Gas prices have been rising this year. Business Insider made an interactive map to compare average prices.
Frederic J. BROWN / AFP via Getty Images
See how your state's gas prices compare to others with our interactive map.
Gas prices have climbed since late February. The US average jumped from $2.98 to over $4.50 a gallon.
Even as the US leads global oil production, prices remain tied to volatile international markets.
How do gas prices in your state stack up against the rest of the country? Business Insider created an interactive map so you can find out.
Gas prices have been climbing steadily since late February.
The national average was $2.98 a gallon on February 27, according to AAA, and surged over $4 by March 31. By mid-May, average prices were over $4.50.
So far, states on the West Coast have been hit the hardest by rising gas prices. California's average first topped $6 a gallon on April 30.
Some central states like Oklahoma and Kansas have consistently posted the lowest averages during the recent spike.
The differences are stark — but prices have risen dramatically in every state.
Analysts say several factors are driving the increase. Blockades by the US and Iran in the Strait of Hormuz — a key global oil transit route that carries an estimated 20% of the world's oil — have raised global supply concerns.
Spring typically brings a switch to a more expensive summer fuel blend, and recent American Midwest refinery outages have also tightened supply in the Great Lakes region.
Those pressures are pushing prices higher even as the US remains the world's largest oil producer, in part because oil is traded globally.
"The same oil we rely on can be moved anywhere in the world, and we're seeing exports at record levels," Patrick De Haan, a fuel market expert at GasBuddy, told Business Insider. "This is capitalism. When you put your house on the market, you're not like 'Yay, I'm going to take the cheapest bid.'"
The supply and geopolitical pressures could keep prices elevated for months — though there's still significant uncertainty. De Haan said there's a "huge range" of potential summer gasoline prices, with a national average of anywhere from $3.50 to $5.50 a gallon.
Adam Jones used AI to edit social media posts for Myer's Bagels. After customers left bad reviews, he took the posts down and apologized.
Adam Jones
Adam Jones owns Myer's Bagels, a Montreal-style bagel shop that tried using AI to edit its social media posts.
Customers decried the use of AI in the shop's Instagram comments and left negative Google reviews.
Jones apologized and said he'd be more careful with AI and social media — though he's not anti-AI.
This as-told-to essay is based on a conversation with Adam Jones, the 53-year-old owner of Myer's Bagels in Burlington, Vermont. It's been edited for length and clarity.
We specialize in Montreal-style bagel production. For us, it's just north of us. For the rest of the country, it's a newer item. We don't proof them as much, so they don't get as airy. They're denser and a little chewier. We also use wood to cook the bagels.
We're about 22 people, and everyone takes the orders, makes the bagels, and does everything.
You always have to recreate the wheel to stay competitive. Social media is where you need to be. I'm not naive to it, but it's not a stronger part of my experiences. So, I'm learning.
In the past, I've had great employees — i.e., college students — who love social media themselves, so they're happy to take it on. It'd be good for a few months, but then the rest of their lives come into place, and it wasn't a full-time job. It would slowly die. We'd get along fine, but then I'd throw another person on, and the same thing would happen.
Why I used AI for our shop's social media posts
I was pointed to a company that helps small businesses with AI. It helps with human resources. I have a paid accountant and bookkeeper. They can give me their analysis, but why not have a third opinion?
This program does many different aspects of business, and one is social media. It was nice. There's a calendar where you can build a whole month of posts to release on Instagram. We're a college town, so graduation is coming up. I'd ask: What's a neat spin on a post about graduation to encourage someone to come in? The program comes up with an idea.
It allows us to import our photos that we've taken, and it uses those, but it can tweak them.Part of the picture was real. I looked at it and said, "Not too bad." Some of the AI bagels they had in the background didn't look like ours, and I would say, "Can you make that smaller and thinner?"
It didn't take long for our audience to realize what we were posting wasn't 100% Myer's. I appreciate their passion for us not doing this the way they'd like.
AI turned an online review for Myer's Bagels into a handwritten note.
Myer's Bagels
One had a picture of our retail bagsfor the grocery store. Those were actually our photos of the bags. The AI went into our Google, Yelp, or old Instagram comments and pulled out a review from Sam. It put our bags in front of a wood fire that wasn't our fire, and took that quote and made a fake handwritten note.
Was it untruthful? Yes. Was it our fire, image, and background? No. For McDonald's or anyone else who's advertising food, it's in a studio. But Sam is a real person. Sam made that quote, and those were our bags.
We have a picture that we'd used before of our baker rolling out the dough. AI took that photo, superimposed it on a nice wooden cutting board, and then put a fire in the background and a kettle to boil bagels.
That's not how we're laid out, and our customers know that. But it was our real bagels, real hands, real dough.
AI generated a new layout of the Myer's Bagels retail store, something customers noticed.
Myer's Bagels
Between those two photos, I bet there were 25-30 comments for each one, plus people replying.
We had some very passionate, adamant people who went to our Google reviews. They left us one star, claiming nothing about the food, just about this situation. No one called us directly; it was more of a response through social media. The pen is mightier than the sword.
There were a few people who went in to defend us, saying, "Look, they're a small business. They just wanted to be more creative." I appreciate those comments, but certainly the negative responses were much more.
Some were very civil, others I would classify as not very civil. I don't judge on that part. The medium allows it, and for us to be on it, you have to be ready to respond or fall victim to whatever you do.
I apologized — but I'm not anti-AI
We got quite a bit of feedback on a few posts. I went in, pulled them, and apologized.
I wasn't married to the idea. It was a test, or the first foot forward. So, I was willing to erase them and say: "We will do better. We've heard." It's just like, if people don't like a bagel flavor and it doesn't sell, we will make a new bagel flavor.
We're not anti-AI. We will continue to use AI in many aspects of the business, because it's what's going to keep Myer's here for another 30 years. Without extra help, I'm not going to be able to keep up with the rhythm and the pace of growth.
AI is not going away. I look at it as a tool to do things better and easier as a business owner.
It's not good for everything, but it's a tool that has value to make Myer's better for my employees and me. Ultimately, it's also good for the consumers, since it keeps prices down and combats inflation. There's a bigger picture here.
For social media, I think there's still a place for it, but I need to tread much more carefully.
Business Insider reporter Stephen Council outside the federal courthouse in Oakland.
Stephen Council/Business Insider
I've reported on CEO Sam Altman since ChatGPT's release, so I was eager to see his trial testimony.
The trial's stakes are sky-high for Altman as he fights to keep control of OpenAI.
Altman delivered his pitch for the company, though he didn't have answers for the hardest questions.
It's tricky business to define Sam Altman's "day job."
He's got the basic task of overseeing OpenAI's executives and staff, as any CEO does. Work doesn't stay at the office, though. Altman has become the public face of the AI boom, responsible for selling a technology that's upending workforces, capturing markets, and costing billions of dollars to build. That's often a high-wire act, and especially so this week — he also had to sell himself.
I watched Altman's testimony on Tuesday in the Oakland courtroom where, for weeks, jurors have weighed Elon Musk's arguments against those of OpenAI and Microsoft. Musk says Altman and OpenAI president Greg Brockman "stole" the OpenAI charity. They responded that the charity actually thrives because of the commercial moves they made.
It's a high-profile, high-stakes trial, and I was keen to see how Altman would perform. Reporting on OpenAI over the last few years, I've watched him schmooze with San Franciscans, chop it up with AI boosters, and hit softball questions from Jimmy Fallon. He posts to X, joins podcasts, writes blogs — it's all part of his gig. Altman is a content creator, a CEO working to spread hype.
So it was no surprise to me that Altman gave the jury an optimistic vision for both his company and for AI itself. The shocking part was how little he did to protect his own reputation.
The softball questions for Sam Altman ran out
Wearing a dark suit, a thin purple tie, and a slightly furrowed brow, Altman was the day's star witness. He followed OpenAI's lawyer William Savitt through anecdotes, details, and admissions for a breezy first hour of testimony. Then, as the attorney wound down his time, Savitt asked Altman about the OpenAI Foundation's work — the actual goals of the nonprofit at the center of all this legal trouble.
Altman was articulate, and, as usual, forward-looking. He talked about plans to research Alzheimer's and to prepare society for economic change. He expressed hope for a better future, saying that the new tech's benefits should spread widely. He said he was still "extremely" enthusiastic about the company's remaining nonprofit oversight and got off a line about Musk trying to "kill" the very charity he was accusing Altman of "stealing."
Do-gooder ideals, plus optimism about healthcare: It was a rebuttal built on Altman's practiced style of pitching AI to the general public.
The cheery rapport stopped there.
Musk's head lawyer, Steven Molo, then kicked off cross-examination by asking Altman whether he's "completely trustworthy."
"I believe so," Altman said.
The feeling in the courtroom had palpably shifted. Molo responded with a hint of incredulity, "But you don't know whether you're completely trustworthy?"
"I'll just amend my answer to yes," Altman replied.
Under Molo's barrage, Altman had an odd response
Attorneys usually try to strip credibility from an opponent's witness. Molo continued peppering Altman with questions aimed at doing exactly that.
The lawyer said Altman had repeatedly been called "deceptive," citing statements from four other witnesses: OpenAI cofounder Ilya Sutskever, former CTO Mira Murati, and former board members Helen Toner and Tasha McCauley.
Gone was the confident, loquacious CEO. Altman deflected on allegation after allegation, often stating that he hadn't heard the testimony or didn't know specifically what had been said.
Once, he made a larger rebuttal, saying there was "a breakdown in trust between me and the board and a difference — a big difference — of opinion." Then he returned to saying he hadn't heard McCauley's testimony.
Finally, Molo asked, "Is it important to you to find out what's going on in this trial?"
"Yes," Altman responded. "Although I also have a very busy day job and have not been able to be here every day."
He added that he "certainly" cares about the trial. A jury has sat for weeks, and the stakes are huge: Musk wants Altman removed as CEO and seeks $150 billion in damages. Most people I spoke with favor OpenAI to win. Still, Altman will need his credibility for the years ahead.
He stepped off the witness stand around midday, after more hammering from Molo. In the back of the courtroom, I was left thinking that the CEO could have — if only for a week — made the trial his day job.
I used Claude Design and Canva to create the same presentation to see how they stack up.
Aditi Bharade
Claude's new model, Opus 4.7, came with a new tool — Claude Design.
The tool allows users to create slide decks, graphics, and marketing collateral.
I tested how Claude Design compares to the tried-and-true Canva for designing a presentation.
For over a decade, Canva has been one of the most popular graphic design platforms. Now, Anthropic has an opportunity to capture some of that market with AI.
Anthropic announced its latest Opus 4.7 AI model on April 16, and a day later announced the launch of Claude Design, a tool which it said allows users to "create polished visual work like designs, prototypes, slides, one-pagers, and more."
Designs created by Claude Design can be exported directly to Canva, Anthropic said in the release.
The new Claude feature comes as the industry warns of a "SaaSpocalypse," in which AI threatens the business models of software companies like Canva, website builder Wix, and management platforms Workday and Asana.
To better understand the SaaSpocalyse fears, I tested how Claude Design and Canva stack up when tasked with designing the same slide deck.
For this test, I asked Claude Design and Canva's AI tool to create an improved version of the Photography 101 slide deck that I made manually in January.
One of the slides in my original presentation deck. It could definitely use some help.
Aditi Bharade
I used the deck while delivering a photography workshop for reporters at Business Insider's Singapore bureau. Let's just say my photography skills are better than my design skills.
I tested Canva AI first
Canva AI's landing page looked like most other vibe-coding platforms.
Screenshot/Aditi Bharade
Canva, like other software companies, is embracing AI to fend off competition from AI. It's rolling out its AI 2.0 feature to all users, turning Canva into an agentic platform that can generate editable designs.
I tested out AI 1.0, then got an early trial also to see how 2.0 stacks up against Claude.
I submitted my prompt, asking for a presentation split into six sections with quizzes to test understanding, and a dark minimalist design with blue accents.
I liked that Canva confirmed the presentation's flow before generating it.
Screenshot/Aditi Bharade
Canva then asked me to choose the audience type — casual, professional, or educational — as well as the deck's style and length. And before designing the slides, Canva presented a rough outline of what each slide would include, which I could edit.
I thought that was neat.
The slides were decent, but required editing
AI 1.0 gave me two designs to choose from. One was an immediate no, with loud elements that clashed and made the text hard to read. The graphics used were also rather simplistic.
I didn't love how the elements overlapped with the text.
Screenshot/Aditi Bharade
The second design was better, but its text boxes overlapped, the text was misaligned, and it was at times minuscule. It also didn't have as many photos as I'd like. I fed Canva another prompt asking it to align and standardize the text boxes, add more pictures, and make it more engaging.
The final product I got from Canva AI was much better than what I had created myself.
Screenshot/Aditi Bharade
The final product was much better than the bad deck I made in January, something I could definitely use for a refresher workshop.
I repeated the same test a couple of days later with AI 2.0. The process was largely unchanged, except that I could manually change elements while chatting with the AI tool — a useful change from 1.0.
Canva AI 2.0 allowed me to use the chatbot and manually edit simultaneously.
Screenshot/Aditi Bharade
The design from AI 2.0 was a bit boring, so I asked it to regenerate the slide deck with more images and a cleaner aesthetic.
It gave me a warning — "This will use a significant amount of your AI usage since I'll need to regenerate each slide" — but I ended up with a slightly more polished presentation.
Canva AI 2.0 redesigned the slides when I asked it to, and I preferred the second set a lot more.
Screenshot/Aditi Bharade
I laughed when the AI got tired of redesigning and told me the redesign was complete, even though four slides remained unchanged. When I prompted it to look again, it hit me with the good old, "You're right, sorry about that!"
It took a couple of prompts to get it to where I wanted, but I was impressed by Canva's AI generation.
We were definitely off to a good start.
Then I tried Claude Design
Next, I turned to Claude Design. I had previously tried generating a logo of my initials with Opus 4.7, and the design left a lot to be desired.
I was initially not too impressed with Claude's design skills.
Screenshot/Aditi Bharade
So I was not expecting a lot when I asked it to create the photography presentation.
To my surprise, Claude asked me in-depth follow-up questions before starting to design. It asked what camera I would be using, the skill levels of participants — beginners, phone-photo experienced, hobbyists — and even the exact shade of blue I wanted as an accent color.
And the deck it generated surpassed my expectations.
I thought the graphics that Claude Design generated were clear and helped explain concepts.
Screenshot/Aditi Bharade
For example, one slide featured a graphic of the "exposure triangle," which illustrated how ISO, aperture, and shutter speed affect a photo's brightness.
Claude fixed errors without any prompting
Claude Design began editing its own work without my input.
Screenshot/Aditi Bharade
But what impressed me the most about Claude Design was that it anticipated my needs.
In the first version, some text boxes overlapped. Without prompting, Claude started identifying and fixing these problems, acting as its own editor without needing any input from me.
The only problem I ran into with Claude Design was when I tried to change an image in one of the quiz slides. The question was about photographing a politician at a podium, but the picture was of a young man sitting in front of a circular window.
I asked Claude to give me a relevant picture.
Screenshot/Aditi Bharade
I asked Claude to change it to a picture of a politician.
Instead, it generated the image of a farmer. I tried again — it gave me a picture of the Senate. Third time, a handshake. At that point, I gave up and asked it to revert to the picture of the guy in front of the window.
Claude didn't really get what I meant by "politician at a podium."
Screenshot/Aditi Bharade
It gave me some insight into the criticisms of Claude's Opus 4.7, with users saying it burns through tokens too quickly and sometimes gives ridiculous answers.
However, it responded better to other prompts. For example, it initially lumped the five editing tips into one slide, but expanded each tip into its own slide seamlessly when I asked it to.
Canva's final product was good, but in my view, Claude Design's was better and required a lot less prompting on my part.
Final thoughts
At the end of building the slide decks, I remember thinking to myself, "I'm never making slides from scratch again."
It was a close competition, but Claude Design edged itbecause it identified its own errors and corrected them without prompting, while Canva needed to be told what to fix.
One consideration — there is a difference in the subscription costs for both platforms. An individual subscription to Claude Pro costs $17 a month, and a Pro Max subscription costs $100 a month and allows higher token usage.
Meanwhile, Canva AI is free for use, for up to "200 Standard AI uses," per its pricing chart. Canva Pro, which offers "10x more AI than Canva Free," costs $18 a month.
And of course, if you can't choose between the two, you can use both together.
"Now, when someone creates an idea or a draft in Claude, they can instantly take it into Canva and turn it into something real where it's fully editable, collaborative, and on brand," Canva's spokesperson said. "From there, it's easy to refine, adapt across different formats, and publish or share it anywhere."
Anthropic did not respond to requests for comment from Business Insider.
Dean Grey, an entry-level software engineer, is participating in a new UBI and upskilling program.
Grey said entry-level jobs have evaporated, and hundreds of applications went nowhere.
Grey said the mentorship and the $1,000 monthly stipend renewed his sense of hope and structure.
This as-told-to essay is based on a conversation with Dean Grey, a participant in a program by AI Commons that aims to upskill and retrain workers who have been displaced by AI alongside a monthly stipend. Grey is an entry-level software engineer who is struggling with his job search.
I never imagined I would become an early participant in what feels like a social experiment for the AI era. But after spending years trying to break into tech, only to watch entry-level jobs evaporate almost as quickly as I got there, that's exactly where I found myself.
I grew up in Massachusetts and graduated from UMass years ago. Before tech, I worked mostly in management jobs, "just sort of trucking along," as I like to joke — though eventually that became literal. I spent about 4 years working in trucking, but I knew I wanted something more stable and better paying, and tech seemed like the future. Shows like "Mr. Robot" and movies like "Hackers" made the industry feel exciting and creative.
So I saved up enough money to make the leap. I enrolled in Hack Reactor, a coding boot camp program, where I learned software engineering. At the time, I thought I was doing everything right.
I trained for a tech career as the market collapsed
After boot camp, I joined a training-to-hire company that promised to prepare people like me for jobs with large clients. The understanding was that we'd move directly into industry work afterward.
By the time my training ended, the market had changed dramatically. Hiring had slowed. Companies were suddenly obsessed with AI. Instead of moving directly into work, many of us were put on waiting lists for months. I waited about five months before eventually landing a contract position with Infosys.
It quickly became clear there wasn't much future there. Our contracts were finite, and there was really no upward mobility. I had put all of my eggs in this basket, and it was disappointing.
When that contract ended, I entered what became one of the hardest periods of my life.
I got a handful of phone screenings and two Zoom interviews that never progressed beyond the first round. At one point, it was exciting to get a rejection because a lot of companies were just ghosting me.
I burned through savings, called in favors, took gig work and temporary jobs, filled out online surveys — basically anything I could do to survive. Unemployment benefits helped for a while, but not enough to build a future.
The UBI program gave me structure and hope
A project Dean Grey is working on with the AI Commons, a high is a chatbot that helps newly unemployed individuals navigate their next steps.
Dean Grey
When the AI Commons approached me about joining a new basic income pilot program for workers displaced by AI, I was mostly grateful not to feel alone anymore.
The program provides a monthly stipend — up to $1,000 — along with technical training, mentorship, engineering projects, and community support.
The project is still so new, and the curriculum is still developing, but it gave me a sense of hope and structure again, which matters more than people realize.
Every morning starts with standups where we discuss what we're building and what's blocking us. I meet weekly with mentors and pair-program with more experienced engineers. For the first time in a long time, I feel like I'm moving forward instead of standing still.
The financial support has also been such a boon and life-changing. Being behind on all my bills has been a crushing stress. But with this basic income, I'm able to focus on improving my résumé, working on projects, attending meetings, and building my network.
The stipend doesn't feel like permission to stop working. If anything, it's the opposite. It gives me enough stability to actually invest in myself again.
Now I'm building an AI chatbot to help laid-off workers
One of the biggest projects I'm working on now is an AI-powered chatbot for people navigating layoffs and unemployment.
The idea is simple: when someone loses a job, they often don't even know what questions to ask first. My chatbot is designed to guide them through the process, explain unemployment laws by state, connect them to resources, and eventually help direct them toward real human support systems.
I don't know exactly what the future of work looks like. Honestly, sometimes it feels bleak. But this program has convinced me that people can adapt if they're given enough support, time, and community to do it.
I hear a lot of criticism of UBI, mostly about how it could erode agency and meaning and might be taken advantage of.
I'm optimistic because, at the end of the day, it might just mean we transition to a new way of finding purpose, if not in what we do for a living, then in how we express ourselves or what we do creatively. I don't think that we'll ever run out of ways to find purpose in life.
"Shark Tank" investor Kevin O'Leary is planning a data center project in Utah.
Michael Buckner/Penske Media via Getty Images
Kevin O'Leary's Utah data center project may create fewer jobs than he initially said.
Wonder Valley's data center job estimates are closer to 4,000 construction roles.
Local residents have opposed the data center due to concerns about water use and air quality.
Kevin O'Leary has said his Utah data center development will create 10,000 construction jobs and 2,000 permanent roles.
There's one problem: It probably won't.
The "Shark Tank" investor may have overstated the data center's hiring potential during the construction phase, Business Insider learned in a conversation with O'Leary Ventures CEO Paul Palandjian.
The accurate estimate is closer to 4,000 new construction jobs over 10 to 15 years — and that isn't a guarantee.
"Look, these numbers are fluid, and they change by the day," Palandjian said, adding that the updated estimate is reflective of "our current thinking on the project."
Dubbed Wonder Valley, O'Leary's combined data center and power plant has the potential to reach 9 gigawatts of capacity. That would make Wonder Valley one of the world's biggest data centers if fully built out.
Data centers generally do not create large numbers of permanent, full-time jobs in local economies. Their impact on local economies is harder to measure, in part because the nascent industry is still being studied. They do come with an enormous demand for temporary skilled labor during the construction phase.
Once a data center of this scale is fully built and operational, the on-site workforce shrinks by an average of 78%, researchers at the University of Southern California Marshall School of Business found.
Using the USC researchers' formula, a more likely estimate for permanent jobs at Wonder Valley is 1,350.
"It's all supply and demand-based, so we've analyzed it for its scaled potential," Palandjian said of his job estimate.
An FAQ sheet for the project, found on Utah Gov. Spencer Cox's website, said that the developer for Wonder Valley has "committed to a projected 2,000 permanent jobs in skilled trade, logistics, IT, and administrative positions to county residents."
Whether Wonder Valley reaches that scaled potential is an open question.
The data center doesn't have a tenant yet, though Palandjian said O'Leary Ventures is in "very early talks" with multiple large tech companies.
Wonder Valley's proposed site occupies a portion of a broader 40,000-acre zone backed by Utah's Military Installation Development Authority, a state agency tasked with raising tax revenues through economic growth.
A final development agreement is still being negotiated and is expected to be signed "in the next month or so," a MIDA spokesperson wrote in an email to Business Insider.
A draft version of the agreement posted on MIDA's website, dated April 24, shows that the data center will receive multiple tax breaks, including a 100% personal property tax rebate and an 80% real property tax rebate, for up to 30 years.
The agreement does not specify a required minimum number of jobs. Utah offers tax breaks to data centers without a job-creation requirement.
The Box Elder County community has expressed fierce opposition to the data center, citing concerns about water use, air quality, and lack of transparency from local officials. Residents of the county filed two referendums on Monday to overturn local officials' approval of the site.
O'Leary has dismissed the protesters, suggesting they were bused in from out of town, and has argued that their claims are based on misconceptions about data centers.
"Think about the number of jobs," O'Leary said in a post on X on Friday.
Thinking Machines Lab CEO Mira Murati and Meta CEO Mark Zuckerberg.
Getty Images
Thinking Machines Lab has lost a third of its founding team to rivals amid fierce AI talent wars.
Meta, OpenAI, and xAI lured the founders of Thinking Machines Lab with huge compensation offers.
CEO Mira Murati's startup has also hired top talent and recently unveiled a new type of AI model.
Thinking Machines Lab has raised billions in capital while assembling one of AI's most elite technical teams. Now it's watching some of that talent walk out the door.
As Big Tech rivals dangle eye-popping pay packages and early employees unlock their first slice of equity, the buzzy, 1-year-old startup has become ground zero for tech's escalating talent wars.
Nearly a third of its 42-person founding team — 13 people — have left since its launch, a Business Insider review of LinkedIn profiles and conversations with four sources found. Those departures include three of its six co-founders.
Thinking Machines Lab's head count has more than quadrupled to over 150 people since it launched, a person familiar with the matter told Business Insider.
The departures have been driven by a mix of factors, though a major one is the huge packages offered by rival AI labs like Meta and OpenAI, as well as the passage of an important milestone in startup compensation: the one-year cliff, four sources familiar with the matter said.
It's not unheard of for early members of a startup to jump ship over time, though Thinking Machines Lab was supposed to be different. It emerged from stealth last year as one of the highest-profile AI labs thanks to its leadership's deep ties to OpenAI. CEO Mira Murati previously served as OpenAI's chief technical officer, while other key figures helped train the early versions of ChatGPT.
Investors eager to back the next breakout AI company flocked to the startup, which raised $2 billion before launching a product.
The startup's reputation as a talent hub has put a target on its back.
Thinking Machines Lab declined to comment.
An opportunity they couldn't refuse
Some of the offers to founding members have reached well into nine figures — hundreds of millions of dollars in cash and stock over several years — according to one defector who shared notes with peers.
"I got an opportunity that I couldn't turn down," the person said.
Meta — which considered buying Thinking Machines Lab last year, The Verge reported — has been the most aggressive poacher, quietly luring away seven founding team members, plus a star AI researcher.
Such was the case of veteran software engineer Joshua Gross. Last year, Gross was deep in the trenches of AI development at Thinking Machines Labs, helping launch the flagship product, Tinker.
Gross "had a blast" shaping Tinker's early vision, he wrote in a LinkedIn post at the time. After helping build and ship the product from "zero to one," he's no longer at Thinking Machines Lab. Wired first reported on details of the compensation packages.
Subject line: $1.5 million cash 'and up'
The race for elite AI talent has become hotter than ever across tech, said Sam Agre, cofounder of recruiting firm People In AI.
While the pool of people with hands-on experience building leading AI models and products is small, it runs deep at Thinking Machines Lab. It's so difficult to poach rank-and-file workers from there that Agre said he has resorted to sending them LinkedIn messages with subject lines touting $1.5 million in cash compensation "and up."
That's more than three times the annual salary for software engineers offered by Thinking Machines Labs on its careers page, which mentions a range of $350,000 and $475,000 a year.
While the nine-figure offers are reminiscent of deals for major athletes, Big Tech firms like Meta covet people who can help them pull ahead in AI, Agre said.
"You don't want to fall behind in that type of arms race," he said.
Meta isn't the only tech giant circling Thinking Machines Lab. OpenAI has hired five founding members, and Elon Musk's xAI has poached one more. OpenAI and Meta declined to comment, and xAI didn't respond to a request for comment.
One-year cliffs opened the floodgates
Meta first began trying to lure away Thinking Machines Lab talent when it launched its superintelligence team last year. It scored a major coup by bringing on cofounder Andrew Tulloch. OpenAI brought on two more of Murati's cofounders, CTO Barret Zoph and researcher Luke Metz.
Then this year, the floodgates opened.
Because the company launched a little over a year ago, many founding employees started hitting their "one-year cliffs" — a standard startup threshold that unlocks the first chunk of equity after 12 months, after which it vests monthly.
Once they get their shares, employees can leave without walking away empty-handed, making it a natural moment for Thinking Machines Lab staffers to entertain outside offers, according to three people familiar with the matter.
One-year cliffs are supposed to help retain talent, though they have become something of an Achilles' heel for startups ever since the AI boom began, said Dan Walter, an independent compensation consultant. He told Business Insider that retention rates in tech startups are the lowest he's ever seen due to the ease of building a startup and rampant poaching.
"The entire model is being challenged," he said.
Some in the compensation industry are now arguing for the cliff to start at five years, not one, to lock in early team members, he added.
Thinking Machines Lab still attracts top talent
Despite the rampant poaching, Thinking Machines Lab has scored wins.
Its CTO — Soumith Chintala, the creator of the open-source AI project PyTorch — left Meta to join Murati's startup. Thinking Machines Lab regularly brings on research talent from Meta, including Kenny Yu, a member of the tech giant's hotshot TBD lab — a small team of AI all-stars launched for its superintelligence push.
Murati's startup has been putting that talent to use. On Monday, it announced that it had built a new type of AI model that interacts with people seamlessly, handles interruptions, and translates languages in real time. The model will launch more widely later this year.
The company is also taking steps to stem the losses. It's hiring a person to build a fresh framework for doling out equity and to implement systems that "attract and retain highly sought-after talent," a job listing says. The position, which pays between $250,000 and $425,000 in base salary, was posted in March.
For all the departures, Murati's reputation as a talent magnet hasn't faded.
When Business Insider asked Carina Hong, CEO and founder of Axiom Math — already valued at $1.6 billion — which company tries to poach most of her researchers, she didn't hesitate: Thinking Machines Lab.
