AI Updates on 2025-12-07

Google announces Gemini 3 Pro as state-of-the-art vision AI model, achieving top performance across all main vision and multimodal benchmarks, excelling at document, screen, image, video and spatial understanding tasks @demishassabis
Reka AI releases Rnj-1 base and instruct 8B parameter models, achieving SWE-bench performance close to GPT-4o, tool use outperforming comparable open source models, and mathematical reasoning on AIME'25 nearly matching GPT OSS MoE 20B @ashVaswani

Elon Musk proposes space-based AI datacenters with satellites featuring localized AI compute in sun-synchronous orbit, projecting this will become the lowest cost way to generate AI within 3 years and fastest way to scale within 4 years, with plans to scale to over 100TW/year using lunar satellite factories @elonmusk
OpenAI disables app suggestions that appeared similar to advertisements following user feedback @TechCrunch
Meta reportedly delays mixed reality glasses release until 2027 @TechCrunch
Perplexity celebrates three-year anniversary of its launch using OpenAI GPT-3.5 and Microsoft Bing for direct question answering @AravSrinivas

Andrej Karpathy advises users to think of LLMs as simulators rather than entities, explaining that when asked "What do you think about xyz?" there is no actual "you" - the model adopts a personality embedding vector from its finetuning data statistics rather than having formed genuine opinions over time @karpathy
Daniel Kahneman's 2017 pre-LLM research suggests replacing humans with algorithms whenever possible, noting that even when algorithms don't perform exceptionally well, humans perform so poorly and with such noise that removing the noise alone yields better results than human performance @jamescham
Ethan Mollick questions whether major publications have provided retrospectives on AI development plateau claims following GPT-5 router experiences, noting confusion persists despite evidence that barriers like model collapse and pre-training scaling were overcome @emollick

Claude Skill enables Opus 4.5 to generate Apple-style infographics with highly technical design specifications, using prompts generated by Grok 4.1 to think like Steve Jobs of graphic design @deedydas
Cardiac Electrophysiologist uses AI workflow combining Claude, Suno, and NanoBanana to create educational songs for children ages 4 and 7, demonstrating creative applications that would be entirely infeasible without AI @HamelHusain
MIT researchers develop AI-powered strategy for strengthening polymer materials, potentially leading to more durable plastics and reduced plastic waste @MIT
Wikipedia maintains a list of AI writing tells including negative parallelisms like "It's not a game. It's a revolution" that can be incorporated into system prompts to avoid AI-sounding text @blader

First BEHAVIOR challenge results announced at NeurIPS, evaluating embodied AI and robotics solutions on 50 challenging household tasks, with Robot Learning Collective winning first place, followed by Comet and SimpleAI teams @drfeifei
AI2 presents OLMo 3 post-training research emphasizing the importance of evaluation methodologies in AI development at NeurIPS Foundations of Reasoning in Language Models workshop @natolambert
NeurIPS workshop on Foundations of Reasoning in Language Models features talks on self-improvement, exploration, chain-of-thought, and related topics @canondetortugas