AI researchers call for a reset

A growing group of AI researchers at NeurIPS say the field needs a fundamental rethink. Despite huge progress, they argue today’s training methods won’t get us to major breakthroughs in science, medicine or real reasoning.

Amazon’s David Luan put it bluntly: “The way we’re training models today will not last.” Others echoed OpenAI co-founder Ilya Sutskever’s recent view that current approaches struggle to help models generalize beyond what they’ve seen.

The core issue is continual learning. Humans learn from one example. Models don’t. They still need massive curated datasets, reinforcement learning, and endless retraining. Researchers at MIT and OpenAI presented early work on “self adapting” models that update themselves with new information, but it’s still experimental.

Not everyone agrees the field needs a reboot. Anthropic CEO Dario Amodei says scaling existing techniques will be enough to reach AGI and argues progress is still accelerating.

Meanwhile, business demand continues. OpenAI expects around $13 billion in revenue this year, Anthropic about $4 billion, and coding tools like Cursor are on pace to surpass $3 billion collectively. Even without new breakthroughs, the current generation of models is powering huge commercial growth.

The tension is clear. Today’s AI is good enough to sell, but many researchers no longer believe it’s good enough to reach what comes next.

Supported by T-Mobile for Business

Handling big data on the job? We’ve got you. T-Mobile’s fast and secure network was made for everything AI.

America’s largest 5G network provides the scale and flexibility to handle it all, from generating, training and trouble-shooting your biggest AI models.

IBM is acquiring Confluent for $11 billion in cash, a major move in CEO Arvind Krishna’s strategy to stay relevant in the AI era by leaning on open-source infrastructure. Confluent, which went public in 2021, will be bought at a 34 percent premium.

The deal follows IBM’s purchases of Red Hat and Hashi as it builds a stack designed to compete with Google and Microsoft. IBM’s stock has more than doubled since Krishna took over. The company is also pushing deeper into quantum computing, betting larger chip clusters will enable new levels of performance within five years.

The Trump administration will allow Nvidia to sell its H200 AI server chips to China, a step up from the weaker H20 chips previously approved. The H200 is one generation behind Nvidia’s newest Blackwell line, but it is still powerful enough to train top-tier models like GPT-5.

The move ends months of debate in Washington over how much cutting-edge U.S. tech China should be allowed to access. It also raises a key question inside China, where officials have pushed major AI labs to rely less on Nvidia and more on domestic chips. Huawei’s hardware is strong for running existing models, but still lags in training new ones, which could make the H200 more attractive.

Nvidia’s China sales have slowed sharply this year, so the decision is a win for the company. Shares rose about 2 percent after the news. The White House had previously pushed for the U.S. government to keep 15 percent of Nvidia’s China revenue.

Supporters of the move, including adviser David Sacks, argue that restricting Nvidia only helps Chinese chipmakers catch up. Critics say it gives China access to technology that could accelerate its AI ambitions.

Instacart is bringing its app to ChatGPT, letting shoppers place orders directly through OpenAI’s Instant Checkout. Payments run through Stripe for now, with Apple Pay and Google Pay support coming soon.

The move adds another major retailer to OpenAI’s growing e-commerce ecosystem. Walmart, Etsy and Shopify have already announced plans to let users shop through ChatGPT, with Target preparing its own integration.

Instacart has been closely linked with OpenAI for years. It was an early partner for tools like Operator, and several OpenAI executives previously worked at Instacart. ChatGPT product head Nick Turley has said deeper app integrations and task completion are top priorities, and Instacart is shaping up to be a flagship example.

The European Commission has opened a new investigation into Google over its AI generated summaries that sit above search results. Regulators want to know whether Google used publisher content without proper compensation and whether creators had a real way to opt out.

The probe also covers how YouTube videos may have been tapped to train Google’s broader AI systems. Google called the investigation a threat to innovation and said it continues to work with publishers and creative industries on the shift to AI.

The scrutiny comes after concerns that Google’s AI Overviews are cutting traffic to publishers. The Daily Mail said its Google search clicks fell by about half after the feature launched. Regulators now want to determine whether Google has been building profitable AI tools on top of work created by others without clear permission.

Campaign groups welcomed the move but say the damage is already unfolding. Many argue that creators and publishers need immediate opt out controls while the investigation plays out.

The Commission says AI can unlock major benefits, but not at the expense of media diversity and a fair creative economy.

<img src=’https://www.vpdae.com/open/9dbf0529.gif?opens=1’ width=’1’ height=’1’>