Good morning. Donald Trump has officially retaken the White House. A slew of executive orders is expected to follow — we’ll be keeping a close eye on his AI-related executive action.

We’ll see what happens next. And then we’ll let you know.

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

In today’s newsletter:

• 🔬 AI for Good: Robotic bugs

• 👁️‍🗨️ Sam Altman wants you to lower your expectations ‘100x’ 

• 💰 Spain grants $155 million in AI subsidies, says country ‘should not fear AI’ 

• 📊 DeepSeek drops open-sourced reasoning model

AI for Good: Robotic bugs

Pollinators, such as bees and butterflies, have been experiencing significant population declines for years now, yet another victim of increasing global temperatures. While some scientists are working on ways to reverse the decline — which largely depend on conservation and biodiversity — others are looking to mechanical solutions. 

What happened: A team of MIT researchers is developing robotic insects that could, perhaps one day, “swarm out of mechanical hives to rapidly perform precise pollination.” 

• Inspired by the natural anatomy of bees, these robot bugs, which weigh less than a paperclip, can now hover for about 17 minutes without losing aerial precision, a major advance in the field. 

• Still, the team has a lot more work to do; they want to continue improving the precision on display here until these robotic insects are able to land and take off from the center of a flower. 

Why it matters: The idea is to enable a future of targeted agriculture through artificial pollination. 

This futuristic approach, however, of farming divorced from the land, doesn’t account for the potential of regenerative agricultural efforts, which, if employed at scale, could restore soil health and boost biodiversity and conservation.

The State of AI Coding Assistants: Insights from 14,000+ Engineers

Are AI coding tools living up to the hype?

Jellyfish analyzed data from over 14,000 engineers at nearly 200 companies to measure the real-world impact of GitHub Copilot. The results? Mixed. While Copilot users were 12.6% faster, they also introduced more bugs. But there's a silver lining - bug backlogs decreased by 27.4%.

Want the full story? Download our new ebook "The State of AI Coding Assistants" to get an in-depth look at Copilot adoption trends, productivity gains, and expert predictions for how AI will reshape engineering in 2025 and beyond.

Sam Altman wants you to lower your expectations ‘100x’ 

We talk often about the rampant hype in this industry. It takes a lot of forms, most often coming directly from the companies and labs building and deploying AI technology. 

• You see it in anthropomorphized chatbot interfaces, in cherry-picked, non-transparent blog posts presented as scientifically validated information and, perhaps most significantly, from the mouths of the executives. 

• Sam Altman and Elon Musk are great examples of the AI hype machine, with both spending the past two years talking loudly and often about the transformative future that’s right on the brink (whether it’s a good future or a bad one, the hype of potential capabilities remains the same, again, not backed by scientific evidence or validation). 

But the hype is getting to be too much even for Altman, who wrote Monday that “Twitter hype is out of control again. We are not gonna deploy AGI next month, nor have we built it.”

He added that “we have some very cool stuff for you but pls chill and cut your expectations 100x!”

Altman’s plea for less hype, please, comes just two weeks after a blog post in which he wrote that “we are now confident we know how to build AGI as we have traditionally understood it … We are beginning to turn our aim beyond that, to superintelligence in the true sense of the word.”

• It also comes shortly after Epoch — the company behind those math benchmarks that OpenAI’s o3 model reportedly smashed through — revealed that the project was funded by OpenAI, and that OpenAI had access to “much” of Epoch’s dataset. 

• As part of Epoch’s contract with OpenAI, other labs reportedly can’t get access to Epoch’s FrontierMath. 

Altman’s statement also comes days after an OpenAI researcher tweeted: “enslaved god is the only good future.” 

This idea of ‘less hype, please’ moved beyond OpenAI last week, when Ali Kani, the head of Nvidia’s automotive division, told AutoCar that fully autonomous cars “will not appear in this decade.”

“It’s a next-decade marvel. We’re not close. It’s super-hard,” he said. 

• Supercharge your organization's AI adoption journey from complex challenges to strategic success with Tines' must-read guide: "Securing AI in the Enterprise." *

• After chaotic weekend, TikTok remains in limbo (Ars Technica).

• Los Angeles under dangerous red flag warnings for extreme fire risk again this week (NBC News).

• Trump casts shadow on Davos (Semafor).

• Trump inauguration live updates: New president prepares to sign Day One executive orders (CNBC).

• Welcome to the era of gangster tech regulation (The Verge).

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.

• The European Commission said Monday that X, YouTube, Meta and other platforms have all agreed to do more to tackle online hate speech as part of a new code of conduct.

• 404 Media reported on GeoSpy, an AI tool that can predict a person’s location based on their photos in seconds. It’s being leveraged by cops, government agencies and stalkers alike.

Spain grants $155 million in AI subsidies, says country ‘should not fear AI’ 

Spanish Prime Minister Pedro Sanchez said Monday that Spain will grant $155 million in subsidies for companies working on developing and deploying artificial intelligence. 

The announcement coincided with the release of a report entitled “Hispania 2040: How Artificial Intelligence Will Improve Our Future.” 

The details: The study was specifically designed to address the most significant challenges the country will face in the coming decades and the specific ways in which AI systems can aid Spain in dealing with said challenges. 

• On display here are challenges of strengthening Spain’s welfare state, achieving and maintaining environmental sustainability, strengthening security and defense, reducing inequality and improving the productive fabric of Spanish society. 

• According to a translation of the document (courtesy of Claude, cross-referenced with my own growing understanding of the language), the report found that even without further technical advances, today’s AI could boost productivity and reduce the administrative burden on teachers and healthcare workers alike, all while enabling better energy production and resource management. 

Specifically, the report said that AI-enabled traffic optimization alone could reduce emissions equivalent to nearly a million cars annually, adding that smart water systems could enable water savings equivalent to “twice the water consumption of Madrid.”

The report did say that any widespread deployment of AI must preserve privacy, data protection and intellectual property, in addition to being wholly transparent, a safeguard against algorithmic bias. The report added that “we will also have to monitor the environmental footprint that AI implementation entails on Spanish soil through research in green algorithms and the development of smaller and more sustainable models.” 

Sanchez is focused on adoption, writing Monday: “Spain should not fear Artificial Intelligence. It should lead its development and adoption. Because, if we use it well, AI will allow us to have better jobs, better public services, and combat climate change.” 

But the report at least tells a story, not of uncritical adoption, but of a seemingly balanced understanding of the scope of the challenge, and the specific manner which could enable positive outcomes here. (I particularly like the acknowledgment of AI’s environmental footprint and methods of mitigation).

DeepSeek drops open-sourced reasoning model

Chinese AI firm DeepSeek launched a new ‘reasoning’ model — the R1 series — on Monday, that, according to DeepSeek, performs on par with OpenAI’s o1 ‘reasoning’ model. But in a significant departure from its Silicon Valley competition, DeepSeek R1 has been open-sourced with an MIT license for distillation and commercialization. 

The model’s output API price is $2.19 per million tokens, cheaper than GPT-4o and significantly cheaper than OpenAI’s o1, which, according to TechCrunch, charges $60 for every 750,000 words outputted. 

The details: DeepSeek explained in a technical paper that the model was trained through large-scale reinforcement learning, which very basically refers to a type of machine learning in which an autonomous agent — in this case, R1 — learns from rewarded behavior while performing trial and error without supervision. 

• According to DeepSeek, R1 represents an attempt to boost large language model reasoning capabilities using “pure” reinforcement learning (RL), without leveraging supervised fine-tuning (SFT) as a preliminary step. 

• DeepSeek said the model is capable of self-verification, reflection and generating long Chains of Thought, a combination which makes R1 a “significant milestone for the research community.” 

“Notably, it is the first open research to validate that reasoning capabilities of LLMs can be incentivized purely through RL, without the need for SFT,” DeepSeek wrote. “This breakthrough paves the way for future advancements in this area.”

DeepSeek researchers were additionally able to distill these reasoning patterns into smaller models, resulting in “powerful” performance in models ranging from 1.5 billion to 70 billion parameters. 

The company wrote that its RL-based approach to its distilled models “yields significant further gains. We believe this warrants further exploration and therefore present only the results of the simple SFT-distilled models here.”

On benchmarks covering coding, math and reasoning tasks, DeepSeek R1 performs either on par with, or slightly above, OpenAI’s o1. 

The paper has not been peer-reviewed or independently validated, and the training data and cost — both in energy and dollars — remain unclear. DeepSeek noted that for all the progress, R1 does have its limitations and challenges, including poor readability and language mixing, areas that it plans to explore more. 

Across a few early and informal tests, AI researcher Ethan Mollick said that the model “is really good, not o1-pro level but surprisingly capable (and small & fast!). Big.” 

Some intrigue: The release comes after DeepSeek released V3 in December, a similarly open-sourced model that performs on par with GPT-4o and Claude 3.5 Sonnet. But training V3 — on older Nvidia H800 chips — cost DeepSeek only $5.6 million, a mere fraction of the hundreds of millions going into prominent models at OpenAI, Google and Anthropic. 

Increasing tensions with China surround this highly competitive landscape, with the U.S. recently enacting an additional round of export restrictions on the hardware needed to build AI models, and the incoming administration expected to take an even harder line against the country. 

“Money has never been the problem for us; bans on shipments of advanced chips are the problem,” DeepSeek CEO Liang Wenfeng said in an interview last year. 

“It looks like China has roughly caught up,” Dan Hendrycks, director of the Center for AI Safety, wrote. “Any AI strategy that depends on a lasting U.S. lead is fragile.”

The cheaply accessible and open-source capabilities of R1 should be enough to make OpenAI sweat a little. DeepSeek might not be on par with o3 — which no one has access to — but if it allows developers to leverage the general capabilities of o1 at a fraction of the cost, I would expect one of two things to happen: either customers will start leaving OpenAI in search of cheaper options, or OpenAI will have to cut prices to be competitive. 

Either way, it’s a blow for the company that has plans to burn around $200 billion before it starts to turn a profit.

🤔 Your thought process:

Selected Image 1 (Left):

• “The lighting looked less ‘designed to impress.’ That lens flare in image number 2, the complete absence of any living thing (or things piloted by living things) on the streets was another giveaway for me.”

Selected Image 2 (Right):

• “Image 1 looked unrealistic with the clearly visibility of windows and cars.”

💭 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 TikTok:

40% of you, quite simply, do not care.

21% of you think the idea of the ban is ridiculous, and 28% of you think the app is awful and should be banned.

Something else:

• “Generally, I'm not a fan of censorship - but we do need to be aware of how data is collected and used - especially with our youth. A bi-partisan group did decide on the ban - and, all politics aside, I would like the information this group had in making their decision. Is there something there?”

Something else:

• “I love the app, almost too much. However, it makes no sense to have what has become a stealthy, massively influential media platform under the control of the CCP. There is precedence. NBC, the NYTimes and others are required to have American ownership. So should TikTok. Yet, this is not without consequences. The alternative is ownership by an oligarch (#Twitter). Even so, for now, we should stick to our principles.”