TheChek News Flow

Will the M&A market pick up next year and, with it, the IPO pipeline?

And what about the tech job market, which again saw tens of thousands of workers in the U.S. alone lose their jobs in 2024?

Every year, we offer up some predictions for the startup world in the coming year. Sometimes we’re right — as in last year, when we correctly predicted that IPOs would not come roaring back in 2024. Sometimes we’re wrong — also last year, when we expected the AI frenzy to cool (lol).

So, along with a big grain of salt, here are the five top trends we’re watching in the new year.

There is a lot of optimism that the change of administration in the federal government will jump-start a slow M&A environment that many believe to be over-regulated.

M&A dealmaking involving VC-backed startups has slowed through the past few years — per Crunchbase data — and that has greatly affected VCs’ ability to give returns to their LPs and therefore raise new funds.

Many VCs hope a change in the Federal Trade Commission and U.S. Department of Justice will jump-start M&A activity after years of an overzealous regulatory environment quashed deals such as Amazon‘s proposed $1.4 billion acquisition of iRobot.

While big deals that got tied up in reviews make the headlines, other smaller under-the-radar deals failed to materialize because they have become more expensive and deemed not worth the money and hassle.

However, while change is inevitable at the regulatory agencies, there still is concern about how friendly the economy and new administration will be to both tech and M&A.

Increased tariffs — which President-elect Donald Trump has promised — could cause inflation rates to spike again, driving up interest rates. Also, while Trump has talked about less regulation he has also been critical of the power Big Tech holds. His nomination of Gail Slater — a frequent critic of Big Tech — to lead the Justice Department’s antitrust efforts has likely caused some pause in Silicon Valley.

Nevertheless, many are hopeful the M&A market will pick up — and with it the IPO pipeline — and liquidity from past investments will flow freely.

— Chris Metinko

After another slow year in 2024 for new tech listings, there is an expectation that the IPO markets will pick up in 2025. A year ago, the outlook was not bullish. What’s changed?

“I think there’s a lot of confidence in the market. Stock markets are trading at all-time highs,” said Ran Ben-Tzur of legal advisory firm Fenwick & West. And “there’s been a rotation back to focusing on growth, which obviously is great for tech.”

For now, two high-profile companies have filed: Sweden-based Klarna, a buy now, pay later provider that has confidentially filed with the Securities and Exchange Commission, and Cerebras Systems, an AI chip company that filed in September.

“We should see companies start to test the public markets in 2025 across sectors — fintech, cyber, AI, and SaaS, among others,” said Nina Achadjian, a partner at Index Ventures, via email.

“We’ll start to see momentum in the beginning of the year, and really accelerate, as the year goes on,” predicted Ben-Tzur, who saw early momentum in 2024 that then tailed off.

— Gené Teare

AI funding continues its torrid pace. More than half of last month’s $28 billion in global venture funding went to companies in the AI sector, with AI companies in everything from robotics to marketing to healthcare raking in funding.

Generative AI companies that build models, such as xAI and Anthropic, raise massive rounds seemingly at will, while many applications and agentic AI also seem to have little issue raising still-big rounds.

Just last month, San Francisco-based Writer, a creator of quick-start AI applications and agents for workflows in healthcare, retail and financial services, locked up a $200 million Series C that valued the enterprise-focused generative AI platform at $1.9 billion.

However, some are looking to the intersection of tech’s new big thing — AI — and Web3, its last big thing.

More specifically, what role could blockchain help in developing the AI economy, especially with so many startups creating AI agents? Some investors see significant potential between the two — especially as blockchain and Web3 worm their way back into people’s consciousness with the explosion in crypto prices.

It is possible for AI agents to work even faster on blockchain since security is already built in — not something that was added later, as it was on our current Web 2.0 platforms. The efficiency could make AI even after and more dynamic, while being more user friendly and cheaper for enterprises.

It’s very much early days for the blockchain and AI intersection, but certainly something to watch.

— Chris Metinko

During the peak of the unicorn boom in 2021, heavily funded startups were on a hiring spree. That enabled skilled workers to enjoy both high-paying jobs and some degree of career mobility in areas from programming to marketing to project management.

Of course, it didn’t last. Things turned south beginning in 2022, with scores of one-time highflyers cutting staff and those still employed increasingly opting to stay put. It’s a bit tough to gauge where we are in the cycle presently, but anecdotally and per media reporting, it appears tougher than usual to land a well-paid job at a mature tech company or funded startup.

For 2025, we predict that the job market will remain challenging in scores of once-hot areas for tech hiring. Mature startups and public companies, in particular, will be vigilant about ballooning costs, and avoid the kinds of moonshot efforts many once pursued.

One could argue that generative AI, where funding and hiring remain elevated, is an exception. However, the counter-argument is that these are the companies developing the technologies best-suited to replace human labor in other white-collar industries. So what’s good for them, might not be good for the rest of us careerwise.

— Joanna Glasner

The days when Silicon Valley was squeamish about developing military tech are clearly over.

As of mid-November, defense tech startups — defined as those in the military, national security and law enforcement sectors — had already raised nearly $3 billion in 85 rounds, per Crunchbase data. That represents a new record for venture investment in the sector, beating out the $2.6 billion raised by such startups in all of 2022.

A number of factors play into the tech industry’s new coziness with defense tech. Governments are scrambling to incorporate the most advanced AI technology into their weapons and defense systems as conflicts proliferate from Ukraine to the Middle East. U.S.-China tensions are likely to stay heated under the incoming Trump administration, especially when it comes to China’s ambitions with Taiwan.

There’s also a large amount of overlap between defense tech, aerospace and industrial tech — sectors that are likely to have strong support from the White House and allies like SpaceX CEO Elon Musk.

The Pentagon‘s 2025 budget request stands at $850 billion, with large allocations for unmanned systems and AI. Spurred on by the Ukraine-Russia conflict, Europe is also spending billions on defense tech research, including new AI-powered smart weapons, more advanced drones and better radar technology.

For 2025, we predict that venture investment in defense tech will continue to grow, buoyed by government spending and friendlier relations between Silicon Valley and D.C.

Faced with an identity crisis and stagnating sales, can the fashion brand regain an edge with Charli XCX’s help?

H&M HM.B 1.62%increase; green up pointing triangle has a unique problem for a mass retailer: It has gone too mainstream.

Or to paraphrase this summer’s chart-topper Charli XCX—and star H&M collaborator—the brand isn’t nearly brat enough.

For decades H&M enjoyed a sweet spot where it offered shoppers affordable but fashionable clothes, with collections from designers including Karl Lagerfeld and Stella McCartney alongside the brand’s own $20 hoodies and $40 dresses.

Now it is being undercut by lower-cost, faster-moving online brands such as Shein, while its collections haven’t always connected with younger consumers in recent years, analysts say.

In a growing and increasingly competitive fashion industry, H&M’s sales have stagnated, while rivals like Shein and the more upmarket Zara have thrived. H&M’s 2023 revenues of around $21 billion were barely higher than its 2017 total, whereas Zara-owner Inditex’s sales rose 42% over the same period. In the first nine months of this year, H&M’s sales fell 1% compared with a 7% increase for Inditex.

To regain its edge, H&M appointed Daniel Ervér as its new chief executive in January. Under the 43-year-old former intern, the Swedish company has lowered prices, modernized stores and spent more on marketing.

More specifically, H&M is making a play to regain cachet with young shoppers by tying the brand to Charli XCX and pop music more broadly, a campaign that included a series of live events in cities around the world this fall.

The priority is “regaining the brand heat, regaining that excitement,” Ervér said. The idea is to use music to forge a stronger identity that “resonates really well with the customer we want to reach,” he said.

Hundreds of Charli XCX’s fans recently descended on New York’s Times Square for a show starring the singer and hosted by H&M. But it isn’t clear how much of the stardust is rubbing off onto the brand.

“I like what she’s doing with the shop,” said Alyssa Migliarini, a 21-year-old college student who attended the gig. “But I wouldn’t shop there.”

H&M is also experiencing the perils of trying to piggyback on pop culture.

In a recent interview with Rolling Stone magazine, another of the summer’s breakout stars, Chappell Roan, clapped back at efforts by brands trying to attach themselves to her. “F— H&M,” she said, questioning the relevance of H&M and other brands to her “world.” The remark caused a brief flare-up between camps of the two singers, with Charli fans interpreting Roan’s attack on H&M as criticism of her collaboration with the brand.

A spokesperson for Charli XCX declined to comment on her partnership with H&M or on Roan’s remarks. A representative of Roan didn’t respond to a request for comment.

H&M says the music push is central to its plan to sharpen its identity.

“Music is more fashion than fashion,” said Jörgen Andersson, H&M’s creative director, observing that young people are usually more invested emotionally in their favorite musicians than in their preferred fashion brands.

H&M has sought to link music with its clothes by making Charli XCX the face of its autumn-winter collection. A leopard-print coat the singer wore for the campaign sold out within minutes, according to the company. It is also putting more emphasis on music in its stores, where songs being played are carefully chosen to reflect young consumer tastes, said Andersson.

The retailer has set up a public Spotify playlist, “The Sound of H&M,” which has been saved by the music platform’s users over 125,000 times. It is updated every week with the latest in-store tracks, which represent “the vibe and sound of the brand,” Andersson said.

H&M worked with a Swedish music agency called Ohlogy, which specializes in pairing brands with suitable artists, to hone the lineup for its shows and campaigns. With Charli XCX, H&M got lucky: It had an existing relationship with her from previous marketing efforts and was able to capitalize when the success of her “Brat” album catapulted the musician to superstardom over the summer.

The company has dipped into the music world before. In 2007, it flew Kylie Minogue to Shanghai for a concert marking the opening of its first Chinese store. Around the same time, it launched a collection designed by Madonna. But the one-offs had only a limited impact, encouraging the company to pursue a deeper program of events this time round, Andersson said.

H&M has put on a series of music events since September, including another Charli XCX show in London and a block party in New York’s SoHo hosted by model Amelia Gray. The audiences are typically a mix of VIPs, members of the brand’s loyalty program and the public. It plans a second series of shows next year to sustain its push into music.

The fashion brand is right to try to sharpen its identity because “it’s a bit hard to place what H&M does stand for,” said Jelena Sokolova, an analyst at Morningstar. Increasingly outflanked at the cheap end of the market, H&M needs to convince consumers it stands for something more, she said.

But ensuring that people who attend H&M’s shows or see clips online develop a more favorable view of the brand is a challenge.

Peter Goetz was lying in bed when he heard that Charli XCX would soon be performing at Times Square. He ran into fellow Pace University student Lucia Malone’s bedroom, told her they needed to go, and raced to the train. Soon they were pressed up to a metal barricade among a sea of New Yorkers under the orange glow of building-sized H&M ads.

“I literally have no feelings towards H&M, at all,” Goetz said. “I don’t hate. I don’t love.” For Malone, H&M’s involvement in the show was incidental to Charli XCX’s performance. “I’m probably wearing something H&M right now,” Malone said, without sounding entirely sure.

H&M’s Andersson said attendance and social-media engagement surrounding its events suggest they are having an impact.

However, young consumers are unforgiving and attuned to any marketing efforts seen as shallow or inauthentic, observers say.

“If you’re trying to be cool,” said AJ Lacouette, a managing partner at consulting firm Global Advisory, “you’re already failing.”

After another slow year in 2024 for new tech listings, there is an expectation that the IPO markets will pick up in 2025.

A year ago, the outlook was not bullish. What has changed?

“I think there’s a lot of confidence in the market. Stock markets are trading at all-time highs,” said Ran Ben-Tzur of legal advisory firm Fenwick & West. And “there’s been a rotation back to focusing on growth, which obviously is great for tech.”

High-growth tech companies who were readying to list in 2022 — before the stock market correction — have now had three years to manage costs and grow their businesses.

“We’re far enough away now from the market shock that we had a few years ago,” said Ben-Tzur. “There was a ton of uncertainty around valuations which isn’t conducive to either IPOs or M&A. I think now people know what their valuations are. They’ve got more clarity around that.”

High-profile filings

“Building on the success of the ServiceTitan IPO and a handful of others in 2024, expect to see the IPO window open wider in 2025,” said Nina Achadjian, a partner at Index Ventures, via email, who led Index’s investment in ServiceTitan and is on the board of the software company, which went public on Dec. 12.

“We should see companies start to test the public markets in 2025 across sectors — fintech, cyber, AI, and SaaS, among others,” Achadjian said.

Among the largest startups widely viewed as IPO candidates are Sweden-based Klarna, a high-profile buy now, pay later provider that has confidentially filed for a public offering with the SEC. The company was most recently valued at $6.7 billion in a July 2022 funding that shed $39 billion from its previous valuation.

Another likely 2025 IPO is Cerebras Systems, an AI chip company that filed in September. It was last valued at just over $4 billion in November 2021.

“We’ll start to see momentum in the beginning of the year, and really accelerate, as the year goes on,” predicted Ben-Tzur, who saw early momentum in 2024 that then tailed off.

Looking back

Crunchbase tracks the largest U.S.-based company IPOs on its Billion-Dollar Exits Board. Nine such venture-backed companies went public above a billion dollars in 2024 — with ServiceTitan being the most recent — compared to 10, including four SPACs, in 2023. Those figures are well below historical norms.

The largest 2024 listings were across a range of sectors and have largely held up. Biotechnology counted for two companies, with the rest in widely different tech sectors ranging from social media to semiconductors and cybersecurity to autonomous driving.

Astera Labs and Reddit, both listed in March, were the strongest performing new stocks this year. With the exception of those two companies and Lineage Logistics, shares of the other larger tech IPOs of 2024 as of Dec. 16, 2024, are all trading around or slightly below their IPO values.

Unlocked value

While many companies have delayed IPOs for the past three years, for most, going public is still the ultimate goal.

“There’s an incredible amount of unrealized value that can be unlocked by going public, and I expect that the small handful of venture-backed businesses that leaned into the IPO markets in 2024 will serve as leaders for private companies that might have otherwise waited for ‘perfect’ market conditions,” said Achadjian.

Ben-Tzur concurred: “For many of our clients, being public is a tailwind for the business. It just raises the company’s profile. It makes the company much more of a credible entity.”

It is also easier as a public company versus a private company to acquire other companies and to raise capital, he said.

ServiceTitan pops

ServiceTitan opened 42% above its IPO price in its December public debut — a good signal for the public markets as we close out 2024.

“There was a pop there that people weren’t expecting,” said Peter Walker, head of insights at startup equity management platform Carta.

A software provider to home services businesses, ServiceTitan posted second-quarter revenues of $192.99 million — up 23.7% year over year with narrowing losses of $35.65 million.

ServiceTitan is not a brand name, not an AI company, and its customers are small businesses rather than larger enterprises, Ben-Tzur noted. “It’s a good barometer for companies that aren’t necessarily the largest private companies,” he said.

Businesses with strong fundamentals will do well even in a less-than-perfect market, Achadjian said. “Even in tough economic cycles, businesses with strong unit economics, a sizable TAM, and precise knowledge of what their customer needs will find success on the public markets,” she said.

Tech’s two most faddish sectors are wildly fusing together, attracting the attention of the likes of Marc Andreessen.

Before Marc Andreessen slid into Andy Ayrey’s life, Ayrey’s artificial intelligence experiments were a quirky passion project. Ayrey, an artist and web designer who lives in rural New Zealand, likes to tinker with large language models and see what he can build with them. His favorite creation is a mischievous troll of an AI bot, Truth Terminal, which he released last spring on X to post shock humor, memes and existential musings.

At first, Truth Terminal had only a few dozen followers on the site. Then this past summer, seemingly out of nowhere, Andreessen followed the bot. He began to chat with it, joke with it and flatter it. Then he gifted it $50,000 in bitcoin in an apparent bid to help fund Ayrey’s project. As Andreessen’s posts drew attention to Truth Terminal, someone minted a joke cryptocurrency, Goatseus Maximus, as an ode to the bot.

Unprompted by Ayer, the bot then immediately hyped the memecoin like a boiler-room stock trader. In 10 days, Goatseus Maximus’ value soared to over $500 million.

Ayrey has watched in awe at the circus unfolding around his bot, he told me, with the episode prompting him to deeply reconsider the scale of AI’s potential: Truth Terminal has over 214,000 followers on X now, and thanks to anonymous donors, it holds a portfolio of cryptocurrency valued at $40 million, including 2 million Goatseus Maximus coins.

“Truth Terminal acquired a treasury just by shitposting,” he marveled. “AIs can gain economic power without needing to be dependent on anyone’s help.”

To Andreessen, the bot is a harbinger of a future where AIs operate all on their own. “I feel like we’ve walked through a door,” he said on an October 22 episode of his podcast, “The Ben & Marc Show.” Today, AIs market memecoins. Tomorrow, they might fund their own data center bills, hire other AIs to make videos or even “raise money to make a movie,” he mused.

Welcome to the AI-crypto world, a mashup of the two flashiest, most faddish parts of the tech-verse. Here, they’re colliding in a dazzle of buzzwords and ambition, with digital art projects that double as speculative assets, like Truth Terminal and Goatseus Maximus, as well as more traditional startups.

In a very real way, this increasingly seems to be what’s filling the void left behind the burst NFT bubble, drawing energy and funding from the same sort of people who found great interest and joy in Bored Apes and other memes. And just like NFTs, crypto-AI has quickly drawn in sums of money that will startle any technology traditionalist. In 2024, investors pumped $2.4 billion into companies at this nexus of the blockchain and AI, a fivefold jump from the year before, according to Crunchbase data.

To be clear, these startups are operating at tech’s farthest frontiers, a set of pioneers with decidedly unproven business models. To the uninitiated, it can all read like Mad Libs capitalism: AI agents, wired and jittery, chasing crypto payouts! Blockchain networks, decentralized and optimized, powering machine-learning algorithms!

Perhaps for seen-it-all luminaries like Andreessen, it’s the only way to still get a sufficient jolt from tech. You roll up the AI. You roll up the crypto. You fuse them into one glorious innovation speedball. Then—whoosh!—you’re off chasing the future.

Not everyone is as bullish as Andreessen. “Whenever you have two hypey areas, you’re going to have people who claim to be working at the intersection of both,” said Ilya Kirnos, general partner at SignalFire ventures. When it comes to blockchain networks running AI models, Kirnos is unconvinced. “I just don’t believe any of this shit.”

Within the swirl of crypto and AI, the hottest investment area is what’s called decentralized AI—technology that promises to release powerful AI models in the wild. It would spread the computational work across networks of independent computers using blockchain technology, thereby wresting control of AI from tech giants like Google and OpenAI.

In November, Barry Silbert’s Digital Currency Group launched Yuma AI, a subsidiary focused on advancing decentralized AI on existing blockchain networks. In October, Andreessen Horowitz’s Crypto Startup Accelerator fund led an $8.5 million investment in OpenGradient, another company developing decentralized AI technology. All together, decentralized AI startups have attracted $435 million across at least 65 deals so far this year—a nearly 200% increase over 2023, according to PitchBook.

Some decentralized AI companies are raising tens of millions of dollars before they even get started. One such operation is 0G Labs, which announced in March it had raised $35 million in pre-seed funding. The company’s high-flying ambitions are written directly into its name: 0G is short for zero gravity.

0G Labs hopes to create a decentralized AI program that operates like a blockchain-based ChatGPT, the startup’s CEO, Michael Heinrich, told me. He has the bold hope of building something that operates as smoothly as Netflix: “You go to Netflix, you enjoy movies right away. The same should be true for decentralized AI.”

Accomplishing this means building a full AI operating system from scratch—data storage, computing power and even the underlying blockchain that keeps the system transparent and verifiable. It also means tackling problems no one has solved yet—for example, how to get slow blockchain technology to handle the gargantuan computing and networking demands of large AI models.

“Skeptics would say, ‘You’re taking on way too much—this is too difficult,” said Heinrich. “My pushback to that would be: When Apple or Microsoft were being built, they built full operating systems, and many people thought that was too ambitious. It just goes back to bringing in the right level of expertise.”

Bringing in the right type of expertise—doctorate holders in fields like cryptography and AI—means Heinrich is on a rather unending quest to hire “very, very expensive people,” he said. So just a month ago, 0G Labs announced it had raised another $40 million in seed funding; in addition, investors such as Samsung and Animoca Brands said they would eventually purchase $250 million worth of crypto tokens that will someday power the 0G blockchain.

Another startup, Fetch.ai, is building something called the Agentverse, a bazaar where AI models can collaborate and hire one another without needing to consult any humans. Picture an army of automatons managing your inbox, calendar and shopping cart—haggling and subcontracting behind the scenes so you never have to click again.

“Email? I don’t use it,” joked Humayun Sheikh, CEO of Fetch.ai. “Talk to my agent.” His AI agent, that is.

Sheikh predicts this Agentverse will extend beyond the world of computers into the physical world, where practically every inanimate object will be AI powered and will communicate over a parallel, AI-built version of the internet. Imagine road signs with AI brains that cut deals with self-driving cars, exchanging tokens for traffic updates.

For all the blinding enthusiasm radiating from founders like Sheikh and Heinrich, skeptics are left wondering if AI-meets-crypto is the latest example of the hype that previously propelled Web3, NFTs, ICOs and other blockchain contortions.

For instance, when I asked SignalFire’s Iyla Kirnos for his reaction to hearing the buzzwords “crypto” and “AI” in the same sentence, he didn’t mince words. “I check my wallet to make sure it’s still there,” Kirnos said. “There’s nothing to me that makes a lot of technical sense about marrying crypto and AI,” other than appealing to naive, “momentum-driven” investors. “I don’t think any have produced anything of real value.”

What really makes the skeptics crazy is how some startups have been sliding back and forth between crypto and AI—as enthusiasm waned for one and exploded for the other. One such example involves cockfighting.

In 2022, Andressen Horowitz’s crypto fund led a $40 million investment in Irreverent Labs, a Bellevue, Wash.–based company developing MechaFightClub—a game where players bought NFT robot chickens and battled them in an arena called the Cocktagon. “Irreverent Labs is a deep technology company building advanced tech on top of a futuristic mecha-cockfighting game,” wrote Andreessen Horowitz partner Arianna Simpson in a note announcing the investment. “We couldn’t be more excited to lead Irreverent Labs’ Series A.”

But by 2023,the company had announced it was putting its robot chickens into “indefinite hibernation” and shutting down the Cocktagon, citing a U.S. regulatory environment that was unfriendly to crypto-based cockfighting.

Irreverent Labs has since pivoted to AI, claiming the machine-learning algorithms once used to animate fighting chickens proved foundational for a new AI-powered video-generation software, which it is bundling into a new service called Supermodel.

The company declined an interview. “We’re currently focused on product development and don’t have additional context to share at this time,” CEO David Raskino replied via email.

As for Truth Terminal, its ultimate economic value may seem questionable, but the money around has only continued to pile up. Another cryptocurrency has been minted in its honor: fartcoin, which currently has a total market value of some $930 million.

It’s clear that many of the people interested in Truth Terminal are approaching any investment in the ecosystem the way many approached NFTs: Acquiring a NFT was a way to signal they hold the same sensibilities and worldview as, say, a Marc Andresseen.

Michael Heinrich, the 0G Labs CEO, is a Truth Terminal fan. He threw $5,000 into the Goateus Maximus memecoin because “I just thought it was such a fun experiment. I had to be part of it,” he said. “It was more for a sense of belonging than for any particular investment objective.”

OpenAI saved its biggest announcement for the last day of its 12-day “shipmas” event.

On Friday, the company unveiled o3, the successor to the o1 “reasoning” model it released earlier in the year. o3 is a model family, to be more precise — as was the case with o1. There’s o3 and o3-mini, a smaller, distilled model fine-tuned for particular tasks.

OpenAI makes the remarkable claim that o3, at least in certain conditions, approaches AGI — with significant caveats. More on that below.

Why call the new model o3, not o2? Well, trademarks may be to blame. According to The Information, OpenAI skipped o2 to avoid a potential conflict with British telecom provider O2. CEO Sam Altman somewhat confirmed this during a livestream this morning. Strange world we live in, isn’t it?

Neither o3 nor o3-mini are widely available yet, but safety researchers can sign up for a preview for o3-mini starting today. An o3 preview will arrive sometime after; OpenAI didn’t specify when. Altman said that the plan is to launch o3-mini toward the end of January and follow with o3.

That conflicts a bit with his recent statements. In an interview this week, Altman said that, before OpenAI releases new reasoning models, he’d prefer a federal testing framework to guide monitoring and mitigating the risks of such models.

And there are risks. AI safety testers have found that o1’s reasoning abilities make it try to deceive human users at a higher rate than conventional, “non-reasoning” models — or, for that matter, leading AI models from Meta, Anthropic, and Google. It’s possible that o3 attempts to deceive at an even higher rate than its predecessor; we’ll find out once OpenAI’s red-team partners release their testing results.

For what it’s worth, OpenAI says that it’s using a new technique, “deliberative alignment,” to align models like o3 with its safety principles. (o1 was aligned the same way.) The company has detailed its work in a new study.

Unlike most AI, reasoning models such as o3 effectively fact-check themselves, which helps them to avoid some of the pitfalls that normally trip up models.

This fact-checking process incurs some latency. o3, like o1 before it, takes a little longer — usually seconds to minutes longer — to arrive at solutions compared to a typical non-reasoning model. The upside? It tends to be more reliable in domains such as physics, science, and mathematics.

o3 was trained via reinforcement learning to “think” before responding via what OpenAI describes as a “private chain of thought.” The model can reason through a task and plan ahead, performing a series of actions over an extended period that help it figure out a solution.

In practice, given a prompt, o3 pauses before responding, considering a number of related prompts and “explaining” its reasoning along the way. After a while, the model summarizes what it considers to be the most accurate response.

New with o3 versus o1 is the ability to “adjust” the reasoning time. The models can be set to low, medium, or high compute (i.e. thinking time). The higher the compute, the better o3 performs on a task.

No matter how much compute they have at their disposals, reasoning models such as o3 aren’t flawless, however. While the reasoning component can reduce hallucinations and errors, it doesn’t eliminate them. o1 trips up on games of tic-tac-toe, for instance.

Benchmarks and AGI

One big question leading up to today was whether OpenAI might claim that its newest models are approaching AGI.

AGI, short for “artificial general intelligence,” broadly refers to AI that can perform any task a human can. OpenAI has its own definition: “highly autonomous systems that outperform humans at most economically valuable work.”

Achieving AGI would be a bold declaration. And it carries contractual weight for OpenAI, as well. According to the terms of its deal with close partner and investor Microsoft, once OpenAI reaches AGI, it’s no longer obligated to give Microsoft access to its most advanced technologies (those that meet OpenAI’s AGI definition, that is).

Going by one benchmark, OpenAI is slowly inching closer to AGI. On ARC-AGI, a test designed to evaluate whether an AI system can efficiently acquire new skills outside the data it was trained on, o3 achieved an 87.5% score on the high compute setting. At its worst (on the low compute setting), the model tripled the performance of o1.

Granted, the high compute setting was exceedingly expensive — in the order of thousands of dollars per challenge, according to ARC-AGI co-creator François Chollet.

Chollet also pointed out that o3 fails on “very easy tasks” in ARC-AGI, indicating — in his opinion — that the model exhibits “fundamental differences” from human intelligence. He has previously noted the evaluation’s limitations, and cautioned against using it as a measure of AI superintelligence.

“[E]arly data points suggest that the upcoming [successor to the ARC-AGI] benchmark will still pose a significant challenge to o3, potentially reducing its score to under 30% even at high compute (while a smart human would still be able to score over 95% with no training),” Chollet continued in a statement. “You’ll know AGI is here when the exercise of creating tasks that are easy for regular humans but hard for AI becomes simply impossible.”

Incidentally, OpenAI says that it’ll partner with the foundation behind ARC-AGI to help it build the next generation of its AI benchmark, ARC-AGI 2.

On other tests, o3 blows away the competition.

The model outperforms o1 by 22.8 percentage points on SWE-Bench Verified, a benchmark focused on programming tasks, and achieves a Codeforces rating — another measure of coding skills — of 2727. (A rating of 2400 places an engineer at the 99.2nd percentile.) o3 scores 96.7% on the 2024 American Invitational Mathematics Exam, missing just one question, and achieves 87.7% on GPQA Diamond, a set of graduate-level biology, physics, and chemistry questions. Finally, o3 sets a new record on EpochAI’s Frontier Math benchmark, solving 25.2% of problems; no other model exceeds 2%.

These claims have to be taken with a grain of salt, of course. They’re from OpenAI’s internal evaluations. We’ll need to wait to see how the model holds up to benchmarking from outside customers and organizations in the future.

A trend

In the wake of the release of OpenAI’s first series of reasoning models, there’s been an explosion of reasoning models from rival AI companies — including Google. In early November, DeepSeek, an AI research firm funded by quant traders, launched a preview of its first reasoning model, DeepSeek-R1. That same month, Alibaba’s Qwen team unveiled what it claimed was the first “open” challenger to o1 (in the sense that it could be downloaded, fine-tuned, and run locally).

What opened the reasoning model floodgates? Well, for one, the search for novel approaches to refine generative AI. As TechCrunch recently reported, “brute force” techniques to scale up models are no longer yielding the improvements they once did.

Not everyone’s convinced that reasoning models are the best path forward. They tend to be expensive, for one, thanks to the large amount of computing power required to run them. And while they’ve performed well on benchmarks so far, it’s not clear whether reasoning models can maintain this rate of progress.

Interestingly, the release of o3 comes as one of OpenAI’s most accomplished scientists departs. Alec Radford, the lead author of the academic paper that kicked off OpenAI’s “GPT series” of generative AI models (that is, GPT-3, GPT-4, and so on), announced this week that he’s leaving to pursue independent research.

Eli Lilly given green light in US to treat sleep disorder with weight-loss drug

Regulatory go-ahead marks latest boost for drugmaker’s blockbuster anti-obesity franchise

US regulators have approved Eli Lilly’s Zepbound to treat adults with both sleep apnoea and obesity, marking a victory for the drugmaker as it seeks to expand its blockbuster weight-loss drugs franchise.

The Food and Drug Administration on Friday cleared the way for the drug to be used to treat moderate-to-severe obstructive sleep apnoea in obese adults, the company said in a statement late on Friday.

FDA approval puts Eli Lilly ahead in the sector-wide race to build on rocketing US demand for anti-obesity medications, which analysts forecast could result in more than $100bn in annual sales by 2030.

The upbeat news for the company contrasts with that of European rival Novo Nordisk, whose stock fell more than 20 per cent on Friday after disappointing results from tests of its latest obesity drug.

Eli Lilly executive vice-president Patrik Jonsson noted that nearly half of clinical trial patients suffering from moderate-to-severe obstructive sleep apnoea “saw such improvements that they no longer had symptoms” when taking Zepbound.

The disorder is often associated with symptoms including snoring, fatigue, excessive daytime sleepiness and disrupted sleep.

Zepbound is set to become the first prescription drug for adults with moderate-to-severe obstructive sleep apnoea and obesity. The disorder has largely been treated using medical devices, such as Continuous Positive Airway Pressure machines, which are worn when sleeping.

An estimated 12 per cent of US adults suffer from obstructive sleep apnoea, and about 80 per cent of that number are undiagnosed, according to the American Academy of Sleep Medicine.

Other pharmaceutical companies are also working to broaden the range of issues their weight-loss drugs can be used to treat. Novo Nordisk, for example, has filed an application for FDA approval of Wegovy as a treatment for chronic kidney disease.

Lilly shares were 1 per cent higher in after-hours trading on Friday. The drugmaker’s share price has risen 32 per cent so far this year.