346: Zuckerberg Finally Finds His People, They Are All AI Agents

Welcome to episode 346 of The Cloud Pod, where the forecast is always cloudy! Hold on to your butts, because Justin, Ryan, and Matt are in the studio today, and they’re ready to bring you all the latest in Cloud and AI news, including the usual: Meta buying social networks, Amazon responding to outages, and OpenAI giving up another version of GPT. Let’s get into it! 

Titles we almost went with this week

• ✍️ Cloudflare Spent $1100 to Rewrite Next.js in a Week
• One Pipe to Rule All Your OpenTelemetry Data
• ☑️ Check Yourself Before Google Wrecks Your Cloud Config
• Copilot Takes Jira Tickets So You Don't Have To
• ‍✈️ GitHub Copilot Agent Joins Your Jira Workflow Uninvited
• When AI Agents Network, Meta Swipes Right on Moltbook
• ️ Sixty Controls Walk Into a Terraform Repository
• One Security Console to Rule All Your Clouds
• AI Ate My Lock-In, and I Feel Fine
• ⛅ Oracle Sees $90 Billion Future Cloudy With a Chance of GPUs
• Your API Has Trust Issues, and We Can Prove It
• Stop Running Three Pipelines Like a Telemetry Hoarder
• From Database Dinosaur to AI Cash Cow
• ☠️ Meta: Target acquired; must kill Moltbook
• Meta saw Moltbook and said, “WE MUST OWN IT AND KILL.”

Follow Up

00:51 Where things stand with the Department of War 

• Anthropic has been designated a supply chain risk to US national security by the Department of War, a designation the company is challenging in court as legally unsound under 10 USC 3252.
• The practical scope of the designation is narrow, applying only to the use of Claude in direct Department of War contracts, not to all customers that hold such contracts or to unrelated business with Anthropic.
• Anthropic has stated that it will continue to provide its models to the Department of War and the national security community at nominal cost, with ongoing engineering support, during any transition period and for as long as permitted.
• The company's two stated exceptions to military use involve fully autonomous weapons and mass domestic surveillance, and Anthropic has clarified these do not extend to operational decision-making, which it considers the military's domain.
• For cloud and enterprise customers, the key takeaway is that existing Claude deployments unrelated to Department of War contracts remain unaffected, though the legal dispute introduces uncertainty into federal procurement pipelines involving AI services.
• We will keep you updated on this in 12-18 months…

AI Is Going Great - Or How ML Makes Money 

01:21 Introducing GPT-5.4

• OpenAI released GPT-5.4 across ChatGPT, the API, and Codex, positioning it as their most capable reasoning model to date. It merges the coding strengths of GPT-5.3-Codex with general reasoning, professional knowledge work, and native computer-use capabilities in a single model.
• The computer-use capabilities are a notable technical step, with GPT-5.4 achieving a 75% success rate on OSWorld-Verified desktop navigation, surpassing the reported human benchmark of 72.4% and up from GPT-5.2's 47.3%. 
• This makes it the first general-purpose OpenAI model with native computer use built in, making it relevant for developers building agents that operate across web browsers and desktop software.
• Tool search is a practical efficiency improvement for agentic API workflows, dynamically loading tool definitions only when needed rather than stuffing all definitions into the prompt upfront. In testing against Scale's MCP Atlas benchmark on 36 MCP servers, this reduced total token usage by 47% with no loss in accuracy, directly translating to lower API costs for tool-heavy applications.
• On the professional work side, GPT-5.4 scores 87.3% on an internal investment banking spreadsheet benchmark, up from 68.4% for GPT-5.2, and achieves 91% on BigLaw Bench for legal document work. The ChatGPT for Excel add-in, launched alongside it, gives Enterprise customers a direct integration path.
• Pricing is higher per token than GPT-5.2 in the API, though OpenAI notes the model's token efficiency should offset costs for many workloads. 
• Batch and Flex pricing remain available at half the standard rate, and Priority processing is available at 2x the standard rate for latency-sensitive use cases.

02:19 Justin - “There’s also been a slew of every cloud provider in the world announcing Chat-GPT 5.4 is now available, and we will not be telling you about all of them, but assume that if you use a different model or different cloud, they probably have it.” 

04:33 Introducing ChatGPT for Excel and new financial data integrations

• OpenAI launched ChatGPT for Excel in beta, an add-in powered by GPT-5.4 that lets users build, update, and analyze spreadsheet models using plain language descriptions. 
• It preserves existing formulas and structure, asks permission before making changes, and links answers to specific cells for auditability. 
• Available now for Business, Enterprise, Edu, Pro, and Plus users in the US, Canada, and Australia.
• GPT-5.4 (also available as GPT-5.4 Thinking) is now live in ChatGPT, Codex, and the API, with OpenAI noting it was specifically tuned on real-world finance workflows, including financial modeling, scenario analysis, data extraction, and long-form research.
• New financial data integrations bring Moody's, Dow Jones Factiva, MSCI, Third Bridge, MT Newswire, and others directly into ChatGPT workflows, with FactSet coming soon. 
• Organizations can also connect proprietary data sources using Model Context Protocol (MCP), centralizing market, company, and internal data in a single interface.
• For enterprise deployments, the Excel add-in supports RBAC, SAML SSO, SCIM, audit logs, AES-256 encryption at rest, TLS 1.2+ in transit, and data residency controls. In Enterprise and Edu workspaces, the feature is off by default and requires admin enablement with custom roles and group permissions.
• ChatGPT for Google Sheets is listed as coming soon, signaling OpenAI is extending this spreadsheet integration beyond the Microsoft ecosystem.

04:49 Justin - “If I were a betting man, I’d also say they’re going to have a PowerPoint version any day.” 

06:13 Meet KARL: A Faster Agent for Enterprise Knowledge, powered by custom RL

• Databricks introduced KARL (Knowledge Agent with Reinforcement Learning), a custom model built using RL techniques to handle grounded reasoning tasks like document search, fact-finding, and multi-step reasoning across enterprise data sources.
• KARL was trained with a few thousand GPU hours using entirely synthetic data. In internal testing, it matched or outperformed Frontier's proprietary models on inference cost, latency, and response quality simultaneously.
• The core technical challenge KARL addresses is hard-to-verify tasks, where there is no single correct answer, making RL reward signal design particularly difficult compared to domains like math or code, where correctness is easier to measure.
• Databricks is now offering a Custom RL private preview backed by Serverless GPU Compute, allowing enterprise customers to use the same RL pipeline that produced KARL to build domain-specific, cost-optimized versions of their own high-volume agents.
• For enterprises running AI agents at scale, this approach suggests that custom RL fine-tuning on smaller models can substantially reduce inference costs compared with relying on general-purpose frontier models, a practical consideration as agentic workload costs grow.
• Interested in checking out the preview? You can find more information on that here. 

07:09 Ryan - “It's kind of a neat idea to provide sort of the pipeline there. I mean, I guess the big cloud providers are producing agent-building platforms and stuff; I wonder how much of this you can follow the path that they use for creating KARL and building your own do