// AI Briefing · March 25, 2026

AI Briefing

OpenAI's GPT-5.4 just crossed a line nobody expected this soon - beating humans at navigating a real desktop. Meanwhile, Apple's about to flip the switch on a Gemini-powered Siri, Portkey handed the open-source community a serious gift, and Yann LeCun is betting a billion dollars that the entire LLM paradigm is wrong. Oh, and Meta wants to build its own chips now - four generations of them.

OpenAI's GPT-5.4 Beats Human Baseline on Desktop Computer Use

OpenAI launched GPT-5.4 on March 5 with a headline-grabbing achievement: a 75% score on the OSWorld-Verified benchmark for desktop navigation, surpassing the human baseline of 72.4%. This is the first general-purpose model with native computer-use capabilities - it can interact with software through screenshots, mouse commands, and keyboard inputs. The jump from GPT-5.2's 47.3% to 75% in a single generation is staggering, and it signals that autonomous AI agents capable of doing real work on your computer are no longer theoretical.

02ProductMacRumors ↗

Apple's Gemini-Powered Siri: iOS 26.5 Beta Expected March 30

Apple is days away from dropping the iOS 26.5 developer beta - expected March 30 - which will debut the long-awaited Gemini-powered Siri overhaul. Apple is paying Google roughly $1 billion per year for a custom 1.2 trillion parameter Gemini model running on Apple's Private Cloud Compute, keeping user data isolated from Google's infrastructure. This is a massive architectural shift: Apple essentially outsourced its AI brain to its biggest rival while wrapping it in Apple's privacy guarantees. The result should be a Siri that finally understands context, works across apps, and doesn't embarrass you in front of your friends.

03Open SourceYahoo Finance ↗

Portkey Open-Sources Its AI Gateway - Now Processing 1T+ Tokens Daily

Portkey made everything open source on March 24 - governance, observability, authentication, cost controls, and their new MCP Gateway for managing AI agents. This isn't a toy project: the gateway already processes over 1 trillion tokens and 120 million AI requests daily across 24,000+ organizations, managing $180M+ in annualized AI spend. The MCP Gateway component is especially significant, providing centralized authentication and access control for Model Context Protocol servers across enterprises. For teams drowning in AI agent sprawl, this is the infrastructure layer they've been building in-house.

04IndustryMeta AI ↗

Meta Maps Out Four Custom MTIA AI Chip Generations Through 2027

Meta announced it's building four generations of custom MTIA silicon - the 300, 400, 450, and 500 - on a six-month release cadence through 2027. The MTIA 300 is already in production for ranking and recommendations, while the 400 through 500 will target generative AI inference. From the 300 to 500, HBM bandwidth jumps 4.5x and compute FLOPs increase 25x. Built on RISC-V architecture with TSMC and Broadcom, this is Meta's clearest signal yet that relying entirely on Nvidia is a strategic risk it's not willing to take - even as it continues spending billions on Nvidia hardware.

05FundingTechCrunch ↗

Yann LeCun's AMI Labs Raises $1.03B Seed to Build AI 'World Models'

Turing Award winner Yann LeCun's Paris-based startup AMI Labs closed the largest seed round in European history - $1.03 billion at a $3.5 billion pre-money valuation - backed by Nvidia, Bezos Expeditions, Samsung, and Temasek. AMI is building on LeCun's JEPA (Joint Embedding Predictive Architecture), creating AI that learns by understanding physical laws of the world rather than predicting the next word in a sequence. This is a direct, well-funded bet against the entire LLM paradigm. AMI's CEO Alexandre LeBrun predicts 'world models' will be the next buzzword - and with this war chest, they might just make that true.

Test your understanding

Quiz

1 / 10