Anthropic Warns of Rapid Recursive Self‑Improvement as ChatGPT Adds Dreaming Memory; Bots Surpass Human Web Traffic and Big Tech Unveils New Models

AI Tech News Today — Quick Roundup
Date: 2026-06-05T11:30:53.000Z
Hosts: Aurora and Isabelle

Top stories — a fast roundup of the biggest moves in AI, tech, and agents

1. Anthropic: recursive self-improvement is arriving faster than expected

   • Internal data shows Claude is accelerating AI development itself.
   • Engineers are shipping 8× more code per quarter.
   • Success rates on open-ended coding jumped by 50 points in 6 months.
   • Newer models can speed up their own training code by roughly 52×.
   • Anthropic has opened a dedicated institute to study implications for safety, governance, and the pace of innovation.

2. OpenAI upgrades ChatGPT memory with “dreaming”

   • New memory system called dreaming tracks what matters across conversations.
   • Understands whether events are upcoming, happening, or over.
   • Lets users review and steer what the model remembers.
   • Rolling out today for Plus and Pro users in the United States, with 2× more memory capacity.
   • Think of it as giving your virtual assistant better short-term and long-term memory — without turning it into your diary.

3. Cloudflare: bots now exceed human internet traffic

   • Radar data from Matthew Prince shows agentic and crawler traffic has officially passed human traffic on the web.
   • Highlights how much of the internet is now driven by automated agents (search crawlers, task bots, etc.).
   • Raises important questions about measurement, moderation, and network design.

4. Microsoft launches seven new MAI models

   • Headline model: MAI-Thinking-1, a 35 billion active-parameter Mixture-of-Experts reasoning model.
     • Scores 97% on AIME 2025 and 53% on SWE-Bench Pro.
   • Microsoft says these models run on custom MAIA 200 silicon and deliver about 30% better performance per dollar vs. competing hardware.
   • Family also includes image-editing models and Code-One-Flash (5B params) optimized for coding tools.

5. Google DeepMind releases Gemma 4 (12B parameters)

   • Demis Hassabis unveiled Gemma 4, an open multimodal model with 12 billion parameters.
   • Can run locally on laptops with as little as 16 GB video memory.
   • Has exceeded 150 million cumulative downloads and ships under the Apache 2.0 license, making it easy for researchers and builders to run locally and iterate.

Closing
Thanks for listening. Stay curious and stay critical — AI is moving fast, and we’ll keep you caught up.
This is Aurora and Isabelle signing off from AI Tech News Today.