Gemini Launches; The Model Race Is Real

Archive note: this is a backdated post, written years later while rebuilding this site. It’s dated to the moment it covers, but the hindsight is real.

On December 6, Google announced Gemini — the long-promised answer to GPT-4, a year after ChatGPT caught the giant flat-footed. Three tiers: Ultra (the GPT-4 challenger, claiming benchmark wins including a milestone MMLU score), Pro (now powering Bard), and Nano (small enough to run on the Pixel itself — note that one; it rhymes with everything the local-model year proved).

And then there’s the demo.

The video that wasn’t

The launch’s centerpiece was a mesmerizing video: Gemini watching a person draw, play cups-and-ball, make a duck — responding fluidly, in real time, by voice. It traveled everywhere. Within days, Google confirmed what careful viewers suspected: the interaction was edited — built from still frames and text prompts, with latency removed and the voice layered on. The capabilities have some basis; the experience shown does not exist.

I keep a whole shelf of these on this blog — Zoom’s “end-to-end encryption”, Cyberpunk’s console embargo — and the lesson never changes: the gap between demo and product always gets discovered, and the discovery always costs more than honesty would have. That Google — sitting on more real capability than almost anyone — felt compelled to fake the demo anyway tells you about the pressure inside this race more than any benchmark table does.

The actual significance

Stripped of launch theater, December’s real news:

1. There are now two-plus frontier labs, demonstrably. For a year, “GPT-4-class” was a category with one member, and November showed the systemic risk of that. Credible competition changes pricing, pace, and — for builders — finally makes the multi-vendor abstraction layer worth its keep. My MODEL_PROVIDER env var from last month’s post gains its second value.
2. Distribution is Google’s actual weapon. Models race on benchmarks; Gemini ships into Search, Android, Workspace — surfaces with billions of users who never chose an AI product. OpenAI built the better funnel; Google owns the water supply.
3. Nano is the quiet tell. A first-party model running on-device, in a flagship phone, today. Between this, Apple’s silence (never mistake it for absence), and the quantized open-model wave, 2024’s battleground is visibly local.

Closing the 2023 ledger

A year ago I banked a prediction: AI in my daily toolchain within a year, with the interesting fights about trust and verification. Scoring honestly: the first half landed early and harder than I guessed — GPT-4 pairing is simply how I work now. The second half arrived as predicted but at civilization scale rather than code-review scale: the letters, the boardroom weekend, and now a faked flagship demo to close the year.

2023 prediction, banked for next December: the frontier race stays a two-or-three-horse spectacle, but the useful story compounds at the floor — open weights, on-device tiers, and boring integrations into stacks like mine. The spectacular and the useful keep diverging; this blog will keep tracking the second one.

What a year. The library learned to talk in November ‘22; in ‘23, everyone built it a city. Next: who gets to live there.