In Today’s Issue:

🧨 OpenAI kills its highly anticipated Sora video generator

🍏 Apple is finally tearing Siri down to the studs

☢️ Microsoft and Nvidia are teaming up to slash nuclear power plant

🧠 Google researchers just solved a massive AI bottleneck

✨ And more AI goodness…

Dear Readers,

Sam Altman just mass-executed one of OpenAI's most hyped products, Sora is dead, compute is being funneled into a mysterious new model called "Spud," and Disney's billion-dollar deal evaporated overnight. Today's issue unpacks that seismic strategic pivot and what it signals about the AI arms race heading into IPO season.

Meanwhile, Apple is finally ripping Siri apart and rebuilding it as a full-blown chatbot with threaded conversations and document analysis, Microsoft and Nvidia are teaming up to fast-track nuclear power plants with AI (yes, cutting permitting time by 92%), and Google just figured out how to shrink AI memory usage by 6x without losing a single drop of accuracy. We've also got Eric Schmidt warning about singularity timelines, a hallucination benchmark where Grok comes out on top, and a quote from an OpenAI collaborator who thinks this year's AI progress will shock people, let's get into it.

All the best,

Kim Isenberg

OpenAI announced the shutdown of the Sora app, thanking creators while hinting at upcoming timelines for data preservation and API changes. The move is reportedly driven by a strategic reallocation of compute power toward training a next-gen model (“Spud”, more on that later), reflecting intensifying competition and prioritization of core AI advancements.

Resources are being funneled into more powerful models, potentially paving the way for Sora-like capabilities to return inside a larger “super app” ecosystem rather than as a standalone product.

Prof. Unutmaz who also regularly works with OpenAI now expects AI advances to accelerate this year by a lot.

Eric Schmidt: Singularity's Arrival, 92-Gigawatt Problem & Recursive Self-Improvement Timeline

Sam Altman just made the boldest bet of his career, and it starts with killing one of OpenAI's most hyped products. The company is shutting down its Sora video app and API entirely to free up computing power for its next major AI model, codenamed "Spud." Altman told staff that pretraining is already complete and a "very strong model" could arrive within weeks, one he believes "can really accelerate the economy."

But the fallout is real. Disney walked away from a planned $1 billion investment tied to a Sora licensing deal struck last December. The deal never closed, and now it won't. Altman is also stepping back from directly overseeing safety and security teams to focus on infrastructure, datacenters, chips, supply chains. Safety moves under CRO Mark Chen; security shifts to Greg Brockman's scaling org.

The message is unmistakable: OpenAI is consolidating around its core AI models and enterprise tools, cutting everything else. Even Fidji Simo's product org gets rebranded to "AGI Deployment." This is a company going all-in on the race against Anthropic and Google, with IPO pressure mounting on both sides.

Why it matters: OpenAI is making its most aggressive strategic pivot yet, sacrificing a flagship consumer product and a billion-dollar partnership to concentrate firepower on its next-gen model. How Spud performs will shape the entire competitive landscape heading into what could be the biggest AI IPO year in history.

Sources:
🔗 https://www.theinformation.com/articles/openai-ceo-shifts-responsibilities-preps-spud-ai-model

They all run their businesses on beehiiv. Newsletters, websites, digital products, and more. beehiiv is the only platform you need to take your content business to the next level.

🚨Limited time offer: Get 30% off your first 3 months on beehiiv. Just use code JOIN30 at checkout.

Start building for 30% off today.

Google just solved one of AI's most annoying bottlenecks, and the solution is almost absurdly elegant. TurboQuant, a new compression algorithm presented at ICLR 2026, shrinks the memory footprint of large language models by at least 6x without losing any accuracy. Think of it like compressing a massive photo album into a tiny thumb drive, except every single picture still looks pixel-perfect.

When AI models process long texts, they store a "cheat sheet" of previously seen information called the key-value cache. As conversations get longer, that cache balloons, eating up expensive GPU memory and slowing everything down. Traditional compression helps, but always sneaks in hidden overhead costs that partially defeat the purpose.

TurboQuant takes a radically different approach using two clever tricks. PolarQuant converts data into polar coordinates - swapping clunky grid-style directions for elegant angle-and-distance descriptions - which eliminates overhead entirely. Then QJL mops up tiny leftover errors using just a single bit. The result? In 4-bit mode, attention computation on NVIDIA H100 GPUs speeds up by as much as 8x compared to the standard 32-bit baseline. No fine-tuning required. No accuracy sacrificed.

The research was led by Google Research's Amir Zandieh and Vice President Vahab Mirrokni, in collaboration with KAIST and NYU. Google already sees this powering Gemini's memory management and supercharging semantic search at massive scale.

This guide distills 10 AI strategies from industry leaders that are transforming marketing.

Outpace your competitors by mastering AI.

Get Your Free Guide