Good morning. Sam Altman says OpenAI has trained up a model that’s good at creative writing, another model that he’s not sure “how/when” it will get released.

But creative writing, unlike math or code, is more than the final product, and is far less measurable than something that is either right or wrong. Storytelling is about how it makes a reader feel. Storytelling is a soulful practice.

Human writing comes infused with intention, fiction especially so. Human storytellers have something to say, and a reason for speaking; LLMs can’t, and don’t.

In other news, though the market slide continued yesterday, some names — Tesla, Nvidia, Microsoft — regained a bit of lost ground.

For your reference:

$SPX ( ▲ 1.56% ) 

$NVDA ( ▲ 2.82% ) 

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

In today’s newsletter:

• 🌎 AI for Good: The hunt for hydrogen

• 🚘 A more ‘interesting’ data source for self-driving cars

• 💻 Interview: Making AI for code generation actually work

🎙️ Podcast: Microsoft Research director on AI for science

In the latest episode of The Deep View: Conversations, I sat down with Dr. Chris Bishop, the director of Microsoft Research AI for Science, to talk about what AI for science really means; how the technology is currently accelerating scientific progress, what kind of AI is behind that push and what impact it all might have.

Check it out below:

AI for Good: The hunt for hydrogen

In the midst of an increasingly urgent search for cleaner energy alternatives, hydrogen is beginning to appear as a possible solution. Though a small industry is beginning to spring up around the search for naturally occurring hydrogen deposits, finding those deposits isn’t an easy task. 

At the moment, companies are borrowing methods from the oil drilling industry in an effort to stumble across hydrogen — but at least one startup is leveraging machine learning to narrow the search. 

The details: Climate Maps, a startup supported by the U.K. government’s Innovate UK Edge and Department for International Trade, is leveraging machine learning models and specialized satellites to map underground hydrogen deposits. 

• The startup is led by the same man — Massoud Maqbool — who started DeepWaters AI, and employs a similar method that was used to find water underground. 

• Using spectral satellites and more than 20 years of satellite data, the team was able to both locate and quantify hydrogen that lives below the ground.  

The impact of hydrogen on the climate isn’t yet well known — some studies have identified the risk of indirect contributions to global warming, though, when pure hydrogen is consumed in a fuel cell, the only byproduct is water, making hydrogen a highly promising fuel type. 

“When we first started, we thought the biggest impact we could contribute to climate change, was by mapping Earth's most polluted locations. However we realized, while important, it wasn't a solution to climate change,” the company writes. “So we mapped Earth's underground hydrogen. We believe hydrogen should benefit the citizens of every country.”

🔥Build AI Agents That Do Everything for You (for free)🤯

2024 was for chatbots & custom GPTs.

But 2025 will be for AI agents.

All companies, agencies & business will integrate agents to improve efficiency & cut costs. And if you don’t learn to build them now, you will be replaced by one.

So grab this chance to learn how to Build your Own AI Agent in just 2 hours with the best in the industry— Matthew Cohn, Founder of FutureFlow AI, who’s revolutionized businesses with AI. 🚀

You will build AI Agents that can:

✅ Streamline operations by automating repetitive tasks, reducing errors, and freeing up resources for strategic initiatives.

✅ Leverage AI agents to analyze large datasets, providing actionable insights that inform business strategies.

✅ Deploy AI-powered virtual assistants to handle customer inquiries efficiently, offering 24/7 support and improving customer satisfaction.

✅ Integrate AI agents as intelligent collaborators, augmenting human expertise and enhancing team productivity.

👉 Hurry! Register now to secure your spot. (FREfdE for the first 100 participants!)

A more ‘interesting’ data source for self-driving cars

Open-source AI platform Hugging Face on Tuesday teamed up with German spatial intelligence company Yaak to launch the “largest open-source dataset for self-driving cars,” a dataset called Learning 2 Drive which boasts 5,000 hours of multimodal self-driving data. 

The details: The dataset was collected over the course of three years by leveraging a fleet of 60 student-driver vehicles. 

• Hugging Face co-founder Thomas Wolf said at a press conference at HumanX that the idea behind this approach is to enable developers to “see how people learn to drive … and how someone teaches someone else to drive.” 

• “In addition to being the largest open-source dataset of multimodal data for training self-driving, it’s also the most interesting, in my opinion, because it has these errors that you’d like to avoid,” he said. 

The dataset is divided into two groups. The first is made up of ‘expert’ actions executed by the driving instructors, and the second consists of ‘student’ actions executed by those less-skilled humans learning how to drive. 

Why it matters: Existing datasets, like those from Waymo, for instance, focus on intermediate planning tasks, such as object detection, which, according to the release, is difficult to scale. But this dataset is focused on “the development of end-to-end learning which learns to predict actions directly from sensor input.”

“The AI community can now build end-to-end self-driving models leveraging the state-of-the-art imitation learning and reinforcement learning models for real world robotics,” the two firms wrote in a blog post. 

The landscape: This all comes amid a steady expansion of certain sects of the self-driving field, with Waymo recently expanding its areas of operation around the Bay Area, even as other players — like Tesla — continue to struggle with unsupervised self-driving. It’s an area that some experts think will always remain bounded by fundamental limitations within the architecture, notably, the propensity of models to hallucinate, and the inability of systems to access a wide enough breadth of training data to cover all possible edge cases across all possible conditions. 

Yaak and Hugging Face plan to conduct real-world, closed-loop testing of models trained on the dataset this summer. 

Webinar: AI Observability and Security for Agentic Workflows

Fiddler AI and AWS share: 

• How to increase visibility into AI agent operations and decision-making processes using AI observability

• Best practices to protect agents from risks using tools like Guardrails

• Key governance considerations for balancing AI agent innovation with security and compliance

👉 Register now and we’ll send you the replay afterward

• AI & heart disease: Caristo Diagnostics said Tuesday that it had received FDA approval for an AI-assisted image analysis tool designed to accelerate coronary artery disease diagnoses. The goal is to expand into hospitals, enabling proactive response rather than reactive treatment.

• Meta’s new chip: Meta has begun testing its first in-house chip designed for AI training, according to Reuters. If the tests go well, Meta intends to begin ramping up production of the chip, something that could lessen its reliance on Nvidia.

• OpenAI will let other apps deploy its computer-operating AI (The Verge).

• Anthropic’s Claude drives strong revenue growth while powering ‘Manus’ agent (The Information).

• What really happened with the DDoS attacks that took down Twitter (Wired).

• Ukraine agrees to US-backed 30-day ceasefire proposal (Semafor).

• Dow drops more than 450 points, S&P 500 posts back-to-back loss over Trump tariff uncertainty (CNBC).

Interview: Making AI for code generation actually work 

Few applications of generative AI seem as real as code generation. 

Though the numbers are a little different from one to the next, a number of reports have identified code generation as one of the major GenAI use cases; this report from last year found that, among enterprises, code generation was the most popular GenAI use case, leveraged by more than half of those enterprises surveyed. 

23% of the tech firms surveyed in this report said they use generative AI tools for code generation, and a GitHub survey of software engineers from August found that 97% of those surveyed reported using an AI tool for coding “at some point.” 

And, last October, Google CEO Sundar Pichai told investors that more than a quarter of Google’s new code was being generated by AI. Anthropic Chief Dario Amodei, meanwhile, just this week made the rather bold prediction that, within the next three to six months, AI “will be writing 90% of the code,” and within the next 12 months, “essentially all” code will be written by AI.

But, as with all things in this field, the line of adoption is a wavy one.

• Studies have found that code generated by AI coding applications tends to be weaker and more bug-prone than human-written code. 

• These studies seem to be panning out in the real world; a recent examination of Cognition’s much-hyped coding assistant, Devin, found a high rate of failure due to inconsistent, unpredictable and time-consuming hallucinations. 

Itamar Friedman, the founder and CEO of AI-coding startup Qodo, has been long awaiting the moment when large language models (LLMs) would become strong enough to handle truly complex software engineering challenges. He’s not concerned just with simple code generation; his focus is on solving for complexity.

“Software runs the world,” Friedman told me. “But the software that really runs the world is the complex software … not the one that builds a blog or website.” 

Those simpler coding challenges, like website building, are today just a few prompts away. But Friedman said that he hasn’t seen a truly significant boost in the generation of the highly complex code that is deployed by enterprises. “The reason is because in the enterprise, it's not just generating code,” he said. “It's a full software development life cycle. You need to make sure things do not break.”

This is a point that was echoed in December by AWS, which said that its developers spend as little as one hour every day actually writing code; the rest of their time is spent reviewing and testing existing code, further evidence for Friedman that code-generation alone isn’t enough of an offering. 

• Qodo, which Friedman started just before ChatGPT burst onto the scene at the end of 2022, offers an AI-enabled coding platform that goes beyond code generation, additionally offering testing and review tools. 

• His focus is on quality, something that, according to Friedman, Qodo achieves through the specific implementation of deterministic guardrails, even and especially when dealing with the greater levels of autonomy presented by so-called ‘agents.’ 

“We believe that the way to reduce hallucination involves the integration of frameworks designed by developers,” Friedman said, adding that developers wishing to write complex code would begin that process by essentially telling their coding agent: “‘hey, this is how we think it should be done,’ guard-railed with different validation points.”

This, he said, is the essence of Qodo: “our entire hypothesis and mission is to bring those guardrails to the developers.”

Friedman said that this approach — of high autonomy combined with highly specific restraints — represents the future of how AI coding tools will be sold. 

🤔 Your thought process:

Selected Image 1 (Left):

• “You have to look at how the human is balanced, where is he looking, what is he doing? Do all the data points make sense? It’s not enough to look perfectly real … it has to look perfectly NORMAL for what a person in that situation would be doing.”

Selected Image 2 (Right):

• “No way! he had a face : )”

💭 A poll before you go

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

We’ll see you in the next one.

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.