Google Launches Gemini Decisions

Google launches Gemini 4 Hydrogen for coding and agent workflows, while Antigravity's Orchestrator can switch models within a session. We separate the confirmed launch from Google's performance claims and ask what a complete, checked task really costs. We also examine IBM's experimental LLMON previews, Salesforce's open speech model and AMD's new memory option, plus Pydantic AI, Langfuse and Gemini eyewear preorders. The episode covers November 12 developments, with research limitations, product eligibility and practical business tests kept explicit. plsautomate.com Chapters 00:00 Cold open 00:47 Today on the show 01:16 Google launches Hydrogen 10:38 IBM draws a line in the prompt 14:36 Salesforce listens for the details 18:20 AMD adds a memory choice 21:59 Quick release notes 21:59 Pydantic AI durable work 22:20 Langfuse gateways and checks 22:34 Gemini eyewear preorders 22:55 The work between the models 25:29 One question for your next trial

Google launches Gemini Decisions for scoring bounded choices, with performance claims that must not be confused with guarantees. We examine the difference between a fast route and permission to act. We also cover AlphaGenome 2 research, AWS Emulator and Qwen4 Max Preview, separating research predictions, local tests and temporary access from production assurances. Quick headlines cover ERNIE 5.5 Flash Preview, LangChain, LiteLLM and Browser Use, all within November 11 in Europe/Sofia. plsautomate.com Chapters 00:00 Cold open 00:35 Welcome 00:55 Main story Gemini Decisions 10:30 AlphaGenome Two 15:52 A W S Emulator 20:59 Qwen Four Preview 24:55 Quick headlines 24:59 Ernie Flash Preview 25:13 LangChain Router Beta 25:26 Lite L L M Fallbacks 25:46 Browser Use Fix 26:03 The bigger picture 28:06 Close

Google expands paid access to Gemini 4 Carbon, with introductory pricing that does not by itself establish the cost of a checked result. We distinguish this rollout from the model's earlier announcement and examine how businesses can test it. We also cover DeepMind's scoped Signal security proofs, the Businesses in ChatGPT beta and Australia's proposed frontier AI safeguards. Quick headlines cover Anthropic's proposed IPO roadshow, FastMCP and Gemini CLI, with confirmed releases kept separate from research and proposals. plsautomate.com Chapters 00:00 Cold open 00:45 Intro 01:07 Gemini 4 Carbon Opens Up 11:13 AlphaProof Code Checks Signal 16:27 Businesses Enter ChatGPT 21:06 Australia Proposes AI Safeguards 26:02 Quick headlines 26:02 Anthropic Roadshow 26:24 FastMCP Fixes 26:39 Gemini CLI Update 26:56 From Access to Evidence 28:58 Close

Meta launches Muse Ember through its API for repeated agent tasks, with pricing that rises above 200K prompt tokens. We separate the confirmed launch from selected performance claims and examine the full cost of a checked result. We also cover Google's new TPU system configurations, Claude Code security fixes, Adobe and Disney's expanded alliance, and confirmed Ollama and Goose updates. The episode covers November 9 announcements and practical tests for business adoption. plsautomate.com Chapters 00:00 Cold open 00:42 Intro 01:08 Meta launches Muse Ember 10:51 Google builds around the chip 15:51 Claude Code closes security gaps 20:55 Adobe and Disney expand alliance 25:28 Ollama reduces memory use 25:49 Goose shows context 26:08 What the stories add up to 27:42 Close

The Financial Times reports that Anthropic took over $3.4 billion of gas generator commitments covering 1.7 GW in late September, not operating generation or the supplier itself. We examine AP's reporting on AI safety staff informally reporting to a White House office, Gurman's unconfirmed agentic Siri roadmap, and confirmed Docling and LiteLLM infrastructure updates with stable and preview channels kept distinct. Quick headlines cover Genesis simulation tools, a Happy Agent preview fix and NVIDIA's board investigation statement, without treating an inquiry as a finding. Each story connects to a practical question about delivery, permissions or evidence for a business using AI. plsautomate.com Chapters 00:00 Cold open 00:41 Intro 01:02 The reported power deal 12:03 Who checks the safety checkers 16:24 Siri and delegated goals 20:48 The plumbing agents need 27:01 Quick headlines 27:06 Genesis simulation update 27:30 Happy Agent preview fix 27:49 NVIDIA board statement 28:12 The bigger picture 30:00 Close

OpenAI says one A I run, costing about one point two million dollars, produced working exploits and candidate patches for 948 Linux kernel privilege escalation bugs. We separate what is confirmed from what is company reported, and ask why only a small share are fixed so far. We also cover Texas voters and the data center backlash, a disputed account of a hacker startup and Anthropic, and Claude credited with finding flaws in the Xen hypervisor. Quick headlines include a cheaper coding model, a hardware gadget designed by A I, and a bank's claim of big savings from self hosted models. Every story is labeled confirmed, preliminary or reported, with one practical takeaway for your business. plsautomate.com Chapters 00:00 Cold open 00:50 Intro 01:21 One run, 948 root exploits 10:28 Texas and the data center backlash 16:04 OpenAI, Anthropic and the hackers 19:27 Claude finds flaws in Xen 23:10 Quick headlines 26:53 The bigger picture 29:16 Wrap up

Anthropic told investors in an amended S one filing that its future models may imminently meet its own definition of dramatically accelerated, A I driven progress, while saying no current model has crossed the line. We also cover new Claude Code limits, malware that asks an A I to rank its victims, hundreds of apps hiding abusive deepfake tools, Goldman Sachs pulling coding tools from most staff, and the Meituan LongCat two point five launch. As always, we label what is confirmed, preliminary, reported or unverified, so you know exactly how far to trust each story. plsautomate.com

Marvell is paying about $115 billion for a company that moves data between AI chips using light. Here’s the AI news from November 2, 2026: • Marvell agrees to buy Lumentum, which makes optical interconnects, for about $115B. Investors weren’t convinced and both stocks fell. • Oracle raises $20B in shares and bonds to fund AI data centers, and its stock drops almost 9% • Anthropic’s updated IPO filing shows $16.96B in quarterly revenue and about $895B in AI infrastructure commitments • OpenAI partners with ADNOC and G42 to bring AI into oil and gas operations in the UAE • Palantir revenue grows 96%, and for the first time its US commercial sales beat its US government sales • German Chancellor Merz calls for Europe to control its own AI stack, from chips to models • The European Central Bank warns that AI valuations are a risk to financial stability The takeaway: Europe wants more control over its AI, and your business can have that too. Choose where your data is processed, and keep sensitive information on infrastructure you control. plsautomate.com

Hackers are now using AI to sort through stolen data and decide which victims are worth targeting first. Here’s the AI news from October 31, 2026: • Cisco Talos finds VOLTA, malware that sends stolen passwords, tokens and financial data to an AI model so it can rank the most valuable victims. If one model refuses, it tries another. • Microsoft releases Mage-VL-Cowork, an open 4B model that operates a computer through both the screen and the terminal. It completes 66% of tasks on OSWorld. • OpenAI reportedly cuts its fundraising target from $30B to $20B • Anthropic’s IPO filing shows about $800B in computing commitments, which analysts say would need huge revenue growth to support • About $173M has been spent on US election ads about data centers as local opposition grows • Meta reportedly built a humanoid robot prototype with Chinese startup Fourier The takeaway: AI makes stolen logins more valuable to attackers. Turn on two-factor authentication, use a password manager, and remove access for tools and people you no longer use. plsautomate.com

Google ran an AI model on a TPU orbiting Earth, and it answered. Here’s the AI news from October 30, 2026: • Google’s Project Suncatcher satellite runs Gemma 4 on a data center TPU in orbit and sends an answer back to Earth • OpenAI, Google, Microsoft and Anthropic back a proposed standard that lets websites tell AI bots whether they can scrape and train on their content • OpenAI launches GPT-6 Rosalind Research, a model for life sciences open only to approved research organizations • SenseNova releases an open 9B model that reuses its own layers to think deeper without adding more parameters • S&P changes Meta’s credit outlook to negative over its huge AI spending • Intel stock drops about 30% after a weak forecast and drags other chip stocks down with it • Foxconn is talking with OpenAI about working together more and already has over 10 humanoid robots in its Texas factories • Xiaomi’s tiny OCR model reads and extracts data from documents and plugs straight into AI agents The takeaway: AI bots are already reading your website. Decide which of them can use your content, and make sure the information you want customers to find is clear enough for AI assistants to pick up. Hosted by the founder of PlsAutomate. We build websites that people and AI assistants can both read clearly, so your business shows up correctly wherever customers search. plsautomate.com