Math needs thinking time, everyday knowledge needs memory, and a new Transformer architecture aims to deliver both
A German research team lets Transformer models decide for themselves how many times they think about a problem. Combined with additional memory, the approach outperforms larger models on math problems.
Qualcomm shrinks AI reasoning chains by 2.4x to fit thinking models on smartphones
Qualcomm AI Research has developed a modular system that brings reasoning-capable language models to smartphones by compressing the models’ verbose thought processes by a factor of 2.4.
Read full article about: As demand for realistic AI training grows, Deeptune raises $43 million to build simulated workplaces
Andreessen Horowitz invests $43 million in Deeptune, a startup that trains AI agents in simulated workplaces.
Deeptune builds simulated work environments where AI agents learn to handle multi-step tasks in software like Slack or Salesforce. CEO Tim Lupo compares the approach to flight simulators for pilots: instead of just learning from text, AI models practice in realistic replicas of workplaces, like those of accountants, lawyers, or software engineers. According to Fortune, Deeptune has already built hundreds of these environments for leading AI labs.
Andreessen Horowitz leading a $43 million funding round signals how seriously the industry takes this training method. Andreessen partner Marco Mascorro told Fortune that AI models are increasingly learning through interaction rather than human-curated data. According to ResearchAndMarkets, the global market for this type of AI training is expected to grow from $11.6 billion in 2025 to over $90 billion by 2034.
Read full article about: OpenAI turns model compression into a talent hunt with its 16 MB "Parameter Golf" challenge
OpenAI challenges researchers to build the best language model in just 16 MB - and uses the competition to scout talent. In an open research competition called "Parameter Golf," OpenAI is asking developers to build the best possible language model under tight constraints: weights and training code combined must stay under 16 MB, and training can take no longer than ten minutes on eight H100 GPUs. Submissions are judged on compression performance against a fixed FineWeb dataset.
OpenAI is putting up one million dollars in computing credits through its partner Runpod. Top performers may get invited for job interviews - the company plans to hire a small group of junior researchers in June, including students and Olympiad winners. The GitHub repository includes baseline models, evaluation scripts, and a public leaderboard. Anyone 18 or older in supported countries can participate through April 30.
The competition for AI talent among big tech companies is more intense than ever. Meta has repeatedly poached top researchers from OpenAI, in some cases offering compensation packages reportedly worth up to 300 million dollars.
OpenClaw-RL trains AI agents "simply by talking," converting every reply into a training signal
AI agents usually throw away valuable feedback from everyday interactions. Princeton’s new OpenClaw-RL framework changes that by turning live signals from chats, terminal commands, and GUI actions into continuous training data. The researchers say just a few dozen interactions are enough for noticeable improvements.
Meta's JEPA architecture outperforms standard AI methods in noisy medical imaging
Researchers have presented an AI model for cardiac ultrasound based on Meta’s JEPA architecture that outperforms common methods such as masked autoencoder or contrastive learning, according to their benchmarks.