Newsletter Archive

Dharma-AI shows Direct Preference Optimization reduces text degeneration in OCR models by 59.4% on average (up to 87.6%) by training on the model's own failure cases, proving DPO's value beyond chatbots for structured extraction tasks.

Hugging Face · 14 min read Research

Alphabet upsized its equity offering to $84.75 billion—up from $80 billion announced two days prior—explicitly to fund ballooning AI infrastructure and compute spending, underscoring the capital arms race among cloud giants.

Bloomberg Tech · 1 min read Market

AWS shows how to combine Supervised Fine-Tuning and Direct Preference Optimization to boost tool-calling accuracy in language models, with SageMaker infrastructure and evaluation methods for comparing fine-tuned variants.

Amazon AWS ML · 15 min read Tools

US industry groups are warning the Trump administration that the AI boom's demand for memory chips is creating critical shortages affecting auto, medical device, and other sectors, pushing for government intervention to expand domestic supply.

Bloomberg Tech · 1 min read Hardware

Nvidia has acquired Kumo AI, a four-year-old startup that built foundation models for business predictions without retraining. The startup's models are already used by DoorDash, Reddit, and Sainsbury's; founders Josifovski, Raghavan, and Leskovec have transitioned to Nvidia roles.

Fortune · 2 min read Market

Trump signed an executive order requiring AI companies to give the U.S. federal government 30-day early access to frontier models before public release, with a classified benchmark to determine which models trigger the requirement. The framework is voluntary but signals direct government oversight of advanced AI development.

Tom's Hardware · 1 min read Policy

The White House issued a new executive order targeting advanced AI innovation and security, establishing federal guidance that will shape how AI systems are developed, tested, and deployed across industries.

Insideprivacy · 3 min read Policy

Reve 2 and Ideogram 4 both launched with breakthroughs in layout control for image generation, ending what was considered a partially AGI-hard problem. Microsoft's MAI-Thinking-1 reports 97% on AIME 2025 and 53% on SWE-Bench Pro without third-party distillation, backed by a rare 109-page technical report detailing training methodology.

Latent Space · 7 min read Industry