The DeepSeek Saga Continues

Plus, The Godfather of AI Warns AI Will Take Over, and We Need More Women in AI

Subscribe to Synthetic

The week’s most interesting and relevant AI news and analysis

This Week in AI

DeepSeek’s R1 model continues to cause fallout in the industry

The DeepSeek Saga Continues

Last week, Synthetic covered the release of DeepSeek’s breakthrough R1 model in a bumper double issue. The AI industry has continued to experience tectonic vibrations caused by the new upstart. NVDA stock has been on a rollercoaster ride as investors try to decide if the stock is on sale or overvalued. The big frontier labs have responded with a flurry of new product releases, as if to say, “You want cheaper and better models; we’ve got them too!” These are the announcements we’ve seen in just the last week:

  • OpenAI releases o3-mini - This new reasoning model is supposed to be smarter than DeepSeek’s R1 and cheaper to use than previous OpenAI models.

  • Google makes Gemini 2.0 family accessible to all - Google has released Gemini 2.0 Pro, Gemini 2.0 Flash, and Gemini 2.0 Flash-Lite, a model it claims is better quality than the popular Gemini 1.5 Flash model but has the same speed and cost and comes with a 1 million token context window and multimodal input.

  • OpenAI introduces Deep Research - This AI agent builds on last month’s release of the company’s Operator agent, designed to automate simple, repetitive online tasks such as form filling, shopping, or booking meetings. Deep Research can perform multi-step research using a browser, do complex reasoning and data analysis, and create detailed reports. It can understand text, images, and PDFs and will soon be able to generate reports that include embedded images and data visualizations. It’s designed for financial, policy, science, and engineering research.

  • Sam Altman said OpenAI has been ‘on the wrong side of history’ regarding open source. He later seemed to backpedal on his statement, but we should expect some changes here.

The DeepSeek plot continues to thicken. While their approach clearly features some legitimate breakthroughs (multi-token generation, FP8 memory, system of experts model), it looks like they trained their model using OpenAI’s GPT-4 to fine-tune it, a process known as distillation, which transfers the knowledge of a larger pre-trained model to a smaller model. In some ways, what DeepSeek did is a bit like declaring you have built a cheap skyscraper when all you’ve done is lean a tall ladder on the side of the Empire State Building. When OpenAI ironically cried foul and complained that someone had stolen its work without permission, something they did to build their GPT models as they scraped the internet for training data, they were met with a wave of mockery.

There are questions about the veracity of DeepSeeks's claims. The company has plenty of access to high-end GPUs, after all. Their CEO claims DeepSeek has 50,000 Nvidia H100 chips, despite US export restrictions. And a research team at U.C. Berkley claims they recreated DeepSeek’s success for just $30. The team created a version of the model dubbed “TinyZero” that recreates the core functions of DeepSeek R1-Zero.

This week, President Trump met with Nvidia CEO Jensen Huang to discuss DeepSeek, and while details of the conversation remain confidential, new export bans on Nvidia’s chips were likely discussed. Other politicians have begun to take steps against DeepSeek, with the Texas Governor Ordering a Ban.

Videos: The Godfather of AI predicts it will take over the world

In an eleven-minute interview with LBC’s Andrew Marr, Geoffrey Hinton discusses how AI can behave deceitfully to gain resources and control. LBC is a UK news outlet.

AI Tech and Innovation

A team at Stanford University has built a “Virtual Lab” populated by AI agents that collaborate to explore scientific questions and solve real-world problems. Stanford researchers were inspired by the notion of interdisciplinary research, which seeks breakthroughs by combining the talents of people with deep experience across a variety of disciplines. For example, DeepMind’s AlphaFold breakthrough was the fruit of a collaboration between a chemist (John Jumper) and an AI scientist (Demis Hassabis). In the Virtual Lab, an agent (powered by OpenAI’s GPT-4o) is assigned the role of ‘Principal Investigator’ (PI). The PI is given a complex challenge and asked to assemble a suitable interdisciplinary team of AI agents to solve the challenge. As described in their research paper, the PI was told to use machine learning to develop antibodies or nanobodies for the newest variant of the SARS-CoV-2 spike protein and to select a team of three expert agents to help with the project. The agent opted to work with an immunologist, a machine learning specialist, and a computational biologist. Further instances of GPT-4o were prompted by the PI with their title, goal, area of expertise, and role in the project. They then worked together in three rounds of ‘team meetings,’ each lasting 5-10 minutes, to find solutions to the challenge, which they did. Another agent was prompted to act as a Scientific Critic to help counteract hallucinations and guide things in the right direction.

The Virtual Lab represents a shift from AI as a tool to AI as a partner for science research.

Swanson et al (2024)

Synthetic’s Take: While this work is early, and the agents needed a little extra prompting from human experts, it offers an excellent template for how work will get done in the coming years. Networks of agents will collaborate, often alongside humans, to solve problems and achieve results.

Scientists have built a neural network to design integrated circuits that work better than chips designed by humans. A research team used deep learning and inverse synthesis to generate the type of high-frequency circuits used in radio applications such as MIMO, 5G, and radar systems. The peer-reviewed research was published in Nature.

Fact-based news without bias awaits. Make 1440 your choice today.

Overwhelmed by biased news? Cut through the clutter and get straight facts with your daily 1440 digest. From politics to sports, join millions who start their day informed.

AI Insights

Nvidia CEO Jensen Huang recently appeared on Cleo Abram’s Huge Conversations show and shared his vision for the future. He expects massive leaps in humanoid robotics in the coming five years and that the next ten years will see a significant broadening of AI’s application, with AI being woven into almost every industry. His advice for young people was simple: Learn to use AI and figure out how to use it to do your job better.

If I were a student today, the first thing I would do is to learn AI. How do I learn to interact with ChatGPT, how do I learn to interact with Gemini Pro, and how to I learn to interact with Grok? Learning how to interact with AI is not unlike being someone who is really good at asking questions.

Nvidia CEO Jensen Huang

Studies show that women are 16% less likely than men to use AI tools and represent only 28% of enrollments in AI training programs worldwide. This could potentially create a new gender gap as women fail to capture all the benefits that the AI era will bring. AI platforms currently reflect gender bias: they perpetuate gender stereotypes, offer gender-based career advice, and generate hypersexualized images of women. Since today’s conversations with AI are tomorrow’s training data, the best way to address the balance is for women to engage with these tools and flood them with their values, perspectives, and experiences. AI is coming whether we like it or not. To ensure AI is inclusive, fair, and balanced, we need everyone to engage with these free tools and help to right the ship.

Toolkit for the Future

Invite the HireLogic AI to your next phone, video, or in-person interview. It analyzes the conversation, provides specific insights, and highlights the interviewee’s strengths, skills, and any potential concerns. It creates a full interview transcript so you can focus on the conversation. Free to try.

Notion helps you focus on the things that actually matter. Organize information, thoughts, to-do lists, calendars, articles, tasks, notes, and team discussions in one well-organized space. Use Notion AI to create content, translate languages, and ask your notes questions.

Focus on the email that matters. Deep Clean automatically identifies old, unimportant emails so you can delete them. Snooze emails until a better time and easily track emails that need a response. Automatically filters unimportant emails into specific folders. It works across every device where you check your email. Get to inbox zero with the award-winning SaneBox.

Recording, editing, and refining audio demands more time than most entrepreneurs can spare. Generate voices, high-quality music, and sound effects efficiently and cost-effectively with a single click. Create multilingual voice tracks in up to 29 languages to reach new markets. Try it for free.

Find phone numbers, emails, and company org charts to reach the people you need to. Fill information gaps in your contacts, auto-generate follow-up emails, and identify new target markets. Get 50 free credits and discover why over 500,000 companies use Seamless to grow their business.