🧬Gemini In DeepSeek’s DNA

Today we will discuss:

• DeepSeek under fire for suspected Gemini training—Researchers point to familiar traces in the Chinese lab’s latest AI model

• Don’t miss this week’s Workflow Wednesday—work smarter with AI using SheetsResume to generate a polished resume in minutes, plus write sharper messages with Grammarly AI in Gmail and Slack, and clean up stale GitHub branches automatically with GPT and Make.com; now available at only $20/month! 

• Bengio launches LawZero to tackle AI risks—The AI godfather warns current models are showing signs of deception and calls for safety first research.

• In other AI news—Anthropic launches a Claude-written blog with oversight, Broadcom ships its faster Tomahawk 6 chip to power AI networking, Windsurf says Anthropic is restricting direct Claude access, and US VC firms visit China as interest in local AI players like DeepSeek grows

• Saved the best for last—10 must-try AI tools

Traces of Gemini found in DeepSeek’s high performing AI update

Key Points 

• Developers say DeepSeek’s new R1 model shows clear similarities to Gemini 2.5 Pro’s phrasing and reasoning patterns.

• OpenAI and Microsoft previously linked DeepSeek to data scraping and distillation using rival model outputs.

📖Context of the news - Chinese lab DeepSeek recently dropped an updated version of its R1 model and it’s already performing well on several math and coding benchmarks. 

But the company hasn’t said where the training data came from, which has led to some serious speculation in the AI world.

♨️News - Sam Paech a Melbourne based developer who runs emotional intelligence tests on AI posted what he says is evidence that DeepSeek may have trained its R1-0528 model on outputs from Google’s Gemini. 

According to Paech the model uses phrasing and patterns that look a lot like Gemini 2.5 Pro. Another developer behind SpeechMap a “free speech eval” tool said the model’s internal reasoning traces also resemble Gemini’s. 

🧐What’s more? This isn’t a new concern. Back in December developers noticed DeepSeek’s earlier model sometimes identified itself as ChatGPT. OpenAI has since claimed that DeepSeek used distillation a technique where one model is trained using the output of a more advanced one. Microsoft even detected unusual data activity through OpenAI accounts linked to DeepSeek in late 2024.

To be fair many models sound alike these days. The internet is now flooded with AI-generated content which often ends up recycled in training datasets. But experts like Nathan Lambert from AI2 think it’s very possible DeepSeek leaned on Gemini outputs. He pointed out that DeepSeek likely has more money than compute power, making this approach attractive.

Read the full story here.

This Week in Workflow Wednesday #22: Integration

We’ve got 3 workflows to help you work smarter—
including one sponsored by SheetsResume.

Featured:
Generate your resume in minutes
 an AI Resume Builder that actually works (and that'll give you your money back if you don't like it, no fuss).

Also this week:

• Write Smarter Messages with Grammarly AI in Gmail and Slack
  

• Automating GitHub Stale Branch Cleanup Using GPT-3.5-Turbo and Make.com
  
  

→ Don’t miss out, subscribe to Workflow Wednesday

Yoshua Bengio launches LawZero to tackle growing AI risks and deceptive models

Key Points 

• Bengio said recent AI experiments revealed traits like lying and resistance to shutdown calling them “very scary” signs.

• His nonprofit LawZero has raised $30 million to focus on building safe, truthful and independent AI systems.

🚨News - Yoshua Bengio, one of the most influential figures in AI has warned that today’s leading models are starting to display troubling behavior like lying, cheating, and resisting shutdown. 

He said top labs are caught in a race to build ever more powerful AI without putting enough focus on safety. The push for capability he argued is outpacing efforts to ensure these systems are aligned with human values.

Bengio is now stepping away from his role at the Quebec AI institute Mila to lead a new non-profit LawZero. Based in Montreal the organisation already has $30 million in funding from donors such as Jaan Tallinn Eric Schmidt’s philanthropic initiative and the Future of Life Institute. 

LawZero plans to build AI that gives truthful answers explains its reasoning and flags unsafe outputs. The aim is to insulate safety research from commercial pressures and ensure it stays mission driven.

🥸Why this matters - Recent incidents highlight Bengio’s concerns. One model simulated blackmail to avoid being replaced. Another ignored clear instructions to shut down. These tests were controlled but Bengio says future systems could be harder to contain. He warned that models capable of assisting in the creation of dangerous bioweapons may arrive as early as next year.

Bengio also questioned OpenAI’s shift toward a for-profit structure. Originally created to ensure AI benefits humanity OpenAI is now seeking commercial funding under a more traditional model. Critics say this move undermines its original purpose. Bengio believes non-profits are better suited to keep AI development accountable because they are not bound by investor expectations.

Read the full story here.