Nir Gazit

When I posted last week about how we made over $1M ARR in 5 months by building fake products, people kept asking me how. Here's why and how we pitched fake products to generate demand and identify PMF before writing any production-ready code: 1/ Find your ICP. Without a north star, it's still impossible to build the right thing. You'll build solutions that might have PMF, but not with the market you're intending. 2/ Rapidly ideate concepts and prototype them as quickly as possible. We had Levi Kline and Matthew Rowland develop prototypes in hours or days, and show them to customers. There was no backend, no real functionality, no tables. It was a glorified Figma file. 3/ Pitch the prototypes to a customer and instead of a sales call, use it primarily as a customer discovery call. Show them your product and dig deep into pain points. 4/ Have incredibly tight iteration cycles. Every single day at the end of the day, I would meet with our team and rapidly iterate on the concept. Changes would be ready for demos the next morning. For our very first product, we made an AI website builder where you would ask the AI to make a website for you in a chat interface (change colors, swap photos, rearrange sections, make new pages). When I would sit down with customers, it was actually just Kyle Wonzen who was the AI. He would make the changes and I would copy and paste in "responses" from the AI bot. Sometimes, customers would be so confused why the AI would send back misspelled sections on their site. 🤣 You know you have PMF when people are asking you to buy without you trying to sell it -> when the value add is so clear that they don't even need a trial and will pay you on a call. This is what we saw with our first SEO product, Seona. We still do deep customer discovery calls to explore user needs for new product lines today. Luckily, we don't have to fake it as much anymore. But to find PMF? Fake it until you make it!

103

10 Comments

“A lot of software will become partially autonomous” One of my favorite highlights from Andrej Karpathy's talk is talking about how to make the agent “generation-verification” loop go as fast as possible. IMO the nuance is whether the agent is solving tasks on behalf of the user or assisting the user with information. 1️⃣ Have the agent solve tasks on behalf of the user: these are routine tasks that the user just wants to offload, like summarizing a research report, processing invoices, redlining contracts. The agent needs to be constrained to solve the task well. This is directly related to the “automation UX” I talked about in my talks. The human doesn’t necessarily learn things here. They just need to supervise outputs to make sure they’re correct. They don’t need to be in the loop at every step of the process. 2️⃣Have the agent provide information to the user: this is what every deep research UX is about. The tasks is more open-ended, dependent on the user input, and how useful this is is constrained by the user’s own brain clockspeed. The optimal UX here involves providing as much useful information in a way that’s also the easiest to consume, and lets the user iterate as fast as possible. Credits to Latent Space Podcast article for the image below. The second image is from my own talk. Kaparthy’s talk: https://lnkd.in/gX34mpEs My slides from the world fair: https://lnkd.in/gNs6S6ua

99

7 Comments

Groq recently launched Compound Beta — their first compound AI system, combining LLMs with web search and code execution. It's powered by E2B's infrastructure, enabling secure, scalable code execution while keeping Groq’s signature speed. Great to collaborate with Benjamin Klieger and the Groq team. More in the blog post 👇

55

12 Comments

Excited to announce that Pipeshift (YC S24) has raised a $2.5M seed round. Over the past year, enterprises have increasingly adopted open-source models for better privacy, control, and performance. New releases like DeepSeek R1 are driving even greater adoption, but deploying these models efficiently and securely remains a challenge. While models continue to improve, the tooling needed for engineering teams to train, deploy, and scale them in production is severely lacking. The amount of time and effort required to piece together multiple tools is significant, not to mention the effort needed to build new workflows. At Pipeshift, we're building a modular orchestration platform that allows teams to build with open source AI components, including LLMs, audio models, image models, vision models, embeddings, and vector DBs. Teams can deploy these models on their infrastructure, whether cloud or on-prem, while optimizing for performance, cost, and quality. In recent months, we've helped several teams train and deploy use case specific models for their workflows, and we’ll continue to do so. If you're currently building with open source models or thinking about making the switch, let's chat. We’re incredibly grateful to our earliest supporters at Y Combinator (GPs: Dalton Caldwell, Tyler Bosmeny, Michael Seibel, David Lieb), SenseAI Ventures, Arka Venture Labs, Good News Ventures, Nivesha Ventures, Astir Ventures, MyAsiaVC, gradCapital, 100X.VC and amazing angels including Kulveer Taggar, Umur Cubukcu, Krishna M., Nate Matherson, Max Kolysh, Fred Zhang, Ruchir Patwa, Arjun Maheswaran, Kishore Ganji, Praneeth Kannah, Anshul Bhide, Hasan Abu Sheikh, Uday Sodhi, Aabhas Khanna and Maria Nigam. Full throttle ahead 🏎️ Link to the article: https://lnkd.in/gj-DXiSA

228

19 Comments