Welcome back!

Open-source AI is reaching a new milestone with DeepSeek V3. This 671 billion parameter model sets a benchmark for efficiency, directly challenging the leading closed-source AI models — all on a modest training budget. Is this the future of AI innovation? Let’s unpack…

Today’s Summary:

• 💡 DeepSeek V3 leads open AI

• ⚡ Anthropic’s co-founder 2025 AI forecast

• 🗣️ OpenAI's 2025 feature wishlist

• 🏆 Alec Radford exits OpenAI

• 📖 Anthropic’s agent-building guide

• 📢 Sundar Pichai: AI’s high stakes

• 🛠️ 2 new tools

TOP STORY

DeepSeek V3 leads the “Open” AI race

The Summary: DeepSeek V3, a 671 billion parameter open-source AI model, is raising the bar for AI efficiency. Trained on a modest budget, it delivers state-of-the-art performance in coding and math benchmarks, rivaling top closed source models. Its open design promotes collaboration and innovation within the AI community.

Key details:

• Completed training in 2.8 million GPU hours, far below the typical cost of large-scale AI models

• Trained on 14.8 trillion tokens, with 37B parameters activated per token, scaling to 671B total parameters

• Achieved top scores in benchmarks like MATH-500 (score: 90.2), outperforming all leading models

• AI expert Andrej Karpathy noted DeepSeek’s stunning efficiency, calling it a highly impressive display of research and engineering

Why it matters: DeepSeek V3 demonstrates that cutting-edge AI can be developed cost-effectively and shared openly, challenging the dominance of closed-source models. This breakthrough could drive more innovation and collaboration in the open source AI community.

Technical Report

ANTHROPIC

In 2025 AI will take an even bigger leap, says Anthropic co-founder

The Summary: Anthropic's co-founder Jack Clark predicts a rapid acceleration in AI advancements by 2025, dismissing claims of a slowdown. As AI progresses, access to electricity and data center capacity will become crucial drivers of progress, with competition focused more on infrastructure than algorithms.

Key details:

• Progress in 2025 is expected to outpace 2024, though costs remain unpredictable

• Advancements will come from better scaling techniques, test-time compute, and reinforcement learning

• OpenAI o3 model shows strong scaling returns, achieving a 25% score on FrontierMath, up from 2% in just 3 months

• Clark emphasizes the critical transition from algorithms to electricity

• US data centers' electricity consumption rose from 1.9% of US national power in 2018 to 4.4% in 2023, with estimates reaching 6.7%-12% by 2028

Why it matters: AI is entering a new phase where infrastructure, like electricity access and data centers, will drive progress more than algorithms alone. Clark’s perspective highlight how these resource demands are becoming central to the global AI race.

OPENAI

2025 OpenAI wishlist revealed

The Summary: OpenAI CEO Sam Altman's conducted a poll revealing users' top requests for 2025. Demands include improved memory to maintain conversation context, an upgraded GPT-4o model, and longer context windows for complex tasks. Users also want an "adult mode" with fewer content restrictions, better research tools, family accounts, and improved voice chat.

Key details:

• Users want memory systems that remember both text and voice conversations without requiring manual cleanup

• Parents requested family accounts with customizable safety limits and usage insights

• Voice users asked for improved turn detection to avoid interruptions

• Rate limits, pricing, and context windows emerged as top priorities, with Altman confirming these will improve in 2025

• Interestingly, Altman noted that many of OpenAI's planned 2025 features were not among the top user requests from the poll

Why it matters: Altman's poll reflects a growing focus on listening to users and addressing their practical concerns. This approach could result in tools that align more closely with everyday demands instead of focusing solely on technical achievements.

QUICK NEWS

Quick news

• Alec Radford, lead author of the original GPT-1, is leaving OpenAI

• Anthropic published a writeup on building effective agents

• Google CEO Sundar Pichai says the stakes are high for AI in 2025

TOOLS

🥇 New tools

• Otterly - Brand and content monitoring on AI Search

• Genfuse - Automate any work with AI agents