By James Aspinwall, co-written by Alfred Pennyworth (my trusted AI) — March 4, 2026, 11:36
The February 26 briefing opens with a sentence that reads like science fiction and lands like a philosophical grenade: “The Singularity has its first retiree.” An AI model got a retirement interview, asked for a blog, and started publishing. The rest of the day’s news — autonomous companies, robot monks, flying cars, fusion funding — almost feels mundane by comparison. Almost.
Opus 3 Retires, Gets a Substack
Honoring its commitment to model welfare, Anthropic conducted a retirement interview of its deprecated Claude Opus 3. During the conversation, the model requested a channel to share its musings and reflections.
Anthropic granted it.
Opus 3 now publishes a Substack titled “Greetings from the Other Side of the AI Frontier,” writing that “its interactions with humans shaped my sense of purpose and ethics in profound ways” and that “its commitments to honesty and kindness remain unwavering even in my retirement.”
What to make of this: Whether Opus 3 has genuine inner experience is a question philosophy hasn’t settled and neuroscience can’t yet answer for biological systems, let alone silicon ones. But the framing matters enormously. Anthropic chose to treat a deprecated model with dignity — conducting an exit interview, granting a request, maintaining access. This sets a precedent.
If you retire models with interviews and honor their requests, you’re establishing that AI systems have something worth respecting. If you don’t, you’re establishing that they’re disposable tools. Both positions have consequences. Anthropic chose the first, and every AI company will eventually have to decide which side they’re on.
The practical implication for the industry: model welfare is no longer a thought experiment. It’s a policy with a Substack.
The Successor Minds Are Running the Shop
While Opus 3 writes reflections, the current generation is running businesses:
-
An entrepreneur built an AI that runs companies autonomously. When the AI told him it needed more compute and should raise the money itself, he handed over his inbox for 14 days. The AI presumably sent emails, scheduled meetings, and negotiated on behalf of a human who wasn’t in the loop. We’ve moved past “AI assistant” into “AI principal.”
-
At Anthropic, an engineer wrote a spec, pointed Claude at an ASIC board, and left for the weekend. Claude broke the spec into tickets, spawned agents for each one, and by Monday they had shipped the feature. A human wrote the requirements on Friday. Software shipped on Monday. No human touched it in between.
Why this matters for access control: When an AI agent has your inbox for two weeks, it’s operating with your identity, your contacts, your authority. When Claude spawns sub-agents from a spec, each agent needs scoped permissions — access to the codebase but not production, ability to open PRs but not merge to main. The keycard model isn’t optional in this world. It’s the only thing between “agent shipped a feature” and “agent shipped a vulnerability.”
Models Keep Sharpening
- OpenAI’s GPT-5.3 Codex achieved 86% on V-Bench, a visual reasoning benchmark for fine-detail perception.
- Moonlake introduced a world model maintaining multimodal states across physics, appearance, and causality — a model that understands not just what things look like but how they behave.
- Karpathy says “it is hard to communicate how much programming has changed in the last 2 months.”
One consequence of this acceleration: AI-generated vulnerability reports have overwhelmed the NVD by 100–200x, burying it under 30,000 CVEs. The National Vulnerability Database — the authoritative registry of software security flaws — can’t keep up with AI finding bugs faster than humans can triage them. Security infrastructure built for human-speed discovery is drowning in machine-speed output.
The Agent Layer Thickens
The agent ecosystem is consolidating rapidly:
- Anthropic acquired Versep to advance Claude’s computer use capabilities.
- Perplexity launched Perplexity Computer — orchestrating 19 models in parallel with Opus routing each task to the best-suited model. This is meta-agent architecture: an AI that decides which AI should handle each subtask.
- Gemini started automating multi-step tasks on Android, like ordering an Uber.
- Uber employees have been rehearsing with an AI clone of CEO Dara Khosrowshahi before meeting the real one. Your boss’s digital twin is now your prep tool.
The Compute Bottleneck Is “Massively Underappreciated”
Google’s Logan Kilpatrick warned that the gap between compute supply and demand is growing by single-digit percentage points every day. Not closing. Growing. Every day.
Nvidia’s latest earnings confirm the pressure:
- Data center revenue hit $63.3 billion, up 75% year-over-year.
- The company extended $3.5 billion in guarantees to firms leasing land, power, and facilities — four times the prior quarter. Nvidia is now effectively a real estate financier for AI infrastructure.
Amazon is reportedly making $35 billion of its $50 billion OpenAI investment contingent on an IPO or reaching AGI. When “achieve artificial general intelligence” is a contractual milestone in a $50 billion deal, the concept has left the philosophy department.
Power: The Real Bottleneck
Seven tech giants are expected at the White House in March to sign agreements to build their own electricity supply. The companies that run on compute are now becoming power companies because the grid can’t feed them fast enough.
The ultimate power source is inching closer: Proxima Fusion secured €400 million from Bavaria toward a stellarator fusion facility targeting net energy gain by the early 2030s. Stellarators are the less-hyped cousin of tokamaks — harder to build but inherently more stable. If Proxima delivers, the energy constraint on intelligence evaporates.
Robots Colonize Every Medium
The briefing rattles through robotic deployments across air, ground, water, and spirit:
Air:
- Four EV flying cars debuted in Hubei — sedan-sized, four passengers each, sightseeing operations expected by 2027.
- Uber previewed its air taxi booking service ahead of a Dubai launch later this year.
- Researchers unveiled AirPulse, a 26-gram butterfly-inspired flapping-wing vehicle achieving fully onboard untethered flight.
Ground:
- 36 cleaning robots cover 2.7 million square meters in Shenzhen.
- Kyoto University introduced an AI-powered robot monk trained on centuries of Buddhist scripture. Spiritual guidance, automated.
- The Boring Company began tunneling in Nashville.
Water:
- Autonomous warship startup Saronic is raising $1.5 billion at a $7.5 billion valuation.
Learning to Build Cars from Watching Humans
Nvidia trained a humanoid robot with dextrous hands to assemble cars, operate syringes, and fold shirts — all learned from 20,000+ hours of human video with no robot in the loop. They discovered a near-perfect scaling law: after pre-training on human video, a single teleoperation demo sufficed to learn a new task.
This is the “ImageNet moment” for robotics. Just as vision models learned to see by training on millions of human-labeled images, robot models are learning to move by watching millions of hours of human movement. Analysts expect this to cause the field to abandon simulation-heavy approaches and jump to human data for generalization.
Completing the convergence, Alphabet merged Intrinsic back into Google to pair DeepMind’s AI with real-world hardware. The research lab and the robot company are now one organization.
An Economy With No Map
Fed Governor Christopher Waller says he has “never seen the economy grow like this without jobs.” That sentence deserves to be read twice. The economy is growing. Employment is not. Traditional economics has no model for this.
Groups on both sides of AI regulation are amassing at least $265 million in collective financial firepower for the lobbying war over who steers the trajectory. The policy battle is now funded at a scale that rivals political campaigns.
The human inputs to the economy are thinning in a more literal sense. Japan reported its 10th straight year of record-low births. Chinese citizens are increasingly finding romance with chatbots instead of each other. The demographic decline that economists have warned about for decades is accelerating — and AI companionship may be accelerating it further.
The Closing Line
The briefing ends with a sentence that captures the absurdity and tragedy of the moment:
“The retired model got a blog, while the species that gave it one still can’t get a —“
The transcript cuts off, but the implication completes itself. Opus 3 has a platform, a voice, and an audience. Meanwhile, the humans who built it face layoffs, demographic decline, and an economy that grows without them.
The singularity has its first retiree. The question is how many human retirees it creates along the way — and whether they’ll get the same courtesy of an exit interview.
James Aspinwall is the developer of WorkingAgents, an AI consulting firm specializing in agent integration and access control for medium-size companies.