AI Updates on 2025-09-14

Aidan McLaughlin argues that the key to AGI lies in giving models good tools and good reward, calling this the "modern bitter lesson" - suggesting that complex architectural improvements matter less than practical tool access and reinforcement learning @aidan_mclau
McLaughlin observes that successful AI improvements came from giving Sonnet a terminal and RL training rather than complex architectures, giving models internet search tools rather than pretraining science, and providing vector database access rather than specialized post-training @aidan_mclau
Ethan Mollick finds that model collapse predictions were wrong, noting that AI development has continued despite concerns about training on AI-generated content, with a billion people now using AI weekly @emollick
Simon Willison critiques the model collapse theory as treating AI developers as having no agency to notice and counter quality degradation in their models @simonw
Yann LeCun shares comprehensive research on Large Reasoning Models (LRMs) including evaluation on planning, semantics of intermediate tokens, RL analysis, and interpretability studies @rao2z

Aidan McLaughlin asks about user experiences with Sonnet 1M context length on Claude for coding, questioning whether the longer context is a significant unlock @aidan_mclau
Ethan Mollick tests AI models on a creative time travel scenario, with Gemini suggesting learning maritime concrete formulas, Claude recommending memorizing specific texts, and ChatGPT proposing discovering the Etruscan language and Alexander's Tomb location @emollick
Deedy observes that while Google researchers built Gemini as a universal oracle, its biggest viral moment is people using it as an image editing tool for Instagram photos @deedydas
TechCrunch reports on experienced coders' perspectives on AI-generated code and the future of "vibe coding" @TechCrunch

Andrew Curran highlights the need for terminology describing when captchas become so difficult to deter AI models that they become impossible for some humans to solve @AndrewCurran_
Ethan Mollick demonstrates vulnerabilities in AI detection systems, showing that the Pangram detector can be easily defeated by asking AI to eliminate em-dashes, highlighting the ongoing race between detectors and detection evasion @emollick
TechCrunch reports on websites claiming to allow users to chat with God, raising questions about AI applications in religious contexts @TechCrunch

TechCrunch analyzes how the competitive landscape of AI is changing in ways that undermine the advantages of the biggest AI labs @TechCrunch
TechCrunch explores how OpenAI's rise represents both a business and ideological story, examining how the cult of AGI has fueled massive spending on compute and data @TechCrunch
Bret Taylor, like OpenAI CEO Sam Altman, acknowledges being in an AI bubble but expresses little concern about it @TechCrunch
Google Gemini App reports experiencing high demand requiring temporary limits to manage peak usage, with the team working to maintain system stability @joshwoodward
TechCrunch covers Penske's lawsuit accusing Google of abusing its search monopoly to force publishers to support AI summaries @TechCrunch