Sumbird - AI news and vibes from Twitter

Google releases updated Gemini 2.5 Pro (I/O edition) with significantly improved coding capabilities, ranking #1 on WebDev Arena leaderboard with a +147 Elo score gain @JeffDean @GoogleDeepMind
The new Gemini 2.5 Pro model (gemini-2.5-pro-preview-05-06) is now #1 across all LMArena leaderboards including text, vision, and WebDev @OfficialLoganK
Meta introduces Meta Perception Encoder, a vision encoder setting new standards in image & video tasks, excelling in zero-shot classification & retrieval @AIatMeta
ServiceNow and NVIDIA announce Apriel Nemotron 15B, a compact AI model built with NVIDIA NeMo and trained on NVIDIA DGX Cloud @NVIDIAAI

Gemini 2.5 Pro achieves 84.8% on the VideoMME benchmark, demonstrating state-of-the-art performance on image and video understanding @JeffDean
Google Research introduces a system using Gemini models designed for high fidelity text simplification that enhances clarity while preserving meaning, detail, and nuance @GoogleAI
Second-order optimization shows promise for more efficient LLM pretraining according to a study on the advantages of Muon @eladgil
BayesFlow 2.0, a Python package for amortized Bayesian inference powered by Keras 3, released with support for JAX, PyTorch, and TF @fchollet

Gemini 2.5 Pro can build interactive web apps, games, and simulations from a single prompt, with significantly improved capabilities for front-end web development, editing, and transformation @demishassabis
Hugging Face releases Open Computer Agent, allowing LLMs to complete tasks using a virtual machine, testing how well current models use a computer to solve everyday tasks @huggingface
Microsoft introduces Claimify, a new method for extracting simple, verifiable claims from LLM outputs that preserves critical context and outperforms past approaches @MSFTResearch
Google launches Simplify feature for iOS that uses AI to make dense text easier to understand @GoogleAI
Computer Use in smolagents launched by Hugging Face, allowing vision models to power complex agentic workflows, especially with Qwen-VL models that support built-in grounding @huggingface
Cursor announces free access for students to their AI-powered coding assistant @cursor_ai
Windsurf introduces Knowledge Base in Wave 8, allowing users to import documents from Google Drive for Cascade to use as context @windsurf_ai

OpenAI's acquisition of Windsurf for $3 billion is moving forward according to Bloomberg reports @AndrewCurran_
Windsurf was acquired for $3B at ~$40M ARR (75x) by OpenAI, while Cursor raised at $9B at ~$300M ARR (30x) @deedydas
Google's improved position in AI is evident as Gemini dethrones previous Gemini versions, signaling that "the dragon woke up" @AndrewCurran_
Chinese AI startups are going to great lengths to not be seen as Chinese, with companies like Genspark presenting themselves as "Palo Alto based" despite connections to China @deedydas
Reinforcement Learning (RL) is very expensive compared to Supervised Fine-Tuning (SFT), but perfect for businesses as it can optimize metrics that matter like sales or customers @alexgraveley
AI lowers the barrier to getting started on anything, but doing great work still requires execution, judgment, creativity, and domain knowledge @paulg

Paul Tudor Jones reported on CNBC that a leading AI developer at a tech conference stated "I think it's going to take an accident where 50 to 100 million people die to make the world take the threat of this really seriously" @AndrewCurran_
MIT Media Lab researchers present TeleAbsence, exploring design principles for how AI could help people cope with loss and plan for how they might be remembered @medialab
Numerous members of the MIT and Media Lab communities will participate in the Venice Biennale with the theme "Intelligens. Natural, artificial, collective," focusing on applying adaptive intelligence to a demanding world @medialab

Hugging Face announced that Nvidia has released Llama-Nemotron, an efficient reasoning model @huggingface
Nvidia open-sourced Parakeet TDT 0.6B, described as the best speech recognition model on Open ASR Leaderboard, capable of transcribing 60 minutes of audio in 1 second @huggingface

MIT researchers developed a new method to make AI models more trustworthy for high-stakes settings by conveying uncertainty more precisely @MIT
Chris Olah investigated whether superposition is a major cause of adversarial examples by training SAEs on adversarially trained models @ch402
Research suggests D-FINE, a real-time object detector faster and more accurate than YOLO with Apache 2.0 license, has been added to Hugging Face transformers @huggingface

Simon Willison released a new llm-video-frames plugin that turns video files into sequences of JPEGs to feed into long-context vision models like GPT-4.1-mini @simonw
Perplexity on WhatsApp provides a convenient way to use AI when in flight, as flight WiFi supports messaging apps well @AravSrinivas
Claude 3.7 Sonnet can now crawl entire websites, extract specific data, and complete research tasks without leaving the desktop app @ycombinator
Google's Veo 2 on the Gemini app allows users to input prompts directly to generate videos, with the model only able to respond in video format @AndrewCurran_
Pulse AI launched Ultra, described as their new hybrid reasoning model and "the most accurate document extraction model in the industry" @ycombinator
Alex 3.0 released with features to automatically compile and fix errors, auto-apply code, add packages, search the web, run terminal commands, and review code with local LLM support @ycombinator

OpenAI announced structural changes: the nonprofit will continue to control the for-profit entity, which will become a Public Benefit Corporation with the same mission @OpenAI @sama
Many companies cannot use Qwen and DeepSeek open models because they come from China, slowing adoption of open models across enterprises @natolambert
Google refreshed its music-generation tools with Lyria 2 for Music AI Sandbox and Lyria RealTime for DJ, producing high-quality 48kHz audio with extensive control over musical attributes @DeepLearningAI
The Keras team released KerasRS, a new library for building recommender systems with easy-to-use building blocks compatible with JAX, PyTorch, TF, and optimized for TPUs @fchollet
Hugging Face introduced the Common Crawl Creative Commons Corpus (C5), a heavily filtered web-crawl dataset containing only Creative Commons licensed documents with 150 billion tokens collected so far @huggingface

Arvind Narayanan discusses two ever-present risks when using generative AI for work: hallucinations/confabulations and deskilling, emphasizing the importance of having a plan to address these risks @random_walker
Stanford HAI reports that Visa's integration of AI into its payment system could lead to consumers facing higher prices and deceptive practices without realizing it @StanfordHAI
A study found that people struggle to get useful health advice from chatbots @TechCrunch
Ethan Mollick shares research suggesting people may be massively underreporting their AI usage in surveys @emollick

AI Updates on 2025-05-06