At some point over the summer, got a little bit smarter.
Sunny, the artificially intelligent chatbot the city of Denver rolled out last spring, wasn鈥檛 roving the Denver interwebs or trolling neighborhood social networks. The genderless bot was just getting faster at what it was designed to do, mainly by remembering what it had already done.
鈥淚t鈥檚 simpler for Sunny to come back quickly with a response if it鈥檚 already gone through the workflow,鈥� said Suma Nallapati, the city鈥檚 chief information officer who鈥檚 in charge of making sure the tech works for city employees and residents. 鈥淔or example, (asking) what is my trash pickup day for this street? If it knows the address and if it knows the pickup day, it鈥檚 easier for it to answer that question the next time around.鈥�
Baby steps, really. Sunny鈥檚 still in a crib to prevent it from escaping into the experimental world of generative AI, the game-changing tech popularized by OpenAI鈥檚 ChatGPT-4 that has machines creating new content. Sunny spent a year in training before a public debut last April, and after the was drafted, vetted and approved.
Humans are still involved in feeding Sunny, to make sure its responses are appropriate, accurate and not, as they say in the AI biz, . Top-ranked restaurants, which asked Sunny about a few months ago, was excised from Sunny鈥檚 memory, since the city isn鈥檛 in the business of specific restaurant promotion. In the past six months, Sunny has handled 20% to 30% of calls and texts to the city鈥檚 311 information service, which eases the workload of the human operators.
In a world where AI is developing so fast, Denver鈥檚 trying to carefully establish itself as a city that welcomes innovation to solve problems but also considers unintended consequences. It even organized next week鈥檚 , touted as the nation鈥檚 first city-led conference on the topic. That鈥檚 in line with the state, where potential privacy, prejudice and piracy issues of AI led lawmakers to in May curbing some AI development. Colorado was first in the country with such a law, though will likely result in revisions.
The Governor鈥檚 Office released last month, which impacts all state agencies, said Amy Bhikha, chief data officer for the Governor鈥檚 Office of Information Technology. Anything with a generative AI component must pass OIT guidelines. GenAI can鈥檛 be used to make decisions 鈥渕aterially impacting individuals or entities.鈥� In other words, vetting applicants for state benefits or funding, or an adverse action, like an arrest or detention, must rely on human decision makers.
OIT is also trying to figure out how tech like ChatGPT, Google鈥檚 Gemini or Microsoft Copilot can help employees in their jobs and assist residents with state services.
鈥淲e have a whole innovation piece that we鈥檙e setting up,鈥� Bhikha said. 鈥淲e鈥檙e doing a trial right now with Google Gemini. We have 150 people across all agencies using it and giving us feedback every week. And then we鈥檙e looking to potentially set up a lab in the future.鈥�
A lab would allow staff to test out available AI tools like Grammarly and Read.ai, to quickly provide summaries and transcripts of meetings.
So far, none of the experiments have the potential to be public facing. But check back in a month, she said.
The speed of AI
AI and tech companies have been around for years in Colorado, with an ecosystem that includes , in hospitals and . Many consumers may not realize they鈥檙e already experiencing it, be it with , getting help from or .
But for some young startups in the genAI space, growth has been lightning fast. Take Adeel Khan, for example. Two years ago, the Denver resident was on a break after more than a decade of teaching and working as a school administrator, including as the founding principal of Conservatory Green High School in Denver. He wasn鈥檛 thinking of starting a company.
Then ChatGPT-4 launched.
鈥淚 started trying to use it for things I would be doing as a teacher and principal and I found a lot of utility (and) I wondered if educators were using it,鈥� Khan said. 鈥淚 had no idea for a software company but I did what a principal would do, which is I went to my old school and I trained teachers on how to use ChatGPT. And in those sessions, (their) jaws were on the floor.鈥�
He鈥檇 follow up weeks later only to find the teachers had stopped using it. They didn鈥檛 have time. They forgot what to do. They never figured out how to ask good to achieve better responses. So Khan started MagicSchool in March 2023 as a digital helper. The service can quickly make a first attempt to generate lesson plans, create math problems or write letters to parents in the same way a teacher鈥檚 assistant would.
MagicSchool raised $2.4 million from local investors last year and went on to in July. Khan now leads the 40-person company from his Denver living room. This week, it reached 3 million users.
鈥淚 think of AI as this really incredible opportunity to (help) educators and to assist them in their work, fight teacher burnout, help them be more efficient and also maybe augment (and) quickly refresh their knowledge on a subject they hadn鈥檛 taught before,鈥� Khan said. 鈥淭he second part is the student tool. We need to prepare students for a future where AI is everywhere. Next month, Apple will implement AI features in the iPhone. They鈥檙e already in Android phones. If we鈥檙e not teaching kids how to use AI in school responsibly, then we鈥檙e just leaving them to their own devices.鈥�
But there are definitely lines MagicSchool won鈥檛 cross: Grades.
鈥淲e鈥檝e made a lot of intentional decisions around this,鈥� Khan said. 鈥淥ne thing we鈥檝e intentionally left out is giving a grade. The grade is a high-stakes decision that could impact a child鈥檚 future. 鈥� The educator is still the center and decider.鈥�
Past investments, future innovations
Venture capital has been paying a lot of attention to AI companies, which scooped up 41% of all U.S. VC investments in the first half of the year, according to PitchBook, a market research firm that tracks venture deals.
When Denver-based Range Ventures invested in MagicSchool in August 2023, the VC consensus was to avoid applications because larger AI companies, like OpenAI or Google, could build their own competitive service, said Adam Burrows, a managing director at the firm.
But there was something magical about Khan鈥檚 startup that made it a better prospect than the customer service startups others were chasing, Burrows said.
鈥淲e looked at the K-12 space and we said, 鈥楬ey, there鈥檚 actually no incumbent here.鈥� AI is allowing something to be built for teachers that has never been built before,鈥� he said. 鈥淥ur view was that AI was actually enabling an entire type of software that didn鈥檛 exist before. So we have been bullish on looking for more situations like that.鈥�
Range Ventures focuses on Front Range companies and continues to invest in young AI startups, most recently , a note-taking administrative assistant for veterinarians. AI isn鈥檛 new to Range Ventures though. It invested about a decade ago in AMP, the robotics and automation company in Louisville that uses AI to sort recycling.
AMP has come a long way from its one-armed robot named Clarke, which used optical sensors, machine learning and AI to out of a pile of moving recyclables.
The company, which has since raised from investors, now builds entire systems to sort recyclable material from plain ol鈥� trash without the touch of a human hand.
鈥淲hat we鈥檝e done with AI is created what we think of as a universal garbage sensor,鈥� said Matanya Horowitz, AMP鈥檚 founder and CEO. 鈥淏ut rather than a robot that picks out aluminum cans, we think of the whole facility as a robot, like a 1,000-ton robot. It鈥檚 awesome.鈥�
Years of training and scanning 150 billion items has helped AMP鈥檚 technology understand the differences between various types of plastics, metals and other recyclables versus actual trash, like chicken bones, diapers and 鈥渁ll the nasties,鈥� he said. Its machines have even been trained to distinguish Starbucks cups (鈥淲e wanted to show that we had no issue getting down to the brand or packing level,鈥� he said).
The AMP One systems are up and running in trash and recycling plants around the country, including at a municipal solid waste facility in .
Like an Amazon warehouse where robots do most of the running around, there are few employees on the floor of an AMP-powered facility. There are still some humans though. AMP calls them data annotators who can correct the machines and teach them the difference between a Starbucks cup and an unbranded coffee cup, for example. Everything is just so much faster.
鈥淭he system can highlight things it鈥檚 uncertain about, giving more leverage to our data annotators,鈥� he said. 鈥淚t鈥檒l say, 鈥楴ot sure about these. Will you help me discriminate them?鈥� The person goes through and makes those corrections and what we鈥檙e finding is intelligent prioritization of the work can make human annotators 100 times more effective.鈥�
While genAI use isn鈥檛 widespread on the floor, AMP engineers use it to code faster.
鈥淎I tools are so powerful. I鈥檝e seen it firsthand. The productivity out of developers is unreal,鈥� he said. 鈥淪ociety will advance so much quicker because software development is about to become somewhere between 10 to 1,000-times more efficient. I think people want to be really thoughtful so that they don鈥檛 slow down the progress and put us at a disadvantage.
Nallapati, Denver鈥檚 CIO, said that鈥檚 part of the purpose behind the city鈥檚 all-day DenAI summit.
鈥淭he cognitive revolution is here and it鈥檚 again characterized by the automation of the cognitive process,鈥� she said. 鈥淎nd we want Denver to build inclusive communities, and we want to make sure that ethical considerations and responsible AI practices are helping us (address) the vast impact on societal problems. We want to take advantage of every resource we have to solve the hard, complex problems we have.鈥�
Tamara Chuang is a reporter for The Colorado Sun. Her work frequently appears on-air at 萝莉少女 91.5 FM and online at 萝莉少女.org. Contact Tamara at tamara@coloradosun.com