February 14, 2026
Beak-to-beak AI smackdown
Gemini 3 Deep Think drew me a good SVG of a pelican riding a bicycle
Google’s bird-on-a-bike sparks cheers, jeers, and ‘rigged’ vibes
TLDR: Google’s new Gemini 3 spit out a shockingly clean pelican-on-bike image in simple SVG code. Commenters are split: some cheer the craft, others say it’s cherry‑picked or in the training set—raising big questions about real progress versus demo magic.
Google’s new Gemini 3 “Deep Think” says it pushes the frontier of brainy tech—but today’s headline act is delightfully silly: it drew a crisp SVG of a pelican riding a bicycle. SVG (Scalable Vector Graphics) is code that makes zoom‑friendly images, and this one looks clean and rideable. The crowd turned the result into a spectator sport. One fan called the bird “pretty fly,” while another joked the AI race is now “benchmaxxing pelican on bike SVGs” like those notorious Will Smith spaghetti videos. Translation: people are using goofy tasks to compare bots—and they’re keeping score.
But the cheers ran into side‑eye. A skeptic argued this isn’t a real test anymore because “it’s probably on everyone’s training set,” hinting the models may have seen similar pelicans already. Another dropped the bomb: “They rigged it,” accusing cherry‑picked demos. Meanwhile, a practical crowd praised that the bike looks genuinely “cyclable” and the bird is “sliced and ready to bbq”—translation: tidy, usable layers. The vibe? Classic AI culture clash: dazzled fans, data‑detectives, and conspiracy‑whisperers. If Google’s Gemini can nail a bird on wheels, what else can it do—and what counts as a fair win?
Key Points
- •Google has introduced a model referred to as “Gemini 3 Deep Think.”
- •Google describes the model as designed to push the frontier of intelligence for science, research, and engineering.
- •The author tested the model by prompting it to create an SVG of a pelican riding a bicycle.
- •The resulting SVG is reported by the author as the best they have seen for this prompt.
- •The article links to the author’s previous collection of similar outputs for comparison.