AI Highlights

My top-3 picks of AI news this week.

OpenAI
1. OpenAI's American Ambition

OpenAI has made three significant announcements this week showcasing its expanding influence:

• Product Development: Released o3-mini, their latest reasoning model that delivers comparable performance to previous models while being 63% cheaper and 24% faster, optimised explicitly for technical domains.

• Strategic Partnership: Signed a groundbreaking agreement with U.S. National Laboratories to deploy their latest reasoning models for scientific research and national security initiatives.

• Enterprise Solution: Launched a new specialised version of ChatGPT designed for government institutions, building on their existing base of over 90,000 public sector users.

Alex’s take: In a Reddit AMA, Sam Altman admitted they're losing ground to Chinese competitors and have been “on the wrong side of history” with their closed-source approach. I hope this year we’ll see the true meaning of “Open” in “OpenAI” come to life.

DeepSeek
2. The DeepSeek Disruption

Chinese AI lab DeepSeek has released Janus-Pro, a new family of multimodal image AI models that challenges industry leaders with impressive performance despite smaller model sizes.

• Benchmark Leader: Their 7B parameter model outperforms OpenAI's DALL-E 3 and other major image generation models on GenEval and DPG-Bench evaluations.

• Efficient Design: Models range from 1B to 7B parameters, achieving high image creation and analysis performance with significantly smaller architectures than competitors.

• Open Access: Released under MIT license on Hugging Face, allowing unrestricted commercial use in the AI image generation community.

Alex’s take: While everyone's focused on building bigger models, DeepSeek's approach reminds me of the “small is beautiful” philosophy. Their ability to achieve DALL-E 3-level image generation performance with just 7B parameters is a wake-up call for the industry. I think we're entering an era where model efficiency, not just raw size, will ultimately be the key differentiator in the AI race.

Alibaba
3. Alibaba's Answer

Alibaba's Qwen team has launched Qwen2.5-VL, a new family of AI models with advanced capabilities in controlling devices and understanding visual content.

• Device Control: Can operate PCs and mobile apps, including booking flights through Booking.com on Android.

• Visual Prowess: Capable of analysing charts, extracting data from forms, and comprehending long-form videos.

• Competitive Edge: Claims to outperform GPT-4o, Claude 3.5 Sonnet, and Gemini 2.0 Flash across various benchmarks, though struggles with the OSWorld computer environment test.

Alex’s take: I wonder if DeepSeek's recent advances prompted Alibaba to make this surprise announcement. Nonetheless, I see general capabilities across device control as the future of AI models. This will enhance productivity by enabling us to move from asking questions back and forth with a chatbot to systems that actually take action for us.

Content I Enjoyed

Dario Amodei: On DeepSeek and Export Controls

Anthropic CEO Dario Amodei dropped a thought-provoking essay this week.

He broke down AI development into three fundamental dynamics:

• Scaling laws (throw more compute at it, get better results)

• Curve shifting (finding clever ways to do more with less)

• Paradigm shifts (completely new approaches, like adding reinforcement learning to the mix)

Something that stood out to me was his insight about the current “crossover point” in AI. We're at this unique moment where multiple companies can produce impressive reasoning models because we're early in the reinforcement learning curve. But Amodei argues this won't last—soon, it'll be a game of who can invest the most in scaling up these systems.

Take, for example, his distinction between DeepSeek's V3 and R1 models. While V3 showed genuine promise in engineering efficiency, R1 simply replicated what OpenAI had already achieved with their o1 model by adding reinforcement learning training.

Amodei points to 2026-2027 as the critical timeline when AI capabilities could hit another inflection point. Export controls will determine whether we end up in a world that is bipolar (including China) or unipolar (just the US and its allies). One where China will have access to millions of chips, or one where they won’t.

Today’s decisions carry significant consequences, particularly in the race for artificial general intelligence (AGI).

Idea I Learned

DeepSeek Drama: Speculation or Substance?

DeepSeek’s Technical Report indicated that it cost $5.6 million in computing power to train V3, their open-source model with capabilities similar to OpenAI’s GPT-4o.

This was accomplished at a fraction of the cost and with fewer resources than the current state-of-the-art models available in the market today.

As a result, when the market started to notice, Nvidia's shares plummeted by 17% on Monday. This ~$600 billion slide in one day is totally unprecedented. It represented the biggest single-day loss of market cap for any public company in history.

Quite the knee-jerk reaction from panicking investor confidence since the Chinese firm not only surpassed the quality of the top AI models from the US at a slither of the training cost but also made it 10x cheaper to actually use the model itself during inference.

However, Scale AI CEO Alexandr Wang and Tesla CEO Elon Musk believed something else was going on. DeepSeek “obviously” had ~50,000 Nvidia H100 chips that they can’t talk about due to US export controls.

With a single H100 chip weighing in at ~$27,000, the 50,000 that weren’t accounted for would increase the actual training cost to > $1 billion.

Yet outside of the cost and capability frenzy, something I think hasn’t been covered anywhere near enough this week is the privacy picture. In China, the National Intelligence Law of 2017 mandates that “any organization or citizen shall support, assist, and cooperate with state intelligence work.” In short, if you’re a Chinese company, you’re legally obliged to hand over any data you hold to the CCP at their beck and call.

What does this mean for you and me?

Be mindful of the questions you ask and the data you share with DeepSeek. Just like if you’re using OpenAI’s ChatGPT, DeepSeek uses your data and responses to help train their model.

What’s clear from this week’s drama is it’s more important than ever before to prioritise education and not blindly follow the blind. AI can be daunting, but if you seek your own truth and look past the headlines, you can leverage yourself and your output using these tools to heights that were never previously possible.

Whether it be DeepSeek, OpenAI, or the other big hitters, the key is understanding how to use these tools responsibly and effectively.

Here's what I recommend:

1. Start with OpenAI's guide to understand how your data is used

2. Try DeepSeek's chat to test capabilities without sharing sensitive information

3. Always sanitize your prompts of personal or proprietary data

Quote to Share

Jim Fan on the democratisation of AI:

Restricting open-source development through compute thresholds was misguided “Silicon Valley hubris.” DeepSeek's success with relatively modest computational resources seems to validate this perspective.

What's particularly interesting is Fan's take on improved scaling efficiency. While some view DeepSeek's capabilities with concern, he celebrates the implications: achieving the same AI capabilities with 1/10th the computational resources potentially means 10x more powerful AI systems using current infrastructure.

2025 should be about accelerating open-source development, not fearmongering AGI.

Source: Jim Fan on X

Question to Ponder

I think the Beatles' “Now and Then” is a fascinating case study of AI's role in music's future.

After diving into the recent coverage, including Paul McCartney's BBC interview and the Grammy nomination controversy, I'm struck by how it perfectly encapsulates the multifaceted impact of AI on music.

Let's start with the listeners. We're entering an era where our favourite artists can continue creating music beyond their lifetime. The Beatles released a “new” song in 2023, 43 years after John Lennon's passing.

For artists, it's a double-edged sword. On one hand, AI tools are changing what's possible in music production. McCartney himself praised this technology, saying it allowed them to “extricate John's voice from a ropey little bit of cassette.” Yet, in his recent BBC interview, he also warned about the darker side—the potential for AI to “rip off” artists and make it impossible for new musicians to make a living.

The recording industry faces its own challenges. The Grammy's decision to allow “Now and Then” to compete for Record of the Year highlights their struggle to define rules around AI in music. They've stated that “only human creators are eligible,” but allow for “elements of AI material.” As Nashville songwriter Mary Bragg puts it, it's a “slippery slope.”

McCartney's stance is an interesting one. Despite using AI to create what he calls “the final Beatles record,” he's actively fighting against proposed changes to UK copyright law that would allow AI companies to use artists' work without proper compensation.

I think the future is about finding the sweet spot between technology and respect for original artistry. The Beatles' journey from analog to AI might just show us the way.

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