Fresh From the Labs

In this final episode of Fresh From the Labs, Shilpa, Kevin, and Jared reflect on their journey at Pioneer Square Labs, share personal updates, and discuss insights gained from their experiences. Jared provides valuable perspectives on the challenges and responsibilities of using AI in coding, while the group revisits their Freshies predictions about AI trends and tools, evaluating their accuracy and relevance. The episode concludes with heartfelt farewells and gratitude for their listeners.

In this episode of Fresh From the Labs, the team says farewell to engineer-turned-founder Jared as he sets off to co -found a startup. But don't worry, he will still be hosting FFTL from the other side.

We delve (kl-note: another AI giveaway! 😅) into the rise of AI-generated resumes and the challenges they pose for recruiters and interviewers. The hosts discuss strategies for detecting synthetic candidates, the ethics of using AI tools to cheat in interviews, and the broader implications of AI on authenticity in the hiring process.

Beyond hiring, we talk about the blurred line between human creativity and AI-assisted content creation, exploring how generative tools can scale personal communication and creativity.

The episode wraps up with some spooky Halloween fun as the hosts share personal stories, decorations, and holiday plans.

In this episode we dive into the latest AI news with a playful twist. Anthropic’s new Haiku model hits the scene, promising lightning‑fast responses and powerful capabilities. We unpack what makes Haiku special and how it stacks up.
Then we explore Atlas, the emerging AI platform shaking up the ecosystem, and discuss how its features could reshape the way we interact with technology.
We also take a look at “vibe coding” tools that let you craft software by describing the vibe rather than the code, and we share our impressions of OpenAI’s new browser functionalities.
Tune in for lively banter, thoughtful analysis, and all the latest from the frontier of AI innovation.

(ed. note: Atlas attempted to edit and publish this podcast end to end. I had to fix a few editing mistakes but otherwise it did it! The description...needs some work, but we'll leave it as a example of Atlas's current capabilities.)

Welcome back to a super special episode of Fresh from the Labs! The team is joined by PSL alum Brian Fioca, now a Solutions Architect at OpenAI, for a deep dive into the latest from Dev Day and the strategies shaping the AI frontier.

The conversation kicks off with AgentKit, OpenAI's new visual workflow for building agents, and the accompanying open-source ChatKit. We discuss how these first-party tools aim to solve the chronic pain of building a high-quality chat experience and remove the undifferentiated heavy lifting of creating evaluation harnesses and tracing systems.

This leads to the billion-dollar question: with OpenAI building so much product, where can startups possibly build? Brian offers an insider's perspective on OpenAI's AGI-focused mission, explaining how products are often stepping stones for research and that the biggest opportunities for startups lie in building domain-specific tools, workflows, and the "picks and shovels" infrastructure that OpenAI won't.

We also explore the other massive Dev Day announcement: the ChatGPT App SDK. Drawing parallels to the Apple App Store boom, the team discusses its potential to become a monumental distribution channel for a new generation of AI-native startups. To cap it off, Brian shares fascinating insights from his work on GPT-5, explaining how models develop "habits" during training, the power of watching "reasoning tokens" to understand a model's thought process, and how advanced prompting is evolving into a deep craft.

This week on Fresh from the Labs, the team dives into OpenAI's Sora 2. More than just an AI video model, it's a full-fledged social network. Kevin shares his shockingly good first experience creating a personalized avatar and generating a video of himself in a 90s rock band, highlighting the platform's surprisingly good sense of humor and the fun of remixing content. This leads to a broader discussion on the massive competitive threat this poses to Meta, why taste and UX are now the ultimate battlegrounds, and whether this explains Mark Zuckerberg's recent talent acquisition frenzy.

Next, the conversation turns to Anthropic's new Sonnet 4.5. The team has had more time to test it, and the verdict is in: while it's fast, it's ultimately not as correct as the latest Codex. We discuss the crucial metric of "time to completed correct code," why Anthropic's "victory lap" announcement feels disconnected from the real-world user experience, and whether the company is getting dangerously squeezed in the model wars.

The team also unpacks the evolution of "vibe coding" as tools like Bolt V2 go pro by adding databases, authentication, and first-party agent integrations. Is this a smart move to capture a larger market, or are they alienating their core non-technical audience and wading into a hyper-competitive space?

Finally, we cover ChatGPT's recent e-commerce and personalization updates. The introduction of an instant checkout feature has major implications for product discovery and merchants, while the new "Pulse" feature aims to deliver a more personalized information feed, consolidating ChatGPT's role as a central user interface. Join us for a jam-packed episode on product showdowns, shifting market dynamics, and the future of creative AI.

This week on Fresh From the Labs, the team kicks things off with the resurgence of autonomous driving. With Waymo cars now spotted on the streets of Seattle and Amazon's toaster-like Zoox shuttles rolling out in Vegas, is the long-promised self-driving revolution finally here? We discuss the two-decade journey from Uber's early arbitrage bet to today's tangible progress, debating user trust, safety, and whether we're truly at a tipping point.

The conversation then pivots to one of the most pressing challenges for any AI founder: pricing. We tackle the immense difficulty of estimating and managing AI costs in a world where token prices for top-tier intelligence remain high. The team explores the evolution of pricing models, from prohibitively expensive early days to today's VC-subsidized landscape, and the pitfalls of confusing credit-based systems that obfuscate true costs.

Jared makes the case for pricing based on value delivered rather than micro-transacting tokens, and the team shares practical, hard-won advice for founders. Learn how to create back-of-the-napkin cost estimates, the importance of running real-world user tests, and what not to do—including the dreaded "unlimited" plan and the bad freemium model. Plus, we cover essential guardrails to prevent runaway costs so you don't get a surprise fifty-thousand-dollar bill. Join us for a drive through the future of transportation and a masterclass in AI unit economics.

This week on Fresh from the Labs, Shilpa, Kevin, and Jared dive into the powerful new Codex CLI release. Based on GPT-5 but custom-trained for real-world agentic use cases, this release confirms their prediction that major labs are now building "first-party agents," moving the key primitive up the stack from the model to the agent itself.

Kevin and Jared share their hands-on impressions and reveal the intricate, spec-driven workflows they've developed to harness its power. They detail their process of prompt looping, using GPT-5 Pro to create detailed specs and then feeding them to Codex, and using automated feedback loops ("check your work") to run agents on complex tasks for hours at a time. The discussion offers a masterclass in systematically building guardrails and support systems around agents to achieve highly performant, real-world results.

The conversation then shifts to the state of the market, sparked by Cognition's massive $400 million fundraise at a $10.2 billion valuation. The team voices their skepticism and questions what the VCs are seeing, debating if it's a strategic play or simply a case of "VC FOMO." This leads to a candid discussion about a recent Wall Street Journal article on the "996" AI founder lifestyle, the pressures of the current hype cycle, and whether that relentless pace is truly sustainable or effective.

Join us for a deep dive into advanced AI coding techniques, the state of AI startup valuations, and a reality check on the pressures of building in a bubble.

This week on Fresh From the Labs, Shilpa, Kevin, and Jared are back to break down the latest in AI. The conversation kicks off with a look at Google's delightfully named new image editing tool, Nano Banana. Kevin shares his hands-on experience, highlighting its impressive ability to flawlessly edit and compose images in ways that feel genuinely useful, marking a potential step-change for image models.

The discussion then takes a serious turn as the team explores the growing sophistication of AI-powered security threats. From hyper-targeted phishing emails to the complex "Singularity Supply Chain Attack" on GitHub, we delve into the core vulnerabilities of today's AI systems. The conversation covers prompt injection, the "lethal trifecta" of security risks, and the immense challenge of sandboxing agents that have access to private data and external tools.

Next, we unpack a fascinating and potentially market-shifting development: the Agent Communication Protocol from the team behind the Zed editor. Jared explains how this new standard allows developers to "bring your own agent"—not just their own model—into their editor. This leads to a deep discussion on the vertical integration of agent stacks, the competitive threat this poses to platforms like Cursor, and how this signals a fundamental shift where the "agent is the new primitive."

Finally, the team tackles Sam Altman's recent comments about being in an AI bubble. Is this the dot-com boom all over again, or are the strong revenues and tangible value created by AI companies a sign that this time is different? Join us for a wide-ranging discussion on everything from creative tools to critical security threats and the strategic battles shaping the future of AI.

This week on Fresh from the Labs, Shilpa, Kevin, and Jared kick things off with a conversion: Kevin has officially joined the Rust cult (beard pending). From there, we dive into three big themes shaping how builders actually ship with AI right now:

• GPT-5 in the wild. A launch that felt…bumpy. We unpack autorouter misfires, sudden model deprecations, and why prompting matters more than ever with thinking/verbosity “knobs.” Kevin compares day-to-day coding performance against Anthropic’s Opus 4.1 and Claude Code—great for devs, less magical for non-technical workflows.

• Agentic browsing vs. reality. Perplexity’s eyebrow-raising $35B Chrome bid sparks a broader debate: is the future in a browser or the OS? Jared’s two-week test drive of Comet delivered slick automations (cart-filling errands) but clashed with classic “just let me Google docs” moments, plus awkward multi-account gaps. We talk cryptographic request signing for agents, potential micro-payments to publishers, and why an “MCP upgrade” path could beat brittle click-automation.

• Enterprise truth serum. A new MIT study claims ~95% of corporate GenAI pilots fail. We break down the why: chained-probability error rates, tool-calling flakiness, procurement drag, and shadow usage of public chatbots. The near-term win? Multiplicative copilots in the tools people already live in (think Microsoft Copilot) over moonshot agents.

Plus: glamping on Orcas Island, off-grid backpacking, and squeezing in those late-summer bike rides. Tune in for hard-earned takes on what’s hype, what’s here, and what’s next.

(created with GPT-5...)

We hit record three hours after OpenAI’s GPT‑5 live stream, with a shoutout to former PSL engineer Brian Fioca on stage, and dive straight into the big takeaway: daily‑driver frontier at a non‑frontier price. Benchmarks look great, but vibes and real‑world performance steal the show—agentic and MMLU jumps, faster tokens, a huge context window that keeps long tasks on track, and noticeably fewer hallucinations. The headline isn’t just capability; it’s capability per dollar. If ChatGPT’s ~700M users wake up to a default upgrade that’s cheaper than prior flagships, the experience gap for “most people” collapses overnight.

We go hands‑on with coding flows that actually changed how we work this week. In Cursor, GPT‑5 felt stateful without much prompt fuss—scratchpads, self‑tracking, and fewer guard‑rail tangles. Kevin details a zero‑to‑working‑product sprint: queue a night’s worth of tasks, have the model generate an implementation guide, a to‑do plan, and a smoke‑test checklist per epic, then wake up to a massive PR across a Rust backend and React desktop app. The new loop is “send, sleep, review, retry”—and now it’s fast and cheap enough to do repeatedly.

We also run a pragmatic face‑off. Opus 4.1 is impressive, but slow and pricey. GPT‑5 felt snappier, delivered stronger results on the same tasks, and hit a price point that makes “try again” affordable. That flips the calculus: unless a model is meaningfully better, cost drags utility, and “intelligence per dollar” becomes the metric that matters. For the first time in a while, a frontier model is both top‑tier and broadly economical.

Beyond OpenAI, the release flood was real: a fresh wave of open‑source options you can run locally (pay once for hardware; inference becomes electricity), Google’s Jules CLI assistant hitting GA, and rumblings of a Cursor CLI. Plus a wow‑moment from Google’s Genie 3: real‑time, prompt‑to‑play worlds with minute‑scale coherence and persistence—paint stays on the wall when you return, water splashes look eerily physical, and the “inception” demo hints at interfaces well beyond chat.

Then the Windsurf saga’s latest twist. After the OpenAI deal reportedly cratered over Microsoft data‑access sensitivities, Google scooped leadership and IP, Cognition acquired the remainder and accelerated equity—and then publicly set a “80 hours, six days” cultural bar, offering nine months’ severance for those opting out. We debate radical candor versus needless PR self‑owns, and what this roller coaster signals for employee expectations, acqui‑hire risk, option value, and term‑sheet protections in a tightening talent market.