wow. just saw The Economic Times newspaper published an article about me 😃 definitely feels so unreal that Sundar Pichai and Jeff Bezos follows me here. @X is truly a miracle. Forever thankful to all of my followers 🙏🫡

Total token volume by model author (Nov 2024–Nov 2025)

https://t.co/hsxNfKRXIU

Anthropic’s new research - Selective GradienT Masking (SGTM) trains an LLM so that high risk knowledge (e.g. about dangerous weapons) gets packed into a small set of weights that can later be deleted with minimal damage to the rest of the model. Most control work today leans on

New paper from top Chinese and US university finds a complete and grand new way to handle token positions in transformers. Introduces GRAPE (Group Representational Position Encoding). GRAPE exactly re-expresses RoPE, ALiBi, and the Forgetting Transformer as special cases of one

Nvidia's GB300 chips fit GB200 racks directly (just plug-and-play), smoothing deployment. And firms using GB300s with vertical integration will scale fast and will win on token cost. ~ Gavin Baker, Managing Partner & CIO of Atreides Management https://t.co/fZEZT5lUEf

⚙️ The Pentagon launched GenAI .mil, its own “bespoke” AI platform and Google’s Gemini is the first model on it. Its an internal platform that brings frontier models to government users at scale. Google says Gemini will be used for tasks such as summarizing policy handbooks,

🔓 OpenAI, Anthropic, and Block formed the Agentic AI Foundation to set open standards so agents built by different teams can work together, with the group hosted by the Linux Foundation. They donated MCP for connecting tools and data between agents from Anthropic, AGENTS. md

Weekly token volume by model type.

Inspirational data point. "Science and medicine doubled life expectancy over the last century."

A great podcast to listen to. Patrick OShaughnessy (@patrick_oshag) with Gavin Baker (@GavinSBaker), Managing Partner & CIO of Atreides Management 📌 Some takeawys - Current AI progress relies on scaling laws, where feeding models more data and computing power reliably makes

"In 3 years, on a bigger phone, you'll be able to run a pruned-down version of Gemini 5, Grok 4, or ChatGPT. ... edge AI is by far the most plausible and scariest bear case." ~ Gavin Baker (@GavinSBaker), Managing Partner & CIO of Atreides Management https://t.co/rypOXM511m

Crrent ChatGPT Enterprise adoption is still heaviest in professional services, finance, and technology.

A core part of Google's Antigravity is self-reporting through Artifacts—As it completes tasks, it will produce task lists, plans, screenshots, and browser recordings that are intended to verify both the work it’s done and what it will do. https://t.co/FY6GvpdIg8

OpenAI is in 'code red' mode for a reason. Google's Gemini now counts around 650mon users, up nearly 50% from 450mn in July. This is not even counting the roughly 2B people who use Gemini indirectly through other Google services like Search. https://t.co/Nqsy1yJxmY

This paper introduces Omega, a system that makes cloud AI agents trustworthy even when the cloud itself is not. The authors start from the fact that real agents today are programs that call language models and tools on a shared cloud, depend on components from many companies,

New Stanford+Linkedin AI paper shows that training with too many step by step examples can quietly hurt reasoning. Chain of thought means the prompt shows short worked solutions before the answer. The authors ask what happens when models train on many explanations but later

The paper measures concrete mistakes in published AI papers and finds they are common and rising over time. At NeurIPS, objective mistakes per paper rise about 55% between 2021 and 2025. Small slips in formulas, proofs, tables, or claims can pass review and later mislead follow

This is so cool for creators who do portraits / products / ads. Single JPEG into a full commercial ad is so possible now.

ImagineArt just launched its Mobile App. Video generation work will be much more smoother now.

🇨🇳🇺🇸 China plans to tightly ration Nvidia’s H200 AI chips at home even after Trump approved exports with a 25% fee, so access remains heavily political. Regulators in Beijing are weighing rules where companies must prove Chinese accelerators are not enough before importing