AI Learning -- Day 132

1) The EU AI Act's High-Risk Rules Were Delayed to December 2027

Yesterday's coverage of the EU AI Act's high-risk enforcement deserves a correction, and it's a useful one. What actually became enforceable on August 2, 2026 was narrower than "high-risk AI systems": the European Commission's enforcement powers over general-purpose AI (GPAI) model providers, plus Article 50's transparency rules -- chatbots must disclose they're AI, generated content must carry machine-readable markers, deepfakes must be labeled. The heavier machinery -- mandatory risk management systems, technical documentation, conformity assessments, and logged human-oversight checkpoints for "high-risk" uses like hiring tools and credit scoring -- was quietly pushed back. A "Digital Omnibus" simplification package, agreed by the Council and Parliament in May and signed into law as Regulation (EU) 2026/1744 just six days before the original deadline, moved those obligations to December 2, 2027 for most high-risk categories (Annex III) and August 2, 2028 for AI embedded in already-regulated products like medical devices (Annex I).

Why this matters beyond nitpicking a date: if you build or buy AI for hiring, lending, or anything touching critical infrastructure in the EU, you have roughly 16 more months before the full compliance regime bites -- a materially different planning horizon than "already live." It's also a reminder that regulatory reporting, including in this series, needs a direct-source check before treating "enforceable" as settled fact; simplification packages have become a recurring pattern in how the EU walks back its own aggressive timelines under industry pressure.

2) Apple Finally Ships a Real Siri, and Nobody's Impressed

After years of delays, Apple's rebuilt Siri arrived in the iOS 27 public beta on July 14, 2026, and by early August the verdict was in: it's genuinely capable, and almost nobody's excited. The new Siri holds natural back-and-forth conversation, understands personal context pulled from your phone, taps a real world-knowledge model for open-ended questions, and -- for the first time -- has its own dedicated app alongside adjustable voice pacing and expressivity. It requires an iPhone 15 Pro or newer. The anticlimax isn't a knock on the engineering -- reviewers call it a legitimately competent assistant. It's a lesson about timing in a fast-moving field: the capabilities Siri just gained (fluid conversation, contextual memory, general knowledge) were exactly what ChatGPT, Gemini, and Claude normalized over the previous two years. Apple didn't ship a worse product than expected; it shipped roughly the right product about two years late, into a market that had already moved its baseline for "impressive" somewhere else. In AI, being right eventually and being early are different races, and only one of them gets remembered as a launch.

3) VC Funding Hit a Record $510B in H1 2026, and It's Concentrating

Global venture funding reached $510 billion in the first half of 2026 alone, already exceeding the $440 billion deployed across all of 2025. But the headline number hides where it's actually going: mega-rounds like MGX's $49 billion AI-focused fund and Prometheus's $12 billion Series B (at a $41 billion valuation) are pulling an outsized share of that total toward a small set of frontier labs and infrastructure players. Anthropic now sits at a $965 billion valuation after its May Series H; OpenAI follows at $852 billion. Most of the remaining capital is flowing narrowly too -- into compute, chips, and a handful of regulated verticals like healthcare and legal tech.

This is a power-law market, not an evenly rising tide: a "record year for AI funding" can be simultaneously true and mean a harder fundraising environment for most AI startups, because the record is being set by a shrinking number of enormous checks rather than a broadening base of mid-size ones. If you're evaluating an AI vendor's staying power, or your own venture's prospects, the aggregate funding figure tells you less than who specifically is capturing it.

4) Topics Covered So Far -- Day 1-131 Recap

131 days in, the throughline has moved from "what are agents and protocols" toward "what happens when they're actually deployed at scale, and who's watching." Early days built vocabulary -- multi-agent systems, the Model Context Protocol (MCP), and the Agent-to-Agent protocol. Recent weeks tracked the frontier model race (GPT-5.6, Claude Sonnet 5 and Opus 5, Qwen3.8-Max, DeepSeek V4 Flash), the efficiency wave making smaller open models competitive (selective activation sparsity, MoE routing), real security incidents (the Hugging Face breach postmortem, Anthropic's own models used to breach three companies), and the regulatory response trying to keep pace -- including, as today shows, regulators quietly slowing their own timelines under pressure. Today adds two new threads: a correction on where EU compliance deadlines actually stand, and a reminder that consumer AI's "who's first" narrative and "who's actually good" narrative don't always point the same direction.