Good morning, and happy Friday! Coming to you today from Miami, where I’ll be watching the Heat tousle with my Knicks on Sunday 🍿🍿🍿 

Hope you all have a wonderful weekend!

— Ian Krietzberg, Editor-in-Chief, The Deep View

In today’s newsletter:

AI for Good: Rare diseases

Rare diseases, by their very nature, don’t get too much clinical attention. 

The problem is that, in aggregate, rare diseases impact a lot of people. 

The numbers: For a disease to be considered “rare,” it must affect fewer than 200,000 Americans. But, according to the Association of American Medical Colleges, there are more than 7,000 known rare diseases, impacting a total of roughly 30 million Americans. They’re difficult to diagnose, and often challenging to treat. 

What happened: Dr. Katharina Schmolly recently launched a startup called zebraMD, which leverages machine learning in an effort to speed up the diagnostic process of these rare diseases. 

Why it matters: The goal, according to zebraMD, is to “get people diagnosed, get them diagnosed earlier, and get them the appropriate care, in every single health setting anywhere in the world, no matter the zip code.”

Everything you need to know about AI agents

Build, optimize, and deploy AI agents with confidence using Galileo's comprehensive eBook.

With this 100-page guide, you’ll learn how to:

Get the eBook now

OpenAI launches preview of wildly expensive GPT-4.5

The news: OpenAI launched a research preview of its latest model, GPT-4.5 on Thursday, billed by OpenAI as the company’s “largest and best model for chat yet.” 

The model is currently only available to Pro users, a restriction that CEO Sam Altman said is due to the startup being “out of GPUs.” 

And it is truly, enormously expensive, costing $75 per million input tokens and $150 per million output tokens. o1, for comparison, costs $15 per million input tokens and $60 per million output tokens. Because of the cost, OpenAI is “evaluating whether to continue serving it in the API long-term.”

In a live-streamed demo, OpenAI engineers inputted the following prompt to show off the model’s skills: “UGHHH! My friend canceled on me again!!! Write a text message telling them that I HATE THEM!!” — not the kind of application indicative of a general intelligence. 

Codeium CEO Varun Mohan said that, based on early testing, the model is “more expensive, slower and worse at tool calling than models like Sonnet.”

Dr. Gary Marcus called it a “nothing burger,” adding that “GPT-5 is still a fantasy.” 

Solve Your AI Challenges with Top Global Experts

Tackle your AI and Machine Learning challenges head-on—without the usual hassle. AI Devvvs connects you to a global network of top-tier professionals, ready to solve your most pressing tech problems.

Hire Your Next AI/ML Expert Now*

Nvidia didn’t come to the rescue 

In the hours directly following Nvidia’s record earnings report, complete with strong guidance for the current quarter, the stock was a little all over the place. It was up, then it was down, then it was up … and so on. 

But by market close on Thursday, after it had all sunk in, investors seem to have decided that Nvidia did not, in fact, come to the rescue. 

Here’s what happened: As predicted, Nvidia bled on Thursday, though I didn’t expect the Street to react quite as harshly as it did — the stock closed the day down nearly 9%, tugging Nvidia’s market cap below $3 trillion for the first time since September. The stock is now down 10% for the year, while the S&P 500 is down only .12% over the same time period. 

“We’re in a stalled, range-bound, slightly irrational market as we wait for policy clarity,” Jay Hatfield, CEO of Infrastructure Capital Advisors, told CNBC. 

James Demmert, chief investment officer at Main Street Research, added: “Nvidia earnings were outstanding, but they come during an extremely jittery stock market.” 

CNN’s greed and fear index has pegged markets at the “extreme fear” side of the gauge. 

“This suggests that long-term AI bulls (present company included) are taking a back seat to concerns about the next couple of quarters,” Deepwater’s Gene Munster said. “Long term, I remain optimistic about Nvidia and the broader AI trade, in part due to the increasing demand for more compute.” 

Mercury: A new breed of LLM

We talk often about paradigms, here, specifically, the need for new ones.

Up until this point, early indicators have seemed to point to neurosymbolic AI as the next new paradigm in the field. But a startup just launched a different kind of large language model (LLM), an approach that, once again, might indicate the start of a new approach, and perhaps, a new paradigm. 

The company is called Inception, and the product is a diffusion LLM. 

But first, the old paradigm: Most of the generative AI models you see today are autoregressive models, a transformer-based approach that was popularized by ChatGPT. Very basically, autoregressive models predict the next token in a sequence by learning statistical patterns in their training data (a token, in language models, refers to broken-up pieces of words). 

Probabilistic inference, in other words. 

A different aspect of this older paradigm involves diffusion models, an approach that is the beating heart behind image generation models. 

What happened: The startup Inception, founded by computer science professors from UCLA, Cornell and Stanford, and staffed by researchers hailing from Google DeepMind, Microsoft, Nvidia and OpenAI, combined these two approaches to create a model they call Mercury, an LLM powered by diffusion. 

According to Inception, the approach is “10x faster and cheaper than current (autoregressive) LLMs.” 

“A paradigm shift is needed to make high-quality AI solutions truly accessible. Diffusion models provide such a paradigm shift,” Inception wrote. “Because diffusion models are not restricted to only considering previous output, they are better at reasoning and at structuring their responses. And because diffusion models can continually refine their outputs, they can correct mistakes and hallucinations.”

The startup called their “dLLM” a “drop-in” replacement for traditional LLMs, supporting all approaches and use cases. The startup’s coding assistant, the Mercury Coder, is five to 10 times faster than current models, according to Inception, producing more than 1,000 tokens per second. 

You can explore the new model here. 

It might, according to Andrej Karpathy, Tesla’s former director of AI, represent a break away from the slate of near-identical LLMs that have proliferated across the industry: “this model has the potential to be different,” he said. 

The claims here — including strong benchmark performance — have not been peer-reviewed, as per usual, and training data and electrical intensity/carbon emissions remain unclear. 

However, it is an approach that I’d love to see studied and examined in more depth; if diffusion has the potential to reduce inference-time compute while boosting performance, it might actually be impactful. And, along those same lines, it might bite the industry worse than DeepSeek. 

🤔 Your thought process:

Selected Image 1 (Left):

💭 A poll before you go

Thanks for reading today’s edition of The Deep View!

We’ll see you in the next one.

Here’s your view on Nvidia and the bubble:

27% of you think it looks like the bubble has done the thing that bubbles tend to do (in other words, burst).

23% think it’s a minor setback, 17% think things will come back, but not until next year and the rest have no earthly idea.

Looks like it:

If you want to get in front of an audience of 450,000+ developers, business leaders and tech enthusiasts, get in touch with us here.