AI Weekly Update - May 26, 2025
Google launches SOTA video generation, Anthropic launches the Claude 4 suite, OpenAI formally hires Jony Ive
what to know for now
🎬 Google I/O rolls out Veo 3 and Gemini upgrades. Gemini Live adds multimodal camera context, tiered subscriptions, and Deep Research reports. Veo 3 embeds text-to-audio video generation across Flow and YouTube pipelines, with wider regional availability. I/O also launched Imagen 4 and AI Mode search. Read more
🤖 Anthropic ships Claude 4 Opus and Sonnet. The two frontier models extend context, tool use, and agent workflows while lifting SWE-bench scores above 70 percent. Opus handles multihour chains; Sonnet gives lower latency and cost. Both reach the Anthropic API, Bedrock, and Vertex AI with unchanged pricing. Read more
🛠️ Jony Ive joins OpenAI hardware venture. OpenAI absorbed Ive’s io Products for $6.5 billion and gave LoveFrom design leadership across form-factor projects. Ive and Sam Altman target “physical AI embodiments” that merge sensors, compute, and model inference in one device line. Peter Welinder runs day-to-day engineering inside the new io division. Read more
🐍 Mistral reveals Devstral coding model. Devstral scores 46.8 percent on SWE-Bench Verified, eclipsing prior open models and rivalling closed GPT-4 variants. Apache-licensed weights run on a single RTX 4090, enabling local agent deployment across private repositories. Mistral also offers API access at Small-tier pricing. Read more
🧪 AI Research of the Week
Probing Semantic Routing in Large Mixture-of-Expert Models
From Intel Labs (Revised May 2025)Jake's Take: The team asks: when a language model picks its “expert” sub-networks, does word meaning drive the choice? They probe six Mixture-of-Experts (MoE) models with two datasets that group tokens by sense. Tokens sharing meaning route to the same experts, the effect peaks in central layers, and scaling the parameter count strengthens the link.
MoE saves compute because each input wakes only part of the network, with proof here that routing follows semantics gives builders leverage. They can freeze or audit experts tied to risk-laden topics instead of scanning every weight, trimming cost (and bolstering oversight). The flip side is that disabling one expert could force the model to basically drop an entire skill. This could set the stage for modular, pay-per-expert licenses.
what to know for later
🍏 Bloomberg leaks Apple AI glasses timeline. Prototype chips derived from Watch silicon enter runs this year, aiming for 2026 mass production. The eyewear will stream translation, navigation, and visual search through on-board models and cameras. Apple escalates field testing to counter Meta Ray-Ban and Google XR demos. Read more
🔎 DOJ eyes Google-Character.ai licensing pact. Antitrust lawyers opened a probe into Google’s licence that embeds Character.ai language models inside Google services. Investigators ask whether the deal dodges merger review thresholds while funnelling talent and IP into Alphabet. Google says regulators may query them and that Character.ai remains separate. Read more
🚀 China lofts first orbital AI supercomputer nodes. Twelve satellites for the Three-Body Computing Constellation carry eight-billion-parameter models and laser links, targeting one peta-ops each. Space cooling and solar power lift throughput toward one exa-ops while bypassing ground bandwidth ceilings. Zhejiang Lab invites international partners for later launches. Read more
🏗️ Oracle commits $40 billion for 400,000 Nvidia GB200s. The chips power OpenAI’s Stargate Abilene hub, rated 1.2 GW and leased to OpenAI for fifteen years. Financing includes $9.6 billion loans from JPMorgan plus equity from SoftBank and MGX. The move reduces OpenAI dependence on Azure. Read more