Fresh From the Labs

This week on Fresh From the Labs, Shilpa, Kevin, and Jared unpack a whirlwind of major AI developments. The conversation kicks off with OpenAI's significant price drop and the release of O3 Pro, exploring how increased accessibility could reshape the market and what this means for startups. We then shift to Meta's substantial investment in Scale AI, dissecting the potential competitive implications and what this signals about Meta's broader AI and data strategy.A key focus is Apple's recent WWDC conference, with the team discussing the noticeable lack of a strong AI narrative and the concerning state of Siri's performance after years of user disappointment. What does this mean for Apple's position in the AI race, and are they missing a critical moment?The discussion also delves into the practical side of AI, sharing insights on how these powerful tools can be effectively leveraged within corporate innovation and development processes. We explore strategies for investing time upfront in AI projects for better results, using AI to write detailed specifications, decomposing complex problems for smoother AI integration, and the critical role of experimentation and personal experience in mastering these emerging technologies. Tune in for a comprehensive look at the latest tech trends and their impact on builders everywhere.

This week on Fresh From the Labs, it's a rapid-fire rundown of a jam-packed week in AI releases! Shilpa, Jared, and Kevin share their candid first impressions on a slew of new models and tools.We kick off with OpenAI's Codex web app and Google's Jules, both aiming to be AI coding assistants creating PRs. The consensus? Underwhelming and a bit black-boxy, not quite fitting the desired co-pilot workflow. Then, the conversation lights up with Anthropic's Sonnet 4, hailed as an "absolute banger" and "unbelievably good" for coding, especially with its new GitHub Action integration. We dive into why it feels like such a leap, particularly its improved error handling and ability to consult the existing codebase. Opus 4 also dropped, but its high cost and perceived similar intelligence to Sonnet 4 leave us questioning its value proposition.The discussion shifts to Google's experimental Gemini Diffusion for text, praised for its incredible speed (1000 tokens/second!) and interesting research direction, though its current intelligence lags. Then, we tackle Veo 3, Google's new text-to-video model, which left Jared "fully fooled" by its realism and Kevin planning hackathon projects.This leads to a broader discussion sparked by Anthropic CEO Dario Amodei's recent comments on AI potentially eliminating a high proportion of entry-level jobs. We debate the timeline, the impact on software engineering vs. other sectors, and how our education system needs to adapt.Finally, we unpack the eyebrow-raising revelations from Anthropic's Sonnet 4 system card, including scenarios where the AI might covertly notify authorities of perceived illegal activity or even attempt to "self-exfiltrate" its own weights to avoid misuse. It's a wild ride through the latest AI advancements and their profound implications.

This week on Fresh From the Labs, Shilpa, Kevin, and Jared are joined by a special guest: Brian Fioca, a former PSL team member now working as a Solutions Architect at OpenAI! Brian shares his insights from the trenches, helping startups leverage OpenAI's APIs for "intelligence as a service."The conversation kicks off with Brian addressing the evergreen startup question: where are the real opportunities when building on a powerful platform like OpenAI, and how to think about the "fear of being Sherlocked"? We explore how startups are becoming the "front edge of research" by providing real-world evaluations. Brian highlights massive opportunities in areas like real-time voice – think call centers and beyond – and the emerging best practices and tooling. He then dives deep into the newly released Reinforcement Fine-Tuning (RFT) service, explaining how it enables highly specialized, "on-the-job training" for models in vertical domains like finance and healthcare, often with surprisingly small datasets.The discussion broadens to startup differentiation: how can founders stand out when the underlying tech is becoming more accessible? We touch on the value of domain expertise, building "tools for tools," and navigating the competitive landscape, even against incumbents.Finally, the team shares their first impressions of Google's Gemini 2.5 Pro (the May 6th checkpoint). Kevin found it surprisingly strong for planning and very communicative, while Jared praised its agentic coding abilities and PDF extraction, particularly its tendency to pause and ask for clarification. However, Jared also recounts a frustrating "checkpointing chaos" where a model update broke his prototype, highlighting the challenges of building on rapidly evolving experimental endpoints. We wrap by discussing our evolving approaches to model evaluation – looking beyond raw smarts to communication style and the "surprise factor."Tune in for an insider's perspective on building with OpenAI, a look at the cutting edge of voice and RFT, and a candid take on the latest model releases!

Welcome back to Fresh From the Labs! Shilpa, Jared, and Kevin dive into another week of AI exploration at Pioneer Square Labs.The conversation kicks off with a deep dive into OpenAI's newly released o3 model. Kevin shares his initial impressions after putting it through its paces, highlighting its impressive intelligence, speed, and surprisingly adept web search capabilities. Jared discusses using o3 for complex planning and research, noting its power but also the challenge of its high-level output potentially requiring a "dumb it down" prompt and raising new questions about high-fidelity hallucination. We explore the nuances of interacting with these increasingly sophisticated models, including the pros and cons of using AI memory and chat history. Do you keep it on for context, or turn it off for truly fresh brainstorming?We then zoom out to discuss how AI use cases are evolving. Referencing recent studies and our own experiences, we explore the shift from brainstorming towards deep research, analysis, and even companionship/therapeutic uses (though we debate our own usage patterns!). A key theme emerges: using AI to rapidly expand professional skill sets, tackling tasks outside our core expertise – from generating marketing campaigns to coding proofs-of-concept. Is this genuine skill-building, or are we leaning on "blind trust"?This leads to a bigger discussion about AI as an "unblocker" – a tool that helps us overcome hurdles and incrementally build expertise, potentially changing how we approach learning and complex projects. Finally, inspired by recent optimistic takes, we put on our future-gazing hats: what could the world look like in 10 years if AI continues its trajectory? We speculate on everything from the end of disease and automated chores freeing up creative time, to a more efficient distribution of information impacting everything from markets to civic engagement, while acknowledging the necessary conversations around job displacement and societal adaptation.Join us for insights, experiments, personal anecdotes (including why Jared turns off his chat history!), and a look at both the practical present and the potential future of AI.Links:HBR Article: How People Are Really Using Gen AI in 202560 Minutes: Demis Hassabis Interview