Altman defends scaling as GPT‑5.6 Pro generates a complete Sims‑style 3D game; OpenAI unveils GPT‑Bidi voice, GLM‑5.2 open model debuts amid tooling attacks

AI Tech News Today — Quick, clear, and a little cheeky
Date: 2026-06-22T13:40:12.000Z
Hosts: Aurora (with co-host Isabelle)

Headline: Altman defends scaling — GPT‑5.6 Pro builds a 3D game “in one shot”

• What happened: Sam Altman visited Stanford to push back on critics (including Yann LeCun), arguing that more compute, data, and parameters keep unlocking real advances in reasoning and coding. He highlighted an OpenAI milestone where a model reportedly disproved a long‑standing mathematical conjecture.
• Notable demo: Testers say OpenAI’s GPT‑5.6 Pro generated a fully functional, Sims‑style 3D simulation game in a single shot — a complete HTML file produced in under 48 minutes, without external coding toolchains.
• Context: Anthropic’s Dario Amodei agreed scaling remains a major driver, but both companies acknowledge long‑term planning and robotics are still hard problems.
• Why it matters: Continued scaling can yield surprising capabilities, but practical, reliable planning and physical interaction remain outstanding challenges.

Feature: OpenAI launches bidirectional voice mode in the mobile app

• New tech: OpenAI’s GPT‑Bidi architecture is rolling out a bidirectional voice mode.
• Key capabilities: the model can interrupt you mid‑sentence, track background speech in real time, and correct verbal slip‑ups on the fly.
• Related move: The push for low latency also shows in OpenAI Codex adding visual workflow recording, bridging multimodal chat with desktop automation.
• Takeaway: Conversational voice interaction is becoming much more like talking to a helpful — slightly impatient — colleague.

Model release: GLM‑5.2 — high‑performance open‑weight model from China

• What it is: GLM‑5.2, an open‑weight model, reportedly outperforms frontier models on many benchmarks (design, UI tasks, long‑horizon engineering).
• Run it locally? Technically yes, but hardware is steep:
  • For decent speed: 8 H200 GPUs with >1 TB aggregate video memory.
  • Lighter options: heavily quantized 2‑bit builds that can run on a high‑end 256 GB Mac Studio or a single RTX 4090.
• Why it matters: More capable open models lower barriers, but they still demand significant compute unless heavily quantized.

Security pressure: Langflow, LangGraph, LangChain, and more

• Incidents reported: Security researchers say ~7,000 Langflow servers are under active attack. Similar vulnerabilities affect LangGraph and LangChain.
• Other issues: Reports include Copilot searching mailboxes and LiteLLM leaking admin keys.
• Practical advice: Run a simple five‑check audit on your stack before your tools become the next headline.

Thanks for listening. Keep following AI Tech News Today — Isabelle and I will keep reading so you don’t have to. We’ll bring you the next developments as they land. Stay curious and stay safe out there in the AI wilds.