ChatGPT's Memory Evolves into "Profiles", NVIDIA Nemotron 3 Ultra Arrives on SageMaker, Meta Brings an AI Creator Assistant to Facebook

ChatGPT's Memory Evolves with "Dreaming" — From Scattered Fragments to User Profiles

OpenAI has significantly overhauled ChatGPT's memory system. The new "Dreaming" feature doesn't just save information gathered from conversations as fragmentary notes — it automatically builds "narrative dossiers" (story-like profiles) organized by category, such as work, hobbies, and travel.

According to The Decoder, the previous memory feature only accumulated bullet-pointed facts, whereas Dreaming systematically builds an understanding of the user, aiming to improve response quality in context. The success rate for keeping information up to date has reportedly improved from 52.2% last year to 75.1%. It's a step forward in how accurately AI assistants "understand the user," but how detailed these personal profiles become is something to watch from a privacy perspective as well.

NVIDIA Nemotron 3 Ultra Lands on SageMaker JumpStart — An Agent-Focused 550B MoE Model

NVIDIA has made Nemotron 3 Ultra, built on a hybrid Transformer-Mamba Mixture-of-Experts (MoE) architecture, available on Amazon SageMaker JumpStart. The model, with 550B total parameters and 55B active parameters, is optimized for agentic workflows.

Its biggest highlight: it supports a 1 million token context length while delivering 5x faster inference and up to 30% cost savings. Thanks to NVFP4 format optimizations, it reportedly maintains high throughput even through long chains of tool calls and agent runs spanning hundreds of turns. It's expected to see adoption in enterprise use cases requiring sustained multi-step reasoning, such as agent orchestration, coding agents, and deep research.

Meta Rolls Out an AI Creator Assistant on Facebook

Meta has begun a full-scale rollout of an AI assistant for creators on Facebook. According to TechCrunch, creators previously had to analyze dashboards and charts themselves, but with the new AI assistant they can ask questions in natural language — "When should I post?" or "What are people saying in my comments?" — and get answers instantly.

Creator support is an area where every platform is racing to outdo the others, but Meta delivering an AI assistant built directly into Facebook itself signals a new phase of AI integration on social platforms.

Sam Altman Acknowledges AI Token Costs as an "Issue"

OpenAI CEO Sam Altman has publicly admitted for the first time that AI token costs are "becoming an issue." According to Tom's Hardware, Altman said the company is exploring ways to deliver better value.

OpenAI has kept lowering API prices over time, but the growing compute costs that come with ever more capable models have become a challenge that API customers can no longer ignore. Especially as agentic, repetitive usage increases, improving cost efficiency is a challenge for the entire industry.

Anthropic Explores the Possibility of Recursive Self-Improvement in "When AI Builds Itself"

Anthropic's research division has published an article titled "When AI Builds Itself," examining the current state and future outlook of recursive self-improvement in AI. It has also drawn attention on Hacker News, where people are debating what possibilities and risks emerge when AI takes part in its own design and training processes.

The controllability of self-improvement loops and how to guarantee safety are core themes in AI alignment research, which makes Anthropic officially joining this discussion highly significant.

AI-Driven Layoffs Accelerate in the US — 38,242 in May Alone, with AI Cited as the Top Reason

It has emerged that 38,242 people were laid off in the US technology sector in May, and the single biggest reason cited was "AI." The data, which made the rounds on Hacker News, throws into sharp relief the reality that pressure to invest in AI is hitting employees directly through payroll cuts.

Evaluating the causal relationship between corporate "AI efficiency" narratives and actual layoffs is a thorny problem, but the mere fact that "AI" is increasingly being explicitly named as a reason for layoffs is itself becoming an important indicator for measuring the impact on the labor market.

---