It felt like a dream come true talking to Logan Kilpatrick last week—a notable figure in the AI space, aka Logan GPT on X, i.e. prior to joining Google Gemini from OpenAI on April 1, 2024—‘April Fool’s Day,’ which is hilarious.
Now, he is simply Logan Kilpatrick, minus GPT.
(Screenshot from a recent chat with Logan Kilpatrick. (Yes, he’s real, not AI-generated!))
The moment felt both exciting and relaxed, with Kilpatrick’s background adding a warm, natural touch. Behind him were a couple of vibrant plants—a leafy Monstera and a small succulent—bringing a sense of calm and simplicity to the setting.
“It feels great, going back to the AI roots,” said Kilpatrick, in an exclusive interaction with AIM, sharing his experience working at Google, alongside Matt Velloso, Google’s vice president of product for AI/ML development, whom he admires the most, besides Jeff Dean, chief scientist at Google DeepMind and Google Research; and Demis Hassabis, CEO of Google DeepMind.
Google Now Ships Faster than OpenAI
The timing could not have been perfect. At the recent All-In Summit 2024, Google founder Sergey Brin mentioned that Google was quite behind a couple of years ago.
“ChatGPT launched, and we were quite behind,” he added, saying how OpenAI pushed Google to accelerate its AI development efforts, and Brin, alongside Larry Page are now full-time at Google unleashing founder-mode, overseeing all the latest advancements, shipping products like there is no tomorrow.
“They helping build Gemini is awesome to see,” said Kilpatrick, who comes from the best of both the AI worlds–OpenAI and Google–and there is no turning back. Ironically, many even quipped, saying OpenAI stopped shipping after Logan’s left the company, and started calling OpenAI a blog post company.
Kilpatrick humbly redirected the conversation, and gave credit to the team, and said: “OpenAI just had that muscle, and now Google is building that muscle.”
Citing OpenAI’s GPT-2, GPT-3 and others, he said they were externalised in various capacities. “The original Gemini model came out only in December, and it hasn’t even been a year,” said Kilpatrick, throwing light at how Google has been shipping its products non-stop.
There is no stopping Google. The company recently introduced three experimental Gemini models to enhance speed, accuracy, and the ability to handle complex prompts. The new models are Gemini 1.5 Flash-8B, Gemini 1.5 Pro, and Gemini 1.5 Flash. Google has expanded Gemini’s capabilities, allowing users to fine-tune models via an API key and integrate them with other projects in Google AI Studio and the Gemini API.
Google recently introduced DataGemma, a new open model that integrates LLMs with real world data from its Data Commons repository, using retrieval augmented methods like RIG and RAG to reduce AI hallucinations and improve the accuracy of generative AI outputs in research and decision making contexts.
When is Google releasing Gemini 2? “That’s the Magic question,” quipped Kilpatrick. However, he said that the company plans to release Veo, Google Search Grounding, Gemini 2.0, and Agents next, tentatively expected to debut in the coming months. “I think it’ll be fun to see what the next six months to a year look like as the trend of unlocking new developments continues,” he added.
OpenAI is Nothing without Google DeepMind
Just when we thought OpenAI wasn’t shipping, Sam Altman and his team took the world by surprise—all thanks to Jimmy Apples patience—decided to drop o1 (internally known as Project Strawberry/Q*), a new language model trained with reinforcement learning (RL) to perform complex reasoning—in just a couple of hours after we spoke to Kilpatrick.
After Transformer—which was first published by Google in the ‘Attention Is All You Need’ paper back in 2017—OpenAI seems to have once again incorporated one of Google DeepMind’s most popular reinforcement learning school of thoughts, pushing Google to accelerate its timeline for releasing their advanced models.
“Anytime something cool and new launches, I always take a few minutes away from my desk to let the moment sink in…,” said Kilpatrick, “now, back to shipping ; ),” congratulating OpenAI team on o1 and o1-mini release.
He further said: “the universe does not want things to be shipped, but you did it.”
Coincidently, a few days ago, Kilpatrick took a jab at critics who claimed that Google lacks innovation, highlighting that the company was the first to ship a 1 million and 2 million context window, a state-of-the-art multi-modal LLM, context caching, and a high-quality small model for developers called Flash. “So yeah, definitely no innovation happening here…..,” he quipped.
Brin also echoed similar views recently and said that they were too timid to deploy the technology for many good reasons. However, Google has now changed its strategy and is focusing on shipping products quickly.
Google DeepMind, FTW!
“I think when I joined OpenAI, it was a 100-person company, and when I left, it was much larger, 1,500. In many ways, coming to Google has felt like going back to the roots of that 200 person startup,” said Kilpatrick. He said that despite Google being big and having a lot of people, the number of people working on Gemini is smaller.
“Obviously, Google is big, and there are lots of people doing it, but I think folks who are actually focused on, like, Gemini, and specifically developer within Gemini, is a much smaller, much more intimate group to me.”
He said it’s like a startup inside the company.
Kilpatrick further revealed that the Google DeepMind team ultimately wants to see the technology in the hands of both developers and the wider world. “They care a lot about making sure that the product teams building on top of the models are kept in the loop with what’s coming.”
“I think a unique differentiator for Google is bringing research and product development as close together as possible. This collaboration fosters the creation of better products,” he added. The same is not the case with OpenAI, as they have stronger internal silos, and with Google, teams collaborate more freely, said Kilpatrick, at the recent interview on ‘The Cognitive Revolution’ podcast.
Further he said AIM that Google is actively taking feedback from the developer community and has been implementing it. “We’ve received feedback on improving multimodal capabilities, and we’ve actually seen some of that feedback come to fruition with bounding boxes, tool use, and function calling. The last piece is code performance, and I think we’ve made significant strides in that area as well,” he concluded.
