\n<\/p>\n
Red rocket takes off from ordered structure of others. <\/span><\/p>\n getty<\/small><\/div>\n<\/div>\n<\/figure>\n How do you take charge of a new AI-native business as we enter an age of AI acceleration? More than a few entrepreneurs are asking these questions, including some of the new grads who just went through commencement lines at MIT and elsewhere a few weeks ago.<\/p>\n In terms of what kinds of advice are most valuable for AI startup leaders, there\u2019s the more general kind of business plan strategy, and then legal considerations<\/u><\/a> and regulatory advice<\/u><\/a>. So it\u2019s important to distinguish between \u201crules\u201d and more malleable first principles advice that founders can use to start out on an even keel.<\/p>\n This April, we had the Imagination in Action event that I help run every year, and a panel addressed these questions, a panel of three business leaders in different verticals where AI applies in a big way. Lily Lyman interviewed Jonathan Tushman of Hi Marley, Jordan Hayashi of Bizen, and Cecilia Liu of Intuitive Motion about how to set up an AI startup the right way, and avoid some of the pitfalls of our brave new world.<\/p>\n<\/section>\n Hayashi explained his company\u2019s approach, where the firm has clients in construction:<\/p>\n \u201cWhat has changed a lot in this era of AI is: we can actually start to meet people where they are,\u201d he said. \u201cAnd for construction, that means people can just whip out their phone while on job sites, and start to collect all of the important context, because everything trickles down from what happens on site. And so if we can capture all of that, then we can help them in all of the work that cascades from that, whether that’s notes or facilitating communication or building out pricing and billing and costing and all of that. And so when it comes to deploying, what matters most is: can these people actually use it, and will they actually use it? And so we’ve been pleasantly surprised when our customers are coming back every day and actually using it.\u201d<\/p>\n Tushman talked regulations. His business deals with insurance claims.<\/p>\n \u201cI’m sure all of my teammates here are SOC compliant,\u201d he said. \u201cI used to think that was the big gold standard of compliance, but: have a top ten insurance company walk in your door and audit you: that\u2019s a whole other level. It has huge implications up and down your stack on how you actually deploy a nondeterministic system. So, it’s a really interesting space.\u201d<\/p>\n Cecilia Liu explained that her firm helps to build AI to handle \u201cdirty tasks\u201d that humans don\u2019t want.<\/p>\n \u201cAI shouldn’t be just doing art,\u201d she said. \u201cIt shouldn’t just write poems and be doing stuff on the software level. It should actually reflect what\u2019s happening in the real world. So long story short, we’re building AI powered robots to tackle the dirty work that people don’t want to do, and what we are really doing is going in from one market segment and one particular workflow, and showing that we can build end to end.\u201d<\/p>\n<\/section>\n The panel discussed some of the obstacles to good AI business.<\/p>\n Tushman mentioned a \u201ctrust gap\u201d that businesses confront.<\/p>\n \u201c(It\u2019s) the understandability of why your agents are making the decisions that they’re making,\u201d he said, invoking the use of audit logs and synthesized test casing. \u201cIt’s hard work building a simple system\u2026 (using) all the stuff underneath it to show that confidence is the real work that our team needs to do.\u201d<\/p>\n Hayashi went over an anecdotal situation where the company\u2019s method failed and the team had to do what an AI agent would normally do by hand.<\/p>\n \u201cWe forced ourselves to use the same tooling, the same UIs, the same everything that we give to our agent,\u201d he said. \u201cAnd so we felt all of the razor edges. They’re not even rough edges. They were straight up razor edges.\u201d<\/p>\n Eventually, he said, the humans got incrementally better at this process.<\/p>\n Liu mentioned some types of human error that can derail processes.<\/p>\n \u201cThere’s just human factors that can disrupt these kinds of insulated environment settings that you kind of expect in the lab,\u201d she said. \u201cHow do we reduce that human error? How do we mostly maintain the conditions over time, and reduce the amount of requirements for a skill to run successfully?\u201d<\/p>\n The gap, she noted, is going to evolve even faster in the AI-accelerated era.<\/p>\n \u201cNow, with foundation models, what’s common sense is that actually even the camera vibrations no longer matter as much,\u201d she said, talking about times that humans can\u2019t touch a camera, for fear of sullying AI work. \u201cThe exact camera angle can change so just like as the technology evolves, be mindful that this gap is also going to change, and (consider) always leveraging the right kind of technology and the right tool out of the toolbox to solve the problems in the best possible way.\u201d<\/p>\n<\/section>\n What creates confidence in a company? Panelists talked about building a brand that\u2019s AI-first.<\/p>\n \u201cIn the early days, your customers are not necessarily believing in the product,\u201d Hayashi said. \u201cThey’re believing in you and the company and the fact that you’ll get there. And so going that extra mile, providing that incredible customer support is what gets you customers for life.\u201d<\/p>\n Lyman built on that by asking panelists about how to build durability and defensibility into systems.<\/p>\n \u201cThe fact that 100 carriers already trust us,\u201d Tushman said, \u201cwe don’t take that lightly, and leaning in there to understand the super-low-level detail expert is how we’re approaching that.\u201d<\/p>\n \u201cSome of our customers are starting to use us as a system of record,\u201d Hayashi said, \u201cbecause that’s where all of the data centers are, and every company out there wants to become system of record, and the only companies that win the right to do that are the ones that actually get use and get people coming back.\u201d<\/p>\n Liu cited what she called a \u201ccommon question\u201d in today\u2019s world of raid acceleration.<\/p>\n \u201cPeople ask: \u2018will this model take over the world?\u2019\u201d she said. \u201cMy understanding is: I firmly disbelieve in zero shot. Zero shot is a research term. It exists in labs, but it doesn’t exist in the real world. There will never be zero shot in the real world. You will always require some level of robot data for post- training or fine-tuning in order to solve the task.\u201d<\/p>\n She explained.<\/p>\n \u201cThis is because the world is built by humans,\u201d Liu added. \u201cOrganizations are run by humans, and they build the set of values and cultures around it, and people then build their own habits, practice and rules around it. Therefore, whichever organization you go to you are trying to sell to, you always have to shape your robot behavior around these practices, the rules, and therefore, the robots always have to learn when they get to the new environment.\u201d<\/p>\n There you have it \u2013 a range of practical points for founders, in a world where there\u2019s so much uncertainty about the right way to do things. Stay tuned.<\/p>\n<\/section>\n<\/div>\nThoughts from an MIT Conference<\/h2>\n
Strategy for AI Business<\/h2>\n
Challenges in AI Business<\/h2>\n
Customer Appeal<\/h2>\n