The model is the first release in OpenAI’s Life Science model series.

OpenAI has announced plans to roll out an early version of GPT-Rosalind, its AI reasoning model designed to support research across biology, drug discovery and translational medicine. 

In a statement on Thursday (16 April), OpenAI explained that on average, it can take up 15 years to move from target discovery to regulatory approval for a new drug in the US, with progress impacted by the difficulty of the underlying science, as well as the complexity of the research workflows.

The organisation said: “Scientists must work across large volumes of literature, specialised databases, experimental data and evolving hypotheses in order to generate and evaluate new ideas. These workflows are often time-intensive, fragmented and difficult to scale.”

Named after Rosalind Franklin, a pioneering figure in the field of DNA, GPT‑Rosalind is now available as a research preview in ChatGPT, Codex and the API for qualified customers through OpenAI’s access programme such as Amgen, Moderna, the Allen Institute and Thermo Fisher Scientific.

GPT-Rosalind is the latest in a series of AI models focused on life sciences applications, with the space becoming increasingly competitive. Last year, France’s Sorbonne University and Qubit Pharmaceuticals announced the “world’s most powerful” AI model for molecular simulation in pharmaceutical chemistry, FeNNix-Biol.

At the time, the research team claimed that FeNNix-Biol’s capabilities are beyond that of Google DeepMind’s AlphaFold, the Nobel Prize-winning deep-learning machine designed to transform our understanding of the molecular biology that underpins health and disease.

OpenAI said: “This is the first release in our life sciences model series and we view it as the beginning of a long-term commitment to building AI that can accelerate scientific discovery in areas that matter deeply to society, from human health to broader biological research. 

“Over time, we expect these systems to become increasingly capable partners in discovery – helping scientists move faster from question to evidence, from evidence to insight and from insight to new treatments for patients.”

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

The original version of this story appeared in Quanta Magazine.

Picture a bizarre training exercise: A group of runners starts jogging around a circular track, with each runner maintaining a unique, constant pace. Will every runner end up “lonely,” or relatively far from everyone else, at least once, no matter their speeds?

Mathematicians conjecture that the answer is yes.

The “lonely runner” problem might seem simple and inconsequential, but it crops up in many guises throughout math. It’s equivalent to questions in number theory, geometry, graph theory, and more—about when it’s possible to get a clear line of sight in a field of obstacles, or where billiard balls might move on a table, or how to organize a network. “It has so many facets. It touches so many different mathematical fields,” said Matthias Beck of San Francisco State University.

For just two or three runners, the conjecture’s proof is elementary. Mathematicians proved it for four runners in the 1970s, and by 2007, they’d gotten as far as seven. But for the past two decades, no one has been able to advance any further.

Then last year, Matthieu Rosenfeld, a mathematician at the Laboratory of Computer Science, Robotics, and Microelectronics of Montpellier, settled the conjecture for eight runners. And within a few weeks, a second-year undergraduate at the University of Oxford named Tanupat (Paul) Trakulthongchai built on Rosenfeld’s ideas to prove it for nine and 10 runners.

The sudden progress has renewed interest in the problem. “It’s really a quantum leap,” said Beck, who was not involved in the work. Adding just one runner makes the task of proving the conjecture “exponentially harder,” he said. “Going from seven runners to now 10 runners is amazing.”

The Starting Dash

At first, the lonely runner problem had nothing to do with running.

Instead, mathematicians were interested in a seemingly unrelated problem: how to use fractions to approximate irrational numbers such as pi, a task that has a vast number of applications. In the 1960s, a graduate student named Jörg M. Wills conjectured that a century-old method for doing so is optimal—that there’s no way to improve it.

In 1998, a group of mathematicians rewrote that conjecture in the language of running. Say N runners start from the same spot on a circular track that’s 1 unit in length, and each runs at a different constant speed. Wills’ conjecture is equivalent to saying that each runner will always end up lonely at some point, no matter what the other runners’ speeds are. More precisely, each runner will at some point find themselves at a distance of at least 1/N from any other runner.

When Wills saw the lonely runner paper, he emailed one of the authors, Luis Goddyn of Simon Fraser University, to congratulate him on “this wonderful and poetic name.” (Goddyn’s reply: “Oh, you are still alive.”)

Jörg Wills made a conjecture in number theory that, decades later, would come to be known as the lonely runner problem.

Advertisement

Courtesy of Jörg Wills/Quanta Magazine

Mathematicians also showed that the lonely runner problem is equivalent to yet another question. Imagine an infinite sheet of graph paper. In the center of every grid, place a small square. Then start at one of the grid corners and draw a straight line. (The line can point in any direction other than perfectly vertical or horizontal.) How big can the smaller squares get before the line must hit one?

As versions of the lonely runner problem proliferated throughout mathematics, interest in the question grew. Mathematicians proved different cases of the conjecture using completely different techniques. Sometimes they relied on tools from number theory; at other times they turned to geometry or graph theory.

Jack Zhang was 34 years old, three and a half years into running a startup, and sitting across from one of the most powerful investors in Silicon Valley. Michael Moritz of Sequoia had invited him to his home — a place with, Zhang recalls, a couple of floors and a view straight to the Golden Gate Bridge — to make the case for selling.

Stripe wanted to buy Airwallex for $1.2 billion. At the time, the Melbourne company had around $2 million in annualized revenue. The math was almost pretty irresistable: a revenue multiple somewhere near 600 times. Patrick Collison, Moritz argued, was a generational founder. The deal would “compound” into something extraordinary. Zhang listened. He walked around San Francisco for two weeks, restless, unable to think straight. At one point, he said yes.

Then he flew nearly 8,000 miles back home.

“I really went deep on what motivates me to build Airwallex,” he said early this week, speaking to this editor from overseas. “I was three and a half years into the business. The business was growing 100 times in 2018. And I only just sort of tasted what it [was like] to be an entrepreneur. And that’s what I’d been dreaming about.”

Two of his three co-founders had voted against the deal, which helped. But he says the clearest signal came from looking at the whiteboard back in his office. The vision was still there, unfinished: to build the financial infrastructure that lets any business operate anywhere in the world as if it were a local company.

That decision is looking increasingly prescient. Airwallex now claims more than $1.3 billion in annualized revenue and is growing at 85% year-over-year. It processes approaching $300 billion in annualized transaction volume. None of it has come easily — and Zhang argues that’s precisely the point.

It’s a conviction that runs a lot deeper than business strategy. Zhang grew up in Qingdao, a port city in northeastern China, and moved to Melbourne at 15 without his parents, barely speaking English, living with a host family. When his family’s finances collapsed, he took on four jobs to get through a computer science degree at the University of Melbourne, according to the Australian Financial Review — bartending, washing dishes, working graveyard shifts at a petrol station, picking lemons on a farm in the school holidays, which he has called the hardest job he ever had. He went on to spend years writing trading code in the front office of an Australian investment bank, a job that paid well and never felt “deeply fulfilling.”

Before Airwallex, he started roughly 10 businesses: a magazine at age 14, a real estate development company, import-export operations running wine and olive oil from Australia to Asia, textiles going the other direction, a burger chain.

He was running a Melbourne coffee shop when the idea for Airwallex took shape. While trying to pay coffee bean suppliers in Brazil, Indonesia, and Guatemala, his co-founder Max Li kept watching payments disappear into correspondent banking systems — flagged and frozen by American intermediary banks enforcing OFAC sanctions rules, sometimes bouncing back weeks after they were sent. “That pushed me to really look at how correspondent banking works,” Zhang said, “how SWIFT works, and how we could build our own global money movement network.”

That’s still the idea, just scaled up considerably. Airwallex now holds close to 90 financial licenses across 50 markets. Zhang estimates Stripe has roughly half that number at best. Getting those licenses has been immensely time consuming — in Japan alone, the process took seven years. In some emerging markets, the company had to acquire shell companies whose licenses were no longer being issued by central banks, then rebuild the technology underneath them entirely.

“You can’t really vibe-code an integration with Mexico’s central bank,” Zhang said. “We have to have a secure room — you have to do a biometric scan just to walk in to access the central bank integration.”

The point of holding these licenses isn’t regulatory window dressing. In Japan, for instance, Stripe and Square can process payments, but they’re required to immediately transfer funds out to the merchant’s bank account. Airwallex, with its fund transfer operator license, can hold those funds inside its ecosystem. That means a customer can issue bank accounts, issue cards, and spend money without it ever leaving the platform.

The foreign exchange economics alone are substantial: a U.S. merchant settling transactions in Australian dollars avoids the 2% to 3% conversion fee that processors like Stripe typically charge to move money back into U.S. dollars — and can use those local balances to pay local vendors, run payroll, and cover digital marketing expenses, all at interbank rates.

“You don’t really operate like a U.S. company anymore,” Zhang said. “You operate like a company with entities around the world, but without needing to physically set up those entities.”

The slow build was intentional, and Zhang has a framework for it that he returns to often: the “path of maximum resistance.” Every license, every bank integration, every local payment rail that Airwallex painstakingly assembled created a layer that makes it harder to compete against. “It took us six and a half years to get to $100 million in annual recurring revenue,” Zhang said. “But after that, it took just over three years to get to a billion.”

The competitive logic, in his telling, comes down to something basic about what it means to own infrastructure versus riding someone else’s. If you don’t control the end-to-end payment workflow and something goes wrong, you can’t access the underlying data to explain it to your customer. You can’t extend new products cleanly on top of someone else’s stack. “Building on top of other infrastructure,” he said, “is simply not scalable.”

For most of its life, Airwallex and Stripe have mostly operated in different geographies, selling to different buyers. That’s changing. As Stripe pushes deeper into international markets, and Airwallex makes its first serious moves into the United States, the overlap is growing.

The buyer for Airwallex has historically been the CFO’s office in Australia and Southeast Asia, where the company is already well-established — finance directors, treasury teams — which puts it in a different sales motion than Stripe, whose customer acquisition has been driven largely by U.S. developers choosing a default starting point for a new company. More than 90% of Airwallex customers land first on a business account product, and payments and spend management follow from there. Over half are using multiple products, says Zhang.

Still, there are challenges that Zhang doesn’t try to downplay. The biggest may be that Stripe is Silicon Valley’s golden child, its privately held shares having minted millionaires across the tech industry. Another is the accompanying brand gap. Airwallex needs to embed itself in the thinking of engineers and developers — not just finance teams — so that founders reach for it instinctively. “Our brand is just not there yet,” he said. “That’s a harder competition to win.”

It’s a competition being watched closely from a variety of vantage points. Sequoia backed Airwallex early — though the deal was sourced through Sequoia Capital China, which has since spun out and rebranded as Hongshan — and remains one of the company’s largest shareholders. The investment firm Greenoaks Capital holds stakes in both companies, too. Zhang shrugged off any suggestion of awkwardness around those overlapping cap tables. The investors, he noted, are betting on a large market.

Still, it brings up the valuation question. Stripe was valued at $159 billion in a February tender offer — up 74% from a year earlier — after processing $1.9 trillion in total payment volume in 2025. Airwallex, assigned an $8 billion valuation in December, is valued at roughly a twentieth of that. But according to Zhang, Stripe’s payment volume is only about six times Airwallex’s, not 20 times. At 85% annual growth and projecting $2 billion in revenue within the next year, Airwallex is closing the revenue gap faster than the valuation gap would suggest.

Whether the market eventually notices is a different question — one that an IPO, which Zhang says is at least three to five years away, would force into the open.

In the meantime, Zhang says he’s focused on longer-horizon targets: a million customers by 2030, $20 billion in annual revenue, average revenue per customer growing from around $12,000 to $13,000 today to roughly $20,000. A suite of AI-powered autonomous finance products — agents that don’t just surface data but actually execute transactions — is rolling out now. The thesis is that a decade of financial data across the entire corporate finance stack, from revenue collection to treasury management to vendor payments and expenses, has created a training set that no competitor can replicate overnight, he suggests.

Now to see if all that hard work is enough to eat into Stripe’s market share. For now, the competition seems to be playing out at a distance. Zhang and Collison were never friends, but they were friendly while merger talks were ongoing years ago. Last year, Zhang and Collison were both at Greenoaks Capital’s annual gathering. They didn’t speak.

Microsoft has confirmed that some Windows domain controllers are entering restart loops due to Local Security Authority Subsystem Service (LSASS) crashes after installing the April 2026 security updates.

The company also warned that Windows admins may encounter this issue when setting up new domain controllers, or even on existing ones, if the server processes authentication requests very early in the startup process.

“After installing the April 2026 Windows security update (KB5082063) and rebooting, non‑Global Catalog (non‑GC) domain controllers (DCs) in environments that use Privileged Access Management (PAM), might experience LSASS crashes during startup,” Microsoft said in a release health dashboard update.

“As a result, affected DCs may restart repeatedly, preventing authentication and directory services from functioning, and potentially rendering the domain unavailable.”

This known issue only impacts organizations using Privileged Access Management (PAM) and is unlikely to affect personal devices that aren’t managed by an IT department. The list of affected platforms includes systems running Windows Server 2025, Windows Server 2022, Windows Server 23H2, Windows Server 2019, and Windows Server 2016.

While Microsoft is still working on a fix, it advised IT administrators to contact Microsoft Support for Business for mitigation measures that can be applied even after deploying the April 2026 update.

Microsoft has addressed multiple domain controller issues caused by security updates in recent years, most recently resolving Windows Server authentication problems in June 2025, which were caused by the April 2025 security updates.

Almost a year earlier, in May 2024, it fixed another known issue that triggered NTLM authentication failures and domain controller reboots after deploying the April 2024 Windows Server security updates.

In March 2024, it released emergency out-of-band (OOB) updates to fix Windows domain controller crashes after installing the March 2024 Windows Server security patches.

Microsoft is now also investigating a separate issue causing this month’s KB5082063 Windows security update to fail to install on some Windows Server 2025 systems.

​On Wednesday, it also warned admins that some Windows Server 2025 devices may also prompt users to enter a BitLocker key after deploying the KB5082063 update.

AI chained four zero-days into one exploit that bypassed both renderer and OS sandboxes. A wave of new exploits is coming.

At the Autonomous Validation Summit (May 12 & 14), see how autonomous, context-rich validation finds what’s exploitable, proves controls hold, and closes the remediation loop.

Claim Your Spot

Shane Wighton of StuffMadeHere spent months poring over his robotic golf club’s algorithms, fine-tuning the improvement to truly understand ball physics. It now allows the thing to completely comprehend the complexities of ball movement and plot paths to overcome notoriously difficult mini golf holes designed to confound even the best players. The cameras installed around the course monitor the club, ball, and cup with laser-like precision at all times, feeding into the raw data that the system utilizes to make choices.

To complete the initial scan of each hole, someone must sit there and carefully drag a ball covered in reflective markings along each wall, ramp, and floor, while another person activates the optical sensor to only focus on the portions he actually needs. This keeps the captured geometry from becoming disorganized and full of errors. All scan data is then sent into a physics engine named MuJoCo. This program does forward simulations of the ball after impact, accounting for each bounce, skid, and roll that the ball makes, all of which is influenced by surface friction and bounciness levels.

However, matching the simulation to reality has proven to be a challenge. To ensure accuracy, he conducted a series of repeated tests, using motion capture recordings of real balls as a benchmark. An automatic solution attempted to get the numbers correct, but Wighton had to go in and manually change things until the virtual bounces matched exactly what the cameras observed in the real world.

Speed became the next issue he had to address. As it stands, a full simulation would take too long to complete while a player swings the club, so Wighton ended up running thousands of possible club angles and swing speeds ahead of time, for every possible ball starting position, and then each successful sequence that ended up in the cup was added to a large database of stored sequences.

When a real person swings the club, the cameras record the motion from the moment they begin the backswing, and the program takes action. It instantly compares the observed path of the club to a database of stored sequences, selects the winning one, and sends a signal to the motor on the club shaft. The motor whips the club head round in less than a second to the exact angle required for that sequence, and because the club head can swivel around a vertical axis without digging into the ground, the adjustment is seamless even during a quick swing.

Bounces on the ball, however, provide a whole new level of complexity. Following a collision with a wall, the ball’s spin might cause it to fly off at an angle or curve. The simulation accounts for this by considering the whole contact dynamics, rather than simply treating it as mirror reflections. Wighton devised a grid of measured points to capture slight slopes and abnormalities on the ground surfaces he dealt with, as they were not always perfectly level. This means that the physics engine may treat the landscape precisely as it is, rather than assuming everything is smooth as silk.

The heat from the lights and bodies in the room causes the camera tripods to expand somewhat, which would otherwise throw the camera’s precision off. To counteract this, he placed certain fixed reference markers in view, allowing the program to detect these little shifts and correct the entire coordinate system on the fly, ensuring that positions remain accurate even in difficult scenarios. Players simply push a button on the grip, swing the club as usual, and see the club head rotate in midair. The ball follows the predetermined course, soars past obstructions, and lands in the cup, even on holes that appear to be engineered to end your winning streak.

Qualcomm’s next flagship chip is starting to take shape as early leaks suggest the standard Snapdragon 8 Elite Gen 6 might be far less of a compromise than expected.

According to regular tipster Digital Chat Station, Qualcomm is preparing both a standard and Pro version of the chipset. But based on the latest details, the gap between the two may not be as wide as in previous generations.

The biggest takeaway is that the standard model is tipped to use a new-generation Oryon CPU, which could be shared with the Pro variant. That’s a notable shift. It hints that both chips will be built on the same core architecture. They will not split performance tiers as aggressively. The main difference, at least so far, comes down to cache, with the standard chip said to feature 6MB of system-level cache. Meanwhile, the Pro model is expected to push higher.

Jay from the Plasma Channel wanted to take cooking off the grid, eliminating gas and those pesky open flames in the process. He pulled off the trick by putting together a portable burner that generates plasma discharges using rechargeable batteries and blasts them directly into a metal pan. Result? Slap this thing down on a table or picnic blanket and you’ll have a sizzling hot meal in minutes, like scrambled eggs or crispy bacon.

Jay’s plasma research resulted in a system of four distinct high-voltage sources arranged in a square formation. Each starts with a spark bouncing between electrodes that are only one centimeter apart. When the spark forms and is pushed up by the heat rising from it, it strikes the pan sitting on top, and voilà! Four of them functioning together means that the heat is uniformly distributed across the bottom of the pan. It’s a 600-watt beast that can cook two complete dinners on a single charge.

Fire-Maple “Fixed Star 1” Backpacking and Camping Stove System | Outdoor Propane Cooking Gear | Portable…

• HIGHLY EFFICIENT – Our industry leading heat exchange technology minimizes wind impact and reduces boiling times by up to 30% compared to traditional…
• COMPACT AND LIGHTWEIGHT – All accessories including stove, bowl and fuel canister fit into the 1 liter cook pot. THIS SET DOES NOT INCLUDE A FUEL…
• INTEGRATED AND EASY TO USE SYSTEM – The 1 liter cooking pot and the stove are one integrated unit, the built in piezoelectric ignitor ignites the…

The batteries power everything, in this case two massive lithium-polymer packs with a total output of twenty amp-hours. That implies the stove does not require a wall outlet to operate. A cooling fan prevents the electronics from overheating during extended operation. The outside shell is held together with strong adhesive and 3D printed elements, and the translucent panels let you to watch the arcs burning while the stove is in action.

The custom coils, however, are the true core of the system. Jay created resin formers, coiled thousands of turns of thin wire on a machine, and then sealed it all up with more resin, making sure to remove all air bubbles. Trapped air would simply generate a rapid flashover, frying the coil. He upgraded the driver circuits to higher-quality capacitors and transistors because the off-the-shelf ones were failing under load.

One of the initial issues was that the system kept burning up the circuit boards because all four drivers were attached to the same ground ring and were essentially battling each other. So Jay realized he needed to rewire each coil output to its own dedicated electrode pair; voila, no more electrical coupling and smooth sailing. Then there was the issue with the stainless steel bolts; at first, they were producing problems because the surface oxides were making poor connections and melting under current, but swapping to brass resolved that quickly.

When you turn this device on, you hear a continuous hum from the drivers and fan, but when it’s at full power, the arcs produce a wilder, louder sound. The plasma channels are stretching and stabilizing. A thermal camera will show you the pan transitioning from cold to cooking temperature in about a minute. Water tests also proved that energy is being transferred: fifty milliliters of water being heated from seventy to one hundred seventy degrees in just over six minutes, even at reduced power. The plasma itself is a warm 6000 degrees Fahrenheit.
[Source]

Looking for the most recent Mini Crossword answer? Click here for today’s Mini Crossword hints, as well as our daily answers and hints for The New York Times Wordle, Strands, Connections and Connections: Sports Edition puzzles.

Need some help with today’s Mini Crossword? It’s the super-long one as always on Saturdays, and a few of the clues are tricky. But if you play all the other New York Times games, 13-Across will be easy. Read on for all the answers. And if you could use some hints and guidance for daily solving, check out our Mini Crossword tips.

If you’re looking for today’s Wordle, Connections, Connections: Sports Edition and Strands answers, you can visit CNET’s NYT puzzle hints page.

Read more: Tips and Tricks for Solving The New York Times Mini Crossword

Let’s get to those Mini Crossword clues and answers.

Mini across clues and answers

1A clue: What people pay Extra for?
Answer: GUM

4A clue: Pre-meal prayer
Answer: GRACE

6A clue: Physicist Bohr
Answer: NIELS

7A clue: Line up a shot
Answer: AIM

8A clue: Photo ___ (P.R. events)
Answer: OPS

10A clue: “Zootopia,” but not “Zoolander”
Answer: PGMOVIE

12A clue: TV show with the initials “TV”
Answer: THEVIEW

13A clue: New York Times game with weaving, interconnected answers
Answer: STRANDS

Mini down clues and answers

1D clue: More bleak
Answer: GRIMMER

2D clue: Dubai’s country, for short
Answer: UAE

3D clue: Nickname of Seth and Evan’s friend in “Superbad”
Answer: MCLOVIN

4D clue: “See you in the mornin’!”
Answer: GNIGHT

5D clue: Fancy term for “noticed”
Answer: ESPIED

7D clue: Many N.Y.C. addresses: Abbr.
Answer: APTS

9D clue: Uses a needle and thread
Answer: SEWS

11D clue: Egg cells
Answer: OVA

Hastings will not stand for re-election to the company board at its June AGM and will instead ‘focus on his philanthropy and other pursuits’.

Netflix co-founder Reed Hastings is to step aside as chair of the company’s board in June, the streaming giant announced in a first-quarter earnings letter to shareholders yesterday (16 April) that caused stocks to fall with its lower-than-expected forecasts for Q2.

For Q1, the company declared revenue of $12.25bn, year-on-year growth of 16.2pc, operating income of $3.96bn at a margin of 32.3pc and net income of $5.28bn, while diluted earnings per share (EPS) was reported at $1.23. Each of these metrics was higher than the forecast given in the company’s Q4 earning report.

However, the letter’s forecasts for the April, May and June quarter – in terms of, for example, revenue ($12.57bn), year-on-year growth (13.5pc) and diluted EPS ($0.78) – are pessimistic in comparison to stock market predictions, according to Bloomberg, which noted that Netflix shares fell by around 9pc as a result of the Q2 estimates.

In delivering its results and forecasts, the quarterly letter also outlined three “areas of focus” for achieving the company’s “goals”: delivering “more entertainment value” to customers, leveraging technology to improve its service and “improving monetisation”.

Netflix is co-led by dual CEOs Ted Sarandos, who earned $53.9m in 2025, and Greg Peters, who made $53.19m last year; departing chair and co-founder Reed Hasting was the CEO for 25 years after Netflix was created in 1997.

The letter referenced the abandoned deal for Netflix to buy Warner Bros, suggesting the price wasn’t “right” – rival Paramount agreed to buy Warner for $110bn in February in a deal awaiting shareholder and regulatory approval.

Sarandos and Peters paid tribute to the departing Hastings in the letter, calling him “a true history-maker”, the entertainment platform’s “biggest champion” and “part of our DNA”.

Hastings co-founded Netflix in 1997 with Marc Randolph and brought it from an online DVD rental delivery service that went public in 2002 into the realm of streaming. He is a board member of Bloomberg, Anthropic and various educational nonprofit organisations.

The shareholders’ letter said he would not stand for re-election to the company board at its June AGM and would instead “focus on his philanthropy and other pursuits”.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.