoped keys Distribution in the chatbot era (GEO) and why Google is underrated Observability is the real bottleneck for production agents Timestamps [00:00:00] Intro and the pitch [00:02:00] Origin story: ebook search engine, game rentals, and the first-ever Stripe payment [00:04:00] Joining Stripe as engineer #5; talent density [00:06:00] Raising the hiring bar, no warm bodies [00:11:00] Executive hiring and reference calls as a growth hack [00:13:00] Why he started Render: ~20% of Stripe's engineers stuck on AWS [00:16:00] Agents as a new kind of application [00:17:00] "We're the application cloud, not the AI cloud" [00:18:00] Workflows, sandboxes, and the consolidated AI runtime [00:24:00] Heroku's decline and the exploding sales pipeline [00:25:00] The agentic moment: when adoption spiked [00:33:00] Security and blast radius [00:50:00] Why SaaS isn't dying specialization [00:58:00] Distribution: from SEO to GEO [01:02:00] Why Google is underrated [01:03:00] Advice for CTOs going all in on agents [01:06:00] Easter egg: a whisper to his 2011 self

#139 Your Future Job Is a Decision Inbox — Max Deichmann Built the Layer That Gets You There // Co-Founder @ Langfuse

Max Deichmann built Langfuse — the open-source LLM engineering platform acquired by ClickHouse — and explains why the engineer of the future isn't writing code, they're reviewing what agents did overnight. Max Deichmann didn't set out to build the observability layer for the AI era. He started with mobile apps, taught himself to code via Harvard's CS50, and ended up in Y Combinator with a SaaS product he wasn't excited about. Then ChatGPT launched, and on a Sunday night at 10 pm, his co-founder asked: "If you just had time, what would you build?" The answer became Langfuse and eventually led to an acquisition by ClickHouse. This episode is a rare, grounded conversation about what building and operating AI agents actually looks like in 2025, from the engineering loop to the 3 am incident, to what the engineer's job becomes when agents are doing most of the execution. Key topics: * Why LLM applications broke traditional observability tools, and what Langfuse does instead * The pre-production → production → evaluation → iteration loop for agent development * Open source as a trust and adoption strategy for dev tools * The ClickHouse acquisition: why they sold, what the half-page doc said, and how it's going * Agentic incident response: copy-pasting alerts into Codex at 3 am, and what comes next * The "decision inbox" engineers are reviewers and decision-makers, not coders * The real state of agents in production: what's working, what's not, and what LinkedIn gets wrong Timestamps: [00:00:00] Intro & guest welcome [00:02:00] Max's nerd origin story CS50 on a beach in Singapore [00:04:00] Why they pivoted to Langfuse: firing customers mid-YC batch [00:06:00] Building the first AI products and discovering the observability gap [00:07:00] What Langfuse actually does: the LLM engineering platform explained [00:09:00] Tracking business AND infrastructure metrics billing via Langfuse [00:10:00] Open source from day one: trust, adoption, and hardening the product [00:13:00] Go-to-market with 1.5 salespeople: how engineers sell to enterprises [00:14:00] The acquisition story: 5 engineers, 40TB/day, and a Series A that became a sale [00:17:00] What it felt like when half the AI ecosystem knocked on their door [00:18:00] Life inside ClickHouse: cultural fit, Tokyo offsite, and what surprised them [00:20:00] Agentic coding in practice: velocity per engineer, what still needs a human [00:22:00] The planning loop: Claude summarising GitHub discussions, RFC → agent → review [00:23:00] The "decision inbox" model: engineers as taste-makers and reviewers [00:27:00] How to build an observability stack for the agentic era from scratch [00:29:00] Agentic on-call: the 3 am Codex workflow and what's coming next [00:32:00] Where Langfuse fits vs. traditional observability agent quality vs. infra health [00:35:00] The real state of agents in production: the non-LinkedIn version Best quotes: "We didn't initially jump on the topic because we thought all the PhD AI people, they are much better at this. We have no idea what's going on, until we figured out nobody has a clue what's going on." — [00:05:00–00:06:00] "We have two guys doing customer support, and we have basically an agent that is doing first-level customer support for us, and I think it's about doing about 10,000 conversations a week. We would never be able to do this type of support with two people." — [00:38:00] "The alert comes in, I wake up at the night, I just take the alert from our Slack, copy paste it into Codex, and we have a skill there with all the context, and then it's just going." — [00:29:00–00:30:00] "I currently think of an email/Linear inbox where an agent tells me, 'Hey Max, we needed to fix this here because this broke.' And then if I want to, I can just dive into it and see all the context within this notification and also take a corrective course, or I just let it go." — [00:41:00]

#138 From Hacker News to W3C: How One Amazon Engineer Accidentally Shaped the Future of AI Browsers // Alex Nahas, MCP-B

How a browser-based fix for an enterprise auth problem became a W3C web standard and what it means for how AI agents will interact with the web. Alex Nahas, founder of MCP-B and initiator of the WebMCP web standard, joins Tobias to explore one of the most underappreciated shifts happening in AI: the browser as the primary runtime for agentic systems. Key topics covered: * What MCP actually is: an RPC framework for calling tools across processes, not the complex protocol it's made out to be * The OAuth problem: why MCP's push towards OAuth locked out most enterprise infrastructure still running on SAML * The WebMCP solution: running an MCP server in client-side JavaScript so agents can use the browser's existing auth context * How a Hacker News post posted under anesthesia got 400 upvotes and caught the attention of Google and Microsoft * Chrome 146 natively supports WebMCP, and what that means for adoption * The chicken-and-egg problem: why website owners won't add WebMCP support until clients support it, and vice versa * Agent identity: why agents don't need their own credentials and can operate as a subset of the user's identity * Real-time bidding for agents: the emerging market where advertisers bid to inject results into agent contexts - THE AGENTIC WEB IN TWO YEARS: HEADLESS BROWSERS, INTENT-BASED INTERFACES, AND AGENTS THAT ONLY SURFACE THE UI YOU NEED. [~04:30] "The browser itself is like the perfect sandbox we've been iterating on for so long now." — Alex Nahas [~06:30] "MCP is just an RPC framework. It's super simple. Basically just a wrapper around API documentation." — Alex Nahas [~13:00] "My first memory coming back was me arguing with people on Hacker News who didn't understand it." — Alex Nahas [~16:30] "Agents don't need their own identity. They can have an identity that's like a subset of the user who spun them off." — Alex Nahas [~32:30] "This reminds him of the dawn of programming — where everyone was just doing things and nobody really knew what they were doing, but people were just trying to figure things out." — Alex Nahas [~43:00] "Believe in yourself." — Alex Nahas