AI Is Moving Deeper Into Medicine, Chips, and Everyday Creation

FEATURED INSIGHT💡

AI-Assisted Rare Disease Research Gets a Real-World Test

Researchers from Boston Children’s Hospital, Harvard, and OpenAI published a study in NEJM AI showing how OpenAI’s o3 Deep Research helped clinicians revisit previously unsolved rare pediatric disease cases. The team reanalyzed 376 de-identified cases that had already gone through extensive expert review, then used the model to surface evidence-linked candidate explanations for specialists to evaluate.

The result: clinicians established diagnoses in 18 cases, a 4.8% additional diagnostic yield. That may sound small at first, but in rare disease work, these were not easy wins. Many of these families had already spent years moving through specialists, genomic testing, and unresolved medical questions. For them, a single confirmed diagnosis can mean closure, clearer care planning, and a better understanding of what happened.

The key point is how the AI was used. The model did not diagnose patients, make clinical decisions, or replace specialists. It acted more like an explanation-first research assistant, connecting clinical features, inheritance patterns, variant evidence, and scientific literature into hypotheses that human experts could investigate. Every diagnosis required expert review, additional testing, clinical confirmation, and lab validation. That distinction matters, especially as AI moves into higher-stakes environments.

Why it matters: rare disease diagnosis is partly a science problem and partly a knowledge-maintenance problem. A child’s genome may not change, but the world’s understanding of genes, variants, and disease links changes constantly. AI-assisted reanalysis could help clinical teams periodically revisit unsolved cases as new evidence appears, making the search for answers more scalable while keeping physicians firmly in control.

Scale AI support on AWS, see how July 9

Customer expectations keep rising. Support budgets don't. On July 9, Fin and AWS are hosting a live executive session on how leading enterprises close that gap: scaling AI-powered support while simplifying how they buy it.

You'll see how to resolve an average 76% of conversations with Fin on AWS enterprise-grade infrastructure, procure through AWS Marketplace to put committed cloud spend to work, and turn the Fin and AWS collaboration into lower support costs. Register for the live session to see how.

Save your spot

ON THE HORIZON 🌅

OpenAI and Broadcom Want to Own More of the AI Stack

OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom “Intelligence Processor” built specifically for LLM inference. Unlike general-purpose accelerators adapted for AI, Jalapeño was designed around the serving patterns behind products like ChatGPT, Codex, the API, and future agentic tools. In plain English: OpenAI wants hardware built for the way its models are actually used.

OpenAI says early testing shows performance-per-watt substantially better than current state-of-the-art systems, although final benchmarks are still pending. The company also says the chip was developed from initial design to manufacturing tape-out in just nine months, with OpenAI models helping accelerate parts of the design and optimization process. If that holds up, it hints at a powerful loop: AI helping design the infrastructure that will run future AI.

Why it matters: inference is where AI reaches users. Training gets the headlines, but inference determines how fast, affordable, and reliable AI products feel in the real world. Faster chips could mean cheaper API calls, smoother coding agents, more responsive assistants, and more dependable access during demand spikes. Jalapeño also signals a broader shift: frontier AI companies are no longer only competing on models. They are competing on the full stack, from chips and networking to products and user experience.

LATEST IMPORTANT NEWS 📰

Anthropic’s Claude Fable 5 returns after federal restrictions. The Trump administration lifted restrictions on Anthropic’s latest Claude models after a weekslong ban tied to cybersecurity concerns. Anthropic says Claude Fable 5 is now widely available again, while its more powerful Mythos 5 model is being restored only to select U.S.-based organizations approved by the federal government. The situation points to a new reality for frontier AI: powerful models are increasingly being evaluated not only for usefulness, but for national security risk, cybersecurity capability, and controlled release.

Meta quietly launches Pocket, a vibe-coded gaming app. Meta appears to be testing a new app called Pocket, which lets users generate small interactive apps and games from AI prompts, then share and explore them in a scrollable feed. The app comes from Meta’s acquisition of the team behind Gizmo, a similar AI game creation platform. It has not been formally announced by Meta, which makes this feel like an early experiment, but the direction is clear: AI creation tools are moving from images and videos into interactive software.

Anthropic is reportedly discussing custom chips with Samsung. Anthropic is exploring a possible custom AI chip collaboration with Samsung, according to reporting cited in the provided material. The company says a diversified hardware stack across Google, Amazon, and Nvidia remains important to its compute strategy, but the broader trend is obvious. OpenAI has Broadcom, Anthropic may look to Samsung, and the AI labs are trying to reduce dependence on a single hardware ecosystem as model demand keeps climbing.

The CRM Behind Every Win

Attio is the agentic CRM that runs the work behind every win.

With agents and automations that build pipeline, chase signals, and move deals forward, Attio orchestrates your revenue work around the clock.

Loved by high-growth startups like Granola, Modal, and Wispr Flow, Attio amplifies what your team can achieve.

Get Started Now

FOR THE TECHNICALLY INCLINED 🛠️

MegaTrain: 100B+ Parameter Training on One GPU

A new paper introduces MegaTrain, a memory-centric system for training 100B+ parameter large language models at full precision on a single GPU. Instead of keeping model parameters and optimizer states persistently on the GPU, MegaTrain stores them in CPU memory and streams each layer onto the GPU when needed. The GPU becomes more of a transient compute engine, while host memory handles the bulk of storage.

The system uses double-buffered execution to overlap parameter prefetching, computation, and gradient offloading, plus stateless layer templates to avoid persistent autograd graph overhead. On a single H200 GPU with 1.5TB of host memory, the researchers report training models up to 120B parameters, along with 1.84x higher throughput than DeepSpeed ZeRO-3 with CPU offloading on 14B models. If the results generalize, this could make large-model experimentation more accessible for labs and teams that do not have massive GPU clusters.

AI TOOL OF THE DAY 🚀

Leapd is an AI platform built to run business operations, engineering, ads, marketing, outbound outreach, and email campaigns around the clock.

Hampton took $440K in planned hires off the calendar

Hampton co-founder Joe Speiser had three roles budgeted: a data engineer, an ops manager, a PM. $440K. He installed Viktor on April 12. Forty-four days later, none are on the calendar, and 18 of his team work with Viktor daily. His VP: we are editors now, not creators.

Get Started for Free