Good morning. I guess it took a few days to sink in, but DeepSeek’s R1 is apparently a big deal.

Wall Street, for one, is not a fan.

But in between massive stock retreats and Twitter conspiracy theories, what we have is a new chapter in the AI race, one that might include price wars and efficiency gains, two things that could prove a boon to adoption.

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

In today’s newsletter:

• 🔬 AI for Good: Molecular dynamics

• 👁️‍🗨️ Devin’s struggles and the unexpected challenge of agentic hallucinations 

• 🚨 Everyone is freaking out over DeepSeek. Here’s why

AI for Good: Molecular dynamics

Generative video models are cool, but, for this team of researchers, the promise has nothing to do with disrupting Hollywood and everything to do with molecular dynamics. 

What happened: Models like DeepMind’s AlphaFold are better equipping researchers to design new drugs by predicting molecular structures. But they aren’t so good when it comes to accurately simulating the physics of those predictions. 

• Molecules aren’t static; they’re in constant motion. The dynamic movement of molecules is therefore a vital part of protein and drug design. And simulating that motion on traditional supercomputers is enormously expensive. 

• So a team of researchers at MIT developed a system called MDGen that is apparently capable of taking a frame of a molecule and simulating its movement through diffusion (just like a generative video clip). 

And unlike previous approaches, MDGen doesn’t need the first frame to predict the next: “you can give the model the first and 10th frame, and it’ll animate what’s in between, or it can remove noise from a molecular video and guess what was hidden.”

Why it matters: It’s an early work in progress, but over time, it could help scientists develop better drugs and medicines for a variety of diseases. 

Sales Teams are Winning More with this AI Meeting Assistant!

AI Meeting Assistants are becoming a critical tool to fix inefficiencies in sales and poor CRM adoption. Studies show that sales reps are wasting 70% of their time on admin tasks that they hate. AI automates these tasks and allows sellers to focus on closing deals.

WINN.AI’s AI meeting assistant automates sales admin and provides real-time coaching:

⏱ Save 6 hours a week for every rep so they have more time to sell.

🦾 In-call AI coaching keeps sellers from missing important questions.

📕 Deel improved playbook adoption by 31%.

💡 Intercom saw 6x higher CRM Fill Rate.

👏 Help sellers reach quota faster and improve their efficiency and performance.

Leading sales teams from companies such as Intercom, Deel, Sendoso, and Churnzero are using WINN to be more efficient, consistent and win more!

Book a quick demo or try it free here

Devin’s struggles and the unexpected challenge of agentic hallucinations 

March 2024 saw the launch of Cognition, a small AI lab with $21 million in funding and a product called Devin, which promised to function as an AI software engineer. Since then, Cognition has made Devin available at a minimum cost of $500 per month. 

What happened: Three researchers at Answer.AI, an AI R&D lab, just spent a month using Devin, which, unlike other assistants, incorporates directly into Slack. Their findings were less than positive. 

• For 20 tasks across three distinct categories — creating new projects from scratch, performing research tasks and analyzing & modifying existing projects — Devin had 14 failures, three inconclusive results and only three successes. A 15% success rate. 

• “The most frustrating aspect wasn’t the failures themselves — all tools have limitations — but rather how much time we spent trying to salvage these attempts,” the team wrote. “Even more telling was that we couldn’t discern any pattern to predict which tasks would work. Tasks that seemed similar to our early successes would fail in unexpected ways.”

Despite the fact that it does work, occasionally, to the three researchers, it’s kind of useless in the end: “tasks it can do are those that are so small and well-defined that I may as well do them myself, faster, my way,” Johno Whitaker, one of the researchers, said. “Larger tasks where I might see time savings I think it will likely fail at. So no real niche where I’ll want to use it.”

Devin’s very promise of autonomous software engineering functioned as its greatest liability. The system would “spend days pursuing impossible solutions rather than recognizing fundamental blockers,” the team wrote. “Social media excitement and company valuations have minimal relationship to real-world utility.”

Cognition didn’t respond to a request for comment. 

The team’s struggles with the unpredictable nature of inconsistent, time-consuming hallucinations have massive implications for this rising tide of AI agents that the sector is pushing. The very promise of automation might well kill the agent before it gains any traction; the reality of hallucination is a constant threat to utility.  

• Character.AI employees, according to The Information, “expressed concerns internally about how the app was affecting the mental health of its users, including teenagers.”

• The agents are coming! The agents are coming! If only Paul Revere was around today … Convergence, a European tech company, launched Proxy, an agent that, it says, outperforms those coming from OpenAI and Anthropic. If I had a nickel for every time I’ve read the word ‘agent’ in the last two weeks …. Convergence called it a “bold win” for the European tech sector.

• Two hundred UK companies sign up for permanent four-day working week (The Guardian).

• Brazil bans Sam Altman's tech firm Tools for Humanity from paying for iris scans (Economic Times).

• OpenAI’s new anti-jobs program (Vox).

• Trump tariffs could raise prices on technology like laptops, smartphones and AI (CNBC).

• Scammers are creating fake news videos to blackmail victims (Wired).

If you want to get in front of an audience of 200,000+ developers, business leaders and tech enthusiasts, get in touch with us here.

Everyone is freaking out over DeepSeek. Here’s why

Volatility is kind of a given when it comes to Wall Street’s tech sector. It doesn’t take much to send things soaring; it likewise doesn’t take much to set off a downward spiral. 

After months of soaring, Monday marked the possible beginning of a spiral, and a Chinese company seems to be at the center of it. 

Alright, what’s going on: A week ago, Chinese tech firm DeepSeek launched R1, a so-called reasoning model, that, according to DeepSeek, has reached technical parity with OpenAI’s o1 across a few benchmarks. But, unlike its American competition, DeepSeek open-sourced R1 under an MIT license, making it significantly cheaper and more accessible than any of the closed models coming from U.S. tech giants. 

• But the real punchline here doesn’t have to do with R1 at all, but with a previous language model — called V3 — that DeepSeek released in December. DeepSeek was reportedly able to train V3 using a small collection of older Nvidia chips (about 2,000 H800s) at a cost of about $5.6 million. 

• Still, training is only one cost of many tied to AI development/deployment; while the costs associated with researching, developing, training and operating both R1 and V3 remain either unknown or unconfirmed, DeepSeek’s apparent ability to reach technical parity at a far reduced cost, without state-of-the-art GPU chips or massive GPU clusters, has a lot of implications for America’s now tenuous position in AI leadership. (Though DeepSeek says it is open-sourced, the company did not release its training data).

Since the release of R1, DeepSeek has become the top free app in Apple’s App Store, bumping ChatGPT to the number two slot. In the midst of its spiking popularity, DeepSeek restricted new sign-ups due to large-scale cyberattacks against its servers. And, as Salesforce Chief Marc Benioff noted, “no Nvidia supercomputers or $100M needed,” a point that the market heard loud and clear. 

What happened: Led by Nvidia, a series of tech and chip stocks, in addition to the three major stock indices, fell hard in pre-market trading early Monday morning. All told, $1.1 trillion of U.S. market cap was erased within a half hour of the opening bell.

• Performance didn’t get better throughout the day. Nvidia closed Monday down 17%, erasing some $600 billion in market capitalization, a Wall Street record. TSMC was down 14%, Arm was down 11%, Broadcom was down 17%, Google was down 4% and Microsoft was down 2%. The S&P fell 1.4% and the Nasdaq fell 3.3%. An Nvidia spokesperson called R1 an “excellent AI advancement.”

• This is all going into a week of Big Tech earnings, where Microsoft and Meta will be held to account for the billions of dollars ($80 billion and $65 billion, respectively) they plan to spend on AI infrastructure in 2025, a cost that Wall Street no longer seems to feel quite so good about. 

It’s hard to miss the political tensions underlying all of this. The tail end of former President Joe Biden’s time in office was marked in part by an increasingly tense trade war with China, wherein both countries issued bans on the export of materials needed to build advanced AI chips. And with President Trump hell-bent on maintaining American leadership in AI, and despite the chip restrictions that are in place, Chinese companies seem to be turning hardware challenges into a motivation for innovation that challenges the American lead, something they seem keen to drive home.

R1, for instance, was announced at around the same time as OpenAI’s $500 billion Project Stargate, two impactfully divergent approaches.

What’s happening here is that the market has finally come around to the idea that maybe the cost of AI development (hundreds of billions of dollars annually) is too high, a recognition “that the winners in AI will be the most innovative companies, not just those with the most GPUs,” according to Writer CTA Waseem Alshikh. “Brute-forcing AI with GPUs is no longer a viable strategy.”

Wedbush analyst Dan Ives, however, thinks this is just a good time to buy into Nvidia — Nvidia and the rest are building infrastructure that, he argues, China will not be able to compete with in the long run. “Launching a competitive LLM model for consumer use cases is one thing,” Ives wrote. “Launching broader AI infrastructure is a whole other ballgame.” 

“I view cost reduction as a good thing. I’m of the belief that if you’re freeing up compute capacity, it likely gets absorbed — we’re going to need innovations like this,” Bernstein semiconductor analyst Stacy Rasgon told Yahoo Finance. “I understand why all the panic is going on. I don’t think DeepSeek is doomsday for AI infrastructure.” 

Somewhat relatedly, Perplexity has already added DeepSeek’s R1 model to its AI search engine. And DeepSeek on Monday launched another model, one capable of competitive image generation.

Last week, I said that R1 should be enough to make OpenAI a little nervous. This anxiety spread way quicker than I anticipated; DeepSeek spent Monday dominating headlines at every publication I came across, setting off a debate and panic that has spread far beyond the tech and AI community. 

Some are concerned about the national security implications of China’s AI capabilities. Some are concerned about the AI trade. Granted, there are more unknowns here than knowns; we do not know the details of DeepSeek’s costs or technical setup (and the costs are likely way higher than they seem). But this does read like a turning point in the AI race.

In January, we talked about reversion to the mean. Right now, it’s too early to tell how long-term the market impacts of DeepSeek will be. But, if Nvidia and the rest fall hard and stay down — or drop lower — through earnings season, one might argue that the bubble has begun to burst. As a part of this, watch model pricing closely; OpenAI may well be forced to bring down the costs of its models to remain competitive.

At the very least, DeepSeek appears to be evidence that scaling is one, not a law, and two, not the only (or best) way to develop more advanced AI models, something that rains heavily on OpenAI and co.’s parade since it runs contrary to everything OpenAI’s been saying for months. Funnily, it actually seems like good news for the science of AI, possibly lighting a path toward systems that are less resource-intensive (which is much needed!)

It’s yet another example of the science and the business of AI not being on the same page.

🤔 Your thought process:

Selected Image 2 (Left):

• “Image 2 shows the limitations of camera focus, the other is perfectly focused on every plane.”

Selected Image 1 (Right):

• “Contrast and combination.”

💭 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 Character’s defense:

A third of you think it’s too soon to tell, but the motion isn’t likely to succeed. 18% think it’s absolute BS, 15% think it makes sense and 21% don’t like it, but can’t fault the approach.

BS:

• “I don't understand why it would work, the first amendment doesn't protect humans telling other humans to kill themselves, that's why cyberbullying is a crime.”

Read: ‘When texts encouraging suicide do not warrant free speech protection,’ by the Boston College Law Review.