π€ AI girlfriends will be the new normal
Dear curious minds,
In contrast to last week, there has been a lot of happening this week in the generative AI space. OpenAI scheduled their spring update event one day before the Google I/O conference. Both events showed many improvements and new releases for their AI models and tools.
In this issue:
π‘ Shared Insight
If She Does Not Live on Your Localhost, She Is Not Your AI Girlfriendπ° AI Update
Big update from OpenAI, especially for free users
Google I/O Conference: Upgrades to Gemini and Gemma Model Familiesπ Media Recommendation
Her: The AI Love Story That Showed Us Our Future
π‘ Shared Insight
If She Does Not Live on Your Localhost, She Is Not Your AI Girlfriend
The latest OpenAI release (summarized in a section below) introduced a more emotional sounding female voice representing their latest chatbot. It may sound strange, but I believe that people will be able to connect emotionally with AI characters. George Hotz, a well-known security hacker and entrepreneur, stated already ten months ago in the Lex Fridman podcast that there is a very high market potential for AI girlfriends and that he likely will work on it as his next project.
A recent post on π made me smile and think:
If she doesn't live on your localhost, she's not your AI girlfriend; she's everybody's.
β Steven Fan
And maybe even more concerning than being everybodyβs girlfriend, your AI girlfriend who is hosted in the cloud might be unavailable or changed without anything you can do about it. The latter happened to users of one of the so far most famous services for virtual AI companions - Replika.
Replika was founded in 2017 with the idea of creating a personal AI to help people express themselves through meaningful conversations. It provides a space for users to share their thoughts, feelings, and experiences. Users interact with Replika in various ways. The free version offers Replika as a "friend," while the paid premium tiers offer roles such as "partner," "spouse," "sibling," or "mentor." In January 2022, they reported 10 million registered users.
Many users have a romantic relationship with their Replika chatbot. An update of the AI guidelines in February 2023 resulted in a change where many users felt that they had lost their loved AI companion. The chatbot responses changed, and especially erotic role-play was prevented. Many users were frustrated and shared their losses and feelings on Reddit and other social media platforms. Three months later, Replika allowed users who created their AI companion before the update to switch to the previous version of their chatbot, as reported by Reuters.
This story shows once more that you cannot rely on any cloud-based models. The company behind these models can change or even turn them off at any time. This is especially devastating if you created an emotional bounding to the service, like many users did with their Replika. The only option where you can be sure that the service will not change is running the model locally, or at least have the ability to do so. This is possible for chatbot with openly released weights and especially the latest Llama 3 release showed that these models are catching up and are not far behind the state-of-the-art closed models from OpenAI, Google and Anthropic which only run in their clouds. For your AI girlfriend or any other chatbot where you intend to share very personal details with, you prefer running it on your own hardware. Besides no one being able to change the service or turn it off, your data is also fully under your control and not used by anyone else.
π° AI Update
Big update from OpenAI, especially for free users
OpenAI announced in its spring update the launch of its newest flagship model, GPT-4o. The model provides GPT-4 level intelligence but is faster and improves on its capabilities across text, voice, and vision.
Free users will also get access to GPT-4o and on top to many other features so far only usable for Plus subscribers: web browsing, data analysis, image discussion, file uploads, GPTs/GPT Store, and memory features. But the number of replies that free users get from GPT-4o will be more restricted.
Voice Mode with natural, real-time voice conversation and the ability to a conversation with ChatGPT via real-time video is shown in several released examples. It will be available in an alpha in the coming weeks for Plus users.
A new ChatGPT desktop app for macOS is being launched, initially only for Plus users. The Windows version is announced to follow later this year. This app allows users to ask questions using a global shortcut, take screenshots and discuss them directly in the app, and have voice conversations.
GPT-4o can also be used via the API, currently as a text and vision model. It only costs with $5 per 1M input and $15 per 1M output tokens, half the price of its predecessor, GPT-4 Turbo.
The GPT2 model which showed up in the Chatbot Arena a while ago was an early the newly released GPT-4o. This was confirmed by William Fedus from OpenAI in a post on π.
Sidenote: With Ilya Sutskever, Chief Scientist, and Jan Leike, co-lead of the Superalignment team, two employees in important positions, decided to leave OpenAI. So far, there is no reason publicly announced. However, Ilya was more or less invisible after his vote on the board to replace Sam Altman as CEO.
My take: The new possibilities to interact with ChatGPT make its usage way more user-friendly. Furthermore, making the latest model and many other features available to free users will help to spread public understanding. GPT4o is really fast and usage via API costs only half the price of the latest GPT-4 Turbo model. Likely this is a sign of its size and there is a bigger model still being trained. The same as for the Llama 3 405B model, which was announced but not released so far by Meta.
Google I/O Conference: Upgrades to Gemini and Gemma Model Families
Googleβs I/O conference took place on the day after the event from OpenAI. There were various improvements and new introductions to the Gemini (cloud) and Gemma (local) model families announced. A summary is given in their blog article.
Their so far strongest model, Gemini 1.5 Pro, with the impressive context size of 1M tokens, got even better. Google states that the new version improved across translation, coding and reasoning.
Furthermore, Gemini 1.5 Flash, a smaller model optimized for faster response times, was introduced. The costs are only $0.35 per 1M tokens.
PaliGemma: Google's first open vision-language model for image captioning, visual Q&A and other tasks which use an image as additional input.
Gemma 2 (launching in June): A 27B parameter model is currently trained, in addition to the previous released 2B and 7B Gemma models. Like the other Gemma models, the weights will be released and with that, you can run it locally.
A new feature named content caching was announced for the Gemini API, to be released in June. This will allow sending parts of the prompt, especially long files, only once to make working with long contexts more usable and affordable.
Google showcased a demo of Project Astra, which is the direct competitor of the new interactive video mode of ChatGPT. At the end of the video, a pair of glasses was used to run the demo.
My take: The extended availability of the improved Gemini 1.5 Pro model is likely the biggest news. Outstanding is the context size of 1M tokens with nearly perfect retrieval. Personally, I look forward to the release of the Gemma 2 model with 27B parameters, as this size should be the largest to fit with a quantization of 4 bits in the 24GB GPU memory of my NVIDIA GeForce 3090.
π Media Recommendation
Her: The AI Love Story That Showed Us Our Future
"Her" is a romantic science fiction drama movie from 2013 (IMDB rating 8.0). It is about a man who forms a deep relationship with his AI voice assistant. To get a better idea of the movie, you might want to watch the trailer.
The movie is currently not included on any of the major streaming platforms. However, you can buy it for a reasonable price on Apple TV and Amazon Prime. A more complete overview of the possibilities to get the movie is given on the WerStreamt.es website.
My take: With the recent release of GPT-4o showcasing its capabilities for fluent conversations, the plot of the movie seems to be closer than ever before. Actually, many people pointed in replies of OpenAI posts on π to this movie. A few people even created a versions of the movie poster showing Sam Altman, OpenAIβs CEO.
Disclaimer: This newsletter is written with the aid of AI. I use AI as an assistant to generate and optimize the text. However, the amount of AI used varies depending on the topic and the content. I always curate and edit the text myself to ensure quality and accuracy. The opinions and views expressed in this newsletter are my own and do not necessarily reflect those of the sources or the AI models.