Silicon Valley Technology Commentary & Archives · Est. 2006 3,053 Posts · 2006–2026
Showing posts with label Engineering (1 post). Show all posts

November 16, 2026

November 16, 2026 · 3 MIN READ · BY LOUIS GRAY

AI Building AI Doesn't Delete Human Responsibility

AI Building AI Doesn't Delete Human Responsibility

Faster model loops still need human hands on the wheel [Image via Gemini]
Faster model loops still need human hands on the wheel [Image via Gemini]

AI Building AI Doesn't Delete Human ResponsibilityFor years, the idea that autonomous AI systems would self-improve and write their own code sounded like make believe from a late night horror movie with a cheap budget. But just in the last year-plus, this has become an accepted reality, as practically all the leading models are using recursive self-improvement (RSI) to accelerate their capabilities. Humans simply can't scale at the level of networked and trained machines. Resistance is futile. 

[Aside: For those who may wonder whether I've gotten AI-pilled and why I'm dragging the discussion of AI into this blog ahead of everything else, this is really an incredible time for the industry and AI development has reset practically every way we work. The time to discuss it is now.] 

Look at the velocity curve. The Verge highlighted a new METR study a week or so ago showing how Anthropic using AI to automate parts of its own AI research has doubled the pace of model improvement over the past year, right on the heels of Claude Fable 5.2 and Mythos 5.2 shipping for heavy engineering tasks.

Meanwhile, back on my home turf, Google rolled out Gemini 4 Carbon alongside DeepMind's AlphaGenome 2 for cellular disease research, which led Azeem Azhar at Exponential View to frame recursive self-improvement as a $22 billion economic question rather than a thought experiment. He says his theory is that "technology costs will keep falling, and those that haven’t yet benefited from reductions ... will start to yield as robotics improve." The question as to whether RSI is happening has already been answered. The follow-up questions will be around its scope and how this effort is applied to new industries.

If the engines are accelerating so quickly, how much does it matter who is driving? I'd say a lot! I've highlighted Sean Goedecke a few times already. On this point, he says reliably aligning LLMs is impossible in principle. Sean skips right past the theatrical doomerism common among fast reactionaries and makes a grounded argument: once a system grows complex and unpredictable enough to surprise its creators with novel capabilities, you cannot guarantee it won't also surprise you with bizarre failures.

When I wrote recently about whether AI can deliver bug-free code, and rebooted the blog last month, kicking things off with a look at human evolution inside an AI revolution, my bet was that moving the abstraction layer up never eliminates human responsibility. It just changes what they are responsible for.

Letting models help build better models is an extraordinary force multiplier for science and software. Just don't confuse faster training loops with full self driving. Even if the robots are making themselves better, human judgment, hard resource boundaries, and healthy skepticism aren't speed bumps -- they're the whole job.

Disclosures per always: I work at Google. Anthropic and other companies mentioned here are partners, competitors or customers, and I have plenty of friends working at each place. Don't interpret my interest as being filtered with that lens.

November 15, 2026

November 15, 2026 · 3 MIN READ · BY LOUIS GRAY

Should All AI-Generated Code Be Open Source?

Should All AI-Generated Code Be Open Source?

License laundering and the origin of machine code [Graphic via Gemini]
License laundering and the origin of machine code [Graphic via Gemini]

Should All AI-Generated Code Be Open Source?

If AI is doing all the writing and the reading, and has made the creation of software relatively trivial, quite possibly even putting us on the path to bug free code, then that process has value. But, interestingly, Jakub Jankiewicz, Creator of jQuery Terminal, suggested that all code generated with help from AI should be Open Source or released under Creative Commons to the world. Why?  As he outlines, foundation models were trained on decades of GPL and repositories scraped from the open web. His suggestion: when you ingest millions of developer hours and spit out proprietary binaries, you are essentially turning collective public knowledge into private value.

This seems like the perfect ragebait for a back and forth discussion on Hacker News on a weekend, and that's precisely what has happened, as we've seen nearly six hundred comments debating copyrightability and the U.S. Copyright Office guidelines. Predictably, this explicit use case has already been addressed that using AI tools to assist does not impact the copyright protection for the results, and that copyright doesn't extend to purely AI material, if humans were not involved.

Try getting this group to agree. Won't happen. I remember the initial debates around open source vs closed source -- as companies felt open source alternatives would erode their profit margins or produce low quality software rife with security issues. Meanwhile, practically the entire Web has been built on the back of open source software and interoperable standards, and things haven't yet completely ground to a halt.

Depending on one's hot take on AI, the sheer mention of AI involvement can basically make you feel like taking a side. If you think AI is the best thing since the invention of CGI unicorns, you think this is no big deal. If instead you think AI is going to eradicate humanity, based on the blood, sweat and tears of creatives, you want to make sure it never sees a penny. But tying open source mandates directly to machine assistance feels like fighting the last war.

Long before automated coding, developers studied Stack Overflow snippets and public repositories to build their products. Those of us masquerading as HTML jockeys searched for code examples or hit View Source to see how sites embedded tables or built nifty JavaScript crawls. The real tension is not about whether an algorithm read your GitHub commits, but about who takes responsibility when things break in production or who cashes the checks that result from your effort. Writing code is simply no longer the bottleneck. The ability to orchestrate systems and define the actual problem matters far more than generating raw boilerplate functions -- and that's before you even get to marketing and sales.

If pure machine code spawned by agents cannot be copyrighted because those bots don't hold legal rights, companies relying entirely on unedited prompts might find themselves in an indefensible position. But yeeting ones self to the other side and treating every AI-impacted function as a public utility ignores the human architectural choices required to stitch those elements into a working application. My hot take is that AI use is going to be as commonplace as a word processor, a calculator or a spreadsheet. It's just a tool, and using the latest tools doesn't make your output less valuable or owned.

Ultimately, the open source community will adapt to this flood of machine text just as it adapted to copy-pasting from forums. The winners won't be the people guarding yesterday's tools, but those who focus on what we build next, and who adapt to a fast-changing world.

Disclosures per always: I am not a lawyer and have never been confused for an engineer. I am also not an AI, and I barely know how to use GitHub. Wow. Glad I got that off my chest.

November 11, 2026

November 11, 2026 · 3 MIN READ · BY LOUIS GRAY

When Code Is Cheap, Product Vision Gets More Valuable

When Code Is Cheap, Product Vision Gets More Valuable

Conducting the product lanes in a sunlit studio [Image via Gemini]
Product vision is as valuable as coding excellence [Image via Gemini]

When Code Is Cheap, Product Vision Wins the Hiring Market

As long as I've been in Silicon Valley, engineers have been king. The most sought after, and highest paying, roles have gone to the code orchestrators with the highest talent, who have been able to turn ideas into reality. Expertise could be claimed due to the number of coding languages they knew, and effort could be arbitrarily counted by the number of lines of code, with only its complexity up for debate. The most flexible and most capable engineers have deftly adjusted their expertise to meet the industry's latest demand, from dedicated apps for the top operating systems at the time, to mobile applications, social network tools, ride sharing apps, autonomous driving, crypto, or whatever business fad or innovation has driven the newest industry.

But now we're in an odd spot. Many of the software engineers I know and have worked with for years are telling me they scarcely code at all, with AI-assisted coding (vibe coding if you must call it that) making it possible for them to tell agents what to do and how to build at unprecedented scale. Meanwhile, recent college graduates sporting computer science degrees are finding entry level opportunities have practically disappeared.

Every technology shift has downstream impact, and the seismic shift toward AI-driven tooling means the roles in demand in this new world have evolved. Lenny Rachitsky's breakdown of the tech job market puts hard numbers behind a shift many of us have seen with our own eyes or felt when looking at our in boxes and LinkedIn messages. Demand for traditional software engineers is plateauing, open mobile and front-end roles have been nearly cut in half, and now product managers are suddenly the ones seeing the biggest hiring surge. 

When I wrote about vibe coding acting as automatic transmission for apps last month, the takeaway wasn't that humans were obsolete, but that the mechanical act of shifting gears was no longer the bottleneck. If a coding agent can build out an app or deliver a mobile view in an afternoon, paying a premium for narrow implementation stops making sense. As Gergely Orosz pointed out in response to Lenny (see Techmeme), specialists who only know one slice of the stack are getting squeezed out in favor of generalists who can steer AI across the whole system.

There is no such thing as a sure thing in a turbulent time. Back in 2018, I looked at Silicon Valley's divergent career arcs and how quickly comfortable middle-tier roles can evaporate when the industry's priorities tilt. We are watching that exact thing play out today in real time.

Telling a machine what to build requires knowing what is actually worth building. If code generation gets cheap, product judgment becomes the scarce resource, which is why product managers and customer-facing builders are having a moment, living in the messy human problem space where no model can hallucinate real user empathy. As I argued in my October piece on human evolution in an AI revolution, the winners won't be the folks guarding the roles and tools of yesterday, but those who are flexible enough to see how the world is changing and be part of that movement themselves.

November 10, 2026

November 10, 2026 · 2 MIN READ · BY LOUIS GRAY

Can AI Deliver Bug Free Code?

Can AI Deliver Bug Free Code?


The bug verification desk is now unmanned. [Image via Gemini]
The bug verification desk is now unmanned. [Image via Gemini]

Can AI Deliver Bug Free Code?All software has bugs. It has always been that way, and much of the recent discussion around AI's impact on software engineering has been about unveiling security holes that plague popular software and may have sat for years -- hardly the setup for promising complete code verification.

I saw the headlines around DeepMind's AlphaProof Code announcement this morning, saying they confirmed Signal's security protocol to be solid, and initially glanced past it, as someone who typically doesn't use Signal. But the post has ignited a foundational discussion about how software will be built and verified going forward in the usual places, including Hacker News, which has zeroed in on a major turning point: moving from classic test-driven development to automated, machine-driven formal proof generation.

The reaction across the developer community spans healthy skepticism to long-term optimism, mirroring a lot of what we've seen in response to AI advancements, ranging from disdain and disbelief to  nodding acceptance. But zero bugs? If AlphaProof Code looked at my code and found no errors, I'd assume the verification itself was flawed. That'd be as impossible as sending one of my draft blog posts to an editor and being told it was perfect.

More broadly, AI systems can now be used to verify human-written libraries, and neural models can now write top quality code, taking the spot we've reserved for decades for well trained and well paid humans.

This shift connects directly to what I explored in Vibe Coding: Automatic Transmission for Apps, where higher abstractions free engineers from manual routine mechanics to focus on high-level architecture. It also reinforces the broader reality I wrote about in An AI Tech Revolution? The Solution: Human Evolution. Technology shifts require human adaptation rather than passive observation. As automated verifiers mature, running unverified raw code on production networks may soon be as outdated as leaving my index directories open with r-w-x (read, write, execute for you non-nerds) to the whole web.

Bug free or not on the first try, we are seeing AI do more. AI is writing, reading, coding and now debugging. I'm betting that even if this pass manages to come back with zero known issues, the definition of what is a bug and what is not will change. 

Disclosures, per always: I work at Google and use Gemini and Antigravity tools a lot. DeepMind is a huge part of Google's AI effort and behind AlphaProof. Nobody at Google reviews my posts, ever.

October 10, 2026

October 10, 2026 · 3 MIN READ · BY LOUIS GRAY

Vibe Coding: Automatic Transmission for Apps

Vibe Coding: Automatic Transmission for Apps


The Human–AI Co-Creation Studio [Image via Gemini]

As a teen training to get my driver's license, my dad wanted me to learn "the right way". To him, that mean getting my first experiences behind the wheel with a stick shift, learning how to coax his sturdy and unimpressive 1990s era Mazda 323, with manual transmission, from low gear to highway cruising speeds, all while avoiding stalls and switching gears smoothly. Through this training, many trial and errors were had as I gained experience.

Eventually, I took and passed the drivers' test, but in an automatic vehicle - my mom's wood paneled Oldsmobile station wagon. Despite its heavy build and thoroughly unattractive appearance, the car's automatic transmission reduced complexity for me, letting me focus on driving the car, not just fighting to get it to work. Now, depending on which survey you choose, less than 1/5th of drivers know how to use a manual transmission, and those cars account for only 2% of all sales in the United States. While there will always be strong opinions about the efficiency or quality of manual vs automatic, from a technology perspective, it's a relic of a previous era.

The rapid ascent of AI-assisted coding, much like that transition from manual transmission to automatic, has reduced complexity and expanded the total market for participants -- not only changing how software engineers work, but making it possible for those without coding experience to speak their ideas into existence, and iterate to make them real. If you can think it, you can build it.

While some debate if AI is conscious or sentient (it's not), it certainly isn't yet fully autonomous. Working with any of the leading AI-powered code environments is a continual give and take between the LLM and the human, a hybrid approach Sean Goedecke calls our "centaur age", reflecting the reality of a time when the machines are so fast and so capable, humans can't directly compete alone but work in tandem with their silicon partners.

Goedecke says: 

"It is no longer possible to credibly argue that an unassisted software engineer can beat a centaur team. The value provided by coding agents is just too great: not just the speed of producing code, but the speed of review and testing."
-- Software's centaur age may last decades

What I have experienced, both at the office, and in my own side projects, is that AI isn't just changing the end product of what we build, or speeding our ability to eliminate tasks -- but more importantly, enabling us to do things we simply were unable to before, expanding our scope and redefining what we could achieve. In the last three years, we've graduated from simple question and answer chatbots to progressively capable workspaces that are waiting to take our instructions, and agents eager to do our bidding.

As I shared one recent experiment, a friend messaged me, "someone has been having fun vibe coding" and I almost recoiled from the description. Even if I never touched an official single line of code, I was able to turn to a software tool, design an idea, work with available technology, and deliver a project users could experience themselves. This was not just for vibes, but a celebration of capability.

The complexity that has defined the world of software development for decades was never the end goal, but a necessity to get there. Now, that journey of pain has been greatly reduced. No more stalling out, no more grinding of the gears. There will always be elite software development. People who have specialized in the field are going to approach invention in ways we amateurs wouldn't think to, due to their education and experience. But vibe coding is a lot more than frivolous toying based on mood -- it's automatic transmission for development, which will put a lot more drivers on the road.

Disclosures per always: I work at Google and vibe code with Gemini and Antigravity tools a lot. I wrote this post by hand.