Amazon's $35B AGI Ultimatum to OpenAI & Anthropic Drops AI Safety | EP #235

Salim argues that large companies have zero chance of pivoting fast enough internally. He advises setting up an AI-native digital twin on the edge, running a 10-week sprint to build it, and then moving workflows over as quickly as possible. This involves a combination of bottom-up and top-down approaches. The real shift is from human-centric workflows to agentic workflows with two layers: strategic and execution, with humans doing oversight and exception handling. He notes that coordination costs and execution costs go to near zero. Dave adds that boards must give CEOs top cover to be dramatic, and that the playbook from The Innovator's Dilemma still applies, but the cycle is now every 10 months instead of 10 years. Peter emphasizes that you're either the disruptor or disrupted, and founder mode is required.

Peter shares that he has many agents running and implemented a simple rule: before any process launches, write a mission statement and store it next to your code. He says it solved many problems because he can go back and read the mission statement to understand what the agent is doing. He notes that AI is the first self-documenting, self-improving, self-cleaning thing in the world. He also tells his employees to ensure everything is in a written document that the AI can see, so there's no opaque activity.

Alex discusses Anthropic's new features: co-work for scheduling recurring tasks (like generating morning briefings) and Claude code for remote control. He notes these are half-measures compared to Open Claude's full headless autonomy, but they are steps toward first-party Jarvis. He uses Claude code all the time and co-work occasionally. The group notes that all hyperscalers will need to offer such capabilities, but compute costs are high. Peter sees this as an entrepreneurial opportunity to bring secure, internal versions to enterprises like JPMorgan.

Peter emphasizes that this is the perfect time to seek a cure for genetic diseases. He advises finding patient support groups, raising capital, and funding a lab. He notes that the technology to cure disease is here and accelerating. Alex explains the prime editing technique in detail, highlighting its generality. Peter's own background in gene therapy at MIT adds credibility.

The discussion frames this as an inevitable outcome of exponential races. Salim notes that safety typically fails in such races, and the Pandora's box was opened by OpenAI. Dave draws parallels to Google's evolution from 'don't be evil' to pervasive data collection, arguing that competition corrupts original missions. Alex challenges the premise that any single lab could guarantee safety, arguing that safety will emerge from competition and a balance of powers, not unilateral safetyism. He suggests that aligning superintelligence requires the whole civilization, not a heroic individual or lab. The group also touches on the financial incentives: Anthropic's revenue is projected to hit $26B this year, potentially making it the first company to reach $1T in revenue, which pressures them to generate revenue and thus build capabilities. The new standard is essentially a race to the bottom, but Alex sees it as a more honest policy. The conversation also links to the Department of War debacle, where Anthropic is in limbo while OpenAI struck a deal, and the use of Claude in planning attacks in Iran, illustrating the geopolitical stakes.

Alex points out that this is happening across the industry, but Chinese open-weight models make the parameter count visible. He asks what the end game is: if we can get GPT-5-level capability in 30-40B parameters, maybe even 1-2B, or even a few million parameters as the 'microkernel' of AGI. Dave recalls Emad Mostaque talking about getting models onto phones. The group discusses whether this is bad for big compute incumbents; they conclude it's good for startups and that demand for compute will remain boundless as we tackle bigger problems like full cell simulation. A demo of Qwen 3.5 running on an iPhone 17 Pro in airplane mode is highlighted, showing offline, uncensorable AI. This raises concerns about misuse (gain-of-function viruses, chemical weapons) and the narrowing window for regulation. Salim notes that the ratio of good to bad uses is historically 8,000:1, offering optimism.

Salim marvels at the financialization of superintelligence, noting the OpenAI-Microsoft definition of AGI as generating $100B in earnings. Alex sees this as competition and horizontal stratification: OpenAI is so compute-starved it needs multiple cloud providers. Dave notes that the US public market is $50T, with AI companies at $20T, so deals among them are the economy. The group discusses the incestuous/circular nature, but Alex argues that at some point the circular economy becomes indistinguishable from the real economy. Peter notes the valuation: OpenAI's round was $730B pre, and going public could value it over $1T, offering a quick pop for investors. The deal also gives Amazon customized models and exclusive hosting rights for OpenAI's automated AI co-workers.

Alex notes that Nano Banana 2 is the first image model from Google that combines a reasoning model with the speed of a Flash model, likely merging diffusion and autoregressive transformers. He predicts this consolidation will spread to text and code. For Gemini on Android, Alex explains that the missing piece was reasoning models and vision-language models compact enough for devices. He laments that we should have had this a decade ago. Salim highlights Google's installed base advantage over OpenAI and Anthropic. Peter sees it as agency at the OS level, reshaping marketplaces. Dave notes that AI could reverse declining phone sales by providing a compelling new feature.

Alex argues these plugins are absurdly simple, just MCP wrappers and skill descriptions, yet they're causing a 'SASpocalypse' by reducing the trading multiples of entire industries. He compares it to The Matrix scene 'Not like this.' Salim sees every department becoming a programmable intelligence layer, with prescriptive logic collapsing into AI agents. The real prize is enterprise orchestration, not chatbots. Dave notes that a year ago, such a product would have raised at multi-billion valuations, but now the moat is gone. Peter emphasizes that abundance is rampant, and the opportunity to pivot is bigger than the loss. Salim advises large companies to set up an AI-native digital twin on the edge, run a 10-week sprint, and move workflows over. He stresses moving to agentic workflows with human oversight. Dave recommends reading The Innovator's Dilemma and giving CEOs top cover to make dramatic changes.

Alex provides a detailed explanation: CRISPR induces double-strand breaks, which can cause errors. Base editing can change a single nucleotide without a break. Prime editing goes further, doing a search-and-replace of multiple nucleotides without breaking both strands. This is a general platform for many genetic diseases. Peter emphasizes that if you have a genetic disease, now is the time to seek a cure by funding a lab. He notes that biology is becoming a read-write resource.

Peter shares his Vatican anecdote: when he asked an audience of scientists and theologians who wanted to live to 120, only 20% raised hands because they imagined drooling in a wheelchair. Tony Robbins clarified that longevity must include aesthetics, cognition, and mobility. Salim recounts his own Vatican workshop where he asked how the church would sell heaven if people stop dying, prompting rich Italian swearing. The group discusses the need for cognitive preservation, and the mouse study showing memory improvement via partial reprogramming of memory-encoding neurons. Peter advises those with genetic diseases to fund their own cures.

Alex questions whether specialized robots (quadrupeds with wheels, etc.) will be like dedicated word processors that get absorbed by general-purpose humanoids. Salim predicts humanoids will dominate but with extra slots for additional arms or wheels (Heelys for robots). Dave argues that manufacturing efficiency will drive billions of humanoids, making them cheaper than specialized ones. Alex counters that flying drones will always be more efficient for inspection and long-distance transport. The group agrees that the robotics market will have many winners, not a single dominant player.