AI can do a lot, but Cyara CEO Sushil Kumar has always been grounded in what chatbots actually are: “It’s nothing but a glorified autocorrect or autocomplete,” he told me. And considering how autocorrect notoriously gets things wrong (I battled with it for years about how my last name is spelled), Kumar says we shouldn’t be surprised when an AI chatbot hallucinates information that just isn’t true.
Still, today’s AI chatbots are a marked improvement over their predecessors that were used just a year ago, and Kumar estimates they’re correct about 95% of the time. This is an admirable accuracy rate, but the problem is AI can’t differentiate between the right answers and the wrong ones users get 5% of the time. And those wrong answers can land in professional writing—like legal filings or research papers—or could lead to flawed research on strategy, finances or planning for a company.
I talked to Kumar about what can be done on the enterprise level to minimize the hallucinations that get in the way of work. An excerpt from our conversation is later in this newsletter.
Until next time.
Samuel Boivin/NurPhoto via Getty Images
Another thing commonly known by all with an interest in tech: AI has a tight—and likely not entirely healthy—hold on the stock market. This week saw a sizable tech stock selloff based on infrastructure cost concerns, as well as fear that chipmaker Micron—seen as a bellwether for AI demand—would not meet quarterly expectations.
When Micron’s earnings report came out after markets closed on Wednesday, the story was the same as other times markets have been jittery about AI. Micron smashed expectations, reporting $41 billion in revenue. It was the best quarter ever for the chip company, beating their prior quarter by 74%, and hitting close to 4.5 times their revenue from a year ago. Early Thursday trading saw most indexes recovering.
Micron is the largest U.S.-based chipmaker, and landed at No. 88 on the 2026 Forbes Global 2000, published this week. Forbes’ Rashi Shrivastava writes that the company is slowly scaling up, ensuring that supply will not exceed demand. In a release about earnings, Micron CEO Sanjay Mehrotra said that he believes the company will stay strong in the future, especially through multi-year strategic customer agreements.
However, investor drama over AI companies is not likely to go away on Wall Street. South Korean memory chip firm SK Hynix, a Micron competitor, filed paperwork this week with the Securities and Exchange Commission for a Wall Street debut as soon as next month.
And another future tech firm is headed to Wall Street: humanoid robot company Agility Robotics. The maker of Digit, a robot that can move and stack heavy containers that has been working in logistics since 2024, will go public through a merger with special purpose acquisition company Churchill Capital Corp. XI, writes Forbes senior contributor John Koetsier. The deal will value Agility at $2.5 billion, and the company expects the deal will net more than $620 million in gross proceeds. Agility plans to use the money for R&D and scaling up.
Agility also teased its planned newest model: Digit v5. This model is designed to be an AI-driven humanoid robot “cobot,” Koetsier writes—meaning it can safely work alongside humans. This new model, set to be released this year, already has $300 million in multi-year orders, and more than 30 customers.
Koetsier writes that Agility’s entry to the stock market will make it one of very few publicly traded pure-play robotics companies worldwide. This designation will provide a new window into the robotics business, letting both potential customers and interested investors peek under the hood at whether the investment is worthwhile.
Fisk University, a small HBCU in Nashville, will give its students a new kind of hands-on tech experience in college, writes Forbes’ Asia Alexander. The school’s $1 billion campus revitalization plan, called “Quantum Leap,” has a $400 million Innovation Center that includes a 70,000-square-foot data center. University President Agenia Walker Clark says about a third of Fisk undergraduates major in computer science, making the tech center a natural fit for the campus. It was first suggested by a Fisk alumna working in Silicon Valley, and Alexander writes the university studied whether a data center could coexist with the campus. Clark said it can be safe and clean—and renting out space in the data center could bring the college much needed revenue.
Cyara CEO Sushil Kumar.
Cyara
At this point, everyone who uses AI knows that it may hallucinate, present invented facts and records, and offer conclusions that don’t make sense. Yet these hallucinations still make it into actual work products, including legal filings and research papers submitted to journals. I talked to Sushil Kumar, CEO of AI-powered CX platform Cyara, about the technical and human solutions to try to keep work products hallucination-free. This conversation has been edited for length, clarity and continuity.
Why are people who should know better allowing hallucinations into their work products?
Kumar: We know there will be hallucinations. The question is: AI does so many things that are good otherwise. For the sake of hallucination, do you stop using it? Second, what is the cost of preventing hallucination? You have to build a very sophisticated agent that has the right context and guardrails: The engine itself doesn’t know that it is hallucinating, so there has to be a guardrail that says, ‘This is where you are wrong.’
Building those guardrails are not easy. They tend to be very customized to a specific use case. Those guardrails exist within the LLMs themselves. They become essentially mini agents. It’s the specific information about your enterprise.
Companies are often stuck between a rock and a hard place: Should I prevent everybody from using this very powerful technology until it is perfection, or should we take a chance and try to educate people that hallucination could be there: You trust, but verify.
The challenge with that approach is if you go to a non-technical person, the information is presented by AI in such a credible manner. It’s hard to go and figure out if this is not right. You almost need to have a sixth sense to know, ‘Hey, this doesn’t look right.’ You have to find more sophisticated people who know how to use AI better.
I tell people: AI can do 60% of your job, but 40% is yours. Twenty percent of that job is making sure that you form the right prompt: not ambiguous, very clear and very surgical in what it wants so that AIi doesn’t take liberties. The other 20% is validating that response. If people take that approach, you can do a lot and overcome hallucination—provided you have a workforce that knows how to use AI with responsibility. That requires people who love AI, but also know AI has limitations.
Do you see a future where there will not be hallucinations because AI has been taught it can’t make up citations or facts?
Yeah, that is the future. Unfortunately, that future may take a little bit longer to get to. Not many people are capable of hosting [their own domain], so you don’t want your data to be training the model. That means the model itself doesn’t get better over interactions in your custom domain. Maybe the guardrails get stricter.
Right now, the models are humongous. Gradually, you’ll see more small- and medium-sized models being developed. It’ll be more possible to host them privately and train them. Then there’ll be a loop every time a user says you made a mistake; the LLM takes that into account to correct it.
It’s not something that I’m holding my breath for. Gradually, the accuracy percentage will definitely increase. When ChatGPT-3 came out, the hallucination rate was 20% to 25%. Now it has come down significantly. But the last mile—going from 98% to 100%—will probably be the hardest. For 2% gain, you will have to make a disparate amount of investment to make sure you have a perfect model.
What advice would you give CIOs who want to make sure people at their company stop allowing hallucinations in their work products?
It’s the combination of tooling, training and education for people to understand what AI is good at and where it needs help.
There’s a lot of buzz about AI agents, and they are surprisingly easy to spin up. But before you allow an army of agents, consider how much human supervision they need in order to keep on task. In the case of agents, less may be more.
It’s been a crazy year, and things are unlikely to settle down any time soon. Your mental health should be one of your top priorities, so when you plan this year’s summer getaway, you should consider a “quietcation,” which gives you a chance to relax and recover.
A member of Congress was caught using Claude to summarize text of a proposed amendment, leaving the words “Claude responded” in the text. Who was it?
A. Rep. Cliff Bentz, R-Ore.
B. Rep. Anna Paulina Luna, R-Fla.
C. Rep. Tracey Mann, R-Kan.
D. Rep. Marlin Stutzman, R-Ind.
See if you got the answer right here.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。