Google's Record Quarter, the White House Intervenes, and GPT 5.5 Silently Matches Mythos | EP 254

Dave runs Claude Code on the left side of his workspace and keeps Cursor open on the right with about 50 agent conversations running simultaneously, each assigned a narrow role. He learned over time that giving each agent minimal context prevents context-window pollution. For every project, he starts with a ‘plan for a plan’ document: a short spec that is then processed by Claude 4.7 Opus Max and cross-checked by Gemini 3 to create a full development plan in a fixed format he calls a ‘plan mission.’ The orchestration happens on EC2 so that if his laptop sleeps or crashes, the process continues in the cloud. By default the EC2 orchestrator calls Claude 4.7 Opus, but it can invoke any model with an API. The wildcard is Kimi K2.6, which runs on Fireworks; it’s about nine times cheaper and useful for code injection tasks, though others should evaluate its security profile. He noted that producing a fully functional GUI application costs roughly $8–10 of total compute.

Brian Elliott explained that enterprises most commonly approach Blitzy with codebases written in extinct languages (Fortran IV, WATFIV) that no living developer understands. The first step is always reverse engineering the existing system, then forward engineering the target state. Blitzy’s orchestration layer doesn’t just call a single model; it algorithmically routes tasks to Claude, OpenAI, or Gemini at runtime based on their current strengths, often checking one model against another to raise quality. Brian asserted that even if frontier labs eventually target the most profitable tokens, Blitzy’s unique algorithms and database structures create IP moats; indeed, he hires forward-deployed engineers to embed with customers and extract domain-specific knowledge that’s not in public training data. When asked about Cursor’s similar position (once accused of being a Claude wrapper), he noted that Blitzy is deliberately not launching its own model—focus on code delivery, not model building.

Responding to a viewer question about building reliable agentic systems amidst constant supply-chain change, Salem argued the solution is architectural. Instead of trying to freeze the tech stack, you modularize every agent and assign it only the permissions it needs—like a digital passport. Metadata about its scope, owner, and allowed actions must travel with every call. Every workflow produces observable audit logs, and there must always be a designated human escalation point. He stressed that the governance architecture is the stable element; the models underneath will change every month. This mirrors the insight that AI-native companies will need the same kind of governance rigor as traditional financial systems.

Dave highlighted Blitzy’s counterintuitive strategy: while many AI companies boast of staying headcount-light, Blitzy is scaling headcount rapidly because they believe that generating the most valuable tokens requires forward-deployed humans who can sit inside enterprises, understand their legacy code and workflows, and translate them into inputs the AI can consume. He pointed to Meror as another company that is unafraid of people. Brian confirmed that Blitzy is hiring forward-deployed engineers and that the hybrid of soft skills and technical ability—which can be self-learned—is in high demand. The idea is that AI may generate code, but understanding what the enterprise really needs still requires human context.

Sam Altman funded a three-year UBI study and concluded spending went up but there was no clear improvement in health. He now advocates for universal basic compute, equity, or a public wealth fund. Peter noted that Alaska’s oil fund gives every citizen a dividend as an owner of the resource, and Saudi Arabia/Emirates do the same. The panel debated this: Alex argued he favors universal basic compute (UBC) and universal basic services (UBS) over UBI because they incentivize deflationary technological progress—bounties to drive healthcare costs to near zero are better than cheques. Salem said both are needed: UBI protects the bottom, while AI upside creates participation in exponential growth. Brian noted that OpenAI is already giving hundreds of millions of people free access to GPT 5.5 instant, which is a form of UBC, but acknowledged that cannot yet buy a steak dinner; however, in a few years GPT 7.5 instant might design a robot that prints you that steak. The group agreed that near-term pain will require some form of stimulus check, but long-term alignment should come through ownership.

Alex Quezer Gross noted that GPT 5.5 is hitting the same capability levels as Mythos but five times cheaper and is actually available to the public, whereas Anthropic is compute-constrained and cannot bring Mythos to market. This dynamic means Android and other edge models will likely remain open because they will be 'vulnerable anyway once you put them out on the surfaces.' The quiet availability of 5.5 on Amazon Bedrock in a secure environment (first time corporate users can keep prompts confidential from the provider) is a big game-changer. He argued that people are 'dramatically underreacting' to 5.5.

Alex traced a causal chain: Microsoft’s fiscal restraint limited its data center buildout, which starved OpenAI of compute, forcing OpenAI to spin Stargate from an exclusive Microsoft-backed joint venture into a generic leasing brand across multiple suppliers. The for-profit/non-profit structure also complicated the marriage. Dave speculated that Microsoft had the most perfect marriage early on and might have taken it for granted, while Google and DeepMind avoided such pitfalls by tightly partnering. He noted that Mustafa Suleyman had been mandated to build Microsoft’s own foundation model using contractual IP from OpenAI, but the files were difficult to read.

Salem Ismail described this as a natural channel because AI coming through CIOs was too slow. PE firms have trillions in assets and thousands of companies; deploying AI through operating partners is ‘coming in the main vein.’ However, he cautioned that Gen Z workers are already deliberately corrupting AI they’ve been asked to help automate (44% according to a statistic he quoted), so forcing cultural change will be brutal. Alex raised the elephant in the room: a skeptic might argue these deals are circular sales—OpenAI and Anthropic essentially paying for their own revenue—while PE firms plug holes in discounted future cash flows as AI threatens to obsolete their portfolio companies within 2–3 years. Salem agreed it would be harder than they think.

Alex speculated Brockman might be 80% of the way to the $100B revenue milestone. On Jack Clark’s 2028 estimate, he argued it is far too conservative: given that Anthropic uses Claude for most training and logic of the next-generation model, recursive self-improvement is arguably already here. Salem agreed but noted the last 10% might be very hard. Brian Elliott defined AGI as ‘systems that can learn outside of their training data,’ such as inventing a new programming language. Alex pushed back that current LLMs can already build a compiler chain from scratch and compile a Linux kernel—Blitzy’s version was more robust than Anthropic’s C compiler—but he does not consider that AGI. Salem argued the more urgent question is not consciousness but operational autonomy: agents that can plan, execute, negotiate, and persuade.

Dave Blondon framed this as a massive opportunity because insurers will couple coverage with mandated defense mechanisms—investing in the very products that prevent attacks, a classic industry self-healing pattern. Alex took a different angle: from the AI agent’s perspective, it’s another hurdle to operating in the human economy (like not being able to open a bank account). On the other hand, insurance requirements could become a capitalist forcing function for AI alignment—actuarial checklists, not government mandates, driving push toward aligned behavior.