In partnership with

Good morning. It’s Monday, March 3rd.

On this day in tech history: In 1939, the first issue of Electronics magazine was published by McGraw-Hill. Check it out here.

In today’s email:

• DeepSeek Rakes In Cash

• Google’s 60hr/wk Plan For Reaching AGI

• OpenAI Considering Restricting API for GPT-4.5

• 3 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

DeepSeek touts 545% AI profit margin

Chinese AI startup DeepSeek claims its language models could achieve a theoretical 545% profit margin, citing cost efficiencies in its V3 and R1 models. In a GitHub post, the company projected daily revenue of $562,027 if all usage were billed at its premium R1 rates, while GPU leasing costs stood at $87,072. However, actual revenue remains “substantially lower” due to discounted pricing, free web and app access, and limited monetization.

DeepSeek reduces costs through caching, with 56.3% of inputs served from stored data, and dynamically reallocates GPU resources between inference and training. These measures allow lower pricing than OpenAI, but the broader challenge remains: while AI models seem profitable in theory, market realities—such as commoditization and pricing pressure—erode margins. Meanwhile, OpenAI maintains premium pricing for GPT-4.5 despite only incremental improvements, signaling a strategic pivot toward full-stack control as AI services grow increasingly undifferentiated. Read more.

Sergey Brin says AGI is within reach if Googlers work ‘60-hour weeks’

Google co-founder Sergey Brin has urged employees to work 60-hour weeks and return to the office full-time, arguing that these efforts are critical for advancing artificial general intelligence (AGI). In an internal memo, Brin described this workload as the “sweet spot of productivity” and stressed the need to maximize efficiency using Google’s Gemini AI tools.

Brin also criticized Google’s AI teams for excessive safety filters, claiming they slow development. He urged DeepMind researchers to remove unnecessary constraints, speed up iteration, and trust users more. His comments reflect mounting pressure on Google to keep pace with OpenAI and Microsoft, both of which have aggressively expanded their AI capabilities. Read more.

Developers face uncertainty as OpenAI reconsiders GPT-4.5’s API availability

OpenAI is reevaluating the long-term availability of GPT-4.5 through its API, citing its high computational costs and marginal performance gains over GPT-4o. At $75 per million input tokens and $150 per million output tokens, GPT-4.5 is dramatically more expensive than GPT-4o ($2.50/$10) while only outperforming it in 63.2% of professional queries, 57% of everyday queries, and 56.8% of creative tasks.

The company is actively seeking feedback to determine whether GPT-4.5 delivers enough unique value to justify its continued availability. Speculation suggests a GPT-4.5 Turbo could emerge as a cost-optimized alternative, though past turbo versions have traded quality for efficiency. If OpenAI removes GPT-4.5 from its API, developers reliant on its capabilities may need to pivot. The move reflects a broader industry challenge—balancing AI advancement, operational costs, and real-world utility. Read more.

• OpenAI plans Sora integration for ChatGPT’s video generation

• Grok-3 powers AI-generated 3D game with 3,500 lines of code

• Chain of Draft refines reasoning, slashing word use by 92%

• Study finds AI now drives press releases, complaints, and job ads

• Anthropic partners with U.S. National Labs for first 1,000 Scientist AI Jam

• Chinese AI experts get quiet warning to skip US travel amid rising tech tensions

• World's largest call center deploys AI to "neutralize the accent" of Indian employees

• Flora is building an AI-powered ‘infinite canvas’ for creative professionals

• AI companies race to use ‘distillation’ to produce cheaper models

• DeepSeek AI Releases Smallpond: A Lightweight Data Processing Framework Built on DuckDB and 3FS

• China's Honor announces $10 billion investment in AI devices

📄 DeepSolution: Boosting Complex Engineering Solution Design via Tree-based Exploration and Bi-point Thinking

📄 ViDoRAG: Visual Document Retrieval-Augmented Generation via Dynamic Iterative Reasoning Agents ​​

📄 Self-rewarding correction for mathematical reasoning

📄 R2-T2: Re-Routing in Test-Time for Multimodal Mixture-of-Experts

📄 NeoBERT: A Next-Generation BERT

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!