Nick Frosst

🚀 Interns in the AI era operate differently. This summer at Adopt, we hired four engineering interns — and they’ve blown us away. They ramped up in 2 days, started shipping by day 3, and since then have: 🛠️ Written production code 🎨 Designed user experiences 📦 Taken product to customers 💬 Even closed a few deals! They’re 19. And they’re not just keeping up — they’re pushing the pace. We hired them in a week. Why? Because they took initiative before we even spoke. Researched our product. Built prototypes. Showed us results. We just had to say yes. AI has been a massive leveler. With the right intent, anyone can learn fast, build fast, and deliver fast. Founders — strongly consider hiring interns. It’s good for the ecosystem, yes. But it’ll surprise you how much value they can drive. They don’t need to unlearn — they’re native to this new world. They use Cursor to code. Build prototypes on Lovable. Automate anything repeatable. They’re not just the future. They’re your unfair advantage — today. 📸 (3 of our 4 summer interns in the pic) #AI #Engineering #AdoptAI

180

6 Comments

Shawn (swyx) Wang wrote last week: “RIP Vibe Coding. [Feb–Oct 2025]” I don’t think it’s dead. I think it’s evolving. > #VibeCoding was great for simple greenfield projects and proof-of-concepts. > #ViableCoding is the next step, making AI useful for real, heavy-duty codebases. In my talk at AI Eng World's Fair 2025, I shared posts from industry leaders exploring this shift: How to move from vibe to viable by • using AI across the entire SDLC (e.g. AI code review, testing, RCA) • tightening integrations across diverse, orthogonal tools (you’ll see more and more of these) • and leveraging a strong context engine to make it all work accurately (critical).

291

15 Comments

The WhyHow.AI team is excited to announce PatientSeek, the first open-source fine-tuned DeepSeek reasoning model for the MED-LEGAL space, designed to run securely and privately on local systems, and trained on one of the largest accessible datasets of patient records. It is purpose-built for MED-LEGAL workflows, focusing on disease and diagnosis identification and correlation reasoning—critical tasks that require the intersection of healthcare and legal expertise. 🛠️ Why PatientSeek? In MED-LEGAL workflows, practitioners often need to make connections between patient records and external events that stand up to legal scrutiny as to correlation and causation. With PatientSeek, these complex tasks become faster, more reliable, and more secure. Key features include: ✅ Human-level comprehension in MED-LEGAL reasoning ✅ Offline, private, and secure operation ✅ Competitive performance at a fraction of the cost of other models ⚙️ How We Built It: Leveraged DeepSeek R1 for reasoning Used Comet for dataset storage and tracking Powered by Amazon Web Services (AWS) Sagemaker (thanks to NVIDIA Inception credits!) Fine-tuned with Unsloth AI and Hugging Face TRL. 💻 Try PatientSeek Today! Download the model here: https://lnkd.in/eA9EF7Uj Get setup instructions here: https://lnkd.in/eqg9iwha Chris Rec Thomas Smoker https://lnkd.in/eXRjxm2e

161

19 Comments

Over the past few months, I’ve been growing a YouTube channel explaining deep technical AI topics in a rigorous but accessible way (RL techniques like PPO and GRPO, knowledge distillation, quantization, etc.). It’s grown to ~20k subscribers, with strong engagement from engineers and ML practitioners. More recently, I’ve started hearing from CTOs and founders who are sharing my content internally to upskill their teams. That inspired me to formalize it into an offering: 👉 Custom AI training for companies 🔵 Workshops for engineering teams new to AI/ML 🔵 Short courses for ML engineers staying on top of new research 🔵 Tailored curriculum based on your org’s product or research domain If your team is investing in AI literacy or applied ML skills, I'd love to hear what you'd find most valuable. YouTube channel: https://lnkd.in/e4PYG5ry

133

15 Comments

Most folks know I’m passionate about AI, but my roots in computational biology run deep. In the 90s, I worked at a Toronto-based startup called Visible Genetics, building DNA sequencing technology with an amazing team. We dreamed of a future where biology and computation would come together to transform medicine. That dream is becoming real. RFdiffusion2 just launched, and it’s a major leap in protein design. It uses all-atom modeling to create proteins that bind to small molecules, metals, and nucleic acids—unlocking new possibilities for drug discovery, diagnostics, and synthetic biology. What’s new in RFdiffusion2: • All-Atom Modeling: Enables precise design of proteins that interact with a wide range of molecules. • Functional Design: Supports binding to targets like heme and therapeutic compounds. • Structural Diversity: Generates a broader variety of complex protein structures with higher accuracy. I feel beyond lucky to be alive and working in AI at time when ideas we once only imagined are now becoming real.

88

8 Comments

OpenAI and Microsoft has accused DeepSeek AI of violating license agreements by training on inputs and outputs from GPT models. It’s ironic because OpenAI has been accused by The New York Times, YouTube, the Toronto Star, and many others of illegally using their data for GPT model training. Despite the legal theater, distillation, the use of one model’s inputs/outputs to train another model with a lower parameter count, could be a major leap forward on getting performant models to run on lower cost hardware. By decreasing the infra footprint of models, there will be a lower barrier to entrance for inference.

42

2 Comments

If you want to fast track to success you got to embarrass yourself. Whether you want to land an internship, get users for your product, or anything else. That's exactly how I got Intern Insider's first 100 users. My co-founder and I were walking through UBC and saw that a hackathon was happening in one of the buildings. So we crashed it and went up to every table one by one, asking students: "Are you looking for a summer internship?" "Yeah?" "We just made this newsletter that sends you 30-50 new internships posted that week, for free!" Then I'd pull out the QR code on my phone and stand there awkwardly so they'd also feel awkward and subscribe. I walked out of that room shaking. It sucked. But we had 100 users. Go where your opportunities are. Ask directly. Stand there awkwardly if you have to. P.S. The same principle applies to asking for referrals, pitching ideas, or making connections. Embarrassment is temporary.

65

5 Comments

I'll admit it—my background is weird for a SaaS founder. A degree in nanobiophysics research doesn't scream "e-commerce entrepreneur."   I started my time at the University of Waterloo as a physics major. And like a lot of people in the field, I was drawn in by the challenge—hard problems, clean solutions, elegant systems.   But when you're deep in physics, you realize the applications that actually help people are few and far between.   So I shifted gears.   I got into nanoscale biophysics, working on disease and drug research, understanding how molecules interact with membranes on an atomic scale.   It felt closer to something that mattered. Positive-sum work. Tangible impact. But I still wasn't seeing the real, tangible effect on people's lives.   Then I founded Smile.io.   The leap from physics to e-commerce isn't typical, but it stemmed from the same drive: wanting to help people and seeing that help materialize quickly.   It was all about time to impact. In physics, you might work years before any real-world application emerges. In e-commerce, we connect with merchants in forums, solve their problems, and watch our solutions go live that same day.   Real customers using what we built. Business owners generating more revenue. Immediate validation.   In physics, professors hand you problem sets. As a founder, your job is solving the next biggest problem—constantly. You must identify which problems matter most.   And that meta-problem of prioritization makes all the difference. If you get it wrong, you waste time. But when you get it right, everything changes.   Every time you pick and solve the right problem as a founder, you make things better for your team, your customers, and their customers.   That’s what building Smile has been: A different set of constraints every day, and a new bottleneck every week.   That’s what keeps me doing this. I always loved solving hard problems. But now, I love what happens when you solve the right ones even more.

43

12 Comments

Unpopular opinion: The RL hype is a red flag for technical depth. When I hear founders or VCs casually dropping "reinforcement learning" in pitches, it's usually code for "we don't actually understand our problem space." Here's the reality check: RL only works when you have: - Perfect simulation environments - Extremely fast iteration cycles - Clear reward functions - Tolerance for massive compute costs That's why it succeeds in games (perfect simulation) and search optimization (Massive compute capacity at Deepmind AlphaGo/Anthropic/OpenAI) but fails spectacularly in the messy, slow-feedback real world. The graveyard of "RL-powered" startups tells the story. They burned through $$$$ trying to apply algorithms to problems that needed simple heuristics or supervised learning. Most of the time, a well-tuned evolutionary algorithm will outperform your fancy RL setup while using <1/10th the compute and delivering results you can actually explain to stakeholders. The next time someone pitches you an RL solution, ask them: "Have you tried solving this with a basic optimization approach first?" Watch how quickly the conversation changes. I am surprised when people are surprised that Direct Preference Optimization (DPO) can exceed PPO-based RLHF [1] or Reflective Prompt Evolution Can Outperform Reinforcement Learning [2]. [1] https://lnkd.in/gCY3bcTQ [2] https://lnkd.in/gQRwyQyM

45

2 Comments

Amazing traction for the inaugural Model Context Protocol (MCP) Hosting Working Group. Kudos to Mathis Joffre for taking the initiative and organizing the first meeting. I saw a couple dozen founders (including Gavin Ching of Toolbase fame), representatives from heavyweight players in the space, including Amazon Web Services (AWS), Google, Oracle, and Snowflake, and even a couple lurking investors =) The Model Context Protocol (MCP) is an emerging open standard that enables AI agents and tools to communicate through structured function calls. As its use grows—especially in local dev tools, cloud services, and enterprise agent frameworks—so does the need for reliable, scalable hosting of MCP servers. However, the core MCP spec focuses on function call semantics and agent interaction, leaving operational concerns like deployment, scaling, transport layers, and multi-tenancy under-specified. The MCP Hosting Working Group (WG) is a proposed community effort to fill this gap by creating shared standards, best practices, and tools for hosting MCP servers in real-world environments. Goals Improve support for transport protocols (HTTP, SSE, WebSockets, etc.). Define multi-tenancy models for secure, shared deployments. Standardize packaging and configuration of MCP servers. Enable runtime-pluggable transport layers. Create mechanisms for server discovery and authentication. Ensure hosting needs are reflected in future MCP spec updates. Roadmap Short-term: Establish governance, prioritize topics, draft best practices. Medium-term: Build reference implementations and pilot real-world deployments. Long-term: Drive adoption, integrate standards into MCP core, and maintain evolving hosting guidance. Participation Open to anyone involved in MCP—hosting providers, developers, infrastructure operators, and community contributors. Join to help shape how MCP servers are deployed and scaled across the ecosystem.

37

16 Comments

Excited to announce our collaboration with Patrick Hsu, Brian Hie, Dave Burke, and the rest of the team at Arc Institute on interpreting Evo 2, their groundbreaking biological foundation model. Evo 2 is an incredible model and represents a significant advance in AI for biology - processing one million DNA letters at once, from bacteria to humans. Through our interpretability work, we've discovered that the model has learned to recognize sophisticated biological concepts, from basic DNA elements to complex protein structures. We're extracting hidden biological knowledge, revealing novel patterns, and advancing genome engineering, all with mechanistic interpretability! The Arc team has been beyond impressive - just an incredible team of scientists and engineers that were a pleasure to work with. More to come on this partnership soon! You can read more here: Technical preprint: https://lnkd.in/eA-jMPQu Interactive visualization tool: https://lnkd.in/eAPswScg Goodfire blog post: https://lnkd.in/ejyAp65Q Arc blog post: https://lnkd.in/e6EJsxmK Shoutout to Myra Deng, Liv Gorton, Nicholas Wang, Nam Nguyen, Tom McGrath, Daniel Balsam, and the entire Goodfire team for their fantastic work here.

642

16 Comments

Startups are changing, but most ecosystems haven’t caught up. At Web Summit #Vancouver, I sat down with JF Gauthier (Startup Genome) to unpack why some #tech ecosystems are accelerating in the AI-native era… while others are stalling out. Check out the rich insights he shared here then watch / listen to the episode 👉🏻 https://lnkd.in/gifNnNph #startups #tech #entrepreneur #entrepreneurship #technews #innovation #vc #ai

19

2 Comments