AI Highlights

My top-3 picks of AI news this week.

DeepSeek
1. DeepSeek Outsmarts the Giants

Chinese AI firm DeepSeek has released its V3 model under an open-source license, marking a significant milestone in AI accessibility and performance.

• Impressive scale: Built with 671B parameters and trained on 14.8T tokens in just two months.

• Benchmark leader: Outperforms other open models and some closed models like Meta's Llama 3.1 405B and OpenAI's GPT-4o in coding competitions.

• Cost-effective: Training cost of $5.5M, significantly lower than competitors, with competitive API pricing ($0.27/M input tokens, $1.10/M output tokens).

• Training quirk: Users have discovered the model occasionally identifies itself as ChatGPT, suggesting its training data may have included significant amounts of GPT-4 outputs.

Alex’s take: What I thought was particularly intriguing about DeepSeek V3 is its dominance across benchmarks. In Codeforces competitions, it's outperforming Meta's Llama 3.1 405B despite being trained on a fraction of the budget. This level of performance from an open-source model, combined with its cost-efficiency, suggests we might be seeing a pivotal moment in AI development where closed-source models no longer hold a decisive advantage.

xAI
2. The $6B bet on Grok

Musk’s xAI has secured a massive $6 billion Series C funding round from prominent investors, doubling the company's valuation to $45 billion.

• Infrastructure scaling: Plans to double their AI supercomputer 'Colossus' from 100,000 to 200,000 NVIDIA Hopper GPUs, initially built in a record time of 122 days.

• Product expansion: Recently launched Aurora (image generation model), Grok 2 (advanced reasoning model), and developer API access for enterprise solutions.

• X integration: Deepening integration with X platform through real-time analysis, web search, citations, and image generation capabilities.

Alex’s take: The speed of xAI's execution is impressive. They built one of the world's largest AI supercomputers in just 122 days, and they're already planning to double it. While most AI labs took years to reach their current scale, xAI seems determined to achieve in months what others did in years.

Unitree
3. Robot Gone Wild

Unitree has unveiled serious upgrades to its B2-W Industrial Wheel robot, showcasing impressive improvements in mobility, endurance, and practical applications.

• Advanced mobility: Combines wheeled and legged locomotion to navigate diverse terrains including stairs, slopes, and rough surfaces with a climbing angle exceeding 45°.

• Enhanced endurance: Achieves a maximum speed of 20 km/h with an impressive 50 km range while carrying a 40kg load, powered by a >2kWh battery system.

• Physical prowess: Capable of jumping 2-meter ditches and continuously climbing stairs of 20-25cm height, while supporting loads of up to 120kg when stationary.

Alex’s take: While the robotics industry often focuses on either wheeled or legged designs, Unitree's hybrid “whegs” approach is particularly intriguing. In search and rescue scenarios, where minutes are precious, this increased mobility could mean the difference between life and death.

Content I Enjoyed

Ray Kurzweil on Living Forever (Almost)

World-renowned inventor and futurist Ray Kurzweil sat down with AT&T Chief Technology Officer Jeremy Legg to talk about the future of technology.

My favourite takeaways from the conversation:

• Exponential growth in computing power: We’ve seen a 75 quadrillion-fold increase in computation per dollar since 1939.

• He predicts AGI will be achieved by 2029.

• The Singularity by 2045: Kurzweil defines the singularity as the point where human intelligence is multiplied a million-fold through merging with AI. This will involve brain-computer interfaces that seamlessly integrate human thought with tremendous computational power.

• Brain-computer interfaces: Kurzweil believes non-invasive brain-computer interfaces will be commonplace by the 2030s, allowing humans to directly access the power of AI.

• Longevity escape velocity: He forecasts that by 2032, advancements in medicine will outpace the aging process, leading to significantly extended lifespans.

This prediction of “longevity escape velocity” by 2032 fascinates me. For every year that passes, scientific advances will add more than a year to your life expectancy. In other words, if you can make it to 2032, you might just keep on living.

This reminds me of my favourite Black Mirror episode, “San Junipero,” where elderly individuals upload their consciousness to live eternally in a simulated reality. Perhaps what once seemed like sci-fi is now becoming closer to sci-fact.

Idea I Learned

Four AI Predictions for 2025

With the New Year rolling through next week, I’ve been scouring the internet for the best predictions of where AI will take us in 2025.

I came across this article from the FT highlighting four predictions that I thought were worth covering in today’s issue.

Firstly, will AI development hit a wall? Mark Zuckerberg highlighted in September that the limit of scaling AI systems “is not going to happen any time soon.” Llama 3 was trained on 10-20K GPUs, Llama 4 will be trained on more than 100K+ GPUs, and Llama 5 will be even more than that.

However, Marc Andreessen recently noted that the “scaling law” of AI, which shows that a model's performance increases as the size of its training data and computational resources increase, is “topping out”. With the emergence of “thinking” models like OpenAI’s o1 and o3, will this take us to a new scaling law entirely? I’m excited to see reasoning models continue to develop in 2025 across narrower applications with domain-specific knowledge.

Secondly, will AI’s ‘killer app’ emerge? We had the “ChatGPT” moment of 2022, making the chat interface mainstream when interacting with AI models. I hope 2025 will bring with it interfaces that integrate AI even more so into our lives. We’re already seeing glimmers of this with Apple Intelligence. It should feel natural and intuitive, not about having separate “AI apps” for the sake of using AI. Instead, humanity should focus on thoughtfully constructing complementary ecosystems that enhance our very own ideas.

Thirdly, will Nvidia’s GPUs still rule the tech world? My guess is in the medium term, we’ll see a continuation of their dominance. In the long term, I’m a massive advocate for smaller domain-specific models at the edge that can connect with each other—an AI for everyone. I don’t think we’ll need 1M GPUs for this.

Finally, will the stock market’s AI boom continue? AI startup valuations are super high in comparison to public comparables. Moreover, Musk recently highlighted “the markets claim to look ahead, but mostly look behind.” If we expect AGI in the next few years, it seems like we have an exciting road ahead. I expect the first $10T company soon.

Quote to Share

Sam Altman’s predictions for 2025:

“Predictions for the three most important technological developments that will happen by 2025:

1) We will get net-gain nuclear fusion working at prototype scale

2) AGI will feel within reach to many people in the industry

3) Gene editing will have cured at least one major disease”

This prediction was made in 2019. Fusion will likely be by the end of the decade, and we’re getting closer to AGI as we transition from reasoners to agents. What’s more, gene editing already has a cure for sickle cell disease. I think it’s incredible to see how far we’ve come—and how much we’ve got to look forward to.

Source: Sam Altman on X

Question to Ponder

The ultimate goal of integrating AI into our society is to propel humanity towards becoming a more advanced civilisation.

Specifically, our aim is to achieve the status of a Type-1 civilisation on the Kardashev Scale, which requires us to harness all available energy on our planet.

AI stands as our most promising tool for achieving this objective.

I thought this video about the Kardashev Scale was a brilliant explanation of what this means for humanity.

If you’ve made it this far, have a great start to 2025, and I’ll see you in the New Year 👍

💡 If you enjoyed this issue, share it with a friend.