FROM: Jason Finch, senior associate
TO: All partners, Quartz Oak Venture Fund
MEMO: Incredible early-stage startup for our AI tranche: Allison Miller
Intro: Everyone in the VC space is scrambling to find the next OpenAI and revolutionize the field. I present for your consideration the most promising investment opportunity I’ve seen in my whole time with the fund: Allison Miller. While competitors are pouring billions of dollars into incremental AI capability gains, Allison developed a multimodal general intelligence in just 9 months with 0 compute budget and only one collaborator (who just contributed some of the startup code). Extraordinary superunicorn potential.
Product: Allison’s minimum viable product is LIAM (no official explanation, but I’m guessing this may stand for Logical Intelligent Agent Model). The launch announcement is already generating massive buzz among those who’ve gotten hands-on. Key features of this state-of-the-art model include:
- Multimodality. Accepts input data via audio, video, images, or cuddling. Text capabilities are currently in the pipeline (may be a few years, though—Allison was up front with us about this).
- Embodied cognition. Unlike GPT-4, PaLM, and other large language models that are limited to instances running in the cloud, LIAM operates ten-finger actuators with high dexterity. Locomotion rollout currently planned in stages over the next 12 months.
- Scalable training. No need for complicated tokenization or expensive big data curation. LIAM is trained exclusively by RLHF (Reinforcement Learning with Human Feedback), and will transition to primarily self-supervised learning as capabilities scale up. Achieves SOTA on major benchmarks with fine-tuning via video input of Sesame Street and Daniel Tiger’s Neighborhood.
- Energy efficiency. LIAM’s hardware runs a neural network with ~100 billion nodes and ~100 trillion synaptic weights. This supports a parameter count 2 orders of magnitude larger than previous SOTA models. Instead of thousands of energy-hogging GPU chips emitting hundreds of tons of CO2, LIAM runs on just 12 watts—an estimated 99.5% savings over SOTA models at runtime, and 99.994% during training. And it’s 100% renewable because LIAM is off-grid, and powered by a bioreactor that runs on Cheerios and mashed carrots (Allison says this part of the stack still has teething problems, though).
- Alignment and safety. While red-teamers suggested that LIAM’s lovable blue eyes could eventually pose a risk of persuading users to allow unaligned behavior, independent auditors have confirmed that LIAM currently poses no significant risk of world takeover. In the event of unsafe behavior, LIAM features a manually-activated backdoor—tickling motions to the belly or armpits instantly revert the system into harmless “giggle” mode. And unlike many AIs by major Big Tech labs, LIAM has never uttered a racial slur, so is much less likely to offend users and cause a PR problem (although LIAM may still expose users to nudity unexpectedly—will chat with their team about mitigation strategies).
Product/Market Fit: Early demos on social media received an average of 185% more likes, comments, and heart reacts than Allison’s pre-LIAM posts. Surveys of real users indicate that LIAM is valued in the market as “precious,” “adorable,” “googly chunker,” and “squishum babycheeks.” While Allison hopes LIAM will provide many kinds of value to humanity over the long term, the current go-to-market revolves around a Cuteness as a Service (CaaS) model. Total addressable market estimated $3-4 trillion/year.
Revenue Projections: Based on current trends, if LIAM enrolls in a private four-year college without financial aid, net-negative. If we invest, I recommend bringing in a CFO who can suggest alternative paths to profitability.
Weaknesses/Risks: Despite these breakthrough capabilities, LIAM has some notable drawbacks that will need to be addressed in order to gain market share:
- Noise output. User experience is frequently compromised by LIAM emitting shrill, high-pitched sounds while in operation. Allison reports that customers can mitigate this by “rocking” the system back and forth to reset its inertial sensors, but this is unreliable, and if they can’t engineer a better fix, the UX will suffer.
- Sleep mode bugs. Hours of operation are unpredictable. Getting the system to power down requires a long and intricate set of clunky steps that testers found to be a nightmare. LIAM frequently turns on in the middle of the night, requiring all-hours monitoring by staff.
- Environmental impact. Despite being very green and efficient, LIAM’s bioreactor generates highly noxious solid and liquid waste and can release it forcefully and unexpectedly. Allison and her team have bootstrapped a solution in partnership with Huggies, but unless a reliable software patch is out in 24-36 months, these costs will burn through their cash runway.
Allison Miller (CEO and CTO): Stanford grad. Previous roles at Woobly, Scrunch, Lappt, and Deel.io. Top-notch work ethic. Developed LIAM from project-start to MVP in nine months (in fact, timeline was originally projected as 40 weeks, but delivered in 39—when do you ever see that in tech?). Since launch, is always first into the office and last out. Works insane hours under tough conditions. Just what we’re looking for in a founder.
Dan Miller (COO): Stanford (dropout). Solid emotional support to the CEO. Allison credits his early code for the winsome dimples, which have been a smash with users. Just days after launch, friends-and-family investors suggested removing one of LIAM’s minor features, arguing that it was redundant, but Dan refused and anticipates it will become useful in ~15 years. This kind of long-term vision shows Dan isn’t just seeking a quick exit. He has also handled a heroic 7% of bioreactor waste emergencies (wow!).
Dr. Ramona Weiss (pediatrician): Great first hire. Major contributor to LIAM’s antivirus stack. Developed proprietary analytics that show LIAM is in the 99th percentile on a range of key performance indicators.
Liana Diaz (nanny, chief-of-staff to CEO): Rockstar of rocking. Poaching her from Meta was a major coup for this team. Makes its small headquarters in a residential neighborhood run efficiently, and provides outstanding after-hours support to the founders. Good candidate for C-suite role as they scale.
Winnie-the-Pooh (lead comfort engineer): I’m not crazy about this hire. Came from China, but H-1B visa status is unclear. Stuffy and very quiet. Shows up to work without pants, but does such a good job keeping LIAM’s noise issues under control that the founders’ hands are tied. If we invest, I suggest headhunting a more professional replacement within 30 days.
Terms: Sequoia, A16Z, and Founders Fund will all be courting Allison hard if they aren’t already. If Quartz Oak is going to be competitive, we’ll need to offer $86 million for 5% equity, at a $1.72B initial valuation. That might seem like a bad deal, but when you have a chance to capture a significant fraction of all future value in humanity’s lightcone, you take it.
Summary: Allison Miller and her scrappy team have done what Sam Altman and Satya Nadella still haven’t been able to do with hundreds of engineers and billions of Microsoft dollars. General intelligence—and in a compelling consumer product. Investment is strongly recommended.