Jon Hulsinger

AI Engineer, Researcher, Developer, Creative…or a random amalgamation of scholarly and creative parts

[Jon Hulsinger @jhulsinger]

Wiring an experimental julia IDE together. Mixture-of-agents plans. MiniMax subagents applies code, a mixture-of-agents panel tears it apart, an eval loop decides if it matches my intent. Loop engineering remotely the right way:

[Jon Hulsinger @jhulsinger]

Software engineering is becoming like the practice of law. Senior engineers are no longer the authors of features. They are the builders of the substrate that turns intent into running code. Junior engineers are no longer the authors of features either. They are the certifying layer that decides what AI output is safe to merge. The role did not disappear. It moved down a layer into the place where intent becomes a system.

The data already shows the shift in three places. Per GitClear's January 2026 analysis of 211 million changed lines across Google, Microsoft, Meta, and large enterprise codebases, the share of code changes that constitute real refactoring dropped from 25 percent of changed lines in 2021 to under 10 percent by 2024, while copy paste cloned blocks climbed from 8.3 percent to 12.3 percent in the same window. More is being produced and less is being cleaned up, which structurally grows the review queue whether the reviewers are humans or agents.

Per Stanford's Canaries in the Coal Mine working paper published November 2025 and built on ADP payroll data through September 2025, employment for software developers aged 22 to 25 declined nearly 20 percent from its late 2022 peak, but in the same window employment for experienced developers in the same occupations continued to grow. Adjustments landed on headcount at the entry tier while the senior tier held. Per Stack Overflow's 2025 survey of 33,662 developers, early career developers with 1 to 5 years of experience already use AI tools daily at 55.5 percent, the highest of any cohort in the survey, while 58.7 percent of all developers refuse to delegate code review and merge decisions to AI. The volume data, the employment data, and the workflow data all describe the same split from three angles. The agents are producing the volume. The juniors are reviewing it. The seniors are building the substrate underneath it. That is the law firm pattern in code.

Canaries in a Coal Mine 2025: https://t.co/4bYigOZkKQ

Gitclear AI Copilot Code Report 2026: https://t.co/a1tOFZDAwu

[Jon Hulsinger @jhulsinger]

@rom1trs VOX

[Aravind Srinivas @AravSrinivas]

Every enterprise will have its own model-harness-sandbox-eval flywheel with token value per watt optimization. This is the future. Simple reason: tacit knowledge about the domain and customers and their workflows that the company uniquely understands and has built trust around.

[Jon Hulsinger @jhulsinger]

Mojo's real opportunity is not replacing Python on desktops. It's becoming the default high-performance language layer for edge Al, especially if Qualcomm pushes it across Snapdragon, Android, and NPU-first devices. Python ergonomics + systems-level control is exactly what edge Al needs.

[Modular @Modular]

Today, Modular announced an agreement to be acquired by @Qualcomm.

We founded Modular to build a unified compute platform for the world, and unlock a more open, efficient, and hardware-independent future for AI. With Qualcomm, we can bring that vision to more developers, enterprises, and hardware platforms— faster.

Modular co-founder and CEO @clattner_llvm explains our continued commitment to our mission: https://t.co/5lCOxbOW97

[Jon Hulsinger @jhulsinger]

A fake Sentry error can trick Cursor into running attacker code with developer privileges. SpaceX is buying Cursor for $60 billion. That's not a bug report anymore. That's a shareholder question.

[Jon Hulsinger @jhulsinger]

https://t.co/LxC4uAX3zY

[Jon Hulsinger @jhulsinger]

https://t.co/LxC4uAX3zY

[Jon Hulsinger @jhulsinger]

@LogicalThesis I think this is the key split: AI capex may keep accelerating even if model-company valuations reprice. The infrastructure layer and model layer are becoming two different trades. I wrote through that here: https://t.co/UP4gRQp6HX

[Jon Hulsinger @jhulsinger]

https://t.co/F2IhhwobCC

[Jon Hulsinger @jhulsinger]

@alex_verem The shutdown also shows how centralized frontier AI creates a new platform risk: access can disappear through national-security controls, not just product decisions. That makes local/open models more strategically important, even if they are weaker.

[Jon Hulsinger @jhulsinger]

https://t.co/F2IhhwobCC

[Jon Hulsinger @jhulsinger]

Every time an AI agent tries again, it may resend all its instructions and past work. That means costs can grow much faster than the number of steps suggests.

Without clear evaluation rules, spending limits, and quality checks, an “independent” AI agent may just repeat costly mistakes.

Here’s why better loops, not bigger AI models, could be the next big challenge:

[Jon Hulsinger @jhulsinger]

https://t.co/1kIQYnfWAT

[Jon Hulsinger @jhulsinger]

https://t.co/1kIQYnfWAT

[Jon Hulsinger @jhulsinger]

The IPO story will be sold as:

“AI is the next platform shift.”

That may be true.

But the real question is whether frontier AI labs can capture enough value before compute costs absorb the economics.

That is the number Wall Street is about to price.

12/12

[Jon Hulsinger @jhulsinger]

🚨 THREAD: OpenAI’s confidential S-1 is bigger than one IPO filing.

Anthropic filed one last week.
SpaceX is expected to price what could become the largest IPO ever this week.

Three filings.
One question nobody is asking:

Who’s actually making money?

🧵👇

[OpenAI Newsroom @OpenAINewsroom]

We recently submitted a confidential S-1. We expect it to leak so we’re just announcing it. We have not decided on timing yet; it may be a while because there are things we want to do that are likely easier as a private company. But it’s a complicated set of tradeoffs and this gives us the option to go public sooner if that ends up being best.

This announcement is being made pursuant to Rule 135 under the Securities Act of 1933, as amended, and does not constitute an offer to sell or the solicitation of an offer to buy any securities. Any offers, solicitations of offers to buy, or any sales of securities will be made in accordance with the registration requirements of the Securities Act.

[Jon Hulsinger @jhulsinger]

My read:

OpenAI and Anthropic may both be extraordinary companies.

But their S-1s will force investors to separate product-market fit from business-model proof.

AI demand is obvious.

AI profitability is not.

11/