• Amazon invests billions of dollars in robots to boost e-commerce efficiency and profitability.
• Back in 2015, the Amazon Picking Challenge tried to spur more research into warehouse automation.
• The competition inspired some of the company's most advanced robots, including Sparrow and Robin.

Amazon is investing billions of dollars in robots to make its e-commerce business more efficient and profitable. This huge initiative started out a lot smaller.

A decade ago, the company launched a competition for university engineering teams called the Amazon Picking Challenge. It called on researchers to design robots for a common warehouse task: Grabbing products from a shelf and putting them in a box.

As a tech reporter, this quirky project intrigued me. At the time in early 2015, Google was testing self-driving cars, a technology that emerged from a similar academic competition known as the DARPA Grand Challenge. What if Amazon was trying to replicate this magic, but with robots rather than automobiles?

Then, a funny thing happened. The Amazon Picking Challenge faded away. It was renamed and only lasted a few years. I chalked this up to another bad call and moved on.

I only thought about this challenge again late last year. That's when Amazon unveiled a next-generation warehouse in Louisiana that has 10 times more robots moving products around and, yes, picking them up with dexterity. The facility processes orders 25% faster and 25% more efficiently, and it will likely be the future of the company's e-commerce operation.

Ten years after the Amazon Picking Challenge, the fruits of this nerdy competition have finally emerged. It follows an uncannily similar timeline to the DARPA Grand Challenge, which started in 2004 and resulted in Google's driverless cars hitting the road roughly a decade later.

So, with the help of Business Insider reporter Eugene Kim, I investigated how Amazon's huge new fleet of picking robots came to be, and how this competition laid the foundation for a new wave of automation that's about to crash over the warehouse and logistics industry.

From pallets to picking

It started with an acquisition. In 2012, Amazon paid $775 million for Kiva Systems, which designed flat robots that zip around warehouse floors. 

This helped move pallets of goods around, but humans still needed to pick items. Getting a robot to spot the correct product in a box, then grab it just hard enough to pick it up, but not damage it — that's incredibly difficult.

This is where the Amazon Picking Challenge came in. Instead of hacking away at this problem itself, the company wanted to focus the broader academic community on the task.

The risk was that any valuable inventions would be out in the public sphere, and Amazon might not directly benefit from them. But the potential gains were much bigger, according to executives and roboticists.

"Amazon doesn't compete with robotics companies," said former Amazon Robotics chief Brad Porter, who runs robotics startup Cobot now. "When facing an unsolved research problem in robotics AI like bin picking, Amazon benefits if anyone solves that problem as long as Amazon can get access to the technology to improve their operations."

"The challenge Amazon was trying to solve was how to motivate researchers to focus on this problem," Porter added. "The Picking Challenge very much succeeded in doing that."

Oreos, Sharpies, and dog toys

The first competition took place over two days in late May 2015 in Seattle, with more than 25 teams from colleges including MIT, Duke, Rutgers, and Georgia Tech.

The contestants had to design a robot that could pick products from a typical shelf found on a Kiva Systems warehouse pod, and then put those items into containers. The picker had to be fully autonomous, and each robot was given 20 minutes to pick 12 target items from the shelves. Contestants had to open-source their creations.

Companies, including ABB, Fanuc, and Rethink Robotics, founded by industry pioneer Rodney Brooks, provided hardware for contestants to repurpose and tinker with.

The products were a preselected set of 25 items commonly sold on Amazon.com, including packs of Oreo cookies, boxes of Sharpie pens, and dog toys.

Some were easier to pick. There were simple cuboids, like a box of coffee stirrers or a whiteboard eraser. Others were trickier. For instance, a box of Cheez-Its could not be removed from the bin without first tilting it, adding another complex step for the robots. Smaller items, such as an individual spark plug, were more difficult to detect and properly grasp.

Vacuum arms and 'catastrophic failure'

Among all 26 teams, a total of 36 correct items were picked, versus seven incorrect items. Another four products were dropped by robots in the competition.

About half of the teams scored zero points, and two teams couldn't get their robots working well enough to even attempt the challenge, according to a research paper analyzing the results.

Problems ranged from the highly technical to the mundane. Some of the same items came packed differently, which made them even more difficult to pick. One team's machine had a vacuum hose that got accidentally wound around the robotic arm.

With each system having hundreds of components, the failure of any one of these could lead to "catastrophic failure of the overall system — as witnessed during the competition," the researchers wrote. 

The main finding from this first Amazon Picking Challenge was that human warehouse workers were a lot better than machines at picking products.

"A human is capable of performing a more complex version of the same task at a rate of ∼400 sorts/hour with minimal errors," the researchers wrote. "While the best robot in the APC achieved a rate of ∼30 sorts/hour with a 16% failure rate."

But the conclusion was hopeful, too: The contest showed that robotics could substantially increase warehouse automation and order fulfillment in the near future.

The competition was renamed the following year as the Amazon Robotics Challenge, and the tasks evolved to be more complex.

Suction and other benefits

Tye Brady, chief technologist at Amazon Robotics, was involved in those later Amazon Robotics Challenges. 

In a recent interview with Business Insider, he said research on robotic manipulation exploded from 2016 through 2018, with many institutions publishing their results and insights. This helped spread valuable knowledge across the industry, speeding up progress. 

At least two professors started graduate-level classes related to Amazon's challenge, and these programs are still churning out experts with valuable practical applied knowledge in robotics, Brady explained. 

"When you get a whole bunch of smart people together in a room and think about focused problems, some great things are going to happen, and that's really what happened," he said. "It inspired a lot of the work that we have today that we see in, for example, our Sparrow and Robin manipulation systems that are real-world products delivering packages inside of our fulfillment centers."

In that first competition in 2015, some robotics teams used grippers that mimicked the way a human hand picks things up. Other teams tried suction instead, with some researchers even strapping off-the-shelf vacuum cleaners to their robots.

Gripping proved more problematic because the robots didn't receive enough information to know when to release or add pressure at the right times. This could result in squashed or crushed products or dropped items.

Sucking the items up so they stuck to the end of robot arms was a more successful approach.

"The idea of high flow suction was novel. Bring your favorite vacuum cleaner and start picking up objects. That was kind of clever," Brady said. "This idea, we used suction inside of our Robin and our Sparrow arms. It's very good."

The boss has noticed

Amazon unveiled Robin, its first robotic arm, in 2021. This machine picks up packages from conveyor belts and places them on other mobile robots called Pegasus.

Sparrow followed in 2023. This was Amazon's first robotic arm to handle individual items rather than packages. It uses computer vision and AI to pick more than 200 million different items from containers and place them into totes.

Amazon CEO Andy Jassy has taken notice. At the AWS re:Invent conference in December, he should have been talking about cloud computing. But he took time away from that subject to wax lyrical about Sparrow. 

"It has to discern which item is which. It has to know how to grasp that item, given the size of it and the materials and the flexibility of that material. And then it has to know where in the receiving bin it can put it," Jassy said. "These are all inventions that are critical to us changing the processing time and the cost to serve our customers."

Wall Street has noticed, too. Morgan Stanley recently estimated that Amazon's warehouse robots could save the company as much as $10 billion a year.

"The big story is we're just getting started," said Brady.

• The rise of AI agents could upend how the internet works.
• The technology has the potential to disrupt Big Tech relationships with consumers.
• AI agents could change consuming content, organizing daily tasks, and making decisions.

In the future, you might not read this column on Business Insider's website or app. Instead, your very own AI agent could read it to you as part of a bespoke smorgasbord of daily content, suggestions, decisions, and actions that make your life easier and more organized.

Silicon Valley is all aquiver about this agentic vision as the third year of the generative AI boom gains momentum.

2025 has been hailed as the year of AI agents—personalized digital assistants that can interact with users, do research, gather information, curate content, and ultimately anticipate your needs and get things done before you even ask.

As usual with techno-futurist predictions, this may not come to pass. Generative AI might not progress as quickly as hoped. However, if the technology becomes capable of such feats, it will upend the internet and could disrupt some of the companies that dominate our current digital world.

"There's an idea we can't seem to shake," Bernstein's Mark Shmulik and Nikhil Devnani, two of the top internet analysts, wrote in a recent note to investors. "If AI agents truly become useful, the internet will go dark."

Websites and apps won't go away; it's just that for many of them, consumers won't visit or see these digital locations directly. Instead, they will access information, content, and widgets through an AI assistant that becomes "the aggregator of the aggregators," the analysts said.

"If it scales and plays out like we think it might, this. Changes. Everything. The aggregators get disaggregated, and much of consumer internet may be structural shorts. Welcome to the Agentic AI era," they wrote. "There's nowhere to hide."

Traveling could become easier

The Bernstein analysts cited an example of flying to New York and needing to get from the airport to the office.

Do you really care whether you take an Uber, a Lyft, a Waymo, a cab, or a generic black car service? Probably not. What you really want is the fastest, cheapest, most comfortable ride into Manhattan.

What if your personal AI agent could sort this whole thing out for you? That would radically change the way the internet works. No need to "Google" anything. You might not even need to take out your smartphone (if we even have phones in this agentic future).  

"The aggregators have control over the supply, but if demand consolidates and gets fulfilled through an AI agent, you may never need to open your rideshare app again!" the Bernstein analysts wrote. 

A new top-of-the-funnel

This could be the ultimate top of the funnel. An AI agent representing each of us would become a powerful new direct connection that tech companies could forge with consumers. All other providers would be funneled through this new digital gate and would likely have to pay some sort of toll—just as Google collects tolls right now on the web through Search ads and Apple collects tolls via App Store fees.

"If you extrapolate these dynamics to their end state, AI Agents could truly disintermediate the aggregators by becoming pseudo marketplaces in their own right," Shmulik wrote. 

Big Tech companies and startups are already furiously jostling for control of this future agentic funnel. 

In late January, OpenAI unveiled Operator, an AI agent system that uses a web browser to take action on behalf of users, such as booking travel reservations and buying them products.

Users can select a specific website through which they want to process their requests, such as OpenTable, or send the request through a search engine like Google. The key here is that the direct relationship is between the OpenAI agent and the user. Previously, this online journey would probably have started with a Google Search. Now, in the future, Google is just one of many services that OpenAI's Operator might choose.

Google is not waiting around to become just another app on someone else's AI agent platform, though. In December, the company showed off Project Mariner, an AI agent that can browse the web and take actions such as clicking buttons and filling out forms. 

Back in October, OpenAI rival Anthropic unveiled a similar tool rolled out as a test feature called "computer use," which enables its AI model, Claude 3.5 Sonnet, to use a computer similarly to humans.

Some Anthropic engineers asked to order enough food to feed a group, and this new AI agent tool selected pizza. Alex Albert, Anthropic's head of developer relations, said the tool navigated DoorDash online, and "about a minute later, we saw Claude decided to order us some pizzas." On Monday, Anthropic launched an updated model, Claude 3.7 Sonnet, with "extended thinking" capabilities.

Other tech companies, including AI startup Perplexity, have similar offerings.

These AI agents may be delivered via new, voice-based devices. Meta has wearable technology such as its goggles and smart glasses. Apple has put its brand of AI on many of its devices and is working on updates to the Vision Pro goggles. Google is busy baking its Gemini AI models into millions of Android phones and Chromebooks.

An even more unbreakable digital connection?

Is this the ultimate top-of-the-funnel technology? Maybe not.

Elon Musk's Neuralink aims to put chips inside humans' brains. That could create an unbreakable bond with consumers, as the chip would read their thoughts, desires, and needs directly from their heads rather than inferring what people want from Google Searches, Apple app behavior, and social media posts.

This might sound like wild sci-fi dreams, though Google executives have talked about this idea for years.  

In 2010, Hal Varian, Google's chief economist, discussed it with The Atlantic, saying a brain implant could be a logical next step for the company's search engine. 

"Now you Google things on your computer. And you Google things on your phone. That's the next stage. And I believe — people may laugh — but I think there will be an implant," he said. 

• OpenAI said this week that DeepSeek may have used OpenAI model outputs "inappropriately."
• OpenAI has been accused of doing the same thing with copyrighted content.
• What's that old saying about karma? Go ask ChatGPT. Or DeepSeek.

In the brave new world of generative AI, there's a moment that everyone will experience: the realization that your original work is being used to train AI models that could be competing against you.

This moment has arrived for OpenAI and its CEO, Sam Altman.

The US startup said this week that the Chinese lab DeepSeek may have "inappropriately" used OpenAI outputs to train new AI models in a process called distillation.

Translation: We think you used our content without permission, and that's not allowed.

For some AI experts, these complaints are hypocritical. OpenAI's success is built on a similar process. The startup has for years collected outputs and data from the internet and used that to train its own models. This includes scooping up copyrighted content and other original work from thousands of companies that have not authorized this use.

In fact, this is what most model developers do, according to Nick Vincent, an assistant professor of computer science at Simon Fraser University who studies how data is used in AI.

"These firms are simultaneously arguing for the right to train on anything they can get their hands on while denying their competitors the right to train on model outputs," he wrote in a blog post this week, adding, "Rules for thee, but not for me?"

DeepSeek = just deserts 

Vincent sees the rise of DeepSeek as the inevitable outcome of a training-data free-for-all where AI companies take whatever content they want and ask for forgiveness later.

This has now backfired on OpenAI, which may be having its own outputs plundered in the name of AI progress. The startup "will struggle to defend itself in the court of public opinion on this," Vincent told Business Insider on Wednesday. "There's a reckoning coming."

He hopes this reckoning will encourage tech companies to create a new system that gives appropriate credit and compensation to content creators. 

"So far, none of the AI labs have seriously thought about this, so DeepSeek is their just deserts," Vincent added.

Fair use just for OpenAI, or for everyone?

High-quality training data is a crucial ingredient of powerful AI models. Many of the companies that created this information want to be paid for providing intelligence to these new products. Tech companies don't want to be forced to pay. This dispute is being fought in court.

OpenAI is being sued by authors who claim the startup is breaking copyright law by using their books to train AI models. The New York Times is pursuing a similar complaint.

OpenAI has also been accused of using YouTube content to train its Sora video-generation model. YouTube's CEO, Neal Mohan, said last year that this would violate the video platform's rules.

OpenAI has denied breaking copyright laws, citing the "fair use" doctrine, which allows unlicensed use of copyrighted works in certain situations, including teaching, research, and news reporting.

So would DeepSeek's use of OpenAI's outputs also constitute fair use?

"Very potentially, yes," Vincent said.

Fair use isn't just for yourself when it's convenient. That would be, well, unfair.

I asked OpenAI about all this on Wednesday and it didn't respond. The startup has partnerships with some companies that authorize the use of their content for AI model training. Axel Springer, BI's owner, struck one of these deals in 2023.

Distillation and karma

How do AI model outputs get scooped up for competitive means anyway?

Distillation is the technical term for extracting intelligence buried in one model and weaving it into a new one, Vincent said. AI godfathers, including Geoffrey Hinton, wrote a research paper about this in 2015 called "Distilling the Knowledge in a Neural Network."

Back then, the researchers described a tamer version of this where a lab or company would take its own old models and use their outputs to cleverly infuse a new offering with more intelligence.

Distilling intelligence from someone else's AI model without permission is frowned upon in some research circles but happens a lot, Vincent said.

DeepSeek's research paper about its new R1 model described using distillation with open-source models, but it didn't mention OpenAI.

"We demonstrate that the reasoning patterns of larger models can be distilled into smaller models, resulting in better performance," the Chinese lab's researchers wrote. 

Since these new offerings began rolling out late last year, some AI researchers have theorized that DeepSeek used outputs from OpenAI's new "reasoning" model, called o1, as synthetic data to improve its own models, such as R1.

In December, when DeepSeek was beginning to wow the AI field, Altman seemed to take a dig at his new rival.

"It is (relatively) easy to copy something that you know works," he wrote on X. "It is extremely hard to do something new, risky, and difficult when you don't know if it will work."

What's that phrase about karma? I can't write it here. If you don't know, go ask ChatGPT. Or DeepSeek.

• DeepSeek's AI models rival top Silicon Valley offerings, excelling in some complex tasks.
• The models use inference-time compute, breaking queries into smaller, manageable tasks.
• DeepSeek's DeepThink mode shows its thought process, solving challenging math problems.

Chinese AI lab DeepSeek recently released AI models that match or exceed some of Silicon Valley's top offerings.

DeepSeek uses an approach called test-time or inference-time compute, which slices queries into smaller tasks, turning each into a new prompt that the model tackles. Each step requires running a new request, which is known as the inference stage in AI.

Business Insider recently tested one of DeepSeek's models using its DeepThink mode, which shows users every step of its thought process.

We gave the model a multistep math problem that was suggested by Charlie Snell, an AI researcher at UC Berkeley.

Snell said DeepSeek performs well on complex math problems that must be solved by thinking for longer and in stages. The researcher picked problems from the American Invitational Mathematics Examination, a challenging test for high school math whizzes.

"I put questions from that into the DeepSeek model," Snell told BI in an interview. "I read through the chain of thought. It was understandable."

A DeepSeek demo

Here's the math problem Snell suggested for BI's DeepSeek demo: "Find a sequence of +, -, /, * which can be applied to the numbers 7, 3, 11, 5 to get to 24, where each of the given numbers is used exactly once."

BI put that prompt into DeepSeek's chat window on its website. The model responded initially by laying out the challenge ahead.

"Alright, so I've got this problem here: I need to use the numbers 7, 3, 11, and 5, and combine them with the operations of addition, subtraction, multiplication, and division, using each number exactly once, to get to 24," it replied. "At first glance, this seems a bit tricky, but I think with some systematic thinking, I can figure it out."

It then proceeded through multiple steps over roughly 16 pages of discussion that included mathematical calculations and equations. The model sometimes got it wrong, but it spotted this and didn't give up. Instead, it swiftly moved on to try another possible solution, then another.

"Almost got close there with 33 / 7 * 5 ≈ 23.57, but not quite 24. Maybe I need to try a different approach," it wrote at one point.

Later on, the DeepSeek model seemed to catch itself repeating a potential solution.

"Wait, I already did that one," the model wrote. "Okay, maybe I need to consider using division in a different way."

After a few minutes, it found the correct answer.

"You can see it try different ideas and backtrack," Snell said. He highlighted this part of DeepSeek's chain of thought as particularly noteworthy:

"This is getting really time-consuming. Maybe I need to consider a different strategy," the AI model wrote. "Instead of combining two numbers at a time, perhaps I should look for a way to group them differently or use operations in a nested manner."

• The Chinese AI lab DeepSeek has rolled out AI models that are a lot cheaper than OpenAI's offerings.
• Its emergence spooked tech stocks Monday, with Nvidia down more than 10% premarket.
• DeepSeek's open-source models challenge OpenAI's proprietary approach.

The cost of using AI models has been plunging as competition intensifies — and Wall Street is spooked about the latest entrant.

Having a leading AI model is nowhere near as special as it was two years ago. These models are mostly trained on data that's publicly available on the internet, so they're not that much different from one another. Standing out in this crowd and charging premium prices is increasingly difficult.

In recent weeks, DeepSeek has taken this pricing dynamic to a new level. The Chinese AI lab rolled out models that are as good as, or better than, the best products from OpenAI, the pioneering creator of ChatGPT.

That's impressive. But what's potentially more disruptive is how cheap DeepSeek's models are.

That disruption hit the markets Monday. The Dutch chip equipment provider ASML fell 11% after the opening bell, and Nvidia fell more than 10% premarket. In Japan, Advantest, TOWA Shibaura, and Ibiden were all down at the closing bell.

Bernstein tech analysts studied DeepSeek's offerings in recent days and found that the Chinese AI lab was massively undercutting OpenAI on price.

"DeepSeek's pricing blows away anything from the competition, with the company pricing their models anywhere from 20-40x cheaper than equivalent models from OpenAI," the analysts wrote in a note to investors on Sunday.

The chart above shows the cost of "tokens," which have become the raw material of generative AI. Chatbots and AI models break down words and other inputs into these tokens to make them easier to process and understand. One token is about three-quarters of a word.

When AI companies are handling prompts and other model inputs and outputs, they often charge users based on a per-token price. DeepSeek's Reasoner model costs $0.55 for every million tokens inputted. OpenAI's o1 model charges $15 for the same number of tokens, according to Bernstein.

The Bernstein analysts also noted that DeepSeek's models are open-source, meaning they are available free to anyone who wants to work with them. That's a contrast with OpenAI, which keeps its top models proprietary and closed while charging relatively high prices for the products.

"The whole thing brings up some very interesting questions about the role and viability of proprietary versus open-source efforts that are probably worth doing more work on," the Bernstein analysts wrote.

But they added, "The resulting Twitterverse panic over the weekend seems overblown."

Edison Lee, an equity research analyst at Jefferies, said it was "difficult to say what the eventual impact will be, but combining our more cautious cyclical view on the chip sector in any case, it may be prudent to reduce exposure to the digital part of the semiconductor supply chain including semicap equipment stocks like ASML."

DeepSeek's newly-released flagship model is now No. 1 on Apple's free apps chart — just ahead of ChatGPT.

• Tesla sold 38,965 Cybertrucks in 2024, according to Cox Automotive estimates.
• The Cybertruck's high price and slowing demand contributed to underwhelming sales figures.
• Despite challenges, the Cybertruck was the top-selling electric pickup, surpassing the Ford F-150 Lightning.

I once had a reservation to buy a Tesla Cybertruck. I canceled it last year and ended up buying a Model 3.

It was December and Tesla was offering juicy year-end incentives that won me over. I picked up the car at a Tesla dealership south of Silicon Valley.

Before driving off, I saw about 20 Cybertrucks lined up in the parking lot. I asked a Tesla employee about these vehicles and he said roughly half of them were spoken for. So about 10 were unsold with three weeks left of the year.

Sales numbers

I don't know what happened to those specific Cybertrucks, but overall sales of Elon Musk's angular, divisive electric pickups underwhelmed in 2024.

Cox Automotive recently released EV sales numbers for the US market. This firm estimates that Tesla sold 38,965 Cybertrucks last year.

In late 2023, I collated some Wall Street forecasts for Cybertruck sales and found that Tesla was expected to sell about 48,500 units in 2024. That was the average of estimates from analysts at Goldman Sachs, Morgan Stanley, UBS, Bernstein, and Wedbush.

So, Tesla missed by almost 10,000 units. I asked Cox about this and Stephanie Valdez Streaty, the firm's director of industry insights, got back to me.

"Demand for the Cybertruck showed signs of slowing down towards the end of the year and was lower than many Tesla enthusiasts and boosters might have expected," she said.

Cox also adjusted its third-quarter Cybertruck sales number to be lower than initial estimates.

"The Cybertruck's current price point — more than $100,000 according to Kelley Blue Book estimates — will make it difficult to achieve the volume sales that were promised prior to launch," she added. Similar electric trucks from Ford and Rivian cost quite a bit less.

A little context

The Cybertruck is unique and has some groundbreaking technology. Some people love the design.

It ended up being the fifth best-selling EV in the US last year, according to Cox. It beat the Ford F-150 Lightning, making the Cybertruck the top electric pickup in terms of sales.

The top EV was the Model Y, followed by the Model 3. Together, those vehicles accounted for more than 600,000 sales in 2024.

So, the Cybertruck it has not been a massive flop, but it hasn't been a runaway hit either.

The outlook

More than 1 million people put $100 down to reserve a Cybertruck after it was first announced in late 2019.

It's unclear how many people canceled their reservations, like me.

"The electric pickup truck market is becoming increasingly competitive, with new models from established automakers," Valdez Streaty said. "This heightened competition will continue to challenge the Cybertruck's market share."

"We are likely through many of the early Tesla Cybertruck adopters. Initial demand was driven by enthusiasts and early adopters," she added. "The focus now is on appealing to a broader market to sustain sales momentum."

• Elon Musk and Jeff Bezos exchanged friendly tweets after rocket launches on Thursday.
• The billionaires have had a long-standing rivalry over space ambitions and more.
• Their recent interactions suggest a possible end to their long public feud.

Elon Musk and Jeff Bezos may no longer be enemies. In fact, a friendship may be budding between the two billionaires, judging by tweets they recently posted on X.

Both tech founders own rocket companies that launched missions on Thursday, with mixed results.

SpaceX's huge Starship spaceship exploded, although its Super Heavy rocket booster successfully returned to a platform on Earth.

"@elonmusk Kudos to you and the whole SpaceX team on the flawless booster catch! Very impressive." Bezos wrote on X.

Earlier in the day, Bezos' Blue Origin launched its 32-story-tall New Glenn rocket into orbit. It lost its booster during the test flight.

"Well done, @JeffBezos, and the Blue Origin team!" Musk wrote.

Is their long feud over?

Followers on X were quick to notice the sea change in the relationship between the billionaires.

Musk and Bezos have been embroiled in a years-long feud.

They've argued over their space ambitions at SpaceX and Blue Origin as they've repeatedly leapfrogged each other for the title of the world's richest person. That's a title Musk currently holds by a massive margin: With Bloomberg estimating his net worth at $439 billion, he's by far the richest person in the world. Bezos is second, with an estimated net worth of $240 billion.

Musk has said Blue Origin has tried to poach talent from SpaceX, and he's taunted Bezos repeatedly about the status of their respective businesses.

In 2015, when SpaceX landed its Falcon 9 spacecraft, Bezos needled Musk on Twitter by suggesting Blue Origin had already accomplished a similar feat.

Bezos has also subtly criticized Musk's mission to colonize Mars over the years. When NASA picked SpaceX for a major contract in 2021, Blue Origin was furious and protested the decision. 

Less than two months ago, they were at it again. In an X post, Musk said Bezos had told people to dump their shares in his companies because Donald Trump was bound to lose the presidential election.

Hours later, Bezos said the claim was "100% not true."

Billionaire "Step Brothers"

This history is what has made Thursday's friendly repartee between the billionaires so jarring.

Musk posted a gif on X from the movie "Step Brothers," with the main characters saying, "Did we just become best friends? Yep!"

"Step Brothers is the perfect meme for @JeffBezos & me," Musk wrote. 

He posted another clip from the same movie, implying his relationship with Bezos was now harmonious. 

Of course, all of this is unfolding on social media, which has its limitations. But one thing is for sure: No matter what's happening behind the scenes or offline, Thursday's exchanges mark a change in Bezos and Musk's online rapport.

• Jack Clark was a reporter at Bloomberg when I was an editor there.
• He told me he was quitting to join OpenAI in 2016.
• I told him that was a terrible idea. The rest is history.

In 2016, Jack Clark walked up to me in Bloomberg's San Francisco newsroom and asked if we could go for a walk. As an editor, it's often not good when one of your reporters makes a request like this.

Sure enough, as we sat on a bench looking over the bay, Jack told me he was quitting to join a nonprofit called OpenAI.

I said this was a terrible idea. OpenAI was less than a year old at the time and was still a relatively obscure AI research group. Its major claim to fame was Elon Musk's (uneven) financial support.

I pressed my case. As a reporter on Bloomberg's Big Tech team, Jack had a pretty stable job. In contrast, OpenAI didn't seem to have much of a direction, and I couldn't see a path for it to become financially sustainable beyond asking Musk for more money. I selfishly also wanted Jack to stay at Bloomberg and keep covering Google and AI, which he was good at.

I thought I was pretty persuasive, but Jack ignored me and left.

"Just read the research papers"

He went on to be an influential expert and advisor on AI safety and related topics, co-authoring several AI research papers. Jack also built one of the most popular AI email newsletters, called Import AI, which researchers widely follow in the field. He still writes this regularly.

He often told me to "just read the research papers" when I asked how to learn more about AI and get better stories about the technology. He was right. There's a lot of valuable information buried in these papers.

Jack stayed at OpenAI for over four years, doing strategy and communications before becoming a policy director. He may have gotten some equity in that startup, but I'm not sure.

Then, in 2020, he left OpenAI and I didn't hear from him for a while. He popped up a few months later as one of seven cofounders of Anthropic, which was started by a bunch of early OpenAI employees.

Cofounders reminisce

Anthropic is now challenging OpenAI at the forefront of generative AI and large language models. It's backed by Amazon and Google, along with several top venture capital firms.

The cofounders got together last month to talk about the start of Anthropic. Jack holds court with his colleagues, who reminisce about the early days.

"I met Dario in 2015 when I went to a conference you were at, and I tried to interview you, and Google PR said I would've read all of your research papers," Jack says to Dario Amodei, CEO of Anthropic, who used to work at Google.

"I think I was writing 'Concrete problems in AI safety' when I was at Google," Amodei replies. "I think you wrote a story about that paper."

"I did," Jack says, with a cheeky smile.

Not his style

Earlier this month, the Wall Street Journal reported that Anthropic was raising money at a $60 billion valuation. Then, Forbes reported that the seven cofounders, including Jack, are set to become billionaires.

I asked Jack about this last week and said I wanted to interview him for a story. 

"Haha, Ali, thanks, but really not my style," he replied.

It's true. Jack is among the gentlest, kindest, and most self-deprecating people I've ever met. He's not classic billionaire material. 

I'm still stunned and trying to process his new situation. What I do know is that Jack's decision to ignore me was a testament to his passion, single-mindedness, and vision.

Back in 2015, when very few people thought about AI, he was obsessed with it and was constantly pushing to write about the technology at Bloomberg. 

Jack knew that AI was important. When his chance came, he took a risk and went for it.

• The AI industry has hit "peak data," OpenAI cofounder Ilya Sutskever said recently.
• DeepMind researchers see outputs from new "reasoning" models as a source of fresh AI training data.
• A new AI technique, known as test-time compute, will be put to the test in 2025.

OpenAI cofounder Ilya Sutskever announced something at a recent conference that should have had the AI industry trembling with fear.

"We've achieved peak data and there'll be no more," he intoned during a speech at the annual Neurips event in December.

All the useful data on the internet has already been used to train AI models. This process, known as pre-training, produced many recent generative AI gains, including ChatGPT. Improvements have slowed, though, and Sutskever said this era "will unquestionably end."

That's a frightening prospect because trillions of dollars in stock market value and AI investment are riding on models continuing to get better.

Yet, most AI experts don't seem that worried. Why?

Inference-time compute

There may be a way to get around this data wall. It's related to a relatively new technique that helps AI models "think" through challenging tasks for longer.

The approach, called test-time or inference-time compute, slices queries into smaller tasks, turning each into a new prompt that the model tackles. Each step requires running a new request, which is known as the inference stage in AI.

This produces a chain of reasoning in which each part of the problem is tackled. The model doesn't move on to the next stage until it gets each part right and ultimately comes up with a better final response.

OpenAI released a model called o1 in September that uses inference-time compute. That was followed swiftly by Google and Chinese AI lab DeepSeek, which rolled out similar "reasoning" models. 

"An iterative self-improvement loop"

Benchmark-based testing of these new models has shown that they often generate better outputs than the previous top AI crop, especially on math questions and similar tasks with clear final answers.

This is where things get interesting. What if these higher-quality outputs were used for new training data? This mountain of new information could be fed back into other AI model training runs to produce even better results.

Google DeepMind researchers published research on test-time compute in August and proposed this technique as a potential way to keep large language models improving through the peak-data wall.

"In the future, we envision that the outputs of applying additional test-time compute can be distilled back into the base LLM, enabling an iterative self-improvement loop," the researchers wrote. "To this end, future work should extend our findings and study how the outputs of applying test-time compute can be used to improve the base LLM itself."

A chat with a test-time researcher

The authors were Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Xu is still at Google, and Kumar spends some of his time at DeepMind, while Lee left to join OpenAI rival Anthropic.

Snell co-wrote the paper while interning at Google DeepMind. He's back at UC Berkeley now, so I called him up to ask what inspired the research.

"I was motivated by some of the things that have been preventing pre-training from continuing to scale, notably the finite supply of data," he told me in a recent interview. 

"If you can get an AI model to use extra inference-time compute and improve its outputs, that's a way for it to generate better synthetic data," he added. "That's a useful new source of training data. This seems to be a promising way to get around these pre-training data bottlenecks."

Satya satisfied

On a recent video podcast, Microsoft CEO Satya Nadella seemed unperturbed and even buoyant when asked about the slowdown in AI model improvement and the lack of new quality training data. 

He described inference-time compute as "another scaling law."

"So you have pre-training, and then you have effectively this test-time sampling that then creates the tokens that can go back into pre-training, creating even more powerful models that then are running on your inference," he explained.

"That's I think a fantastic way to increase model capability," Nadella added, with a smile. 

Sutskever also mentioned test-time compute as one possible solution to the peak-data problem, during his Neurips talk in early December.

Test time for test-time compute

2025 will see this approach put to the test. It's not a slam-dunk, although Snell is optimistic. 

"Over the last three years or so, it seemed more clear," he said of AI progress. "Now we're in this exploratory mode."

One open question: How well does this test-time compute technique generalize? Snell said it performs well with questions where the answer is knowable and you can check it, such as a math challenge.

"But a lot of things that need reasoning aren't easy to check. For instance, writing an essay. There's often no straight answer on how good this is," he explained. 

Still, there are early signs of success and Snell suspects outputs from these types of reasoning AI models are already being used to train new models. 

"There's a good chance that this synthetic data is better than what's out on the internet," he said.

If outputs from OpenAI's o1 model are better than GPT-4, the startup's previous top model, then these new outputs can in theory be reused for future AI model training, Snell explained. 

He shared a theoretical example: Say o1 gets a 90% score on a particular AI benchmark, you could take those answers and feed them to GPT-4 and get that model up to 90%, too. 

"If you have a large set of prompts, you could get a bunch of data from o1 and create a large set of training examples and pre-train a new model on them, or continue training GPT-4 to be better," Snell said. 

A TechCrunch report from late December suggested that DeepSeek may have used outputs from OpenAI's o1 to train its own AI model. Its latest offering, called DeepSeek V3, performs well on industry benchmarks. 

"They were probably the first ones to reproduce o1," Snell said. "I've asked people at OpenAI what they think of it. They say it looks like the same thing, but they don't how DeepSeek did this so fast."

OpenAI and DeepSeek didn't respond to requests for comment. 

• Generative AI is transforming technical tasks, making them accessible to non-experts.
• AI tools like v0 and Julius AI streamline processes such as web development and data analysis.
• Vercel's CFO uses generative AI tools to become a "quasi-coder."

The AI boom has added trillions of dollars to tech company valuations. Is it living up to the hype?

In some real ways, the answer is yes. This is especially true when it comes to the technical plumbing of modern companies. These are tasks that often go on behind the scenes and are either unknown or taken for granted by most non-technical people.

Generative AI burst onto the scene in late 2022 with OpenAI's release of ChatGPT, a chatbot that answers many questions and creates realistic and convincing content.

Since that flashy launch, this new form of AI has quietly begun to transform more mundane jobs and processes, such as web development, data analysis, legal research, and code writing.

At Vercel's Next.js conference in San Francisco earlier this year, the event was packed with young developers who were using AI models and tools to streamline hundreds of these technical tasks. This stuff has mostly been run by human technical employees. Now, that's changing in major ways.

"All the power was previously behind a gate guarded by programmers who were paid hundreds of thousands of dollars a year. Now, these capabilities are available to all," said attendee Rahul Sonwalkar, founder of Julius AI, a startup that's using AI models to automate data analysis.

Saving on legal fees

It's not just startups. A good friend who's an executive at an investment fund used ChatGPT recently to research a legal issue.

The chatbot helped him understand a lot of the background, including relevant laws and other rules. 

When he met with his law firm, he was able to jump past the basics and get to the meat of the task more quickly. This is important when attorneys can cost $500 to $1,000 an hour. 

My friend estimates this initial AI-powered research saved his investment firm $50,000 to $70,000 in legal fees and roughly 60 to 80 hours of work time over 2 months.

20x more code at Google

At Google, generative AI is upending how the internet giant creates products.

Another old friend of mine has worked at Google for well over a decade. He recently described how he's writing 20 times more software code than he used to, thanks to generative AI tools.

He starts in the usual way, by typing in some initial code. Then the AI autocompletes much of the rest.

The technology sometimes autocompletes in the wrong direction — essentially misunderstanding his intentions. He still needs the technical skills to spot these occasional mistakes. But fixing it is pretty straight forward: He goes back to where his own code ended and types a bit more of his own work. Then the system adjusts and completes the task accurately.

A CFO becomes a 'quasi-coder'

Vercel CFO Marten Abrahamsen is no professional coder. But even he's experienced the benefits of generative AI making technical tasks more accessible.

He cited Vercel's v0 service, which lets anyone type in English language requests and responds with code and outputs such as brand new websites.

"I can't do complex coding, but I can type in English and v0 creates what I want. This turns me into a quasi-coder," Abrahamsen said.

The CFO said this tool helps him get ideas in front of more technical colleagues quicker, and ensures the nascent products are in better shape at the pitch stage.

Vercel's goal is to use generative AI to increase "iteration velocity" by automating a lot of the technical blocking and tackling so developers can spend more time on the creative parts of their jobs, he explained.

"Making developers much more productive with generative AI — investors and Vercel are quite bullish on this. That's a very interesting new use case for AI," Abrahamsen told me in a recent interview.

Creating a website in 2 minutes or less

I tried v0 myself on Friday. It took about 45 seconds to create a website based on this simple request: "Make me a website that looks like Business Insider."

Vercel's v0 system responded in English with the steps it would take. Then, on the right hand side of the page, it swiftly pumped out the required software code and previewed the new website in less than a minute. Here's a look:

I asked for a little tweak: "Make the background more blue and add photos."

v0 responded with a similar English language answer, followed by more code generation and an updated site.

I then asked to make the top of the site blue and the system added that in maybe 20 seconds.

I could go on, but you get the point. I can't code at all, and I made a relatively solid website in about 2 minutes with v0.

2 million lines of code a day

Julius AI is taking a similar approach to automate data-analysis tasks. The service is used by scientists, marketing folks, hedge fund analysts and anyone else who needs to interpret a lot of data and isn't an expert at pulling such insights from mountains of information. 

The online tool can ingest data in many forms, including Excel tables and PDFs, or via APIs and databases. You can drag and drop these into an open window and ask questions in plain English. Julius AI then taps into various AI models to spot correlations in the data and generate insights in seconds via charts and text outputs. 

The service automatically generates the software code needed to do this analysis, and makes that available to re-use on other projects. This also helps users go back and check how the outputs were created. 

Julius AI has about 2 million registered users and has pumped out more than 7 million data visualizations so far, according to Sonwalkar, who notes the service writes roughly 2 million lines of code a day.

"It would take an army of human coders to do that," he said. "A good engineer who's focusing on a good day can put out about 1,000 lines of code."

Quantitative hedge funds use Julius AI to create financial models from the data they drop into the tool. One model might factor in currency changes and how that impacts other parts of the world, such as oil and gas prices, for instance. 

One Julius AI customer is a hedge fund with seven employees who are finance experts. 

"Normally this firm would also hire a quantitative programmer to create financial models for data analysis," Sonwalker said. "AI does this in seconds now, without the need for a programming expert." 

• OpenAI's o1 model was hailed as a breakthrough in September.
• By November, a Chinese AI lab had released a similar model called DeepSeek.
• On Thursday, Google came out with a challenger called Gemini 2.0 Flash Thinking.

In September, OpenAI unveiled a radically new type of AI model called o1. In a matter of months, rivals introduced similar offerings.

On Thursday, Google released Gemini 2.0 Flash Thinking, which uses reasoning techniques that look a lot like o1.

Even before that, in November, a Chinese company announced DeepSeek, an AI model that breaks challenging questions down into more manageable tasks like OpenAI's o1 does.

This is the latest example of a crowded AI frontier where pricey innovations are swiftly matched, making it harder to stand out.

"It's amazing how quickly AI model improvements get commoditized," Rahul Sonwalkar, CEO of the startup Julius AI, said. "Companies spend massive amounts building these new models, and within a few months they become a commodity."

The proliferation of multiple AI models with similar capabilities could make it difficult to justify charging high prices to use these tools. The price of accessing AI models has indeed plunged in the past year or so.

That, in turn, could raise questions about whether it's worth spending hundreds of millions of dollars, or even billions, to build the next top AI model.

September is a lifetime ago in the AI industry

When OpenAI previewed its o1 model in September, the product was hailed as a breakthrough. It uses a new approach called inference-time compute to answer more challenging questions.

It does this by slicing queries into more digestible tasks and turning each of these stages into a new prompt that the model tackles. Each step requires running a new request, which is known as the inference stage in AI.

This produces a chain of thought or chain of reasoning in which each part of the problem is answered, and the model doesn't move on to the next stage until it ultimately comes up with a full response.

The model can even backtrack and check its prior steps and correct errors, or try solutions and fail before trying something else. This is akin to how humans spend longer working through complex tasks.

DeepSeek rises

In a mere two months, o1 had a rival. On November 20, a Chinese AI company released DeepSeek.

"They were probably the first ones to reproduce o1," said Charlie Snell, an AI researcher at UC Berkeley who coauthored a Google DeepMind paper this year on inference-time compute.

He's tried DeepSeek's AI model and says it performs well on complex math problems that must be solved by thinking for longer and in stages. 

He noted that in DeepSeek's DeepThink mode, the model shows users every step of its thought process. With o1, these intermediate steps are hidden from users. 

"I've asked people at OpenAI what they think of it," Snell told BI. "They say it looks like the same thing, but they don't how DeepSeek did this so fast."

OpenAI didn't respond to a request for comment. On Friday, the startup previewed an o1 successor, called o3. Francois Chollet, a respected AI expert, called the update a "significant breakthrough."

Andrej Karpathy, an OpenAI cofounder, praised Google's new "Thinking" model for the same reasoning feature.

"The prominent and pleasant surprise here is that unlike o1 the reasoning traces of the model are shown," he wrote on X. "As a user I personally really like this because the reasoning itself is interesting to see and read — the models actively think through different possibilities, ideas, debate themselves, etc., it's part of the value add."

A DeepSeek demo

Snell shared a multistep math problem with Business Insider, which we used to test DeepSeek for ourselves: 

"Find a sequence of +, -, /, * which can be applied to the numbers 7, 3, 11, 5 to get to 24, where each of the given numbers is used exactly once."

BI put that prompt in DeepSeek's chat window on its website. The model responded initially by laying out the challenge ahead.

"Alright, so I've got this problem here: I need to use the numbers 7, 3, 11, and 5, and combine them with the operations of addition, subtraction, multiplication, and division, using each number exactly once, to get to 24," it replied. "At first glance, this seems a bit tricky, but I think with some systematic thinking, I can figure it out."

It then proceeded through multiple steps over roughly 16 pages of discussion that included mathematical calculations and equations. The model sometimes got it wrong, but it spotted this and didn't give up. Instead, it swiftly moved on to another possible solution. 

"Almost got close there with 33 / 7 * 5 ≈ 23.57, but not quite 24. Maybe I need to try a different approach," it wrote at one point. 

After a few minutes, it found the correct solution. 

"You can see it try different ideas and backtrack," Snell said in an interview on Wednesday. He highlighted this part of DeepSeek's chain of thought as particularly noteworthy:

"This is getting really time-consuming. Maybe I need to consider a different strategy," the AI model wrote. "Instead of combining two numbers at a time, perhaps I should look for a way to group them differently or use operations in a nested manner."

Then Google appears

Snell said other companies are likely working on AI models that use the same inference-time compute approach as OpenAI.

"DeepSeek does this already, so I assume others are working on this," he added on Wednesday.

The following day, Google released Gemini 2.0 Flash Thinking. Like DeepSeek, this new model shows users each step of its thought process while tackling problems. 

Jeff Dean, a Google AI veteran, shared a demo on X that showed this new model solving a physics problem and explained its reasoning steps. 

"This model is trained to use thoughts to strengthen its reasoning," Dean wrote. "We see promising results when we increase inference time computation!"

• Suchir Balaji helped OpenAI collect data from the internet for AI model training, the NYT reported.
• He was found dead in an apartment in San Francisco in late November, according to police.
• About a month before, Balaji published an essay criticizing how AI models use data.

The recent death of former OpenAI researcher Suchir Balaji has brought an under-discussed AI debate back into the limelight.

AI models are trained on information from the internet. These tools answer user questions directly, so fewer people visit the websites that created and verified the original data. This drains resources from content creators, which could lead to a less accurate and rich internet.

Elon Musk calls this "Death by LLM." Stack Overflow, a coding Q&A website, has already been damaged by this phenomenon. And Balaji was concerned about this.

Balaji was found dead in late November. The San Francisco Police Department said it found "no evidence of foul play" during the initial investigation. The city's chief medical examiner determined the death to be suicide.

Balaji's concerns

About a month before Balaji died, he published an essay on his personal website that addressed how AI models are created and how this may be bad for the internet. 

He cited research that studied the impact of AI models using online data for free to answer questions directly while sucking traffic away from the original sources.

The study analyzed Stack Overflow and found that traffic to this site declined by about 12% after the release of ChatGPT. Instead of going to Stack Overflow to ask coding questions and do research, some developers were just asking ChatGPT for the answers. 

Other findings from the research Balaji cited: 

• There was a decline in the number of questions posted on Stack Overflow after the release of ChatGPT.
• The average account age of the question-askers rose after ChatGPT came out, suggesting fewer people signed up to Stack Overflow or that more users left the online community.

This suggests that AI models could undermine some of the incentives that created the information-rich internet as we know it today.

If people can get their answers directly from AI models, there's no need to go to the original sources of the information. If people don't visit websites as much, advertising and subscription revenue may fall, and there would be less money to fund the creation and verification of high-quality online data.

MKBHD wants to opt out

It's even more galling to imagine that AI models might be doing this based partly on your own work. 

Tech reviewer Marques Brownlee experienced this recently when he reviewed OpenAI's Sora video model and found that it created a clip with a plant that looked a lot like a plant from his own videos posted on YouTube. 

"Are my videos in that source material? Is this exact plant part of the source material? Is it just a coincidence?" said Brownlee, who's known as MKBHD.

Naturally, he also wanted to know if he could opt out and prevent his videos from being used to train AI models. "We don't know if it's too late to opt out," Brownlee said.

'Not a sustainable model'

In an interview with The New York Times published in October, Balaji said AI chatbots like ChatGPT are stripping away the commercial value of people's work and services.

The publication reported that while working at OpenAI, Balaji was part of a team that collected data from the internet for AI model training. He joined the startup with high hopes for how AI could help society, but became disillusioned, NYT wrote. 

"This is not a sustainable model for the internet ecosystem," he told the publication.

In a statement to the Times about Balaji's comments, OpenAI said the way it builds AI models is protected by fair use copyright principles and supported by legal precedents. "We view this principle as fair to creators, necessary for innovators, and critical for US competitiveness," it added.

In his essay, Balaji disagreed.

One of the four tests for copyright infringement is whether a new work impacts the potential market for, or value of, the original copyrighted work. If it does this type of damage, then it's not "fair use" and is not allowed. 

Balaji concluded that ChatGPT and other AI models don't quality for fair use copyright protection. 

"None of the four factors seem to weigh in favor of ChatGPT being a fair use of its training data," he wrote. "That being said, none of the arguments here are fundamentally specific to ChatGPT either, and similar arguments could be made for many generative AI products in a wide variety of domains."

Talking about data

Tech companies producing these powerful AI models don't like to talk about the value of training data. They've even stopped disclosing where they get the data from, which was a common practice until a few years ago. 

"They always highlight their clever algorithms, not the underlying data," Nick Vincent, an AI researcher, told BI last year.

Balaji's death may finally give this debate the attention it deserves. 

"We are devastated to learn of this incredibly sad news today and our hearts go out to Suchir's loved ones during this difficult time," an OpenAI spokesperson told BI recently. 

If you or someone you know is experiencing depression or has had thoughts of harming themself or taking their own life, get help. In the US, call or text 988 to reach the Suicide & Crisis Lifeline, which provides 24/7, free, confidential support for people in distress, as well as best practices for professionals and resources to aid in prevention and crisis situations. Help is also available through the Crisis Text Line — just text "HOME" to 741741. The International Association for Suicide Prevention offers resources for those outside the US.

• Generative AI tools have made it easier to create fake images, videos, and audio.
• That sparked concern that this busy election year would be disrupted by realistic disinformation.
• The barrage of AI deepfakes didn't happen. An AI researcher explains why and what's to come.

Oren Etzioni has studied artificial intelligence and worked on the technology for well over a decade, so when he saw the huge election cycle of 2024 coming, he got ready.

India, Indonesia, and the US were just some of the populous nations sending citizens to the ballot box. Generative AI had been unleashed upon the world about a year earlier, and there were major concerns about a potential wave of AI-powered disinformation disrupting the democratic process.

"We're going into the jungle without bug spray," Etzioni recalled thinking at the time.

He responded by starting TrueMedia.org, a nonprofit that uses AI-detection technologies to help people determine whether online videos, images, and audio are real or fake.

The group launched an early beta version of its service in April, so it was ready for a barrage of realistic AI deepfakes and other misleading online content.

In the end, the barrage never came.

"It really wasn't nearly as bad as we thought," Etzioni said. "That was good news, period."

He's still slightly mystified by this, although he has theories.

First, you don't need AI to lie during elections.

"Out-and-out lies and conspiracy theories were prevalent, but they weren't always accompanied by synthetic media," Etzioni said.

Second, he suspects that generative AI technology is not quite there yet, particularly when it comes to deepfake videos. 

"Some of the most egregious videos that are truly realistic — those are still pretty hard to create," Etzioni said. "There's another lap to go before people can generate what they want easily and have it look the way they want. Awareness of how to do this may not have penetrated the dark corners of the internet yet."

One thing he's sure of: High-end AI video-generation capabilities will come. This might happen during the next major election cycle or the one after that, but it's coming.

With that in mind, Etzioni shared learnings from TrueMedia's first go-round this year:

• Democracies are still not prepared for the worst-case scenario when it comes to AI deepfakes.
• There's no purely technical solution for this looming problem, and AI will need regulation. 
• Social media has an important role to play. 
• TrueMedia achieves roughly 90% accuracy, although people asked for more. It will be impossible to be 100% accurate, so there's room for human analysts.
• It's not always scalable to have humans at the end checking every decision, so humans only get involved in edge cases, such as when users question a decision made by TrueMedia's technology. 

The group plans to publish research on its AI deepfake detection efforts, and it's working on potential licensing deals. 

"There's a lot of interest in our AI models that have been tuned based on the flurry of uploads and deepfakes," Etzioni said. "We hope to license those to entities that are mission-oriented."

• Vercel said it added Steffan Tomlinson to its board.
• Tomlinson is the CFO of Stripe and has experience taking tech startups public.
• He used to be CFO at several other tech companies, including Palo Alto Networks and Confluent.

Vercel, an AI startup valued at more than $3 billion, just bulked up its board with the addition of a finance executive who has experience taking tech companies public.

Stripe Chief Financial Officer Steffan Tomlinson will serve as a director on Vercel's board, the startup said on Tuesday.

Tomlinson was previously CFO at several other tech startups, guiding Palo Alto Networks, Confluent, and Aruba Networks through the IPO process.

Stripe, one of the world's most valuable startups, has long been mentioned as an IPO candidate. Vercel is earlier in its lifecycle, however the AI startup has been putting some of the early pieces in place to potentially go public someday.

"Steffan's experience leading developer-focused companies from startup to public markets makes him an ideal addition to Vercel's Board of Directors as we continue to put our products in the hands of every developer," Vercel CEO and founder Guillermo Rauch said.

Last year, Vercel tapped Marten Abrahamsen as its CFO. He's been building out Vercel's finance, legal, and corporate development teams and systems while leading the startup through a $250 million funding round at a $3.25 billion valuation in May.

"Steffan's financial expertise and leadership experience come at a pivotal moment for Vercel as we scale our enterprise presence and build on our momentum," Abrahamsen said.

GenAI growth

The generative AI boom has recently powered Vercel's growth. The startup offers AI tools to developers, and earlier this year it surpassed $100 million in annualized revenue.

Vercel's AI SDK, a software toolkit that helps developers build AI applications, was downloaded more than 700,000 times last week, up from about 80,000 downloads a year ago, according to NPM data.

The company's Next.js open-source framework was downloaded 7.9 million times last week, compared to roughly 4.6 million downloads a year earlier, NPM data also shows.

Abrahamsen said they are building a company to one day go public, but stressed that there's no timeline or date set for such a move. 

Consumption-based business models

At Stripe and Confluent, Tomlinson gained experience with software that helps developers build cloud and web-based applications — and how these offerings generate revenue.

"Steffan's track record with consumption-based software business models makes him the ideal partner to inform strategic decisions," Rauch said.

Vercel is among a crop of newer developer-focused tech companies that charge based on usage. For instance, as traffic and uptime increase for developers, Vercel generates more revenue, so it's aligned with customers, Abrahamsen told Business Insider. 

Similarly, Stripe collects a small fee every time someone makes a payment in an app. Confluent has a consumption-based business model, too.

This is different from traditional software-as-a-service providers, which often charge based on the number of users, or seats. For instance, Microsoft 365 costs a certain amount per month, per user. 

Tomlinson also has experience working with developer-focused companies with technical founders, such as the Collison brothers who started Stripe. 

• Suchir Balaji, a former OpenAI researcher, was found dead on Nov. 26 in his apartment, reports say.
• Balaji, 26, was an OpenAI researcher of four years who left the company in August.
• He had accused his employer of violating copyright law with its highly popular ChatGPT model.

Suchir Balaji, a former OpenAI researcher of four years, was found dead in his San Francisco apartment on November 26, according to multiple reports. He was 26.

Balaji had recently criticized OpenAI over how the startup collects data from the internet to train its AI models. One of his jobs at OpenAI was gather this information for the development of the company's powerful GPT-4 AI model, and he'd become concerned about how this could undermine how content is created and shared on the internet.

A spokesperson for the San Francisco Police Department told Business Insider that "no evidence of foul play was found during the initial investigation."

David Serrano Sewell, executive director of the city's office of chief medical examiner, told the San Jose Mercury News "the manner of death has been determined to be suicide." A spokesperson for the city's medical examiner's office did not immediately respond to a request for comment from BI.

"We are devastated to learn of this incredibly sad news today and our hearts go out to Suchir's loved ones during this difficult time," an OpenAI spokesperson said in a statement to BI.

In October, Balaji published an essay on his personal website that raised questions around what is considered "fair use" and whether it can apply to the training data OpenAI used for its highly popular ChatGPT model.

"While generative models rarely produce outputs that are substantially similar to any of their training inputs, the process of training a generative model involves making copies of copyrighted data," Balaji wrote. "If these copies are unauthorized, this could potentially be considered copyright infringement, depending on whether or not the specific use of the model qualifies as 'fair use.' Because fair use is determined on a case-by-case basis, no broad statement can be made about when generative AI qualifies for fair use."

Balaji argued in his personal essay that training AI models with masses of data copied for free from the internet is potentially damaging online knowledge communities.

He cited a research paper that described the example of Stack Overflow, a coding Q&A website that saw big declines in traffic and user engagement after ChatGPT and AI models such as GPT-4 came out.

Large language models and chatbots answer user questions directly, so there's less need for people to go to the original sources for answers now.

In the case of Stack Overflow, chatbots and LLMs are answering coding questions, so fewer people visit Stack Overflow to ask that community for help. This, in turn, means the coding website generates less new human content.

Elon Musk has warned about this, calling the phenomenon "Death by LLM."

OpenAI faces multiple lawsuits that accuse the company of copyright infringement.

The New York Times sued OpenAI last year, accusing the start up and Microsoft of "unlawful use of The Times's work to create artificial intelligence products that compete with it."

In an interview with Times that was published October, Balaji said chatbots like ChatGPT are stripping away the commercial value of people's work and services.

"This is not a sustainable model for the internet ecosystem as a whole," he told the publication.

In a statement to the Times about Balaji's accusations, OpenAI said: "We build our A.I. models using publicly available data, in a manner protected by fair use and related principles, and supported by longstanding and widely accepted legal precedents. We view this principle as fair to creators, necessary for innovators, and critical for US competitiveness."

Balaji was later named in the Times' lawsuit against OpenAI as a "custodian" or an individual who holds relevant documents for the case, according to a letter filed on November 18 that was viewed by BI.

If you or someone you know is experiencing depression or has had thoughts of harming themself or taking their own life, get help. In the US, call or text 988 to reach the Suicide & Crisis Lifeline, which provides 24/7, free, confidential support for people in distress, as well as best practices for professionals and resources to aid in prevention and crisis situations. Help is also available through the Crisis Text Line — just text "HOME" to 741741. The International Association for Suicide Prevention offers resources for those outside the US.

• Modern factories, supply chains and Amazon have turned 'stuff' into a commodity. 
• The same inevitable supply-and-demand dynamic could wash over us again with generative AI.
• The ultimate outcome may be a new limited-edition luxury item: Humans. 

"Live experiences are the new luxury good," Kevin Hartz said in 2013 when Eventbrite, the ticketing startup he cofounded, got a big new funding round.

By that point, modern factories, supply chains, and Amazon had boiled down "stuff" to a commodity. You can now buy an overwhelming variety of tennis shoes, or spatulas, or sweatpants online. This abundance has taken much of the satisfaction away from purchasing physical things. This is why experiences, which by definition are finite, became more valuable.

There are only a few opportunities to see Taylor Swift on stage, versus the availability to purchase more than 20,000 kinds of tennis shoes on Amazon. So the price of Eras tickets soar, and shoes are cheap.

The same inevitable supply-and-demand dynamic is about to wash over us again with large language models and generative AI.

The ultimate outcome could be a new limited-edition luxury item: Humans. 

Unlimited content vs 'finite resources'

AI models can now automatically generate text, software code, medical diagnoses, images, voices, music, video, and lots more. The barriers to using this technology are falling away quickly. Anyone can fire up ChatGPT, GPT-4, DALL-E and other tools to produce an almost unlimited quantity of content.

This should be a boon to society. Many tasks will be completed more efficiently, making products and services more affordable and accessible, as venture capitalist Marc Andreessen recently explained.

There will be a reaction though: In a world of machine-generated abundance, human-centered services and experiences will become increasingly rare, valuable, and therefore desirable.

"The world's information is being turned into 1s and 0s and all this is being commoditized," Hartz told BI. "What can't be commoditized is finite resources like real estate, travel, seeing the sunset on Mediterranean, or surfing in Fiji. These are the luxury goods of the power elite."

Cooks, tutors, and robo-advisors

The more that AI automates restaurants, the more we'll want personal chefs such as John Barone, who cooks five days a week in the home of a wealthy Silicon Valley couple.

As AI tutor bots proliferate in education, the richest will pay for more exclusive access to the best human tutors for their kids.

The more robo-advisors handle our money, the stronger the urge of the wealthy to recruit savvy human experts to manage their family offices.

A new flood of automated emails

Email marketing is a simple example that some technologists are already worried about.

Generative AI tools are making it much quicker and easier to write marketing copy. The end result will be a flood of new emails that will overwhelm recipients and make them even less likely to open the messages.

"And our own machines will read those AI automated sales emails," Hartz quipped.

So, either your marketing email won't reach the humans you're trying to engage, or another AI bot will open it and you'll never be quite sure who read the message. A hand-typed email from a real human will be, relatively speaking, a rare and beautiful thing (complete with typos).

AI tutors versus small classrooms

AI models are beginning to revolutionize education, according to Sal Khan, the founder of Khan Academy. His organization has been working with OpenAI models to coach students in powerful new ways and help teachers develop class plans.

The gold standard throughout history has always been to have a personal tutor, and AI models can help personalize the education experience to bring some of this curated approach to more students, he explained during a No Priors podcast earlier this year.

"We don't have the resources to give everyone a tutor," he said during the podcast. "A generative AI tutor supporting students. That's going to be mainstream in 3 to 5 years," he added.

Pricey schools and a personal carpenter

And yet, Silicon Valley's top private schools, where many tech execs send their kids, are all about getting access to human teachers in small group settings. 

Castillja in Palo Alto highlights a student to faculty ratio of 7 to 1. Nueva, a Silicon Valley school for gifted kids, promises a similar ratio. The Menlo School in Menlo Park says it has a student-teacher ratio of 10 to 1 in the upper school.

These institutions cost $58,000 to $60,000 a year and I don't see any drop-off in demand among the tech elite. They're still jostling to get their kids into these bespoke, human-centered learning environments.

One persistent, apocryphal Silicon Valley story illustrates this point. On weekends, one tech billionaire has been known to hire a personal carpenter to hand-make wooden toys for their kids build and play with.

Who manages the money?

What about when it comes to managing fortunes amassed by successful tech entrepreneurs? The wealthiest rely on talented financial advisors who are hired directly to oversee this money in family offices.

Bill Gates has his own private investment firm, Cascade, which has been run by money manager Michael Larson since 1994. Elon Musk's family office, Excession, has been run by a former Morgan Stanley banker called Jared Birchall for years.

Using AI for trading has been tough so far. AI models are trained on masses of data from the past. When new situations arise, they struggle to adapt quickly enough.

Even quantitative hedge fund firms, which use machine learning and other automated techniques, rely on humans. Two Sigma, a famous quant firm, is for the first time exploring ways to add traders who rely on their human judgment to make money, Bloomberg reported recently.

"The major challenge with using things like reinforcement learning for trading is that it's a non-stationary environment," AI researcher Noam Brown said on the No Priors podcast in April. He's worked on algorithmic trading strategies in the past and was a researcher at Meta before recently joining OpenAI.

"So you can have all this historical data but it's not a stationary system," he explained, referring to how markets respond swiftly to world events and other developments.

Part of the problem relates to what he calls sample efficiency. Humans are good at learning quickly from a small amount of data, while AI models need mountains of information to train on.

"Humans are very good at adapting to novel situations," he added. "And you run into these novel situations pretty frequently in financial markets."

Social media bots vs. martial arts

AI is making social media increasingly machine-driven, too. Soon, human content creators will be vying for attention with content generated by AI models.

Last month, Meta CEO Mark Zuckerberg unveiled more than 25 new AI assistants with different personalities that use celebrities' images. Users will be able to interact with these bots on Meta's platforms in the future.

In a recent podcast, he described this new supply-and-demand situation well, saying human creators can't keep up with demand from followers.

"There are both people who out there who would benefit from being able to talk to an AI version of you," Zuckerberg explained. "You and other creators would benefit from being able to keep your community engaged."

So Meta will make an AI version of celebrities that can post constantly. Again, this will be infinite. And actually interacting with the real human celebrity will become more rare and valuable.

Meanwhile, when Zuckerberg is relaxing outside of work, he spends some of that time pursuing a very human pastime: Rolling around with other humans in martial arts contests.

Medical models and human doctors

AI models, such as Google DeepMind's Med-PaLM 2, are becoming incredibly good at answering medical questions and analyzing x-rays and other health data. But when wealthy parents have really sick children, they will still seek out the smartest doctors in the relevant fields of medicine.

You can see this in Silicon Valley's embrace of medical concierge services that provide special access to doctors and other human health specialists.

One Medical succeeded by offering better access to human doctors, and Amazon ended up buying it for almost $4 billion.

"We're inspired by their human-centered, technology-forward approach," an Amazon executive said when the deal was announced.

'Utility, value and signaling'

Hartz, a venture capitalist who now chairs Eventbrite's board, says successful technologists will continue to spend heavily on human experiences. But he says this depends on the activity and the motivations behind different actions.

He breaks this into "utility, value and signaling."

Many standard, common situations can be handled by software bots or even physical machines. Repetitive tasks at work and some educational functions are examples of these utility-type solutions.

In other situations, users will get more value from having machines handle the work, so humans can focus on more valuable tasks. If you're a well-paid machine-learning engineer, it will be better to have a robot clean your house so you can focus more on your job, he explained.

And then there will still many situations where humans will want to enjoy their success and signal the fruits of their achievements. And these activities will increasingly focus on finite human resources and experiences, Hartz said.

"You can't put on headset and pretend to be in Fiji," he added.

• The DOJ proposed banning Google from paying for search distribution deals.
• Google's search dominance relies on distribution, not just technology.
• Investors worry Google's market share could drop if distribution deals end.

The online search business is not about technology. It's about distribution.

The US Department of Justice made that clear Wednesday when it proposed fixes for a judge's earth-shaking ruling that Google is an illegal monopolist.

The DOJ's remedies cut to the heart of how Google distributes its search engine and how that broad reach is key to the company's dominance of this crucial and lucrative market.

The government's suggestion that Google be forced to sell Chrome initially grabbed the headlines. But, on Thursday, the potential crackdown on all distribution deals caught investors' attention.

The US government's lawyers said Google should be banned from offering "anything of value for any form" of search distribution. That especially includes Apple, but also covers any other partner or company, with limited exceptions, according to the DOJ's executive summary.

ISI Evercore internet analyst Mark Mahaney called this distribution crackdown "draconian" and said investors were surprised by the severity of the proposals. Google shares dropped 5% on Thursday.

The reason for this concern is that the online search business is not really about the quality of the technology. The edge comes from massive distribution and the huge volume of user queries that come with such a broad reach.

When people use Google to search on the web, the company monitors what results they click on. It feeds these responses back into its search engine, and the product gets constantly better. For instance, if most people click on the third result for a particular query, Google's search engine will likely adjust and rank that result higher in the future.

This self-reinforcing system is very hard to compete against. This is how the DOJ put it on Wednesday:

"Search engines rely on user data to improve search quality — an outcome that drives more users to a search engine. Users attract advertisers, and advertising dollars fund general search engines, creating a perpetual feedback loop that further entrenches Google."

One of the few ways to compete is to get more distribution than Google and pull in the extra queries and click-behavior data.

For many years, Google has paid to lock down most major sources of distribution. The most famous deal is with Apple. Google pays the iPhone maker about $20 billion a year to be the default search engine on Apple's mobile devices.

If the search business was actually about the quality of Google's technology, why does it have to pay Apple $20 billion a year? That question is at the heart of the DOJ's case, and Google has never been able to answer it properly. Because it keeps paying Apple.

If Google search technology is so great, the company shouldn't have to pay for distribution. People would just flock to its search engine all by themselves.

We could soon see a real-world test of this.

If the judge in this case agrees with the DOJ, then these payments will end — not just with Apple, but with any other third-party source of online distribution for Google's search engine.

This may have freaked investors out on Thursday. They know that the search business is mainly about distribution, and Google may not be able to do this now.

In a worst-case scenario, Google could lose a material slice of the US search market, according to Mahaney.

"We believe Google's default search placements via contractual agreements represent 50%+ of Google's US search queries," he estimated on Thursday.

If half of Google's US search queries go away, that could threaten the self-reinforcing cycle of user click data improving its results.

Suddenly, Google Search may not be so uncatchable.

Google's top lawyer, Kent Walker, said the DOJ's proposals would "break" the company's search engine and "deliberately hobble people's ability to access" the service.

Google gets to propose its own remedies on December 20.