I’ve spent the week at one of the world’s top AI research conferences. Here’s what is sticking with me

I am reporting Eye on AI this week from Vancouver, where I’ve been hanging out with thousands of PhD-level AI researchers at the International Conference for Machine Learning (ICML), one of the top annual gatherings for AI talent from elite universities, Big Tech labs, and AI startups. 

I’ll be honest: It’s humbling to be here, surrounded by professors, postdocs, and industry researchers who casually drop references to mathematical proofs and thermodynamics metaphors into everyday conversation. There are thousands of posters, papers, and presentations—far too many to meaningfully absorb, even if I fed them all into ChatGPT. My brain feels like it’s running at max capacity as I try to make sense of talks titled “Controlling Underestimation Bias in Constrained Reinforcement Learning for Safe Exploration” and “Discrete Flow Matching for Graph Generation.”

Still, there’s something electric about being in the room where the future of AI is being debated, defined, and maybe even redirected. I’m a big believer in getting out of my comfort zone—and into a beginner’s mindset—especially in a space where today’s theories might become tomorrow’s technologies. As I wrap up my time at ICML, here’s what’s sticking with me: 

The AI talent wars are in full swing. Meta’s extraordinary, ongoing hiring spree, which has thrown tens and even hundreds of millions of dollars at elite AI researchers lured from OpenAI, Anthropic, Google DeepMind and others, is the talk of ICML. Some clearly see it as over-the-top, with one London-based researcher telling me it is creating “a bubble.” Others envision Meta as a dream gig—one Swedish PhD student, showing off an armful of swag from various exhibition hall booths, including a pair of blue IBM socks—said he could only hope to land an interview there one day. Either way, Big Tech had recruiters working overtime, with private after-hour events for candidates sprinkled out at venues near the Vancouver Convention Center. 

There’s no better place to ask questions. One of the best things about hanging out with really smart people is picking the brains of really smart people. It helps that AI researchers are typically incredibly kind about explaining their work to a curious journalist. In a vast expo hall filled with seemingly-endless rows of poster presentations, you never know when you’ll happen upon a well-known Stanford University professor perfectly happy to spend 20 minutes chatting about AI model behavior and ethical guardrails. How valuable are one-on-one meetings here? How about a machine learning pioneer waxing philosophically about the old days (which are only 10-15 years ago)? Or a former Google Brain researcher patiently explaining the ins and outs of Transformer models? Priceless. 

Scaling up reinforcement learning (RL) is all the rage. That’s not my line—it’s from former Tesla AI chief Andrej Karpathy on X—but RL was everywhere at ICML. (RL, or reinforcement learning, is a training method where AI learns by trial and error, to maximize some reward. That could be points in a game, or the number of “thumbs up” grades an AI model’s outputs receive from a human evaluator.) Now, researchers are taking reinforcement learning techniques and pushing them to much larger scales to train or fine-tune today’s big language and multimodal models. With more data, more compute, and harder tasks, you (hopefully) get models that work to reason better, follow instructions more reliably, and behave more safely in messy, real-world, enterprise settings. 

Lots of researchers are ready to live their founder dream. There is the startlingly young duo out of Princeton building multimodal medical foundation models. The PhD-level intern at Waymo using Pokemon games to stress-test large language models and AI agents on strategy. The ex-Google and ex-OpenAI researchers aiming to leapfrog current AI tech. With VCs walking the halls during the day and offering open bar events at night, it was easy to view this research conference as partly a hotbed of current and potential founders. 

I’m ready to head back to New Jersey, with plenty of new story ideas and sources in tow. With that, here’s the rest of the AI news.

Sharon Goldman
sh************@*****ne.com
@sharongoldman

AI IN THE NEWS

Scale AI cuts 14% of workforce after Meta investment, hiring of founder Wang. Just weeks after Meta invested $14.3 billion in Scale AI and hired founder Alexandr Wang, the AI startup is laying off 200 full-time employees—about 14% of its workforce, according to CNBC. In a memo to staff on Wednesday, interim CEO Jason Droege, who recently took over for Wang, said the company had scaled its generative AI efforts “too quickly” and had built up “excessive bureaucracy.” Despite the cuts, Droege emphasized that Scale remains well-funded and well-resourced. “These changes will make us more nimble—enabling us to react more quickly to shifts in the market and customer needs,” he wrote in the memo viewed by CNBC. “This structure will allow us to better serve the customers we have today and win back customers that have slowed down work with us.”

OpenAI to take cut of ChatGPT shopping sales in hunt for revenues. According to the Financial Times, OpenAI is planning to take a cut of product sales made directly through ChatGPT as it ramps up efforts to turn the chatbot into an ecommerce platform. Currently, ChatGPT surfaces product listings with links to external retailers. But according to multiple sources familiar with the plans, OpenAI now aims to integrate a full checkout system—allowing users to complete purchases within ChatGPT. Merchants who fulfill orders through this system would pay OpenAI a commission. The company also announced a partnership with Shopify in April as part of its broader push into AI-driven shopping.

Former top Google researchers have built a new kind of AI agent—and they think it could be a step toward superintelligence. Their startup, Reflection, just unveiled Asimov, an agent that learns how software is built not just by reading code, but by digesting everything around it: emails, Slack messages, documentation, and project updates. According to Wired, the goal is to create a more capable AI assistant—and to teach models how humans actually build things, one decision at a time.

FORTUNE ON AI

Elon Musk released xAI’s Grok 4 without any safety reports—despite calling AI more ‘dangerous than nukes’ —by Beatrice Nolan

Commentary: The companies laying off staff for AI today will regret it in five years —by Alexandra Ebert

Delta moves toward eliminating set prices in favor of AI that determines how much you personally will pay for a ticket —by Irina Ivanova

CEO of $14 billion AI firm Perplexity says the secret to success is ‘sleeping with that fear’ that your competitor will steal your idea —by Preston Fore

AI CALENDAR

July 22-23: Fortune Brainstorm AI Singapore. Apply to attend here.

July 26-28: World Artificial Intelligence Conference (WAIC), Shanghai. 

Sept. 8-10: Fortune Brainstorm Tech, Park City, Utah. Apply to attend here.

Oct. 6-10: World AI Week, Amsterdam

Dec. 2-7: NeurIPS, San Diego

Dec. 8-9: Fortune Brainstorm AI San Francisco. Apply to attend here.

EYE ON AI NUMBERS

75.6%