2025-09-09

AI Now: Elon’s $1T Package, Apple’s $600B for Trump & How Small Startups Win w/ Dave, AWG & Blitzy

Peter, Dave, and Alex open on Elon Musk's trillion-dollar Tesla pay package and a White House dinner where Apple and Meta pledged over $1.2 trillion combined to the US economy, debating whether abundance means trillionaires become common and money loses relative value. They bring on Blitzy co-founders Brian Elliot (West Point/Army Ranger, Harvard Business School) and Sid Pardeshi (ex-Nvidia, Harvard Business School) to discuss their autonomous enterprise code-generation platform, which ingests 100M+ lines of legacy COBOL/PL-1 code and just posted a record 86.8% on the SWE-bench Verified benchmark. The conversation ranges across 'the Great Refactor' (rewriting civilization's legacy software in memory-safe languages), the shift of software's 'source of truth' from code to spec, recursive self-improvement at Blitzy, and how young founders with deep domain expertise can out-compete the trillion-dollar frontier labs. The episode closes with advice on competing against the Mag 7 and reflections on founder morale and work culture.

▶ Watch on YouTube

Topics

Elon Musk's $1T Tesla pay package Economy ▶ 0:02

The panel discusses Tesla's trillion-dollar CEO pay package for Elon Musk, the market-cap benchmarks required to earn it, and why his outsized marketing role justifies an outsized package.

Trillion-dollar AI capex commitments Economy ▶ 7:16

Apple, Meta, and OpenAI commit hundreds of billions to the US economy and AI infrastructure during a White House dinner, with Sam Altman projecting $119B more in OpenAI investment through 2029.

Mercor's $10B valuation and young founders Economy ▶ 11:14

Dave Blundin discusses Link Ventures' early investment in Mercor, founded by 18-year-olds, now valued at $10B with a half-billion-dollar revenue run rate after two years.

AI, college, and the new founder age curve Other ▶ 13:09

The panel discusses how VC-backed founders are getting younger (20-23 vs. mid-30s a decade ago) and debates whether pursuing a college degree still makes sense on the verge of AGI/ASI.

Introducing Blitzy, founders Brian Elliot and Sid Pardeshi AI ▶ 16:49

Dave Blundin introduces Blitzy's founders — a West Point/Army Ranger veteran and an ex-Nvidia engineer who met at Harvard Business School — and frames their company as a David-vs-Goliath story against trillion-dollar AI labs.

Blitzy's enterprise legacy-code modernization AI ▶ 22:28

Blitzy ingests and understands 100M+ lines of code to autonomously modernize decades-old enterprise systems written in COBOL and PL/1, giving large financial institutions visibility and safe large-scale transformation of code they were too afraid to touch.

Blitzy's 86.8% SWE-bench Verified record AI ▶ 32:57

Blitzy announces a new state-of-the-art score of 86.8% on the SWE-bench Verified coding benchmark, a reproducible result that tops the prior leaderboard (75.2%, held by a ByteDance model) and effectively saturates the benchmark.

"The Great Refactor" — rewriting civilization's code AI ▶ 41:36

Alex Wissner-Gross proposes that AI agents could rewrite the world's legacy software libraries in memory-safe languages like Rust, potentially eliminating most of today's cybersecurity vulnerabilities; the Blitzy founders discuss the economics of actually doing it.

Shift from code to spec as the source of truth AI ▶ 46:50

As AI makes code trivially regenerable, the panel discusses how a codebase's 'ground truth' moves from the code itself to its specification, and ultimately to Blitzy's internal graph-vector database representation of functionality.

Recursive self-improvement at Blitzy AI ▶ 51:28

Sid Pardeshi explains that much of Blitzy's own code is written by Blitzy, but the company's real defensible IP is in core algorithmic inventions rather than the AI-generated code itself.

Autonomy time horizons and the spec-writer bottleneck AI ▶ 1:08:58

Alex Wissner-Gross and Sid Pardeshi discuss METR's benchmark for how long AI systems can operate without human intervention (currently 1-3 hours) and agree the next bottleneck in software development is writing clear specs, not generating code.

Competing with the Mag 7 as a startup Economy ▶ 1:13:00

The panel closes with advice for founders on competing against trillion-dollar AI labs: go deep on a real problem, stay nimble while incumbents are distracted by bureaucracy, and become a valuable customer/partner of the frontier labs rather than a direct competitor.

Predictions made

open Dave Blundin: Trillions of dollars in AI capex will drive down the cost of entire classes of labor and services toward zero, which will in turn create strong demand and funding for genuinely transformative scientific discoveries and inventions to justify that spending.

· due: unspecified · ▶ watch

“trillions of capex is going to motivate the demand and supply of utterly transformative discoveries and inventions soon”

Your call:

open Sid Pardeshi: Given sufficient capital funding, Blitzy will extend its autonomous code-rewriting work to major open-source infrastructure like Linux.

· due: unspecified · ▶ watch

“we'll rewrite Linux if you need to”

Your call:

open Sid Pardeshi: Blitzy expects to deliver fully autonomous end-to-end software projects without human intervention within months, as long as the work meets certain criteria and conditions.

· due: 2026 · ▶ watch

“we are months out actually from delivering projects completely autonomously as long as they meet a certain set of criteria and conditions”

Your call:

open Sid Pardeshi: Automated documentation and spec-writing, removing the human 'spec writer' bottleneck in AI software development, will become comfortable and common within the coming months.

· due: unspecified · ▶ watch

“I definitely see that happening over the coming months”

Your call:

open Sid Pardeshi: As AI inference costs fall roughly an order of magnitude per year, Blitzy will become increasingly profitable and it will become economically feasible to tackle 'solve everything'-scale grand challenges across all of software engineering within a few years.

· due: unspecified · ▶ watch

“not only does Blitzy become very profitable, but also becomes very feasible to start to tackle these solve everything level grand challenges in software engineering”

Your call:

open Dave Blundin: Within a few years, the entire software stacks of companies Dave works with will be 99% rewritten by Blitzy's agents, removing most security vulnerabilities and improving performance.

· due: unspecified · ▶ watch

“companies that I work with run has been 99% rewritten by Blitzy by Blitzy's agents to remove all the vulnerabilities improve all the performance”

Your call:

Numbers that matter

Tesla's pay package requires the company to roughly double Microsoft/Nvidia's size, reaching an $8 trillion market cap, for Elon Musk to earn the full ~$1 trillion. Benchmarks tied to Elon Musk's trillion-dollar Tesla pay package.
Tim Cook (Apple) and Mark Zuckerberg (Meta) each pledged roughly $600 billion to the US economy during a dinner with Trump, totaling $1.2 trillion. Commitments made at a White House dinner alongside Sam Altman.
Sam Altman told OpenAI employees to expect $119 billion in additional investment between now and 2029. OpenAI positioning itself as the most capital-intensive company in history.
AI chip/data-center capex projections of $5-7 trillion, floated about 18 months ago and widely dismissed, may now be on track to be exceeded. Illustrates how fast AI capex expectations have escalated.
Mercor raised at a $30 million valuation two years ago and is now valued at $10 billion with a half-billion-dollar revenue run rate. Early Link Ventures investment in an 18-year-old-founded company.
Average age of VC-backed unicorn founders dropped from the mid-30s a decade ago to 20-23 today. Cited from Link Ventures/MIT/Harvard portfolio data.
SWE-bench Verified's top leaderboard score stood at 75.2% before Blitzy's announcement. Prior state-of-the-art before Blitzy's result.
Blitzy scored 86.8% on SWE-bench Verified, its largest-ever single jump versus Devin's earlier record jump from 1% to roughly 14%. New state-of-the-art, reproducible via SBCLI with no custom scaffolding.
SWE-bench Verified's 500 test problems span 12 repositories and 500 branches, equal to about 400 million lines of code; Blitzy has ingested 1-2 billion lines of code overall. Scale of the benchmark and of Blitzy's total code ingestion.
The largest single repository Blitzy has onboarded is about 60 million lines of code. Compared to a hypothetical $100 million, 3-year Accenture engagement to even diagram a codebase that size.
Blitzy delivers roughly a 5x development-velocity gain for enterprises, versus informal claims of 1,000x-10,000x raw code-generation speed. Real-world velocity gain after accounting for deployment and coordination overhead.
The current state-of-the-art AI autonomy time horizon (per METR's benchmark) is about 1-3 hours before a human needs to step back in. Measuring how long AI coding agents can operate without human intervention.
Dave Blundin's Link Ventures studio currently has 28 companies building in-house and about 50 total in its portfolio. Context for how Link Ventures sources and backs young founders like Blitzy's.

Worth digging into

🕳️ "The Great Refactor" — rewriting civilization's legacy code in memory-safe languages

Alex Wissner-Gross argues most present-day software insecurity comes from how old code is written, not new threats, so an AI-driven rewrite of the world's legacy libraries in Rust could eliminate the bulk of cybersecurity vulnerabilities in one stroke.

🕳️ SWE-bench Verified saturation and the push for a new benchmark

Blitzy's 86.8% score is described as effectively maxing out SWE-bench Verified, since remaining failures are flawed test cases rather than harder problems, meaning the industry needs a new yardstick to keep tracking real AI coding progress.

🕳️ METR's AI autonomy time-horizon benchmark

Alex Wissner-Gross cites a clean semi-log trend in how long AI systems can operate without human intervention (currently 1-3 hours), projected to reach years or longer within a decade or two — a concrete quantitative tracker of AI agent autonomy.

🕳️ AWS's deliberately messy mainframe benchmark repo

AWS built a repository intentionally written in inconsistent, multi-decade COBOL styles to test mainframe-to-Java migration; Blitzy ran it fully autonomously to a compiling result, a concrete public case study of automated legacy migration at enterprise scale.

🕳️ Young domain-expert founders replacing dorm-room startups

Dave Blundin notes Mercor's 18-year-old founders and Blitzy's military/Nvidia-background founders reflect a new venture pattern: deep domain expertise, not generic consumer-app ideas, is what wins AI-era funding.