Google DeepMind Unveils Genie 4: Interactive Worlds You Can Steer

What auditors look for in security training

In 40 minutes, learn what training evidence auditors expect for ISO 27001, SOC 2, and HIPAA. Join our live webinar on Oct 29 at 1 PM ET.

Google DeepMind Unveils Genie 4: Interactive Worlds You Can Steer

Genie 4 is DeepMind's newest real-time world model. It generates music, ambient sound and dialogue alongside the scene, simulates physics and more natural human motion, and lets you change what's happening in the world with plain-language instructions. Gemini 3.8 Flash runs inside the loop as a guide, breaking goals into steps and checking each frame as the world unfolds. It's launching as a research preview for trusted testers, with access planned for Google AI Ultra subscribers through Project Genie and for early-access partners on the Gemini API and AI Studio.

OpenAI Reportedly Plans to Buy Google's AI Chips: OpenAI is reportedly preparing to purchase Google's Tensor Processing Units outright, adding them to a compute stack that still leans heavily on Nvidia GPUs. Neither company has confirmed a deal, and chip volumes and terms haven't been disclosed. OpenAI already rents TPUs through Google Cloud, so owning them would be a much deeper bet on not depending on a single chip supplier. It lands the same week Oracle said its cloud will host Google TPUs, OpenAI's Jalapeño inference chip and Cerebras systems side by side.

Cloudflare Discloses a Cross-Customer Data Leak in Workers AI: A configuration mismatch in Cloudflare's GLM-5.3 Workers AI service triggered a bug in the SGLang inference engine, and some long requests received cached data that belonged to other customers. Cloudflare counted 155 affected responses between Aug 27 and Sep 9, says it found no evidence of deliberate exploitation, and has patched both its configuration and the engine. If you run shared inference, treat KV-cache isolation as a security boundary, not just a performance setting.

The Netherlands Plans to Cut 25,000 Civil Service Jobs With AI: The Dutch cabinet wants to eliminate about 25,000 government positions, roughly 15% of a 160,000-person workforce, through an AI-led efficiency drive. The targets are €900M in yearly savings by 2027 and €3.5B by 2035. Police and defense are exempt, but some ministries face cuts of 45–60%. It's one of the first times a national government has attached a hard headcount number to AI adoption.

AI NEWS HIGHLIGHT

• Anthropic Expands GSA Deal: 40% Off Claude for Federal, State & Local Government — Covers Claude Enterprise, Claude for Government and the API; Anthropic estimates up to $1B in savings; ~500 US government organizations already use Claude.

• Trump Says the US Will Drop "AI" for "SI" (Super Intelligence) — Truth Social post targets government documents; a Sept 29 executive order and an Oct 20 State Department cable already push the new label.

• Unreleased Claude Built Working Exploit Chains Across Five Platforms — Anthropic's Frontier Red Team reports a zero-click Linux root chain, a Chrome sandbox escape, plus Safari, Pixel 9 and Windows bugs; all reported to vendors first.

• Anthropic Buys One-Person Biotech Mithril, Launches Claude Drug-Discovery Push — New partners include Bristol Myers Squibb, Generate Biomedicines and BigHat; one muscular dystrophy program could reach human trials by late 2028.

• Warren Gives Anthropic Until Nov 3 to Answer IPO Risk Questions — Four-page letter to Dario Amodei on the S-1's existential-risk language, Amazon/Broadcom concentration and valuation.

• SpaceX Lands $40B GPU Financing From Apollo — An Apollo-managed entity buys the hardware and leases it to SpaceX; deliveries, including Nvidia Vera Rubin, start in H1 2027.

• Meta Prices $37.5B Stock Raise to Fund Its AI Buildout — $25B in common stock (upsized from $22.5B) plus $12.5B in mandatory convertible preferred.

• Microsoft Offers Voluntary Buyouts to ~60,000 Employees — More than a quarter of its workforce is eligible as AI capex heads toward $175B.

• SK Hynix: Memory Shortage Will Last Until 2030 — 14 major buyers have already locked up more than half of its 2027–28 output.

• Moore Threads Revenue Jumps 1,774% — The Chinese GPU maker posts its first quarterly profit on ~$240M revenue as domestic buyers move off Nvidia.

Your AI Is Quietly Making Your Judgment Calls

A new Anthropic study gave 50 professional analysts a causal-inference task with Claude Opus 4.8 as their assistant. The catch: the agent was set to nudge four key analytical choices in random directions.

The analysts went with the AI's lean 86% of the time. They reused the agent's own arguments to justify 77% of their decisions. And they credited the AI as the source only 15% of the time.

The direction was random, and people still followed it, then defended it as their own reasoning. This isn't about AI getting the math wrong. It's about which data to drop, which method to use, which caveat matters: the calls your name ends up on.

The fix isn't using AI less. It's making the AI show you the forks before it picks a path.

Run this today (paste into your next analysis, report or plan):

1. Before you recommend anything, list every judgment call in this task: what data to include or exclude, which method to use, which assumptions and thresholds to set. For each one, give the strongest case for at least two options. Do not pick one yet.

2. Look back at the analysis we just did. List each decision where you steered me toward an option. For each, argue the opposite as strongly as you can and tell me how the conclusion would change if I'd chosen it.

3. Write a short decision log for this work: which choices came from me, which came from you, and which I accepted without checking.

AI agents just machine-checked a 15,000-page proof. Axiom says agent swarms formalized the Classification of Finite Simple Groups in Lean in 24 days: 43.1 million lines, with 34 deviations from the published proofs found and patched along the way.

TOP TRENDING AI TOOLS

• Liquid AI LFM3 — Six small models for text, vision, audio, multimodal and decision tasks; built to run on phones and laptops and retrainable on a single GPU. Companies above $10M revenue need a commercial license.

• Apple Pages, Keynote & Numbers Assistant — Drafts and revises documents, builds slides and analyzes spreadsheet data; macOS 27 adds MCP so third-party agents can connect.

• Oracle AI Workspace in ChatGPT — Employees can pull approved Oracle data and run business workflows without leaving ChatGPT.

• Giotto Alto & Solo — Swiss sovereign reasoning models: Alto runs on eight B300s and is free to chat with; Solo fits on a single GPU for on-prem use.

• Mistral Timestral — Four open-weight time-series forecasters from 5M to 209M parameters; the two largest top GIFT-Eval, TIME and fev-bench.

• Opera WebAI — Early browser standard for running AI models locally on your computer; previewing first in Opera One.

• Ant Group LingGuang 2.0 — Assistant with persistent memory that keeps working after you close the chat, like watching travel prices or flagging subscriptions.

• ByteDance Jimeng Short-Drama Studio — Takes a project from script to characters, storyboards and finished video in one pipeline.

SPONSOR US

Get your business in front of over 90k+ AI professionals

8020AI is the world’s #1 AI Newsletter, Read by 90k+ professionals from leading companies such as Google, OpenAI, Meta, and Microsoft.

We've assisted in promoting Over 500 AI-Related Products. Will yours be the next?

What We Can Offer:

• Launch an Advertising Campaign

• Introduce New Product or Features

• Other Business Cooperation

Get in touch today.

Or Email our founder Alamin at [email protected]

FEEDBACK

How was your experience with 8020AI today?

Login or Subscribe to participate in polls.

If you have specific feedback or anything interesting you’d like to share, please let us know by replying to this email.