The AI Week That Changed More Than It Looked Like It Did

The Programmer's Dilemma: Is Software Engineering About to Shrink?

Former Harvard professor and Google engineer Matt Welsh has argued that large language models will automate enough coding work to meaningfully reduce the number of programmers the industry needs. He's not talking about tomorrow — his estimate runs four to fifteen years — but the direction, in his view, is clear.

It's worth pausing before accepting that framing wholesale. Predictions about tech roles being eliminated have a poor track record. Low-code platforms were supposed to make developers obsolete. They didn't. Instead, they shifted what developers spend their time on and opened new categories of work. The same pattern played out with database tools, cloud infrastructure, and automated testing.

So why does this moment feel different? Partly because LLMs don't just automate a narrow slice of the job — they compress the full loop from idea to working code. That's a qualitative shift, not just a faster version of the same thing.

The broader social math is also worth noting. Cheaper, faster software creation is genuinely good for people who need software but can't currently afford to build it. The disruption, if it comes, won't be costless for engineers. But it probably won't look like elimination either.

The real question isn't whether AI changes software engineering — it's how fast and how deeply.

The Inventor Who Grew Tired of His Own Invention

In 2017, eight researchers at Google published a paper called "Attention Is All You Need." It introduced the transformer architecture that now underpins almost every major AI system you've heard of. Eight authors. One paper. An industry remade.

Here's the strange part: Llion Jones, one of those eight co-authors, has co-founded a Tokyo-based startup called Sakana AI — and its mission is, in part, to explore what comes after transformers. Google has since made a strategic investment in the company, which is its own kind of irony.

Sakana's research pursues bio-inspired approaches to AI. One of its projects, the Continuous Thought Machine, departs from standard transformer architecture in meaningful ways. Another, called the AI Scientist, has produced autonomously generated research that has gone through peer review — a concrete, if contested, benchmark for what AI systems can now do. Their ALE-Agent has also shown competitive performance in coding contests.

Only two of the original eight transformer authors are still at major AI labs. The others have scattered to startups, academia, and independent research. That's not a scandal — it's just how fast this field moves.

When the people who built the dominant tool are actively questioning it, that's worth paying attention to.

Big Tech's Realignments and the Debates Nobody Expected

Apple is reportedly planning to announce Siri features powered by Google's Gemini. Read that sentence again. Two companies that have spent years competing fiercely on hardware, software, and services are now, apparently, sharing AI infrastructure. It signals something real about the limits of building everything in-house — even for a company with Apple's resources.

Meanwhile, ChatGPT has been pulling content from sources like Grokipedia, highlighting the increasingly tangled questions around AI data sourcing. The era of clean, self-contained AI ecosystems — if it ever truly existed — looks to be over. Even the largest players are cross-pollinating. For users, this means more capable features. It also means harder questions about where the data comes from and who is accountable when something goes wrong.

The cultural debates are catching up just as fast. Pope Leo XIV publicly cautioned against "overly affectionate" AI chatbots — and while that may sound easy to dismiss, it reflects a concern that psychologists and ethicists have been raising in less quotable language for some time. The worry isn't that people use AI. It's that some people are beginning to prefer it to human interaction.

A recent Gallup poll found that 12% of American workers use AI daily, with the highest concentration in tech, finance, and education. That's not a majority. But it's not a rounding error either.

On the creative side, Comic-Con and several science fiction competitions have moved to ban AI-generated entries. The argument isn't always about quality — it's about authenticity. What counts as human work is now a live policy question, not just a philosophical one.

The most interesting AI debates right now aren't technical — they're about what we actually want from these machines.

What to Watch Next: Habits, Voices, and the Tools Worth Your Time

One underused application of AI is surprisingly simple: using it to warm up your thinking before the day's noise sets in. Instead of opening email first thing, try asking an AI to surface one small detail worth noticing today, or to push back on an assumption you've been holding. It shifts your mental mode from reactive to curious. Rotating the prompt weekly helps avoid the habituation that makes any habit go stale. Running the same prompt with a team can surface unexpectedly different perspectives from people you thought you knew well.

On the tools front, several voice synthesis models are worth watching. Qwen3-TTS is designed for local use. PersonaPlex-7B allows customizable voice personas. Inworld TTS-1.5 prioritizes expressive speed. Chroma 1.0 offers near-instant voice cloning. Taken together, they suggest that voice AI is moving from a demo feature to something closer to daily infrastructure — faster than most forecasts predicted.

Also worth noting: a humanoid robot fighting tournament called REK1, which is either a publicity stunt or a genuine stress test for real-world robotics depending on your tolerance for spectacle. APEX-Agents is testing AI on professional tasks in ways that go beyond abstract benchmarks. Stanford's AI4ALL program is introducing ninth graders to hands-on AI work, which matters more for the long run than most product launches.

OpenAI's new Codex tools ship with built-in misuse safeguards — a sign that responsible deployment is becoming a competitive differentiator, not just a compliance checkbox.

The tools that look like novelties today have a habit of becoming defaults before anyone quite decided to make them one.