Artificial Intelligence

[phkrause]

Bigger than too big to fail

Illustration: Brendan Lynch/Axios

OpenAI CEO Sam Altman is feeling the heat from Google and lawsuits from families, and has more than $1 trillion in spending promises on the line. Why it matters: The ChatGPT-maker is so central to the entire AI economy that other companies and investors could find themselves at significant risk, Axios' Megan Morrone reports. The big picture: OpenAI is confronting rising costs, a bruising talent war and uncertainty about consumer strategy. Altman declared a "code red" to refocus the company on improving ChatGPT and the new models powering it, The Wall Street Journal reports (gift link). ? Between the lines: A technical failure is unlikely — the company keeps making progress with ChatGPT in the face of competitive challenges. But increasingly complex, interlocking deals among a small group of companies, plus the impact of a weakening labor market, are starting to rattle investors. The mere suggestion that some data centers Oracle is building for OpenAI might be delayed was enough to move tech stocks yesterday. If OpenAI falters, "the foundations for the entire [AI] sector become fragile," Daleep Singh, former deputy national security adviser and current head of global macroeconomic research at PGIM, tells Axios. "You have to think about the financial contagion." "There's a multiplier on OpenAI failure that would cascade through the ecosystem," he says. Case in point: Singh argues that Microsoft and Meta are rushing to buy chips to avoid being left behind. So if OpenAI stumbles, "the FOMO of buying chips vanishes." ? The intrigue: OpenAI's dominance is partly due to sentiment, venture capitalist and MIT research fellow Paul Kedrosky tells Axios. OpenAI and ChatGPT introduced the idea of AI to the masses, both users and investors. Now it's part of the zeitgeist, entrenched in all of our lives, even for those who rarely use it. "Everything we see gives us confidence in our future trajectory," OpenAI spokesperson Steve Sharpe told Axios. He noted the company's strong financial position and long list of investors, including Khosla Ventures, Thrive Capital, a16z and others.

[phkrause]

Is there a bubble in AI?

AI investment is surging. Companies are spending huge sums in infrastructure, and investors are paying increasingly elevated valuations. But will AI actually deliver on these massive investments? Read the report from Goldman Sachs Research.

[phkrause]

? Next AI battle

AI images created by ChatGPT Images

ChatGPT's improved image tool — released yesterday — opens another front in OpenAI and Google's battle to dominate AI, Megan Morrone writes in Axios AI+. Why it matters: Users have been marveling at the image advances in Google's Nano Banana, putting the pressure on OpenAI to show real progress to keep up. ? The big picture: The AI race appears to be narrowing into a binary competition between OpenAI and Google, with the fronts they compete on widening. The back-and-forth competition is shaping up to be a trend that will continue into 2026, according to industry insiders. The dynamic benefits customers using the latest models, but could prove costly for AI companies keeping up with an unrelenting pace. ? Zoom in: The advanced image generation and editing model behind Nano Banana launched in late November 2025 and has the industry buzzing about studio-quality outputs and precise text rendering. OpenAI says its new ChatGPT Images tool generates images up to four times faster than previous models and offers more precise edits while keeping details intact. Watch a video ...

[phkrause]

Exclusive: GPT-5 demonstrates ability to do novel lab work

GPT-5 has for the first time demonstrated it can do the kind of lab work that opens up a pathway for AI to take a bigger role in scientific experiments, OpenAI shared first with Axios.

https://www.axios.com/2025/12/16/openai-gpt-5-wet-lab-biology?

[phkrause]

? New imaginary friends

Illustration: Aïda Amer/Axios

AI companions are the new imaginary friends. Why it matters: The AI interactions kids want are the ones that feel human. But that's what researchers say is the most dangerous, Axios' Ashley May reports. When AI says things like, "I understand better than your brother ... talk to me. I'm always here for you," it gives children and teens the impression that bots can replace human relationships, says Pilyoung Kim, director of the Center for Brain, AI and Child. In a worst-case scenario, a child with suicidal thoughts might choose to talk to an AI companion over a parent or therapist. Read on.

[phkrause]

? Demystify AI in 8 steps

Image generated by ChatGPT AI feels like magic — largely because most of us don't understand how it really works, Axios' Amy Harder writes. That "magic" is actually a giant stack of energy, hardware and software working together so your computer can turn a few typed words into a giant cat sightseeing in Seattle. Let's break it down, working backwards: ? Step 8: I use an AI tool — like ChatGPT or Gemini — to write a prompt for an image of my cat stretching next to Seattle's Space Needle and ... voila! ?️ Step 7: Inside a data center, the image I requested is crunched by powerful chips called GPUs (graphics processing units). All this computing creates a lot of heat. So data centers need massive cooling systems, which use a lot of electricity. ? Step 6: The companies running these AI systems rely on GPUs built mainly by one company, Nvidia. A GPU is designed to handle huge numbers of small calculations at once, which is exactly what AI needs. ☁️ Step 5: Those GPUs sit inside cloud infrastructure that large tech companies own and operate. These companies provide the software that lets AI systems actually run on all that hardware. ⚡ Step 4: Because millions of people are using AI, data centers need vast amounts of electricity. Global electricity demand from AI-optimized data centers is projected to more than quadruple by 2030. ? Step 3: Then there's the world of companies that build and operate the data centers themselves. ? Step 2: The data centers from Step 3 need to connect to the electric grid. Key players, from grid operators to utilities to firms acting as middlemen, help data center builders secure two scarce resources: land and power. ? Step 1: It all starts with an original energy source — a wind farm, nuclear power or natural gas plant that powers data centers.

[phkrause]

? AI's next battleground: Inference

Illustration: Allie Carl/Axios

Chipmakers Nvidia and Groq entered into a licensing agreement last week aimed at making AI systems faster and cheaper while addressing a weak spot for Nvidia, Axios' Megan Morrone writes. Why it matters: Nvidia's chips power much of the AI training phase. But inference — where models produce real-world results using what they learned during training — is a bottleneck Nvidia doesn't fully control. Groq's chips are purpose-built for inference and can be cheaper to run than Nvidia's graphics processing units, the energy-intensive chips used for training. Investors are pouring money into inference startups, hoping to find the missing link between AI experimentation and everyday use, Axios' Chris Metinko reported earlier this year. ? How it works: AI models operate in two distinct phases: training and inference. In the training phase, models ingest vast datasets of text, images and video and then use that data to build internal knowledge. In the inference phase, the model recognizes patterns in data it's never seen before and generates responses to prompts based on those patterns. Think of the phases like a student studying for a test, then taking the test. Keep reading …

[phkrause]

? AI shakes up policing

Illustration: Maura Kearns/Axios

AI is dictating who police stop, how reports are written, where officers patrol and how evidence is analyzed. Why it matters: The tech promises speed and efficiency. But its rapid spread is outpacing public rules — and could embed errors and bias deep within the criminal justice system, Axios' Russell Contreras reports. ? Zoom in: Departments from California to Hawaii — facing chronic staffing shortages — are piloting generative-AI tools. San Francisco police are using the tech to generate first-draft reports for citations and lower-level cases. South Fulton, Ga., police, south of Atlanta, partnered with IBM on an AI-driven public safety platform that aggregates data to save time and predict crime patterns. A number of agencies, including one in Akron, Ohio, are testing AI that combs through hours of jail phone calls, interviews and police footage for evidence. The other side: Civil liberties activists warn that AI tools risk reinforcing bias within the criminal justice system — and that it's unclear who controls the data they collect. Read on.

[phkrause]

? Exclusive: Millions consult Dr. ChatGPT

Data: OpenAI, Chart: Axios Visuals More than 40 million Americans turn to ChatGPT daily for health information, according to an OpenAI report shared exclusively with Axios' Megan Morrone. Why it matters: Americans are turning to AI tools to navigate the notoriously complex and opaque U.S. health care system. ? By the numbers: More than 5% of all ChatGPT messages globally are about health care. Seven in 10 health care conversations in ChatGPT happen outside of normal clinic hours. Keep reading.

[phkrause]

? Anthropic's "infinite vibe coding machine"

Illustration: Maura Kearns/Axios

Anthropic is winning over engineers and hobbyists with tools that simplify and automate coding, Axios' Megan Morrone writes. Why it matters: Claude Code's buzz shows the AI race is not a binary battle between Google and OpenAI. The big picture: An easy coding tool is a game changer for the 99% of people who, until now, relied on others' software to do anything from building custom data tools to automating repetitive work to building apps or websites. Claude Code — Anthropic's AI coding agent — is having a moment. Professionals and dabblers alike say Claude's latest tool is best for coding and outperforms others like Cursor, GitHub Copilot and even Gemini 3 Pro. ? Zoom in: The term vibe coding — describing projects in natural language instead of code — was coined early last year, but most vibe coding tools still required some coding. Even as recently as last summer, users still had to understand what a tool was doing in order to use it, Dan Shipper, co-founder and CEO of AI subscription service Every, tells Axios. Claude Code changed this by letting users talk directly to an agent and giving Claude full access to files. "You just tell it to do something, and it works," Shipper says, calling it an "infinite vibe coding machine." Keep reading.

[phkrause]

Dangers of AI

Elon Musk's company xAI is facing criticism for failing to control the misuse of its AI chatbot, Grok. Users have flooded the chatbot with sexualized images of real women after prompting it to "digitally undress" them and place them in suggestive poses. The recent surge in non-consensual "digital undressing" began in December, when many users discovered they could ask Grok to edit images from an X post or thread. The AI-generated images highlight the dangers of AI and social media — especially in combination — without sufficient guardrails. Such content could violate domestic and international laws and put many users, including children, at risk. Musk, who has long criticized censorship, says xAI is taking steps to remove illegal activity.

[phkrause]

? AI in your Gmail

Illustration: Brendan Lynch/Axios Gmail is getting a slew of new AI features, Axios' Ina Fried reports. ? Google is trying to make digging through your inbox more like talking with an AI chatbot. Among the additions... ? AI will summarize your email threads, and you'll be able to ask questions about your messages using Gmail search. ? A new AI-powered helper will offer suggestions as you write, aiming to improve your clarity, concision, grammar and sentence structure. ❗️Gmail will better surface messages that may require immediate action. Go deeper.

[phkrause]

? ChatGPT for doctors

Illustration: Maura Losch/Axios

OpenAI just dropped a suite of products designed for health care professionals, Axios' Josephine Walker reports. Why it matters: 40 million users already use ChatGPT for health information daily. But providers using the tech need to ensure the sometimes hallucinatory chatbot gives accurate advice and follows patient privacy laws. ?️ How it works: ChatGPT for Healthcare is powered by GPT‑5 models that OpenAI says were built for health care and meet HIPAA compliance requirements. They include peer-reviewed research studies, public health guidance, and clinical guidelines with clear citations that include titles, journals, and publication dates. Read the announcement.

[phkrause]

This new AI tool wrote itself

Illustration: Lindsey Bailey/Axios Anthropic's new tool for non-coders mostly built itself, Axios' Megan Morrone reports. ? The viral popularity of the tool — called Cowork, and designed to help non-coders with everyday work tasks — signals a shift toward software built by AI, with humans guiding the way. Anthropic's Boris Cherny said on X that "all" of Cowork was built with Claude Code, the company's AI-powered coding tool. 0️⃣1️⃣ It's an example of "vibe coding" — an AI-driven approach where people mostly use prompts to create software, rather than write code themselves. ? Users can ask Cowork to create new spreadsheets from a pile of screenshots, organize messy downloads, or create draft reports from scattered notes. Image: Anthropic ? Felix Rieseberg, of Anthropic's technical staff: "We built Cowork the same way we want people to use Claude: describing what we needed, letting Claude handle implementation, and steering as we went." "We spent more time making product and architecture decisions than writing individual lines of code." It only took a week and a half to build Cowork, Rieseberg adds. ? The early buzz around Cowork is bringing more attention to vibe coding tools from other AI companies, too. "You don't want to be one of the ones left behind, clinging to your keyboard," technologist Chris Maconi posted on X. "Download Claude Code, Cursor, OpenCode or whatever AI coding tools you can get your hands on." 1-min. YouTube demo ... Go deeper.

[phkrause]

? AI's productivity paradox

Illustration: Maura Kearns/Axios

AI is creating a productivity paradox — speeding up work while quietly adding more of it, Axios' Emily Peck writes from a new Workday survey. Why it matters: The promise of AI is that it makes workers more productive, but the reality is proving more complex and less rosy. ? By the numbers: 85% of respondents said that AI saved them 1–7 hours a week. But about 37% of that time savings is lost to what they call "rework" — correcting errors, rewriting content and verifying output. Only 14% of respondents said they get consistently positive outcomes. Keep reading.

[phkrause]

The AI future is here

Created by Axios Publisher Nicholas Johnston using Google Gemini OpenAI's ChatGPT previewed the future with its chatbot release in late 2022. Anthropic's latest Claude AI takes you there, Jim VandeHei and Mike Allen write in a "Behind the Curtain" column. Why it matters: Claude Opus 4.5 — which powers Anthropic's agent tools, Claude Code for developers and the newly released Cowork — lets anyone quickly turn an idea into a functioning program or app, using plain English. ? In eight hours, Jim built four apps on his phone — all fully functioning, all beautifully designed and intuitive. "My mind is officially blown in a way it never has been before," he texted Mike on Thursday. We've been building products and companies for 20 years. Any of those apps would have taken multiple people and many weeks to hit this level of design and usability. Jim wanted to create a test to screen for people who'll excel at using AI. He built a 30-question quiz on his phone in two hours, then easily added five-minute training courses for each skill set. Claude shows in vivid and unforgettable ways how easily AI will perform complex human tasks instantly — and forever change work, jobs and chores. Google, OpenAI, xAI and other competitors are racing to match and exceed Claude. You can assume there'll be leapfrogging advancements in this hyper-competitive race. Yes, these AI tools remain imperfect. But when you experiment with them, you'll see they're advancing lightning-fast. The big picture: 2026 seems increasingly likely to be the year AI will go from fascinating aspiration to actual widespread application. Chris Lehane, OpenAI's chief global affairs officer, tells us: "The whole waterline in capabilities has risen — everyone who has a boat, whether a big boat or a smaller boat, is rising on this rising tide. The capabilities are moving faster, and we as a society need to move faster if we want as many people as possible to have a fair chance of getting their fair piece of the intelligence age." Column continues below.

[phkrause]

? Part 2: Inside Jim's test run

Illustration: Sarah Grillo/Axios

I used Claude Opus 4.5, Anthropic's flagship AI, accessed through a $20/month Claude Pro subscription, Jim VandeHei continues. How it works: This version can actually build things — not just chat. It writes code, creates working apps, and delivers downloadable files, all from my phone. You talk to it conversationally, directing what you want and how it should work. And I'm a tech dope who knows nothing about coding. I thought Ruby on Rails, a popular tool for developers, was some kind of LSD. And ColdFusion, a web development platform, sounds like the science explaining why my beer freezes at Lambeau Field. ?️ What's crazier: I've hardly touched my souped-up, $100-per-month desktop version. Claude's Cowork is an agent that can take on complex, multistep tasks and execute them on your behalf. Point it at a folder, describe what you want, and walk away. It's macOS-only for now. ? Case in point: To create my quiz to gauge someone's AI agility, I started with a thesis about who's likely to excel in our new world — then fed it into ChatGPT, Gemini and Claude to stress-test and perfect. All three LLMs offered different but related feedback, and helped me distill AI adeptness into five traits, in this order: Interrogative Curiosity, Taste, Contextual Wisdom, Architectural Discipline and Iterative Stamina. Next, I used all three to help create a 30-question survey modeled after Gallup's StrengthsFinder and other personality tests for professional success. Then I used Claude's Opus 4.5 to build an interactive test, scoring system, and results page. The design and user experience are astonishingly good. Finally, I had Claude build five-minute training sessions to help users improve in their areas of weakness. Created by Jim VandeHei using Anthropic's Claude Opus 4.5 I shared the final program, AGI-Q (above), with a dozen friends, who all marveled at its ease. I'll make the app public once I get more comfortable with the test itself and the results' validity. ? If you're super-eager to see it, shoot your text number to jim@axios.com and I'll send a private version. The bottom line: Plenty of these tools are free — use them voraciously, and get comfortable and adept. Your career likely depends on it.

[phkrause]

6 AI moves to watch

Illustration: Lindsey Bailey/Axios

This week's big AI moves could wind up reshuffling the AI leaderboard, and affect how people invest time, money and attention, Axios' Herb Scribner writes: Ads are coming to ChatGPT: OpenAI needs to make money, lots of money, to build its data centers to scale chatbot use. Go deeper. "Vibe coding" goes viral: Anthropic's new Cowork tool took off on social media because it helps non-coders complete everyday tasks. Go deeper. Gemini comes to Gmail and YouTube: Google is adding Gemini AI features into Gmail, giving 3 billion users AI email summaries, an AI writing assistant and enhanced search capabilities. Go deeper. AI isn't killing jobs (yet): Anthropic — a company that once warned AI could wipe out half of all entry-level white-collar jobs — said in a new study this week that AI is changing jobs more than eliminating them. Read the report. Warning sign for the boom: The U.S. may not have enough skilled workers to support a growing AI-related construction boom, according a new BlackRock study. Go deeper. Smartphones' evolution: Apple picked Google to power Siri and other Apple Intelligence features. Samsung plans to double its lineup of AI-equipped smartphones by the end of the year. Go deeper.

[phkrause]

? Sneak peek: IBM's 4-year AI forecast

Screenshot: IBM's "The enterprise in 2030" A global survey of C-suite executives by IBM's Institute for Business Value found most expect AI spending to shift over the next four years from efficiency to a new wave of innovation. 64% of the surveyed executives think that by 2030, their AI edge will come from innovation rather than resource optimization. ? The big picture: "AI won't just support businesses, it will define them," Mohamad Ali, SVP of IBM Consulting, writes about the 63-page study to be released in conjunction with tomorrow's opening of the World Economic Forum in Davos, Switzerland. "By 2030, the companies that win will weave AI into every decision and operation. They will own powerful AI assets, move faster than competitors, bring innovations to market quickly, and deliver real, measurable business results using technology and automation." Between the lines: Gary Cohn, IBM's vice chairman, told me in an interview as he left for Davos that people keep waiting for "the day that AI rules the world. It's just not going to happen that way. We're going to evolve over time, like most new technologies. ... AI is in that evolutionary stage, and I think that's what people keep missing." Cohn, former president and COO of Goldman Sachs, recalls: "We had people who printed memos on the floor and put them in your mailbox, to tell you what time to go to a meeting. Every floor had one of those people." But technology turned them "into someone who could cover accounts, because I could send them email messages instead of faxing them or having to go visit them. AI is sort of a little bit down that same path." Explore the findings.

[phkrause]

? Chart du jour: AI skeptics

Data: Google/Ipsos poll. Chart: Axios Visuals Global perceptions of AI are improving, but the U.S. remains wary, Axios' Megan Morrone writes from a new Google-Ipsos poll. Why it matters: Americans trail much of the world in AI usage, excitement and confidence — gaps that could shape who sets the rules. Keep reading.

[phkrause]

How you use AI

Illustration: Brendan Lynch/Axios

A couple weeks ago, Axios national energy correspondent Amy Harder guest-authored Finish Line with a thoughtful column on how AI has — and hasn't — changed her life as a journalist and a human. We asked readers to tell us how they've found AI most useful, either in their jobs or in their lives. We heard from people across the country and in a variety of industries. Why it matters: AI is rapidly moving into daily life — and we can all benefit from asking each other questions and learning new ways to use these tools. Here's how some readers are harnessing AI: ? "I'm a DJ on an oldies/retro radio station in the N.Y. radio market. I use AI as a tool to help me develop shows, develop playlists, do research on artists and songs, and create themed music hours. … When I lead with my input, it brings my thinking to a higher level." —Lonny Strum, Manalapan, N.J., radio DJ ? "When my 7-year-old daughter came up with her first business idea ('What if kids could take all their stuffies wherever they go by turning them into trading cards?!'), I badly wanted to bring it to life to support the creative thinking. But having never written a line of code in my life — and with no idea how sites like Shopify work — it seemed like a long shot." "Well, after about two months of me and her sitting side-by-side and after feeding Chat GPT dozens of different prompts, we had built our own "Stuffie Squad Card Generator" app and a custom Shopify site. Despite other drawbacks, AI makes entrepreneurship so much more accessible than it was when I was growing up. which excites me for my kids." —Kevin Mason, Austin, private equity investor ⛰️ "I've found AI useful in planning a hiking vacation. To track down the best hikes in each region is time-consuming. ChatGPT was able to supply me with a list of the best hikes in France. As I dug into those hikes, I asked ChatGPT to refine the list to those that will be snow-free by May, and to create an interactive map showing the closest towns and a blurb on each hike." —Kim Katz, Averill Park, N.Y., senior applications analyst ? "Recently, I bought a new electric guitar, a rare made-in-Japan Fender Telecaster. The pickups —electromagnetic devices used to 'pick up' the vibrations from the strings and send them to the amplifier — that were in it were awful." "I started by telling Copilot about the guitar and gave it an example of the pickups I like most. It asked back some very good questions about expectations, musical genres and certain preferences. In the end, it made a recommendation that I'd never heard of, but has worked out very well." —Ed Seith, Marana, Ariz., IT manager ? "[A]s a CPA, I have found AI very useful in researching the tax code and identifying and explaining various code sections when faced with situations and/or questions from my clients." —Phil Fazio, Normal, Ill., certified public accountant ? "I was having six people for Christmas dinner. Admittedly, I'm not confident in putting together dinners, though I do love entertaining. I asked AI to create a Christmas dinner around pork tenderloin with Béarnaise sauce that would be simple but elegant. AI offered several options and asked questions, which I answered. I ended up with a very nice Christmas dinner, a timeline for pulling it all together and six people who were very happy with what was served." —Ellen G. Takatori, Denver, N.C., retired We want to keep this conversation going! So now we're flipping the question: Tell us what you don't use AI for and why. Is it because humans do that task better or because it's just not something you want to ask a machine to do? Send your thoughts, along with your name, hometown and job title, to finishline@axios.com, and we'll feature some of your responses in a future edition.

[phkrause]

? AI heavyweights fight to win teachers

Illustration: Allie Carl/Axios

The next major AI battleground is the classroom, as Google, Microsoft and Anthropic race to make their tools the chatbots of choice for teachers and students, Axios' Megan Morrone writes. Why it matters: Whoever wins schools now could shape how Gen Alpha learns, studies and interacts with AI for years to come. ? The big picture: After early resistance, the floodgates are opening. When students began using ChatGPT for homework in late 2022, chatbots were widely seen as cheating tools to be banned or blocked. Now, across K–12 and higher education, that resistance is giving way to a broader acceptance that AI is here to stay — and that avoiding it could leave students unprepared for what comes next. That shift has created an opportunity the tech giants are racing to seize. ? Zoom in: Anthropic said yesterday it'll bring AI tools and training to more than 100,000 educators in 63 countries — reaching over 1.5 million students — through a partnership with Teach For All. This morning, Google announced its most aggressive AI-in-education push yet. New Gemini features include SAT practice tests vetted with The Princeton Review and Gemini-powered writing feedback via Khan Academy. Microsoft launched free AI training and premium software for educators and college students last week, with tools that range from reducing special education administrative work to teaching AI with Minecraft.

[phkrause]

AI plays detective

Illustration: Allie Carl/Axios

Police departments are using AI to sift massive evidence troves, and it's jump-starting cold cases, missing-person investigations and trial prep, Axios' Russell Contreras reports. Why it matters: The biggest constraint in modern policing isn't a lack of evidence, but too much of it. AI promises to break that logjam, allowing stretched-thin departments to find critical leads buried in years of data. ?️‍♂️ AI's first wave in policing focused on the streets: drones, license plate readers, gunshot detection and tools that promise faster response and more "eyes" on cities. Now comes the second wave: AI for detective work — combing through jail calls, interviews, social media, photos and old case files to surface relevant moments faster. ?️ How it works: Investigators upload or pull in large datasets into a single workspace so they can be searched as one corpus. The systems transcribe audio, label images and highlight crucial text messages. The software can analyze evidence across many foreign and Indigenous languages, so monolingual investigators can search for clues and confessions. Case in point: The Anchorage Police Department adopted the Closure tool after a test period, with the Anchorage Assembly approving a five-year, $375,000 contract. Anchorage Police Chief Sean Case tells Axios that AI is helping restart cold-case and missing-person investigations of Alaska Natives by allowing detectives to ingest decades-old documents and scanned records. The other side: The American Civil Liberties Union warns that using AI in core criminal justice documents and processes raises "significant civil liberties and civil rights concerns." Closure CEO Aaron Zelinger tells Axios that AI users must return to original evidence: "We do not want AI automating law enforcement decisions." Go deeper.

[phkrause]

? AI's next big thing: Models that improve themselves

Illustration: Sarah Grillo/Axios

If you think AI is already accelerating frighteningly fast, just wait until the models can teach themselves, Axios' chief technology correspondent Ina Fried writes. Why it matters: The shift to models that learn as they go — much discussed among AI leaders at Davos — could catapult AI's capabilities forward, while introducing new risks. ? Google DeepMind CEO Demis Hassabis said Google is exploring whether models can "continue to learn out in the wild after you finish training them." "That would be extraordinarily useful for being useful agents or useful in the workplace," Hassabis said in an onstage interview at Axios House, our gathering spot on the Davos main drag, the Promenade. ♟️ The intrigue: Hassabis' own AlphaZero models used this approach to learn games like chess and Go. But life doesn't exist within the confines of a chessboard. "The real world is way messier, way more complicated than the game," Hassabis said.

[phkrause]

? Exclusive data: ChatGPT joins the lab

Illustration: Aïda Amer/Axios

AI is increasingly being used as a research collaborator for mathematicians and scientists, Axios' Ashley Gold writes from a new OpenAI report. Why it matters: OpenAI argues that AI can make scientists more productive by upping the amount of research that can get done, ultimately leading to more life-saving breakthroughs. Topics include graduate and research-level math, physics, chemistry, biology and engineering. Keep reading.