AI Updates on 2025-12-14

OpenAI releases GPT-5.2 Pro with extended thinking capabilities, showing significant improvements over 5.1 Pro comparable to the jump from o1 Pro to o3 Pro @MParakhin
Google announces realtime speech-to-speech translation powered by Gemini, now available in Google Translate and coming to developers early next year @OfficialLoganK
Gemini 2.5 and Gemini 3 Pro demonstrate improved performance on various reasoning tasks, with Gemini 3 Pro achieving the highest score of 9.1% on CritPt physics reasoning benchmark @mark_k

AI has made it possible for founders to craft perfect pitches at scale, making it untenable for VCs to rely on inbound cold emails alone, fundamentally changing how startups break through to investors @TechCrunch
Current code review tools are inadequate for AI-generated code, with developers needing to know the original prompt, human corrections made, and clear marking of unmodified AI-generated sections @GergelyOrosz
A team of strong software engineers who care about code quality and maintainability outperforms teams using powerful AI coding agents mindlessly, as AI tools tempt developers to push verbose, less maintainable code @GergelyOrosz
Staff engineers report that AI enables them to ask questions more freely without fear of judgment, leading to faster learning compared to traditional team dynamics where senior titles discourage basic questions @GergelyOrosz
Future AI systems in 10-15 years will be 4-5 orders of magnitude more energy efficient than current AI, with hardware becoming the main deployment bottleneck rather than power @fchollet
Datacenters in space are not economically viable, being 50-100x more expensive than ground-based nuclear or renewable-powered datacenters when considering launch costs, maintenance complexity, and high-bandwidth communications @fchollet

AI-generated disinformation is already being used to spread false narratives, with fabricated backstories and names being created for real people involved in news events, demonstrating the immediate threat to information integrity @Nrg8000
Sergey Brin admits Google under-invested in transformer architecture it invented because the company was too scared to release chatbots that say dumb things, allowing OpenAI to scale compute and run with the technology @slow_developer
Getting accurate answers from current AI is compared to tricking a habitual liar into telling the truth, requiring users to back the system into the right corner or provide the right prompts @paulg

JustHTML, a new Python library with no dependencies, was built mostly by coding agents over a couple of months, comprising 3,000 lines of code that parses HTML according to HTML5 specification and passes 9,200 html5lib-tests @simonw
A 17-step guide demonstrates using VS Code agent mode with Claude 3.7 Sonnet, Gemini Pro 3, and Claude Opus to build production-quality code, showcasing serious engineering rather than vibe coding @simonw
Codex team adds experimental support for skills that combines well with GPT-5.2, enabling fine-tuning of Qwen3-0.6B to achieve +6 improvement on HumanEval benchmark @thsottiaux
Comet Assistant is moving compute toward fast lightweight models that can potentially run locally, enabling deeper analysis on any article, video, or website without switching context @AravSrinivas

GPT-5.2 Pro scores 0% on CritPt, a research-level physics reasoning benchmark designed to test expert-grade theoretical physics reasoning, while Gemini 3 Pro achieves the highest score of 9.1% @mark_k
All recent AI models now correctly solve the surgeon riddle on first try, demonstrating progress in handling gender bias in reasoning tasks @emollick
Open models year in review identifies DeepSeek R1, Qwen 3 Family, and Kimi K2 Family as top performers, with predictions that scaling will continue and the open-closed frontier gap will remain roughly the same on public benchmarks in 2026 @natolambert
Stanford's Foundation Model Transparency Index shows industry transparency collapsing from 58 to 40.69, with only IBM and Writer maintaining transparency while others reduced disclosure @JesseDLandry