# 🎅🤖 ThursdAI - 2025 a year of AI in review - Recap of the most notable AI updates for the past year! Source: https://www.youtube.com/watch?v=F1RtlhyErmg Recap page: https://rapidrecap.app/video/F1RtlhyErmg Generated: 2025-12-25T19:33:03.389+00:00 --- ## Quick Overview The year 2025 in AI was defined by massive acceleration, highlighted by the explosion of agents, especially CLI agents starting with Claude Code, significant advancements in multimodal models like Google's V3 video generation, and the open-source community making massive strides with models like DeepSeek R1 and Qwen 3, which frequently outperformed larger proprietary models, leading to a paradigm shift where proprietary AI's dominance was closely challenged. **Key Points:** - The most important AI release of 2025, according to multiple co-hosts, was Claude Code, marking the start of the 'year of the agents' and the explosion of CLI tools. - DeepSeek R1 shattered the open-source landscape in January, reportedly looking like it beat foundational models and causing a temporary $560 billion loss on the NASDAQ for Nvidia due to its perceived low training cost ($5 million). - Google's V3 video model, released in May, was 'earthshattering' for video generation, featuring native audio with perfect lip sync and marking the crossing of the uncanny valley. - Anthropic's Claude 4 Opus and Sonnet were live-dropped in May, becoming the first models to cross 80% on Swagbench verified, demonstrating hybrid reasoning and instant response capabilities. - OpenAI's GPT-4.5 release in February was rumored to be a large, unsuccessful release, contrasting with the success of models like Claude 3.7 Sonnet, which was mind-blowing for coding capabilities. - Qwen 3, released in May by Alibaba, was an 'insane release' for open source, with its 4 billion parameter dense model beating the 72 billion parameter Qwen 2.5 model on multiple benchmarks and offering a runtime toggle for Chain of Thought on demand. - Andre Karpathy coined the term 'Vibe Coding' in February, signifying the shift where developers talked to AI agents for side projects instead of manually reading code, coinciding with the rise of agentic tools like Cursor and Claude Code. **Context:** This episode of ThursdAI, hosted by Alex Vulov and co-hosts Wolfram, Yam, and LDJ, serves as the 2025 yearly recap, covering the massive wave of AI releases throughout the year, starting from Q1 in January. The discussion features guest Halman Kramer, CEO of Daily, to discuss key developments, particularly focusing on agents, model performance benchmarks, and shifts in the competitive landscape between proprietary labs and the open-source community. ## Detailed Analysis The recap of 2025 reveals an intense acceleration in AI, starting with Q1 being an 'earthquake' fueled by DeepSeek R1 challenging proprietary models and OpenAI's initial agentic chip, Operator. February cemented the 'year of the agents' with the release of Claude Code and Claude 3.7 Sonnet, which excelled at frontend design, while Andre Karpathy popularized 'Vibe Coding.' Q2 saw major milestones: May was highlighted by Google's V3 video model seemingly killing the uncanny valley with perfect lip-syncing and the release of Claude 4 Opus crossing 80% on Swagbench, alongside Alibaba's Qwen 3 offering unprecedented open-source performance improvements, such as a 4B model outperforming a previous 72B model. June focused on the price war heating up, with 03 dropping prices significantly, making intelligence cheaper for startups, though Meta's Llama 4 release was considered a fumble, leading to massive hiring acquisitions like Alex Wang to bolster their standing. Throughout the year, key themes included the propagation of reasoning across models, the rise of multimodal reasoning (like 03's tool calling), and the increasing blurring of lines between proprietary and open-source capabilities, making it a highly competitive and transformative year. ### Q1 Highlights (Jan-Mar) - DeepSeek R1 shattered open source expectations, nearly impacting Nvidia stock - OpenAI released Operator, the first agentic chip, evolving into Atlas - March saw Gemini 2.5 Pro and the adoption of MCP, while OpenAI's GPT-4.1/4.5 naming was confusing. ### Agent & Coding Revolution - Claude Code's success stemmed from being internally dog-fooded by Anthropic, functioning as a thin wrapper with a great system prompt - Vibe Coding emerged as the new paradigm for side projects, driving adoption of tools like Cursor. ### Q2 Highlights (Apr-Jun) - April featured OpenAI's 03/04 mini reasoning models with tool calling and Meta's disappointing Llama 4 release - May brought Google V3 video (uncanny valley crossed) and Claude 4 Opus/Sonnet hitting 80% on Swagbench - June saw 03 price drops initiating a price war, while Meta made major acquisitions following Llama 4's struggles. ### Voice & Multimodality - OpenAI separated its GPT real-time speech-to-speech model into steerable transcription and voice generation LLMs in March - GPT-4o's native image generation fueled 'Giblmania 2.0' due to conversational iteration - Voice agents saw a shift from GPT-4o to Gemini 2.5 Flash for production due to compelling updates. ### Open Source Dominance - Qwen 3 was a top open-source release, featuring a runtime toggle for Chain of Thought on demand and significant performance gains over larger predecessors - DeepSeek R1 caused a market downtrend by demonstrating high performance at low training costs, signaling proprietary AI was no longer unchallenged.