AI Updates on 2025-05-14

Google DeepMind introduces AlphaEvolve, a Gemini-powered coding agent for algorithm discovery that can design faster matrix multiplication algorithms, find new solutions to open math problems, and make data centers more efficient @GoogleDeepMind
OpenAI makes GPT-4.1 and GPT-4.1 mini available directly in ChatGPT, with GPT-4.1 mini replacing GPT-4o mini @OpenAI
Stability AI releases Stable Audio Open Small, a 341M-parameter text-to-audio model optimized to run entirely on Arm CPUs, enabling on-device audio generation on 99% of smartphones @StabilityAI
Hugging Face releases Wan2.1, a model excelling in Text-to-Video, Image-to-Video, Video Editing, Text-to-Image, and Video-to-Audio @huggingface
StepFun AI releases Step1X-3D, an open 3D generation framework with 4.8B parameters (1.3B geometry + 3.5B texture) under Apache 2.0 license @huggingface
Meta FAIR releases Open Molecules 2025 (OMol25) dataset and Universal Model for Atoms (UMA) for molecular discovery and modeling atom interactions @AIatMeta

AlphaEvolve applied to over 50 open problems in mathematical analysis, geometry, combinatorics and number theory, rediscovered state-of-the-art solutions in 75% of cases and improved upon previous best solutions in 20% of cases @GoogleDeepMind
AlphaEvolve found a simple code rewrite that removed unnecessary bits in TPU design, validated by TPU designers for correctness, representing Gemini's first direct contribution to TPU arithmetic circuits @AndrewCurran_
AlphaEvolve sped up the FlashAttention kernel by 32% and found improvements in pre- and postprocessing of kernel inputs and outputs, resulting in a 15% speed up @AndrewCurran_
Meta FAIR and the Rothschild Foundation Hospital partnered on a large-scale study revealing striking parallels between language development in humans and LLMs @AIatMeta
Meta releases Adjoint Sampling, a scalable algorithm for training generative models based on scalar rewards @AIatMeta

Anthropic launches a bug bounty initiative to stress-test an updated version of their anti-jailbreaking system before public deployment, in partnership with HackerOne @AnthropicAI
Gemini Advanced now connects with GitHub, allowing users to generate/modify functions, explain complex code, ask questions about codebases, and debug by importing code from public or private repositories @GeminiApp
Perplexity announces integration with PayPal and Venmo for commerce features including shopping, travel, voice assistants, and their upcoming agentic browser called Comet @perplexity_ai
Google brings Gemini to Wear OS, Android Auto, Google TV, and Android XR, while making Gemini Live's camera and screen sharing features free for all Android users @demishassabis
Y Combinator launches Storyboards, a tool that turns scripts into full storyboards with shot-level control and character/scene consistency @ycombinator
Amjad Masad announces Percival, an AI agent that can evaluate and fix other AI agents, outperforming SOTA LLMs by 2.9x on the TRAIL dataset @amasad

BigTech jobs (Google, Microsoft, Apple, Tesla, Meta, Nvidia, Palantir) show zero growth in the last 3 years, contributing to difficulty for CS majors to find jobs, with companies potentially leveraging AI to grow without hiring @deedydas
Kaggle partners with Hugging Face to enable direct use of Hugging Face models in Kaggle Notebooks, along with discovering linked public code examples @kaggle
Databricks acquires serverless Postgres startup Neon for $1B, representing a rare unicorn exit in the current tech market @deedydas
Andrew Ng announces new course on Model Context Protocol (MCP) in partnership with Anthropic, teaching how to build AI apps that access tools, data, and prompts using the standardized protocol @AndrewYNg

OpenAI introduces the Safety Evaluations Hub, a resource to explore safety results for their models that will be updated periodically as part of efforts to communicate proactively about safety @OpenAI
Anthropic notes that some future models may require the advanced "AI Safety Level 3" protections outlined in their Responsible Scaling Policy @AnthropicAI
Paul Graham suggests that AGI would mean the end of prompt engineering, as moderately intelligent humans can figure out what you want without elaborate prompts, and we can use the care needed to construct prompts as an index of how close we're getting to AGI @paulg