Why Medical AI is Garbage, Realistic Perspectives on DeepSeek Models, Understanding Reasoning Models, and More

An AI engineer's must-reads for 1-31-25

Here’s this week’s must-reads and must-knows for anyone interested in AI engineering. Make sure to support the authors of the resources. A huge thanks to all supporters of Society’s Backend! If you want an extended reading list, more resources, and extra articles, you can support Society’s Backend for only $2/mo.

If you’re interested in learning AI/machine learning, I’ve created a roadmap to walk you through prereqs and ML fundamentals you should know entirely for free. Check it out here.

If you missed last week’s must-reads, check them out here:

Events you should know

• DeepSeek sent the internet into a frenzy over inexpensive, open AI models that rival frontier models. DeepSeek is a smaller Chinese company. This has implications in both nationalistic AI and the number of researchers needed to develop models. This was way overblown online, but is still important.

• Humanity’s Last Exam is a test for AI designed to test the limits of AI knowledge at the frontiers of human expertise. It aims to be the most difficult public AI test and to comprehensively test reasoning. Current top models score less than 10% on it.

• The Stargate Project is a $500 billion AI infrastructure initiative announced by President Trump to build data centers for AI training/inference across the US. The goal is to cement the US at the lead of global AI development and stimulate economic growth.

• OpenAI releases o3-mini, a new small model at the frontier of reasoning. It marks another milestone for AI achieving better scores on common reasoning topics such as science, math, and coding at a lower cost and latency than previous models.

Resources you should read

What Is an AI Engineer? (And How to Become One)

AI engineers develop applications and systems that use artificial intelligence and machine learning to enhance business efficiency and decision-making. The field is rapidly growing, with a projected job growth of 23% and an average salary of over $108,000 in the U.S. Individuals seeking to become AI engineers should focus on acquiring technical skills in programming, statistics, and machine learning frameworks, with many learning through online courses or professional certificates.

Source

Why Most Medical AI Is Garbage—And Why No One Cares

By

Sergei Polevikov

Most medical AI technologies are ineffective due to underlying data issues and a lack of transparency in the healthcare system. Experts argue that rather than relying on AI, the focus should be on practical automation that improves efficiency without replacing healthcare professionals. The hype around AI in healthcare often masks deeper problems and financial motives, leading to skepticism about its true value.

Source

On DeepSeek and Export Controls

DeepSeek, a Chinese AI company, has developed a model that performs comparably to older US models at a lower training cost, but it is not a game-changer in the AI landscape. The ongoing trend shows that both US and Chinese companies will continue to invest heavily in training smarter AI models, consuming any cost savings to achieve greater intelligence. Export controls have not significantly hindered DeepSeek's ability to access the necessary chips, as they have managed to acquire resources comparable to those of US AI labs.

Source

Why reasoning models will generalize

By

Nathan Lambert

New reasoning models are expected to generalize beyond their initial applications in coding and math, enhancing their performance across various tasks. These models utilize "chain of thought" reasoning, processing information step-by-step, which allows them to better manage complexity and allocate compute resources effectively. As development progresses, reasoning models may outperform traditional models in many unexpected areas, leading to significant advancements in AI capabilities.

Source

DeepSeek: Frequently Asked Questions

By

Charlie Guo

DeepSeek, a Chinese AI company, has gained significant attention by releasing cost-effective models that perform comparably to industry leaders like OpenAI. Its recent success has sparked concerns among major tech firms and led to a historic drop in Nvidia's stock. While DeepSeek’s advancements raise questions about the future of AI development, it remains to be seen if it signifies a shift in dominance from the US to China.

Source

DeepSeek Lecture (1/28)

By

Tom Yeh

Tom Yeh hosted a public lecture on DeepSeek on January 28, 2025. The lecture focused on understanding the inner workings of the DeepSeek model, emphasizing algorithms like Multi-head Latent Attention and Mixture of Experts. This is the first in a series of interesting lectures that include hands-on learning of machine learning concepts.

Source

Mixture-of-Experts (MoE) LLMs

By

Cameron R. Wolfe, Ph.D.

Mixture-of-Experts (MoE) models enhance the efficiency and performance of large language models by introducing sparsity, allowing for a larger number of parameters without increasing computational costs. These models use a routing mechanism to select a small subset of experts for processing each token, which helps balance load among the experts. Recent MoE models like DeepSeek demonstrate impressive performance improvements and training efficiency, making them competitive with traditional dense models.

Source

What Math do you need to be Good at AI

By

Devansh

As AI evolves, different roles in the field require varying levels of mathematical understanding. Non-technical individuals should grasp basic concepts, while ML engineers need practical skills, and AI researchers require deep mathematical knowledge. Focusing on key mathematical principles helps all personas effectively engage with AI technologies.

Source

AI Revolution: Why This Is The Best Time To Start A Startup

The current advancements in AI technology create a unique opportunity for new startups. Entrepreneurs can leverage AI to innovate and solve problems more efficiently. Now is an ideal time to launch a business that harnesses these powerful tools.

Source

Introducing Gemini 2.0: our new AI model for the agentic era

Gemini 2.0 is Google's latest AI model, designed to enhance multimodal capabilities and improve user assistance. It introduces features like native image and audio output, advanced reasoning, and tool use for more effective interactions. Users can now access Gemini 2.0 Flash, which will be integrated into various Google products soon.

Source

How artist Yinka Ilori is using AI to bring his vision to life