Apple's privacy-first Siri, NVIDIA's 4-bit training breakthrough, and graduates boo AI at commencement

Picture a room full of people who just spent four years and tens of thousands of dollars on an education - and the keynote speaker tells them AI is going to reshape everything they just trained for. The boos that followed Eric Schmidt's commencement address yesterday capture something real: the gap between how the tech industry talks about AI and how everyone else is starting to feel about it. Today we've got that story, plus a genuine technical leap from NVIDIA, a privacy play from Apple, and a slow-motion infrastructure crisis in the UK.

NVIDIA's new NVFP4 pretraining methodology was validated on a hybrid model architecture. What two model types does that hybrid combine?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

Apple is taking a different approach to the AI assistant wars, and privacy is the weapon. According to Bloomberg's Mark Gurman, the revamped Siri arriving in iOS 27 will let users auto-delete chat histories - with options to keep conversations for 30 days, one year, or indefinitely.

This is a direct contrast to how competitors handle things. Most AI assistants retain conversation history to improve personalization - Apple is flipping that trade-off and banking on the fact that users increasingly want control over what sticks around.

Apple has been openly criticized for falling behind on AI features compared to Google, OpenAI, and Anthropic. The auto-delete feature looks like a strategic pivot - rather than competing on raw capability (where Apple is behind), it's competing on trust and data control. For enterprise users and privacy-conscious consumers, that's a legitimate differentiator. Whether it's enough to close the gap on features is another question.

Read the full story →

NVIDIA has published details on a new 4-bit pretraining methodology built around its NVFP4 microscaling format, and the scale of the validation is what makes this notable. They trained a 12B parameter hybrid model on 10 trillion tokens - officially the longest publicly documented 4-bit pretraining run ever attempted.

The technical approach combines several components: selective BF16 layers for stability, 16x16 Random Hadamard Transforms applied to weight gradient inputs, 2D weight scaling, and stochastic rounding on gradients. The result is downstream accuracy that closely matches full-precision training - but at a fraction of the compute cost.

Here's the plain-English version: training AI models requires enormous amounts of numerical computation. Reducing the precision of those numbers from 16-bit to 4-bit means you can fit roughly 4x more computation into the same hardware. The challenge has always been that lower precision introduces errors that compound over training. NVIDIA's methodology appears to solve that problem at scale - 10 trillion tokens is not a toy experiment.

For labs and developers trying to train large models without NVIDIA's own budget, this is significant. Lower-precision training that actually works means smaller organizations can potentially run longer training runs on less hardware. Follow AI infrastructure developments like this closely - they tend to quietly reshape what's possible.

Read the full story →

More than 100 new UK data centers plan to burn gas to generate electricity - and some may do so permanently. The driver isn't negligence; it's the National Grid. Operators face a years-long wait to connect to the grid, so they're applying for gas connections as a workaround.

The numbers are stark: requests for gas connections by data center operators amount to more than 15 terawatt hours per year. British officials acknowledge this directly threatens the UK's climate targets, with one describing it as raising an "interesting question." That's a remarkably understated response to a significant problem.

This story is part of a broader pattern we've been tracking. The AI infrastructure buildout is colliding with energy systems that weren't designed for this level of demand. In the US, we covered xAI's gas turbine controversy recently. The UK version of this problem is more systemic - it's not one rogue operator, it's the default solution for over 100 facilities. If governments want to hit climate targets while building AI infrastructure, they need grid expansion to move faster than it currently does.

Read the full story →

Former Google CEO Eric Schmidt delivered a commencement address at the University of Arizona last Friday, and when the speech turned to AI, the graduating class responded with boos. Repeated ones. Schmidt acknowledged the anxiety in the room - students about to enter a job market that is visibly being reshaped by automation aren't exactly primed for AI optimism.

The Verge notes this is part of a broader trend: AI has become a genuinely fraught topic at commencement speeches in 2026. Tech leaders who built careers championing disruption are now facing audiences who will live with the consequences of it. The reception Schmidt received wasn't a fluke - it reflects something that polling has been showing for months.

This connects to something worth watching: the public trust gap around AI isn't just about abstract policy concerns anymore. It's personal. Graduates who spent four years studying are looking at AI tools that can approximate parts of their training, and they're skeptical of being told to embrace it. The industry would be wise to take that reaction seriously rather than treating it as a communications problem to solve.

Read the full story →

Vercel Labs released something genuinely interesting this week: Zero, an experimental systems programming language designed specifically so AI agents can read, repair, and ship native programs without needing a human to interpret compiler errors.

The core idea is clever. Standard programming languages produce error messages optimized for human developers. Zero instead emits JSON diagnostics with stable error codes and typed repair metadata - structured output that an AI agent can parse and act on directly. It also enforces capability-based I/O at compile time and compiles to native binaries under 10 KiB.

Most AI coding tools today work by bolting agents onto existing languages that were never designed with agent workflows in mind. Vercel is proposing something more fundamental: rethinking the language itself so the agent is a first-class participant in the development loop. It's experimental, but it's a window into where AI coding tools may be heading. If you're a developer building agent-powered workflows, this is worth watching closely.

Speaking of building fast - if you need to spin up a website for a project or client without touching code, 60sec.site is an AI website builder that turns a prompt into a live site in under a minute. Worth a look if you want to move quickly.

Read the full story →

The answer is B - Mamba and Transformer! NVIDIA's NVFP4 methodology was validated on a 12B hybrid Mamba-Transformer model trained on 10 trillion tokens. Mamba is a state-space model architecture that handles long sequences more efficiently than standard Transformers, making the hybrid design particularly interesting for long-context tasks.

Today's grad boo story got me thinking - how do the people around you actually feel about AI? Are your friends, family, or colleagues excited, skeptical, or just tired of hearing about it? Hit reply and tell me what you're seeing on the ground. I read every response and these answers genuinely shape what we cover.

That's it for today - see you tomorrow with more from the rapidly evolving world of AI. If you found this useful, share it with someone who'd appreciate a no-nonsense daily briefing. And for more coverage, head to dailyinference.com.