The road to AGI is paved with Silicon Valley acquihires; Google is training Veo on YouTube content; AI search optimization is gaining momentum; women are lagging behind on AI; EU's waffle on AI

How can people in entry jobs level up in the age of AI; workers in the UK told to embrace AI; tensions between OpenAI and Microsoft boil over; AI companies reimagined as Wizard of Oz characters

Jun 20, 2025

∙ Paid

If you want to see capitalism’s version of online dating, watch a giant tech company acquihire an AI startup. In one swipe right, they get the cap table, the code base and most importantly, all the caffeinated PhDs who keep the best snacks in a secret drawer for themselves.

For Big Tech it feels like an efficient, almost inevitable process. For me, it’s a red light flashing on the road to AGI; and here’s why.

Let’s start with the echo chambers of Silicon Valley. The Bay Area has a huge problem with not invented here (NIH) syndrome, going back to the days where actual silicon was being designed and manufactured there. (As an aside, the media is incredibly guilty of this too, as it propagates—and sometimes participates in—the echo chamber by obsessively covering every trivial detail or minor update as if it’s breaking news, refusing to accept innovation can happen elsewhere or founders can build successful companies outside the San Francisco postcode.)

Acquihired teams are prized exactly because there’s a belief they will fit right in. That’s code for shared resumes (OpenAI, Google DeepMind, Stanford, MIT), libraries (PyTorch), and assumptions. It speeds execution but squeezes out heterodox thinking, the very stuff breakthroughs are made of. If AGI really demands fresh cognitive maps, building a monoculture (or let’s be honest, a cult) of credentialed disciples won’t cut it. I like to call it the gradient descent to groupthink risk: everybody converges to the same local minimum, and genuine novelty never appears.

Then there’s the cultural graft rejection. Having worked in hyperscale-big tech, small-big tech and small-small tech, I can confidently say that startups run on adrenaline and informal consensus, scaleups run on product market fit, and hyperscalers run on compliance checklists and quarterly earnings calls.

The moment the badge printer spits out a Big Tech logo, incentives flip: equity upside shrinks, promotion ladders loom, and risk tolerance flat-lines. Many star hires quietly head for the exit before year two, often starting the next startup Big Tech will feel compelled to buy. That revolving door is an expensive churn, not a flywheel. Of course, it works for the VCs who cash in but does it actually lead to AGI-type innovation?

Thirdly, we have the dilution of accountability. When talent comes pre-packaged, managers stop investing in their own benches. Internal AI researchers see the writing on the wall (*cough, FAIR, cough*): the only way to get budget is to quit, raise a seed round, and sell back to the mothership. In corporate finance that would be called share buybacks at the top. In HR, it’s just lazy. Over time the acquirer forgets how to grow engineers the old fashioned way, leaving it dependent on a pipeline it doesn’t control.

More recently, this type of ecosystem has also been forced to learn how to navigate a growing network for antitrust tripwires. US and EU regulators have discovered a new hobby: adding “killer acquisition” case files to their reading stacks. A few years ago, it used to be that Big Tech could go on a shopping spree largely unbothered. Now, snapping up every promising AI outfit isn’t possible any more; it’s an engraved invitation to DOJ and DMA scrutiny. An AGI-scale investigation could freeze integrations for months, stretching technical debt and demoralizing both sides.

Then you have the problem with ROI. To build its new superintelligence lab, Meta is making offers that are rumored to top $100 million—of course, most of that goes towards options and hiring bonuses rather than annual salaries. Even so, that might be defensible for a 10x productivity gain, but most AGI roadmaps still resemble science projects, not revenue streams. Paying late-stage prices for pre-product talent is the human capital equivalent of buying Nvidia B200s at eBay scalper prices: sometimes necessary, never optimal.

Finally, Silicon Valley’s strength is its modular, competitive mess. If the exit path for every wannabe frontier model maker is a single buyer auction, outside investors eventually stop funding them, and the funnel dries up. The result: fewer weird bets, fewer independent checks on safety, and a public that must trust a shrinking number of labs to align intelligence superior to theirs. Concentrated power plus opaque research agendas is how you get policy backlash or worse, technical blind spots with planetary blast radius.

You may agree or not with acquihiring as an operational strategy but hopefully there’s no denying of a simple reality: when you buy the team, you buy the baggage. Startups pivot fast because nobody is wedded to legacy code; Big Tech’s stack, by contrast, is a fossil record. Merging the two often means rewriting from scratch or, worse, kludging bridges that slow everything down. In-house M&A experts will tell you the dirty secret: half the acquihired code ends up deprecated before the first internal demo. Paying for IP you can’t use used to be an old Wall Street tradition, but now it’s bleeding into AI.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

Keep reading with a 7-day free trial

Subscribe to Computerspeak by Alexandru Voica to keep reading this post and get 7 days of free access to the full post archives.