one year on
Google unveils Gemini AI model as Bard gets Gemini Pro
The search giant launches Gemini and says Bard will use Gemini Pro, while Hacker News commenters trade examples of strong answers and occasional hallucinations.
Google DeepMind today launches Gemini, and says Bard will use Gemini Pro starting today.
Early user reports are mixed. On Hacker News, where the announcement drew over 2100 points and 1600 comments, some users reported impressive results. One commenter noted that Gemini correctly identified an impossible TypeScript feature and linked to a relevant GitHub issue, while another found it gave a better technical answer than ChatGPT for a sandbox directory question. However, others reported hallucinations and pointed out that Google’s demo video showing Gemini fluidly tracking a cup shuffle was staged: the company confirmed it used still images with carefully crafted prompts.
The launch positions Google as a direct competitor to OpenAI, but questions remain about how the model will perform at scale. The staged demo has already sparked debate about whether Google is overselling capabilities, reminiscent of criticisms leveled at other tech demos in the past.
Reported that the model correctly identified an impossible TypeScript feature and linked to a GitHub issue.
Reported that Bard suggested mounting a filesystem for a sandbox directory problem, which they found better than ChatGPT's sticky-bit suggestion.
Pointed out that Google's demo video showing Gemini's fluid real-time interaction with a cup shuffling task was staged using still images with carefully crafted prompts.
One year later — open only if you can handle spoilers
Gemini would later evolve into a family of models including Gemini Pro, Ultra, and Flash, with ongoing improvements. The staging controversy was quickly overshadowed as the model gained adoption in Google products and third-party applications, though it never unseated GPT-4 as the community's default benchmark for language model quality.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy