ChatGPT’s Smarter Memory, Self-Improving AI, and Apple’s Siri Reboot

FEATURED INSIGHT💡

ChatGPT’s Memory Learns to Keep Up

OpenAI is rolling out a new memory architecture designed to make ChatGPT better at carrying useful context across conversations without relying only on explicit “remember this” instructions. The system, called dreaming, synthesizes information from past chats, identifies relevant preferences and constraints, and updates memories as circumstances change.

OpenAI reports major gains over earlier versions. Factual recall improved from 41.5% in 2024 to 82.8% in 2026, preference adherence rose from 31.4% to 71.3%, and performance on time-sensitive memories increased from 9.4% to 75.1%. Users can also review a memory summary, correct outdated information, and tell ChatGPT which topics should or should not influence future responses.

Why it matters: More capable memory turns ChatGPT into a better long-term assistant for ongoing projects, recurring decisions, and personalized workflows. For businesses, it also raises important questions around user control, data governance, and how persistent context should be managed across customers and teams.

Attio is the AI CRM for high-growth teams.

Connect your email, calls, product data and more, and Attio instantly builds your CRM with enriched data and complete context. Whether you’re running product-led growth or enterprise sales, Attio adapts to your unique GTM motion.

Then Ask Attio to plan your next move.

Run deep web research on prospects. Update your pipeline as you work. Find customers and draft outreach emails. Powered by Universal Context, Attio's intelligence layer, Attio searches, updates, and creates across your data to accelerate your workflow.

Ask more from your CRM.

Ask Attio

ON THE HORIZON 🌅

When AI Starts Helping Build Its Successor

Anthropic says Claude is now performing a growing share of the engineering and research work used to improve Claude itself. As of May 2026, the company reports that more than 80% of the code merged into its codebase was authored by Claude, while engineers were merging roughly eight times as much code per day as they did in 2024.

The shift extends beyond coding. Anthropic says Claude can optimize experiments, investigate technical failures, and run multi-agent research workflows with limited human intervention. In one internal task, Claude improved model-training code by roughly 52 times, while another group of agents independently proposed and tested experiments around weak-to-strong supervision.

Humans still choose the broader goals, judge which findings matter, and decide what should be built next. But as execution becomes faster and increasingly automated, the bottleneck moves toward research judgment, review, safety, and verification. If AI systems eventually become capable of choosing valuable research directions as well as executing them, recursive self-improvement could move from theory toward a working development loop.

LATEST IMPORTANT NEWS 📰

Apple Finally Gives Siri Its AI Overhaul

Apple unveiled Siri AI at WWDC 2026, with a beta planned for later this year. The rebuilt assistant can hold longer conversations, use current world knowledge, interpret information displayed on the screen, and draw from emails, calendars, contacts, and other device data when completing tasks. A new “Write with Siri” feature can adapt drafts to the way a user normally communicates with a particular person, while integrations across Dynamic Island, Spotlight, macOS, and watchOS position Siri as a system-wide AI interface rather than a basic voice-command tool.

US Directive Forces Anthropic Models Offline

Anthropic says a US government export-control directive forced it to disable access to its Fable 5 and Mythos 5 models for every customer because foreign nationals could no longer be allowed to access them, including Anthropic employees. The government reportedly cited a potential cybersecurity jailbreak, while Anthropic argues that the demonstrated technique was narrow, revealed only previously known vulnerabilities, and provided capabilities already available through other models. The company is complying while challenging the technical basis and transparency of the decision, offering an early example of how government action can abruptly turn model availability into a platform risk for developers and enterprises.

Amazon Considers Selling Trainium Chips Directly

AWS is exploring whether to sell Trainium AI chip racks to outside data-center operators, taking its custom silicon business beyond the AWS cloud and into more direct competition with Nvidia. Amazon CEO Andy Jassy has estimated that the chip operation could represent a roughly $50 billion annual business if treated as a standalone supplier. Demand may be the immediate constraint: existing Trainium capacity has reportedly sold out quickly, future capacity is already heavily committed, and Amazon would need access to additional manufacturing supply before it could serve both cloud customers and external chip buyers at scale.

One AI employee. Engineering, finance, growth, ops.

Last week Viktor opened 14 pull requests, closed two month-end books, drafted a board update, deployed three landing pages, and triaged 600 support tickets. From inside Slack and Microsoft Teams. 20,000+ teams now run this way.

Get Started for Free

FOR THE TECHNICALLY INCLINED 🛠️

SANA-Sprint Pushes Image Generation Into Real Time

NVIDIA researchers introduced SANA-Sprint, a diffusion model capable of generating a 1024 × 1024 image in approximately 0.1 seconds on an H100 and 0.31 seconds on an RTX 4090. The model reduces generation from 20 inference steps to between one and four while outperforming FLUX-schnell on the reported FID and GenEval benchmarks. Its hybrid distillation method combines continuous-time consistency distillation, which aligns the smaller model with its teacher, with latent adversarial distillation to improve single-step fidelity.

SANA-Sprint also supports ControlNet generation at roughly 0.25 seconds on an H100, enabling users to manipulate structure and receive visual feedback almost immediately. That latency changes the interaction model for generative design. Image creation can begin behaving more like editing, where users adjust prompts, layouts, or controls and see the result as part of a continuous creative workflow.

AI TOOL OF THE DAY 🚀

Lovable lets you build and iterate on websites and software products by describing what you want through a conversational interface.

PRDs by voice. Bug reports by voice. Ship faster.

Dictate acceptance criteria and reproductions inside Cursor or Warp. Wispr Flow auto-tags file names, preserves syntax, and gives you paste-ready text in seconds. 4x faster than typing.

Try Wispr Flow free