TheChek News Flow

Even as liquidity from exits remains stalled, it hasn’t stopped some firms from raising massive new growth funds.

The most recent example is Peter Thiel’s Founders Fund, which has closed a $4.6 billion late-stage venture fund, according to a filing with the U.S. Securities and Exchange Commission.

The new fund, Founders Fund Growth III, had been reported back in December as being about $3 billion.

The San Francisco-based firm is best known for its investments over the years which include Airbnb, Palantir and Stripe, and recently invested heavily in defense tech — including co-leading Costa Mesa, California-based Anduril Industries’ huge $1.5 billion Series F along with Sands Capital Ventures that valued the company at $14 billion.

News of the new fund comes just as several companies halt their IPO process and the M&A market remains murky as stock market tumult and threats of a possible trade war make many investors uneasy.

The stall in exits has been a cause for concern among VCs, as DPI — distributed to paid-in capital, or the capital paid to funds’ LPs after exits by those funds’ portfolio companies — remains their top issue.

Despite that, Founders has been quite active this year, per Crunchbase data. In the first three-and-a-half months, the firm has made 20 investments into startups.

Late last year, Founders led a $600 million round for Crusoe Energy Systems that valued the energy firm at $2.8 billion.

Apple details how it plans to improve its AI models by privately analyzing user data

In the wake of criticism over the underwhelming performance of its AI products, especially in areas like notification summaries, Apple on Monday detailed how it is trying to improve its AI models by analyzing user data privately with the aid of synthetic data.

Using an approach called “differential privacy,” the company said it would first generate synthetic data and then poll users’ devices (provided they’ve opted-in to share device analytics with Apple) with snippets of the generated synthetic data to compare how accurate its models are, and subsequently improve them.

“Synthetic data are created to mimic the format and important properties of user data, but do not contain any actual user generated content,” the company wrote in a blog post. “To curate a representative set of synthetic emails, we start by creating a large set of synthetic messages on a variety of topics […] We then derive a representation, called an embedding, of each synthetic message that captures some of the key dimensions of the message like language, topic, and length.”

The company said these embeddings are then sent to a small number of user devices that have opted in to Device Analytics, and the devices then compare them with a sample of emails to tell Apple which embeddings are most accurate.

The company said it is using this approach to improve its Genmoji models, and would in the future use synthetic data for Image Playground, Image Wand, Memories Creation and Writing Tools as well as Visual Intelligence. Apple said it would also poll users who opt in to share device analytics with synthetic data to improve email summaries.

U.K. startup Marshmallow has blown up over the years by using innovations in data science to build car insurance policies for immigrants and other consumers who have been overlooked or priced out of traditional insurance. Now, with a million drivers insured and a profitable annual revenue run rate of $500 million, Marshmallow has raised a fresh $90 million to expand.

Marshmallow plans to use the funding to move into financial services, as well as more insurance products that it hopes will appeal to a population of people that — despite the chilling effects of Brexit — is growing.

“We think of migration as a huge opportunity,” CEO Oliver Kent-Braham said in an interview. He noted that in the U.K., there are more people coming out of the workforce than there are going in, with 1.2 million migrants recorded coming to the U.K. in 2024 alone. “We need migration to put more people into work, and we want to help people move and integrate into the U.K.”

In Marshmallow’s view, that integration comes with being able to drive your own insured vehicle and soon, the startup hopes, buying home insurance and taking out loans.

Marshmallow plans to launch its first lending product later this year, Kent-Braham said, en route to building a “one-stop shop” for everything financial and insurance that a new arrival to the U.K. might need to adjust to life.

This round is roughly split 50-50 between equity and debt, according to Kent-Braham, and it is coming at a valuation of just over $2 billion. To put that into context, Marshmallow last raised funding at $1.25 billion in 2021.

The startup has seen considerable growth on the business front in that time. In 2021, Marshmallow had insured just 100,000 people. Now, in cities like London, the 1-million insured number is bolstered with a pink outdoor ad campaign that is hard to miss.

Portage Capital is leading the round, with participation from BlackRock and Columbia Lake Partners. Previous backers of the company have included Passion Capital, Investec and Scor. Marshmallow has raised around $220 million to date.

Notably, the new round has been in the works since at least January, and Kent-Braham noted that one part of the equity was convertible debt raised in 2023.

Marshmallow’s funding is coming at a complex moment for insurance startups in Europe.

On one side, there is the grim story of WeFox.

Backed by SoftBank, Omers, Salesforce and dozens of others, WeFox’s valuation rose to as much $4.5 billion by 2023. Just two years later, after years of losses and complications in its distributed/broker-based business model, WeFox has fallen on hard times. The company has been selling off parts of its business and picking up lifeline financing to stay afloat.

Yet there are also some brighter signs of insurtech startups building more sustainable businesses. And those that can demonstrate a strong technology story are getting attention from investors.

Just last week, Ominimo — a new startup out of Poland — picked up a major strategic investor that invested $10 million at a valuation of over $200 million. It was Ominimo’s first time raising outside money after becoming profitable while bootstrapped. Like Marshmallow, the startup started off with car insurance and is rethinking actuarial formulas and using AI to make new inroads into risk prediction.

While data science and AI are quickly becoming table stakes for insurance startups, there are other details about Marshmallow that set it apart from the pack and even some of its bigger competitors (like the price-busting mega-retailer Tesco).

The ideas of inclusivity and diversity that underpin how Marshmallow is approaching its target customer base run deep at the startup.

Kent-Braham co-founded London-based Marshmallow with his identical twin, Alexander, and David Goaté. The twins really do look a lot alike. “You could actually be talking to Alexander right now!” Oliver joked when we spoke for this story. More seriously, though, the startup is an underrepresented rarity in another way, too.

It is one of what appears to be only two “unicorn” startups in the U.K. from a Black founder, the other being WorldRemit. The statistics are not hugely encouraging outside the U.K. either; one 2024 study found that across the U.K. and the U.S., only 3% of startups with valuations of over $1 billion have Black founders.

At a time when diversity, equity and inclusion programs are being dismantled in the U.S., it’s notable that Marshmallow’s investors see particular strength precisely because of its diverse leadership.

“This is a very strong founding team,” Devon Kirk, GP and co-head of Portage Capital Solutions, said in an interview. “We think that financial services benefits from different perspectives and leaders coming up with innovative solutions to address those needs.”

Even as artificial intelligence has made strides in summarizing research papers or solving mathematical problems, professionals in many fields still believed humans would be in charge of coming up with ideas for new discoveries.

Now AI is getting good at such brainstorming, too.

OpenAI is preparing to launch new AI models as soon as this week that can connect the dots between concepts from different fields to suggest new types of experiments involving anything from nuclear fusion to pathogen detection, according to three people who have tested the models but are not authorized to speak about it.

The Takeaway

• New AI aims to resemble inventors like Nikola Tesla who blended information from multiple fields

• OpenAI believes it can charge $20,000 per month for doctorate-level AI

• A gap remains between ideas AI can generate and the scientists’ ability to verify them

If the upcoming models, dubbed o3 and o4-mini, perform the way their early testers say they do, the technology might soon come up with novel ideas for AI customers on how to tackle problems such as designing or discovering new types of materials or drugs. That could attract Fortune 500 customers, such as oil and gas companies and commercial drug developers, in addition to research lab scientists.

The apparent improvements highlight the benefits of AI models focused on reasoning, which the ChatGPT maker debuted in September. Reasoning models perform better the more time they can spend processing answers, and they excel in problems with solutions that can be verified objectively, such as math theorems. OpenAI last year shifted its research effort to reasoning as traditional methods of improving AI slowed.

The progress of such software helps explain why OpenAI believes it could eventually charge upward of $20,000 per month, or 1,000 times the cost of a basic ChatGPT subscription, for AI that can replicate the work of doctorate-level researchers.

The ability of OpenAI’s upcoming models to synthesize new ideas represents one of several pillars the firm is developing to match or outperform humans at “most economically valuable work,” otherwise known as artificial general intelligence.

OpenAI says it has trained its reasoning models to include a large base of knowledge across multiple domains, such as biology, physics and various types of engineering.

What makes the soon-to-be-released reasoning models unique is that they can compute an answer or suggest an idea using information from multiple fields—physics and engineering, for instance—simultaneously, two of the people who tested the models said. Most scientists, in contrast, must collaborate with experts from other fields to come up with similar answers or ideas, these people said. That’s a time-consuming process.

In that sense, the AI aims to resemble the kind of inventors who blend information from multiple fields, such as Nikola Tesla and Richard Feynman, whose knowledge in physics, engineering and math drove discoveries in electrical devices and quantum mechanics, respectively.

OpenAI’s newer reasoning models aren’t available for purchase yet but are already powering features in ChatGPT such as deep research, which browses the web to compile research reports.

With that technology, scientists can direct the AI to read publicly available literature in various scientific domains, summarize the experiments researchers have already conducted and suggest new approaches that haven’t been tried yet, said a person who has tested them. (An OpenAI spokesperson did not have a comment for this article.)

Saving Time

What highlights the potential of the new reasoning models is the fact that even the less-advanced reasoning models already available have been a game-changer for scientists.

For instance, Sarah Owens, a molecular biologist at Argonne National Laboratory in Illinois, was curious whether a statistical technique from ecology science could help her look for pathogens in wastewater.

In February, she used OpenAI’s o3-mini-high model, which is commercially available, to design a small-scale study of whether she could use an ecological technique known as occupancy modeling to predict the presence of pathogens even if they don’t show up in a particular water sample.

Trying to design the study without AI “would have taken days,” she said. “To have this information synthesized in this way really saved a lot of time.”

Similarly, Massimiliano Delferro, a chemist at Argonne who conducts research on plastic waste recycling, asked o3-mini-high whether he could use a certain technique to break down plastic waste more efficiently, and to design an experiment to test that.

He said the model returned instructions for an experiment, including a range of temperatures and pressures to use, significantly faster than he could design the experiment himself.

“That’s when I went from a skeptic to excited,” he said.

In another example, scientists at an “AI Jam Session” in February at Argonne were impressed with the ability of o1-pro and o3-mini-high to determine the potential environmental effects of building power plants and mines in specific geographic regions, according to an attendee who used the models.

“Really, where we’re headed is towards models that spend a lot of time thinking really hard about important scientific problems, and that’s something that I hope over upcoming years will make all of you 10 times or 100 times more effective,” OpenAI President Greg Brockman said at another February “AI Jam Session” event, held at Oak Ridge National Laboratory in Tennessee, which OpenAI co-hosted with the Department of Energy for 1,000 scientists from nine federal research labs.

OpenAI has said it would give several national labs private access to a reasoning model for their research, hosted on a supercomputer at Los Alamos National Laboratory.

Still, in many cases a gap will remain between the ideas AI can generate and the scientists’ ability to verify the answers.

For instance, the upcoming reasoning models can suggest how strong a laser should be to generate a certain amount of energy when it hits a capsule of fuel, but a scientist would still need a simulator or other software to test how promising the suggestion is, said a person who has tested it.

Testing With Robots

Suggestions involving chemistry or biology might require testing in a physical lab. Robots could be used to automate the experiments, scientists say, but such systems could take a long time to develop.

In these examples, combining reasoning models that can suggest hypotheses with AI agents that can access simulators or robots to test them would be key to accelerating new discoveries, these scientists say.

Some AI and robotics firms, including startup FutureHouse and Alphabet-owned Isomorphic Labs, say they are using similar approaches to automate biology research and drug discovery, respectively. In January, Sam Rodriques, CEO of FutureHouse, posted a photo of a humanoid robot on X that would eventually be able to help the startup run experiments in a physical lab.

Today’s commercially available AI agents operate primarily through computer browsers or other software applications, and they remain imperfect. They aim to automate complex tasks, from coding to processing human resources and IT queries from employees.

OpenAI has publicly released Operator, a computer- and browser-using agent, which still makes mistakes and struggles when navigating complex sites, some users say.

To improve Operator, OpenAI plans to collect data from the people who use it, filter out examples where the agent fails to complete the task and train the agent on the remaining examples, according to a person who has worked on the product.

This process, known as reinforcement learning from human feedback, is a bedrock of how OpenAI and other firms improve AI to learn tasks beyond the data used to train it.

Before the advent of reasoning models, if a traditional conversational AI model “discovered a new theorem that had never been solved, it would have been penalized because [the new theorem wasn’t] in its training data,” said David Luan, head of Amazon’s AGI SF Lab and a former head of engineering at OpenAI.

Allowing agents to “play” and try different approaches in an environment such as a browser and rewarding them when they complete a task can help them learn new skills, he said.

OpenAI, for its part, has been developing an advanced software coding agent it hopes can eventually automate the work of AI researchers and engineers, including generating code for their experiments related to AI models.