AI News This Week: Grok 3, Perplexity AI, DeepSeek R1 & More

February 2025 has been a turning point month for artificial intelligence, with significant advancements from both big players and new startups. From Elon Musk's xAI launching the unbelievably powerful Grok 3 to ex-OpenAI founder - Mira Murti, starting new companies. The AI ecosystem keeps changing at breakneck speed. Meanwhile, mainstream media outlets are figuring out their relationship with AI tools, and businesses are struggling with issues around global AI development. This series of updates really shows how AI is evolving through expert reasoning models, more intense research methods, and fresh approaches to human-AI collaboration. At the same time, access, security, and ethical issues continue to be the top industry agendas.

So, let’s quickly check out the latest happenings in the AI industry this week that you truly can’t miss!

1. Elon Musk's xAI Releases Grok 3

Elon Musk's AI company xAI has released its latest flagship model, Grok 3, featuring significant improvements over its predecessor. Developed using approximately 10 times more computing power than Grok 2, with training on an expanded dataset including court filings, Grok 3 represents a family of models including a smaller "mini" version optimized for speed. 

• The release introduces two specialized reasoning models, Grok 3 Reasoning and Grok 3 mini Reasoning. 

• These models are designed to analyze problems methodically and perform self-verification before providing answers. 

According to xAI, these models outperform competitors like OpenAI's o3-mini on multiple benchmarks. The Grok app will soon gain voice capabilities, with enterprise API access following shortly thereafter. 

Musk plans to open-source Grok 2 once Grok 3 reaches stability, continuing his approach of releasing previous versions as newer iterations mature. Premium+ X subscribers ($50/month) will receive prioritized access to Grok 3, while additional features will be available through a new SuperGrok subscription plan costing approximately $30 monthly or $300 annually.

2. Perplexity Deep Research: Capabilities and Applications

Perplexity AI launched its Deep Research feature in February 2025, transforming how users access specialized information through AI-powered analysis. 

• The system employs test time compute (TTC) expansion, mimicking human cognitive processes by dissecting complex queries, conducting dozens of autonomous web searches, evaluating hundreds of sources, and synthesizing findings through probabilistic reasoning. 

• This methodology allows the AI to reconcile contradictory information and prioritize authoritative sources, achieving a 21.1% score on the "Humanity's Last Exam" benchmark—significantly outperforming GPT-4o (3.1%) and DeepSeek-R1 (8.5%). 

Unlike competitors like Google and OpenAI, which offer similar functionality, Perplexity provides free access to deep research with five daily queries for non-subscribers, while Pro users ($20/month) receive 500 daily queries. Though fast and reasonably accurate, testing reveals that perplexity still makes minor factual errors requiring expert verification. Users can enhance results through specific query phrasing, source weighting, and format directives, with reports generated in 2-4 minutes and exportable as PDFs or markdown documents.

3. Mira Murati Launches Thinking Machines Lab

Former OpenAI CTO Mira Murati has launched Thinking Machines Lab, a new AI startup focused on developing multimodal systems that emphasize human-AI collaboration rather than fully autonomous AI. Announced on February 18, 2025, the company already employs 29 people, with Murati serving as CEO. 

• The leadership team includes other former OpenAI executives, including Barrett Zoff (former VP of research at OpenAI) as CTO and John Shulman (OpenAI co-founder who briefly worked at Anthropic) as chief scientist.

•  The startup aims to create flexible, adaptable, and personalized AI systems applicable across all fields.

Thinking Machines Lab has committed to open science principles, pledging to regularly publish technical blogs, papers, and code to advance collective understanding of AI. The company believes sharing research benefits both the public and its internal research culture. Currently recruiting additional talent, Thinking Machines Lab is seeking product builders, machine learning experts, and research program managers to join their team.

4. The New York Times Embraces AI Tools

The New York Times has authorized its product and editorial teams to use AI tools, according to a staff email that introduced an internal AI summary tool called Echo. The newspaper shared a suite of approved AI products alongside editorial guidelines for their appropriate use. 

• Editorial staff are encouraged to leverage AI for suggesting edits, brainstorming interview questions, and research assistance, but are explicitly prohibited from using AI to draft or significantly revise articles or to input confidential source information. Future AI applications at the Times may include digitally voiced articles and translations. 

• Approved AI platforms include GitHub Copilot for coding assistance, Google's Vertex AI for product development, NotebookLM, selected Amazon AI products, and OpenAI's non-ChatGPT API through a business account. 

This embrace of AI technology comes despite the Times' ongoing lawsuit against OpenAI and Microsoft for allegedly violating copyright law by training generative AI models on the publisher's content without permission.

5. Perplexity Offers Access to DeepSeek R1 with Enhanced Security

Perplexity and You.com have begun offering access to DeepSeek AI's powerful language models while addressing security and censorship concerns associated with the Chinese-developed technology. 

• Perplexity now hosts DeepSeek R1 exclusively on US/EU data centers, ensuring user data never reaches Chinese servers.

• Free users receive three Pro-level queries daily, while $20/month Pro subscribers gain unlimited access. You.com similarly offers both DeepSeek V3 and R1 through its $15/month Pro tier (normally $20). 

• Both platforms claim to have mitigated some of the censorship built into the models, particularly when connecting them to external web sources. 

You.com co-founder Bryan McCann explained that their implementation provides three access methods: through a proprietary "trust layer" with web source access, without web sources (allowing exploration of the models' native capabilities), or through customized prompting with user-defined instructions. Testing revealed that DeepSeek models' responses on political topics varied significantly depending on whether web sources were included, with models appearing less restricted when grounded in public sources.