Nova update: launches, round and community
April was the kind of month where looking back, you wish you had spread things out a bit more. Nova Recruiter launched publicly. The seed round extension closed most of its allocation in under 24 hours. Nova 111 went live across 8 countries simultaneously, adding 888 new members to the community, many of them joining Nova for the first time.
The round extension of €1.5 million reserved a dedicated allocation for Nova members. Within 24 hours of opening, interest crossed €150k. Despite some friction from the platform being primarily in Spanish, the response from an international community was notable. The remaining allocation opened to the broader network shortly after.
For the Nova 111, this was the first time eight lists launched at the same time, including the first ever UK edition. The winners receive lifetime membership and access to all features, and a group of them will join the Nova Experience in October, a three-day retreat in the Ticino region of Italy.
Coming from nowhere, being able to enter the Oxford campus to celebrate talent and merit is something I look forward to.
Ramón Rodrigáñez, Co-founder, Nova Talent
On the product side, the campaign module received a major update, giving users much more control over frequency, channel prioritization and email follow-ups. The MCP integration moved from internal use to a public launch. And the gamification tier system, including the ability to redeem points directly on the platform, went live after months of replacing a cumbersome Typeform process.
What we learned launching Nova Recruiter
Any product launch comes with uncertainty. The day before, you genuinely do not know if it will go quietly or go very loud. Nova Recruiter landed somewhere in between: strong early traction, a few hundred new users in the first month, and a set of clear lessons about what worked and what to do differently next time.
What worked
The video quality made a difference. In 2025, a launch video that stays somewhere between a founder talking to camera and a real product demo does not cut through. You need to either show the product clearly or do something genuinely memorable and unexpected. The Nova Recruiter video leaned into quality production and it showed in engagement relative to competitors who launched similar tools weeks later.
The LinkedIn strategy was well executed. The video did not go viral on X but performed strongly on LinkedIn, which is where the ICP actually lives. Product Hunt brought additional visibility: fifth place on the day, with OpenAI's image generation model taking the top spot on the same morning.
The most creative idea of the launch, and arguably the one that generated the most pure impressions, was the team changing their LinkedIn job titles to list themselves as early adopters of Nova Recruiter. LinkedIn's algorithm treats job changes as high-priority signals without reading the content. The result was that those update posts reached audiences far beyond the official launch announcement.
My job change post had more reactions and more shares than the main launch video. People were like, what is going on? And that attention spilled directly onto the product.
Ramón Rodrigáñez
The lead magnet that outperformed everything came after the launch: a free course by CPO Christian on how to become an AI recruiter, integrating Nova Recruiter into a full Claude workflow. It reached 858 comments and drew exactly the right ICP: founders and recruiters. One free course out-distributed weeks of launch activity.
What to work on
Free tier conversion is the current focus. Nova Recruiter has been generous on the free offering and users are engaging, but upgrading to paid is not happening at the rate needed. Tightening the free-to-paid journey, testing paywalls, and making the value of the premium tier clearer is the next chapter.
Testing the paperclip autonomous company framework
Paperclip is an open-source framework that lets you run what it describes as an autonomous company. You act as the board, define goals, and an AI CEO hires and directs agents to execute tasks. It runs locally in your browser. The concept is genuinely interesting. The reality, at least right now, is more complicated.
Andrea tested it with a real side project: a tool to help high school students match with the right university program based on their qualifications and preferences. His honest assessment after running it:
What worked
- Interesting framework for structured autonomous tasking
- Good for understanding agentic delegation concepts
- Open source and locally run
What fell short
- Output quality was not there without heavy human feedback
- Credits burned extremely fast on stronger models
- Lacked product taste and judgment for anything UX-related
- One evening with Claude alone produced better results
My feeling at the end was that one evening with Claude myself I would have done ten times better. The CEO and its agents were not able to deliver to the level I needed. The criteria, the taste, the judgment: it's not there yet.
Andrea Marino, Co-founder, Nova Talent
The framework is worth watching. The gap between the concept and the execution is mostly a model capability problem, not a framework problem. As models improve, agentic company frameworks like this will get dramatically more useful. For now, treat it as a preview rather than a production tool.
OpenAI enters services. Why now?
OpenAI raised $4 billion to launch a services and forward deployment arm, attracting investment from McKinsey, BCG and Accenture among others. Anthropic made a similar move slightly earlier, building a services partner ecosystem with PwC, Deloitte and Accenture while also maintaining their own direct services capability.
The strategic reasoning is straightforward. Enterprise AI adoption is far slower than startup adoption. Startups iterate weekly. Large organizations run multi-year procurement cycles. Closing that gap requires human relationships and custom implementation work, not just API access.
There is also a lock-in argument: a deeply integrated workflow built around a specific model is harder to replace than an API key. Andrea is skeptical of how durable that lock-in really is.
What OpenAI ultimately provides is a key and a model. That is less defensible than software. Tomorrow you change the API key to Anthropic or Google, run the evals again, and rebuilding the use case does not take that long.
Andrea Marino
Ramón's counter is that the real lock-in is not in the product, it is in the relationship. If McKinsey or Deloitte is the firm that builds the client workflow, and they are aligned exclusively with one provider, the human relationship becomes the moat. The client does not switch models. They stay with their advisor, and their advisor stays with the model.
Both views have merit. The longer-term question is whether the best service firms will accept exclusivity at all. The incentive to stay flexible is strong when model rankings shift monthly.
The HTML vs markdown trick
This one is practical and immediately useful. An Anthropic engineer shared internally (and it has since spread) that when you need to stay in the loop with long AI outputs, asking for HTML instead of markdown is dramatically better.
Why HTML beats markdown for long AI outputs
- Markdown documents are hard to read at length. HTML can be visually rich, with layouts, sections, graphs and sliders
- Anthropic engineers reportedly use this approach to absorb complex AI output faster and give better feedback
- HTML lets you imagine the actual product or interface, not just read a description of it
- The model takes slightly more tokens to produce HTML, but your comprehension and decision quality go up significantly
Try it yourself
Next time you ask Claude or any LLM to produce a document, specification, funnel breakdown, or structured plan, ask for HTML output instead of markdown. Open it in a browser. Notice how much faster you can actually read and respond to it.
Andrea used this approach while rebuilding Nova's HubSpot funnels with the team. Instead of creating a Notion doc, he asked Claude to produce an HTML version of the funnel layout. The result looked like an actual interface and made feedback and iteration significantly faster.
GEO in action: what Tuyo got right
Tuyo is a Spanish home insurance company doing things differently. They are digitally native, straightforward in their product, and recently became one of the first companies in Spain to enable direct purchases through ChatGPT. You can talk to ChatGPT, describe what you need, and buy a Tuyo policy without leaving the conversation.
What makes their approach worth noting is the GEO dimension. Generative Engine Optimization is the practice of making your brand show up well in AI-generated answers, not just in traditional search results. Tuyo's market share in Spanish home insurance is small. But ask almost any AI in Spain which home insurance provider to consider, and Tuyo comes up in the top two or three responses consistently.
If you check their market share versus their positioning on AI search, it's crazy. Very low market share. But ask any AI in Spain about home insurance, and Tuyo is always there.
Andrea Marino
The tool they are using to track and improve this positioning is built by a Nova member, the former General Manager of Klarna in Spain. The tool shows you how you rank in AI-generated responses across different queries in your category, which content blocks you need to appear in, and what to produce to improve your position. As AI assistants increasingly replace search for product discovery, this is the kind of optimization that will define brand visibility.
Cerebras and the chip race
Cerebras, the chip company founded in 2015, is reportedly targeting a $26 billion IPO, one of the largest in recent history. The company is over $500 million in annual recurring revenue and has a notable compute agreement with OpenAI worth $20 billion.
Their technical edge is specific: for training very large models, they compress both compute and memory into the same silicon chip. Nvidia's GPU architecture separates these two components. For large model training, Cerebras claims their approach delivers faster performance. It is a narrow but valuable advantage in a market where every percentage point of training speed has economic significance.
The broader point Andrea draws from the Cerebras story applies across the AI stack. The companies making the most money in this cycle are not the model companies. They are the chip companies. Competition at the infrastructure layer is good for everyone who builds on top of it, and the fact that a credible Nvidia challenger is approaching a $26 billion valuation suggests that competition is finally arriving.
Takeaways from this episode
- The best GTM idea of the Nova Recruiter launch was a side idea, not the main plan. Always keep generating, even after you think the playbook is set
- A free course built around your product can outperform a full launch campaign if it genuinely teaches something useful to your ICP
- Autonomous AI company frameworks like paperclip are worth watching but not yet reliable for production work. The gap is model judgment, not framework design
- OpenAI and Anthropic are both entering services to close the enterprise adoption gap. The real moat is in human relationships, not model lock-in
- Switch from markdown to HTML for long AI outputs. It takes more tokens to produce but saves far more time in comprehension and feedback
- GEO is real and already creating competitive advantages. If you are not thinking about how your brand appears in AI search, start now
- The chip layer is where the most durable margins in AI will sit. Cerebras's $26 billion IPO target signals that the Nvidia monopoly is being challenged seriously
- Disengagement at work is rising alongside AI adoption. The productivity gains are real for some, but the majority of workers still see AI as a threat, not a tool
Join the network that ships, builds and stays curious.
Nova connects over 25,000 top professionals across 80+ countries. It's merit-based, curated, and built for people who want to be around the most driven talent in the world.
