AI Unicorns Raising Funds Too Fast, Investors Begin to Reflect
-
The 'world's first AI programmer' Devin, which recently went viral online, is not only taking the jobs of coders but also the money of investors.
According to the Wall Street Journal, on March 31, Cognition Labs, the startup behind Devin, is in talks with investors to raise funds at a valuation of up to $2 billion. Existing shareholders of Cognition Labs, including Founders Fund, are negotiating the latest round of financing.
The fundraising speed is astonishing. When announcing the completion of its Series A funding in March, Cognition's valuation was only $350 million. At that time, Cognition received support from tech industry leaders such as Peter Thiel's venture capital firm Founders Fund, Tony Xu (founder of DoorDash, the largest food delivery platform in the U.S.), and former X executive Elad Gil, raising a total of $21 million. On the blue banner at the top of its official website, Cognition proudly showcases its impressive Series A achievements. Image source: Company official website
If this round of financing is successfully completed, the valuation of this young company will skyrocket, increasing sixfold compared to the previous round. It's worth noting that Cognition's products have not yet generated significant revenue, and the product developed last year is currently only available by invitation.
Devin made its debut, astonishing everyone with its technology. Devin has recently become a sensation in the AI community. In the announcement on March 12th, Cognition proudly introduced this 'overtime king': always ready to serve programmers, completing tasks and delivering results.
"Devin can plan and execute complex engineering tasks requiring thousands of decisions, recalling relevant context at each step, continuously learning, and correcting errors."
With a single instruction, Devin can run an entire development project end-to-end. The official demo videos showcased Devin's versatility and self-sufficiency: learning new technologies, building and deploying end-to-end applications, finding and fixing bugs, training and fine-tuning its own AI models... Patrick Collison, co-founder of Stripe, exclaimed that Devin isn't just for "show"—it gets real work done.
In the authoritative SWE-bench evaluation, Devin successfully solved 13.86% of the problems, surpassing Claude 2, Llama, and GPT-4. Before this, the best performer, Claude 2, could only solve 4.80% of the problems, less than half of Devin's success rate.
Some programmers feel distressed, fearing they may no longer be able to support their families in the future.
Fred Ehrsam, co-founder of Paradigm and Coinbase, remarked that it's the first time he's seen AI undertake complex tasks—breaking them down into steps, completing them independently, and even presenting the process. "It's already at a level where it can take human jobs," he emphasized.
Devin has even been listed on the freelance platform Upwork to take on jobs and successfully completed the assigned tasks. At first glance, Cognition AI's team is composed of academic elites. Their website proudly showcases:
Members have collectively won 10 gold medals in the International Olympiad in Informatics (IOI). The team includes leaders and builders working on cutting-edge applied AI, with experience at companies like Cursor, Scale AI, Lunchclub, Modal, Google DeepMind, Waymo, and Nuro.
CEO Scott Wu believes that teaching AI to program is actually a profound algorithmic challenge. He describes the process as "almost like a game I've been playing in my mind for years, and now we're finally able to turn it into an AI system."
While technical details remain undisclosed, he hinted that Devin's excellent reasoning capabilities are thanks to a combination of large language models and reinforcement learning techniques. Devin became a model programmer but got flamed on Reddit, the American equivalent of an online forum. Under posts like 'Fxxk you Devin,' programmers gathered—angry and frustrated.
The top-voted comment reads, 'This is unbearable. These people are taking completely unnecessary risks, burning money without a second thought.'
One bold user predicted, 'This thing will definitely be trash in the future. Everyone knows it deep down but just won’t admit it!' 'If they really admitted it, they’d get roasted, lol.' Are the harsh criticisms from 'Tieba' users justified?
The investment frenzy sparked by generative AI continues to soar to astonishing heights.
In December last year, French AI model star company Mistral AI reached a valuation of $2 billion, approximately seven times higher than its summer funding round. In recent weeks, two-year-old AI search startup Perplexity secured a $1 billion valuation in another funding round.
But behind these sky-high valuations lies a cost quagmire currently faced by AI startups. Investors have recently started to reflect: pouring so much money into AI startups does indeed have an element of gambling—but setting aside my concerns, are you actually capable?
Excluding general-purpose model companies like OpenAI, AI startups face significant hurdles in R&D. They must either build foundational models from scratch or develop on existing ones—both of which come with daunting costs, whether in training or inference.
At the recent AI Ascent 2024 organized by Sequoia Capital, it was shared that last year, AI companies collectively spent $50 billion on Nvidia GPUs, generating only $3 billion in revenue. When it comes to the model inference application stage, the demand for computing power becomes even more staggering than during training. Moreover, with chips being in high demand, many AI startups may not even secure access and must pay premium prices to rent AI chips and cloud computing services. Investors might have been able to provide support initially, but they later found that AI startups are struggling to become self-sustaining.
For instance, Artifact, which focused on AI-powered personalized news recommendations, peaked early but failed to sustain user growth and shut down in January this year, eventually being acquired by Yahoo. According to TechCrunch, Artifact struggled to expand its user base outside the U.S. After its launch, downloads plummeted, indicating it failed to attract a mainstream audience.
More often, the issue lies in retaining users. Sequoia also pointed out that the operational traffic of AI companies like OpenAI lags significantly behind other internet applications. Metrics such as the ratio of daily active users to monthly active users and first-month user retention still have considerable room for improvement. AI companies' services have not yet become daily-use products for consumers. Conversely, rapid user growth may instead overwhelm these companies.
Today, when encountering "freebie-seekers" in AI startups, capital has become more cautious—they want to see your long-term value proposition and demand sustainable business models. According to a Crunchbase report earlier this year, Niko Bonatsos from the renowned fund General Catalyst stated that some AI startups' valuations are undoubtedly inflated, and their investors may already be regretting their decisions.
The "cooling down" trend became evident last year. Research firm CB Insights' "State of AI 2023" report shows that global funding for AI startups in 2023 decreased by 10% compared to 2022, with investment deals hitting an industry low since 2017. Those top-performing AI startups have the backing of deep-pocketed tech giants, which allows these companies to continue growing. For instance, Microsoft, behind OpenAI, invested $1 billion in 2019 and followed up with over $10 billion in 2023. Recently, there have been reports of the two collaborating on a $100 billion AI supercomputing project. At the end of last month, Amazon also invested an additional $2.75 billion in AI startup Anthropic, fulfilling its investment commitment from last year.
However, the current trend shows that the Matthew Effect in resource allocation is becoming increasingly pronounced. For future AI startups, how many more funding miracles like Cognition can we expect?