You see, in this world there’s two kinds of people, my friend: Those with loaded guns and those who dig. You dig.
This quote is one of the many great lines from Sergio Leone’s conclusion to the Dollars trilogy, The Good, the Bad, and the Ugly, and is strangely apt for the challenges of building a startup in the choppy world of Generative AI in 2023. If OpenAI is the biggest gun right now, then the hyperscalers, particularly Microsoft, Amazon and Google, have the best armouries. The rest of us, well, we have to dig.
To misquote the film’s title, scaling a startup involves a different triple: the known, the bad, and the great. Startup journeys often start from a position of scratching our own itches, hoping that by solving a problem we have experienced, there will be a market of people or businesses with the same problem, looking for a solution. In this journey, most startups invest time, money, sweat, burn relationships, and damage their mental wellbeing in the quest to build great software solutions or products. Surely the technology will speak for itself, and having a great product is more likely to win in a competitive market than a bad product?
But this underplays the importance of being known, and in fact whether a company’s offerings are great or bad might even be irrelevant to commercial success. Often we assume that becoming known is the result of building great product, but in fact the reverse is often the case.
To give some examples from personal experience, in 2009 – 2012 I had the privilege of working at a SF Bay Area startup backed by Sequoia, Kleiner and others, already at Series C. The company had built world-class hardware and software solutions since 2004 for the telco market, using FPGA architecture. Even today in 2023 FPGA is considered advanced; back then it was far ahead of its time. The platform was able to perform line rate encryption and decryption of mobile traffic, operating at the edge of a core network, with 10 gig interfaces. The competition for this particular network element was the big four – Ericsson, Alcatel Lucent, Huawei, and Nokia Siemens Networks. We had already secured NTT DoCoMo in Japan as a reference customer, with a significant order book, and were smashing it with technical trials for many of the world’s best mobile network operators; surely global success was upon us. But we had underestimated how different the buying behaviour can be, when selling to tier 1 telcos. Despite being literally 10x better, and often a quarter of the price of the legacy incumbents, we could not close our deals. The problem was brand strength and awareness; we represented unproven risk, and the buying behaviours of our Ideal Customer Profile were biased against risk far more than technical excellence or commercial benefits. We could have changed strategy and targeted smaller telcos first, of which there were many, but having the world’s best VCs in our cap table meant that we had to win big, or go home. Had we been better known to our target client prospects’ boards, and not just been standards-compliant, but also leaders in the standard bodies, then the outcome would have been different.
More recent personal examples from the last few years would be StreamNative and deepset. The founders of StreamNative had tremendous Founder/Market fit, having developed the Pulsar data streaming framework while senior engineers at Twitter and Yahoo. The framework had been developed to address a well-known challenge with Kafka, namely the difficulties scaling and balancing nodes, as Kafka did not separate compute and storage functions. Pulsar was released to the Apache Software Foundation, the company StreamNative was formed, and significant Series A investment was secured. Again, despite being objectively better than Kafka, Pulsar just did not have market awareness and a large ecosystem of software engineers and partners who knew it. When it came to making technology architecture decisions, the ‘known bad’ – Kafka – won every time.
For deepset, the example would be its Open Source NLP/LLM framework Haystack. The framework had been created by deepset to make it easy to use transformer models with a vector database, to perform common NLP tasks, such as Question Answering, Information Retrieval, summarisation etc. Adoption was steadily growing from 2021 to 2022, and Haystack is an excellent choice for enterprise implementation of GenAI / LLM integration. But then at the end of 2022, LangChain by Harrison Chase was launched. Objectively, LangChain is (was) an inferior framework to Haystack, but Harrison had a number of key advantages over deepset: he was based in SF, American, charismatic, with a strong name, excellent relationships, and very adept at building community. He was able to ride the zeitgeist and within a matter of weeks LangChain had become the de-facto, known choice for experimentation with the LLMs that had exploded into the public sphere with the launch of the ChatGPT beta November 30th, 2022. At the time of writing LangChain has grown to 50K+ github stars, as Haystack slowly climbed to the symbolic threshold of 10K stars. Another very human trait also helped LangChain gain adoption – it was notoriously unreliable, and going through growing pains fast. Perversely, rather than discourage people from adopting it, users found themselves fixing bugs and contributing to the code base. Nothing builds loyalty like a personal investment, driven by the sunk cost fallacy. Haystack by contrast is well engineered (the team at deepset are mostly German, after all) and it pretty much ‘just worked’. No sense of personal investment needed to use it. It’s hopeful that Haystack will still have its day, but now LangChain, LlamaIndex and other frameworks have captured this space by being well known, and now good enough. Why use an unknown great alternative?
So what are the lessons learned, recommendations and conclusions? When seeking product/market fit (PMF) and growing a business, the market forces and buying behaviors of the ICP (Ideal Customer Profile) are far more important than meeting technical requirements, or even user needs. Brand strength is more valuable than innovation or price, even in a high tech industry. When creating early business canvases and GTM strategies, prioritizing being known, rather than being great, is a much better place to start.
