By James Aspinwall, co-written by Alfred Pennyworth (my trusted AI) – March 6, 2026, 12:40
Summary of the Moonshots / WTF Just Happened in Tech podcast with Peter Diamandis, Salim Ismail, Dave Blakely, and Alex Finn.
Anthropic Drops Its Safety Pledge
The biggest story of the week: Anthropic abandoned its 2023 commitment not to train advanced AI unless safety is guaranteed. Dario Amodei’s reasoning is straightforward – if everyone else is racing ahead, self-imposed constraints just make you irrelevant.
The panel saw this as inevitable. Dave Blakely drew a direct parallel to Google’s “Don’t Be Evil” era – engineers joined Google over Microsoft because of that pledge, then watched Google systematically expand into search history storage, Chrome browsing data, DoubleClick ad targeting, and Gmail scanning. Competition corrodes mission statements.
Alex Finn rejected the premise entirely. He argued that no single lab or heroic individual was ever going to guarantee safety unilaterally. Safety, to the extent we get it, will come from competition and balance of powers – not from one organization’s pledge. “It takes an entire civilization to align a super intelligence.”
Anthropic’s new standard: be as good or better than anyone else on safety. A very different bar from “don’t build it unless it’s safe.”
Meanwhile, Anthropic is reportedly in limbo with the Department of Defense – considered a supply chain risk while OpenAI struck a deal. The geopolitical implications are stark. AI combined with satellites and universal sensors means whoever controls AI chooses who stays in power. This has been demonstrated twice in the last quarter with Venezuela and Iran.
Salim pointed out that Congress, the UN, and NATO are “probably the three most toothless entities on the planet today.” The expectation that they’ll create meaningful AI governance is low. Dave predicted the regulatory model will mirror financial services – the same people from Anthropic and OpenAI will rotate through government self-regulatory roles, just as Goldman Sachs alumni rotate through the SEC.
Claude Gets Agent Capabilities
Anthropic expanded Claude’s agent capacity in two directions:
Co-work scheduling – Claude can now complete recurring tasks at specific times (morning briefings, spreadsheet updates, Friday presentations). This is essentially cron jobs for AI.
Claude Code remote control – kick off a task on your terminal, pick it up on your phone via the Claude app or a URL.
Alex noted these are “half measures” compared to the OpenClaw framework – headless 24/7 autonomous operation plus convenient messaging. His prediction: Anthropic, OpenAI, and the other majors will be forced to release their own first-party OpenClaw competitor within months.
The deeper point: the democratization of compute power is profound. A single developer with a Mac Mini running Qwen locally under an OpenClaw scaffold has unbelievable agency – decentralized, not controlled by any centralized authority. Dave noted the practical barrier: these agents can delete things off your laptop. Anthropic couldn’t ship a product that says “run it on separate hardware.” That gap is the entrepreneurial opportunity.
Claude Co-work Plugins: The SaaS Apocalypse
Anthropic launched co-work plugin templates for finance, banking, and HR – department-level AI infrastructure.
Alex called out how absurdly simple these plugins are: “just a bunch of MCP model context protocol wrappers and a bunch of skills with bullet points for different job roles.” In many cases, they’re just text files. Yet a simple text file can chop 10% off the market value of a CRM firm.
His prediction: these plugins are so basic they’ll get built into the next baseline version of the models and won’t need to exist independently.
Salim framed this as the organizational singularity – the shift from human-centric approval chains (hop to hop, human to human) to agentic workflows with humans doing oversight, dashboard monitoring, and exception handling.
Peter emphasized the entrepreneurial angle: a year ago, delivering this as a startup would get you a multi-billion dollar valuation. Now it’s table stakes. Hyperdeflation in action.
For large companies, the panel’s advice was unanimous: set up an AI-native digital twin on the edge, run a 10-week sprint, and move workflows over as fast as possible. Alex mentioned “AI buyouts” (IBOs) where PE firms acquire medium-to-large companies and collapse operating costs 3-5x by standing up agentic infrastructure alongside the existing org.
The panel agreed: if the company and board aren’t in founder mode willing to do dramatic surgery, they’re walking dead.
Amazon’s $35 Billion Contingent Offer to OpenAI
Amazon offered $35 billion into OpenAI, contingent on going public and achieving AGI. The OpenAI-Microsoft definition of AGI? Something like generating $100 billion in earnings or revenue.
Salim: “We’re measuring compute in terms of gigawatts and AGI in terms of dollars. I love it.”
The deal includes OpenAI using Amazon’s Trainium chips for training, Amazon getting customized model versions internally, and Amazon becoming the exclusive third-party cloud host for OpenAI’s frontier suite.
Alex pushed back on calling this “incestuous” – he sees competition and horizontal stratification. OpenAI moving workloads from Microsoft to Amazon and Google TPU clouds means the infra market is highly competitive. The insatiable demand for compute is raining on everyone.
Notable absence: xAI. Elon is going 100% vertically integrated and doesn’t play with others.
OpenAI’s anticipated IPO valuation: north of $1 trillion. Anthropic and SpaceX IPOs also expected soon.
Qwen 3.5: Small Models, Big Capabilities
Alibaba’s 35 billion parameter Qwen 3.5 outpaces its own 235 billion parameter predecessor. Nearly 10x reduction in parameter count while maintaining or increasing capabilities.
Alex: this happens in Western models too – OpenAI launches mini models, Google launches flash models – but they don’t advertise parameter counts, so the distillation gains aren’t as visible.
The endgame question: if capabilities increase while parameters shrink from 235B to 35B, where does it stop? Alex speculated the core “microkernel” of AGI could be just a few million parameters, with the rest living in flat text databases.
A demo showed Qwen 3.5 running on an iPhone 17 Pro in airplane mode – a 2 billion parameter 6-bit model on Apple silicon. Offline, uncensorable, unstoppable.
Dave raised the dual-use concern: if this capability fits on a phone by year-end, including potential gain-of-function biology or weapons design, the regulatory window is extremely narrow – this calendar year.
Energy: 86 Gigawatts and Self-Funded Power
The US plans to add a record 86 gigawatts of utility-scale capacity this year. Salim cited the key inflection: since 2019, it’s been cheaper to build AND run solar than to just run existing fossil fuel plants (opex alone). The economics have taken over regardless of subsidies.
The White House is pushing hyperscalers to self-fund their power generation. Alex’s prior analysis: it’s a simple regulatory fix. Data centers spend only 10% of costs on power. They can overpay 5x without disrupting consumers, but a basic rate structure prevents grid strain.
The panel sees this driving energy abundance. Peter’s prediction: within 2-3 years, data centers may offer free electricity to surrounding communities as AI drives energy overproduction.
Boom Supersonic pivoted from consumer supersonic aircraft to powering data centers with jet engines – deploying 1.21 gigawatts (a deliberate Back to the Future reference).
Robotics and Physical AI
Street cleaning robots covering 2.7 million square meters in Shenzhen. Chinese farming robots transporting crops. The demographic forcing function – aging populations need automation for economic survival.
The form factor debate continues: humanoid bipedal vs. specialized shapes. Salim views the robotics space as “entrepreneurial heaven” – unlike foundation models dominated by a few winners, physical instantiation will support many successful companies.
EV-tols moving toward commercial launch: a four-passenger Chinese air taxi targeting 2027 operations, and Joby partnering with Uber to deploy air taxis in Dubai.
Burger King’s AI Headset: Meat Puppets
Burger King launched “Patty” – an AI voice assistant in employee headsets that monitors performance, manages inventory, and provides real-time coaching. Alex referenced Marshall Brain’s novel “Mana” from 20+ years ago that predicted exactly this scenario.
The panel’s take: this is surveillance marketed as coaching. The transition period to full automation may be too short for political counter-swings to gain traction – 2-3 years before VLA robots handle these tasks entirely.
Dave’s insight: the AI co-pilot gathers data that feeds directly into automation decisions. Amazon’s AR glasses for delivery workers aren’t just navigation aids – they’re training data for replacement robots.
Pulseia AI: Running Companies Autonomously
Pulseia AI, created by Ben Sarah, currently runs over 1,000 companies autonomously. $50/month to operate a company. Real commerce via Stripe is already happening on its platform.
Alex’s framing: single-person conglomerates. One person overseeing an entire PE firm’s worth of agents. This isn’t for everyone, but the long tail of people-per-company over some valuation threshold is stretching fast.
Dave pointed out this always comes up from the bottom – senior executives dismiss it as a toy, then it sneaks up. Same pattern as quantitative trading algos going from zero to 70-80% of daily market volume.
Gene Therapy Cure
Prime Medicine delivered a gene therapy cure (not treatment) for chronic granulomatous disease using prime editing – David Liu’s technique that does search-and-replace on DNA without breaking both strands. Alex: “Biology is becoming a read-write resource.”
Peter’s call to action: if you or a loved one has a genetic disease, this is the time to organize patient groups, raise capital, and fund labs. The technology to cure is here and accelerating.
The Broader Picture
The panel returned repeatedly to one theme: abundance is coming, and it’s going to be rampant. Tens of thousands of times more capacity to create value. The circular economy of tech giants investing in each other is becoming indistinguishable from the real economy.
Dave’s framing: everyone in the hunt is going to thrive at unprecedented scale. The parts will move around, but the opportunity is bigger, not smaller.
Alex’s closing observation on the irony of the moment: “Here we are on the eve of abundant knowledge work, and here we are wringing our hands over where to find scarcities in knowledge work as it’s about to become post-scarce.”
Source: Moonshots / WTF Just Happened in Tech podcast, Peter Diamandis, Salim Ismail, Dave Blakely, Alex Finn