Alex O'Byrne

10 steps of how funding worked in 2025 (A post nobody asked for): Step 1: Say you’re building an AI copilot. Doesn’t matter for what. Carpets? Bicycles? Pickles? Just add ‘copilot.’ Step 2: Whisper the word “moat.” Very quietly. Step 3: Tell the VC you’re raising a “tiny” pre-seed of $4M. They love the word “tiny.” Step 4: Hire an advisor who once ate lunch near OpenAI HQ. Step 5: Put “ex-FAANG” on your deck. Doesn’t matter if it was an internship. Or a tour. Step 6: Say you’re “pre-revenue by choice.” Make it sound strategic. Like intermittent fasting. Step 7: Build a waitlist. Your product doesn’t need to exist. Just the waitlist. Step 8: Tell VCs you’re not fundraising. Watch them become extremely interested. Step 9: Get term sheets with 48-hour deadlines. Decline them politely. Return two weeks later. Ask if it’s still available. (It will be.) Step 10: Announce your round on TechCrunch. Congrats! Your AI copilot for pickles is now “category-defining.”

923

173 Comments

We’ve been jamming with Y Combinator’s Summer 25 batch, and have been blown away by the quality and pace of the teams. What jumps out: -Infra > interface: some excellent teams are building in the execution fabric, trust rails, or evaluation loops. They’re tackling latency budgets, anti-bot arms-races and SOC-2 checklists: the plumbing that lets agents graduate from demos to enterprise workloads. -Data moats by design: the strongest vertical plays lead with proprietary claims corpora or HIPAA-grade call graphs. -Hard-tech is back: Flywheel’s tele-ops excavators and Verne's one-hour-training vision stack prove YC is funding robots again, but with a software-first margin profile. -Founders who’ve lived the pain: Ex-Cleartax PMs fixing GST reconciliation. Ex-Goldman post-trade leads building zero-trust for agents. Wet-lab scientists automating plate handling. Why we’re excited: -These teams are solving the reliability, audit, and compliance gaps so agents can be boringly dependable, instead of asking if they will work. -Several are already in production with paying customers pre-Demo Day. -Defensibility comes from data moats, regulatory grounding and real infrastrucutre. You can see the agentic web moving from prototype to platform. Looking forward to seeing some of you tomorrow at Demo Day.

213

10 Comments

AI music platform Suno just hit $150M ARR. That’s 4× growth in just 12 months. Here’s what this means for the AI music industry 👇 For context: Suno is an AI music platform that lets anyone create full songs, lyrics, and vocals from a single prompt. In one year, it has nearly 4x to $150M ARR with margins above 60%. And with 44M monthly website visits (SimilarWeb) and over 5M paying users (estimated as per current ARR), it is building real retention. This is product-market fit at scale and makes Suno one of the fastest-growing consumer AI apps since ChatGPT. Here’s what this signals for the industry: • Revenue is moving from labels to platforms • Creativity is moving from professionals to everyone • Ownership models will need to be rebuilt Suno is also in talks with major record labels to settle copyright disputes. Once finalized, those licensing deals could make it the first fully compliant AI music platform. Why it matters: > AI music is entering mainstream adoption > There is now proof of scale and monetization > Strategic buyers are positioning around audio The M&A signals are clear: → Spotify is building AI voice and moderation tools → OpenAI is moving into multimedia creation with Sora All are chasing the same goal: control over the next interface of human-AI interaction. That interface is sound. ----- Our team at FE International is more bullish than ever on the rise of AI-first businesses. If you're a founder curious about your company’s true value, get a free valuation. (Link in comments) 👇

82

46 Comments

a16z: the AI ideas founders will regret ignoring in 2026 📢 This year’s list feels different. AI is moving past copilots and experiments toward systems that act on their own, agent-native infrastructure, and products that replace entire workflows. When software starts executing instead of assisting, whole categories reset. These are patterns a16z partners are already seeing across real startups and deployments. If you’re deciding what to build or how to position for the next wave of AI, this matters. I broke down the 15 AI ideas shaping 2026 and why they matter now. Full article 👇 https://lnkd.in/e_zaX_xB

154

26 Comments

Deep *startup* research 🧐 Austin Johnsen from Zapier shared in a post a few weeks ago the prompt he uses to do diligence on M&A targets (https://lnkd.in/gNTgBNde). It's a great prompt, and now Scout in Harmonic lets you save prompts like this and fire them off to do deep research on any startup, with much richer startup data than would be available from any general LLM or deep web research tool. Please share any thoughts/feedback that we can incorporate to make this even more powerful 😁 .

97

10 Comments

Universal Commerce Protocol isn’t just a technical standard — it’s a distribution unlock. For marketing and sales teams, UCP: * Shortens the path from attention to revenue * Reduces platform dependency * Enables AI-native selling * Makes commerce composable and scalable early Startups that design for UCP early will be easier to sell, easier to buy, and easier to grow in a world where commerce happens everywhere, not just on your website. Ecommerce brands without websites, selling through AI agents, are already happening! Huge thanks to our amazing mentors Bahar Bozdogan and Monish Datta for the fantastic UCP workshop on Friday with our portfolio companies. There is so much to build and so many opportunities for startups there! :) Entrepreneurs Roundtable Accelerator Remarkable Ventures

36

4 Comments

The big labs are expansive in their product ambition, especially since foundation models have largely improved in lockstep - in order to compete with them you have to do things they won’t which are: - building a very rich software ecosystem around a primitive - orchestration across multiple models - going insanely deep on product and growth for a narrow vertical domain

58

3 Comments

We just launched a new LLM that was fine tuned to reduce false positives in watch-list screening. The headliner results beat all expectations - ◼️ 98% accuracy in name degradation testing (vs ~80% for others) ◼️ True positive to false positive ratio improved from 100:1 to 3:1 ◼️ 100% recall during testing across X PEP/sanctioned individuals ◼️ 20X reduction in false positives per 1000 checks At a high level, we trained the models to understand the context surrounding each potential hit. Then gave them thousands of historical examples of known analysts' resolutions to learn the difference between true positives and false positives. Our own Joshua Linn wrote a deep dive article explaining their process. Link in comments.

149

5 Comments