Anthropic files IPO with $895B committed.

Oracle announced plans to raise $20 billion to fund its AI infrastructure expansion, splitting the total evenly between $10 billion in stock and $10 billion in bonds. The market immediately balked at the capital requirements, sending Oracle shares down nearly 6%.

This marks a shift in investor patience. For the last few years, announcing a massive capital expenditure for AI was a guaranteed stock bump. Now, the bill is coming due, and shareholders are getting nervous about the endless appetite for server farms and power contracts.

Oracle is trapped in a classic arms race. They have to spend the $20 billion to remain relevant as a cloud provider for AI workloads, even if it tanks their stock in the short term. The cloud giants are playing a game of chicken with their own balance sheets, and the only guaranteed winner is the supply chain selling them the shovels.

Researchers at MIT CSAIL managed to significantly improve a small model's math benchmark score from 7.9% to 48.1%. The catch? They did it by editing just 0.016% of the weights in GPT-6 Astra, entirely bypassing the need for a full retraining run. The paper is not yet peer-reviewed.

This is a fundamental shift in how we handle model performance. Right now, fixing a foundational flaw usually means throwing millions of dollars of compute at a new training run. If you can achieve a 40-point benchmark jump by tweaking a fraction of a percent of the network, the economics of model maintenance change overnight.

If this technique scales, the labs that figure out weight-space programming will run circles around those still relying on brute force. It turns model debugging from a sledgehammer into a scalpel.

Microsoft filed a patent for an AI system designed to rewrite what a speaker says during a live video feed. The technology alters the audio and simultaneously reshapes the speaker's lips and face so the edit appears natural. The patent suggests use cases ranging from live translation and filler word removal to outright censorship.

This goes far beyond standard deepfakes. By embedding this capability directly into live broadcast pipelines, Microsoft is opening the door to a world where you can never be entirely sure if the words coming out of a live speaker's mouth are their own, or an algorithmic correction applied mid-stream.

For enterprise communications, this will be sold as a polish tool. For the rest of us, it means the end of trusting live video. The companies that build cryptographic verification for live streams just got a massive validation of their business model.

Ali Ghodsi, co-founder and CEO of Databricks, has publicly stated that custom model training holds little value. According to reports, Databricks is now pivoting its focus toward serving existing models rather than pushing clients to build their own from scratch.

This is a stunning reversal. Just a year ago, the enterprise narrative was that every Fortune 500 company needed its own bespoke foundation model to protect its data moat. Now, one of the primary beneficiaries of that hype cycle is admitting that the economics and outcomes just do not justify the effort.

The era of every brand training a custom LLM is dead. The winners are the inference providers and the closed-source giants; the losers are the consultancies that billed millions for custom training runs that ultimately underperformed standard off-the-shelf APIs.

Privacy-focused company Proton canceled its plans to re-interview nearly all of its staff regarding their AI usage. The reversal follows intense backlash from both employees and users, sparked by reports that CEO Andy Yen was pushing AI adoption despite internal security concerns.

When a company built entirely on the premise of privacy and security tries to force-feed AI tools to its workforce, the cultural rejection is immediate. Engineers are acutely aware of the data leakage risks associated with enterprise AI adoption, and forcing them to justify their AI usage in re-interviews was a heavy-handed misstep.

Proton leadership lost this round. It is a clear warning to management teams everywhere: you cannot mandate AI enthusiasm, especially when your core product is built on paranoia.