G'day, this was one of those weeks where the AI world moved faster than most teams can keep up with. Anthropic went to war with the Pentagon, Karpathy showed us what autonomous research looks like, and NVIDIA painted a $1 trillion picture at GTC. Here's everything that matters, distilled into one read.
TL;DR
- Anthropic vs The Pentagon: Anthropic sued the Trump administration after the Pentagon labelled it a supply chain risk for refusing to allow Claude in mass surveillance or autonomous weapons. Nearly 150 retired judges filed in support.
- The Karpathy Loop: Andrej Karpathy open-sourced autoresearch, where an AI agent autonomously ran 700 experiments in 2 days, found 20 optimisations, and improved training efficiency by 11%.
- NVIDIA GTC 2026: Jensen Huang unveiled the Agent Toolkit (adopted by Adobe, Salesforce, SAP and 14 others), the Groq 3 LPU chip, and projected $1 trillion in compute orders through 2027.
- GPT-5.4 Goes Live: OpenAI shipped its most capable model with a 1.05M token context window, 33% fewer false claims, and record knowledge work benchmark scores.
- ARC-AGI-3 Launches March 25: Francois Chollet's new interactive reasoning benchmark drops this week, the first major format change since 2019, testing agentic reasoning through never-before-seen environments.
- AI Reshapes the Workforce: Oracle plans to cut up to 30,000 jobs to fund AI data centre expansion. Block (Square, Cash App) eliminated 40% of its workforce.
Anthropic vs The Pentagon: When AI Safety Meets National Security
The biggest story in AI this week wasn't a model launch. It was a constitutional showdown.
Anthropic, the company behind Claude, filed two federal lawsuits against the Trump administration after the Pentagon designated it a "supply chain risk", a label typically reserved for foreign adversaries. The trigger: CEO Dario Amodei's refusal to let Claude be used for mass surveillance of US citizens or in fully autonomous weapons systems.
The Pentagon's position is blunt: the military should be able to use technology for "all lawful purposes" and won't let a vendor "insert itself into the chain of command." President Trump amplified the dispute on Truth Social, directing federal agencies to stop using Anthropic's technology.
But support for Anthropic has been remarkable. Nearly 150 retired federal and state judges filed an amicus brief. Microsoft, staffers from competing AI labs, and major industry organisations have all weighed in.
Why this matters for business: If you're using AI in your organisation, this clash is setting precedent for who defines the boundaries of AI use. For Australian businesses, your AI vendor's governance posture is now a material procurement consideration, not just a nice-to-have.
The Karpathy Loop: 700 Experiments, Zero Humans
Andrej Karpathy, OpenAI co-founder and former Tesla AI director, dropped something genuinely significant this week: a 630-line Python script called "autoresearch" that ran 700 autonomous experiments in two days, with no human intervention.
The system is elegantly simple. An AI agent gets a training script and a fixed compute budget (5 minutes on one GPU). It reads its own source code, forms a hypothesis, modifies the code, runs the experiment, and evaluates results. Keep if better, revert if not. Repeat 700 times.
The results: 20 additive optimisations that transferred perfectly to larger models, dropping the "Time to GPT-2" benchmark from 2.02 to 1.80 hours, an 11% efficiency gain. In a No Priors interview, Karpathy admitted he hasn't written a line of code since December.
The code is open source on GitHub. His vision: massively parallel agents collaborating asynchronously on different optimisations.
GTC 2026: NVIDIA Bets $1 Trillion on the Agent Era
NVIDIA's GTC conference (March 16-19, San Jose) was the industry's clearest signal yet that 2026 is the year AI agents go enterprise.
The headline: the NVIDIA Agent Toolkit, an open-source stack for building autonomous AI agents. It includes OpenShell (a runtime with security guardrails), the AI-Q Blueprint (topping DeepResearch Bench accuracy), and Nemotron (open models for agentic reasoning). Adobe, Atlassian, Cisco, Salesforce, SAP, ServiceNow and 11 others are already building with it.
Jensen Huang also unveiled the Groq 3 LPU, NVIDIA's first chip from its $20b Groq acquisition, shipping Q3 2026. And in the week's most eye-catching projection, Huang said compute orders between Blackwell and Vera Rubin will reach $1 trillion through 2027, doubling the previous $500b estimate.
Autonomous driving also featured prominently: Uber launching NVIDIA-powered fleets across 28 cities by 2028, with BYD, Nissan, and Hyundai building Level 4 vehicles on the platform.
Quick Hits
- GPT-5.4 launched: OpenAI's new flagship (March 5) with 1.05M token context window, 128K max output, and 33% fewer false claims. TechCrunch
- ARC-AGI-3 drops March 25: Francois Chollet's new interactive reasoning benchmark tests agentic reasoning through never-before-seen environments. ARC Prize
- Visa launches Agentic Ready: A program for banks to test AI agent-initiated payments, with 21 partners including Barclays, HSBC, and Santander. Visa
- Apple's Siri overhaul delayed again: The AI-powered relaunch (powered by Google's Gemini in a $1b/year deal) may slip past the March iOS 26.4 target. MacRumors
- Oracle plans 20,000-30,000 layoffs: To redirect $8-10b toward AI data centre infrastructure. Cuts could begin within weeks. Bloomberg
- Claude Code Channels: Anthropic shipped the ability to connect Claude Code to Discord and Telegram, letting you message it directly to write code. VentureBeat
- AI compute crunch incoming: SemiAnalysis CEO Dylan Patel told Dwarkesh Patel that combined hyperscaler CapEx will hit $600b, translating to ~50 gigawatts. Dwarkesh Podcast
- Diamandis launches $3.5m XPRIZE: The Future Vision XPRIZE (backed by Google) rewards optimistic sci-fi films. Submissions opened March 9. Fortune
The ClickedOn Take
Here's what struck us most this week: the gap between "AI as a product" and "AI as infrastructure" is closing fast.
When NVIDIA launches an agent toolkit with 17 enterprise partners on day one, when Visa is enrolling 21 banks in an AI payments program, and when Karpathy demonstrates that a 630-line script can replace two days of a researcher's work, we're not talking about experimental technology anymore. We're talking about operational infrastructure.
For marketers and business owners in Australia, the practical implication is this: the companies that will win over the next 12 months aren't the ones asking "should we use AI?" They're the ones asking "which AI workflows can we automate this quarter?"
At ClickedOn, we've been building AI-first into our own workflows, from GEO (Generative Engine Optimisation) for clients who want to show up in AI-generated answers, to using AI agents for campaign analysis and content production. If you're unsure where to start, our free AI Visibility Audit will show you exactly how your brand appears in AI-generated search results.
Tool of the Week
Autoresearch by Andrej Karpathy →
A minimal Python script (~630 lines) that creates an autonomous AI research loop. Give it a training script and a GPU, and it continuously forms hypotheses, modifies code, runs experiments, and keeps improvements, with zero human input. Even if you're not training language models, the architecture is a blueprint for how autonomous agents can optimise any iterative process.
Sources This Week
Axios, CNN, NPR, The Hill, Lawfare, Fortune, VentureBeat, The New Stack, NVIDIA Newsroom, CNBC, TechCrunch, ARC Prize, Visa, MacRumors, Bloomberg, Dwarkesh Podcast