🍓 Rumors of OpenAI’s Newest Model

Sam Altman’s recent post on X, featuring an image of strawberries with the caption “I like summer within the garden,” has sparked a frenzy of speculation about OpenAI’s rumored recent foundation model, code-named “Strawberry.” This seemingly innocuous post has excited the AI community, suggesting potential confirmation of OpenAI’s next major development, possibly the anticipated GPT-5. This follows whispers of a more ambitious project, codenamed “Q*,” aiming for AGI—a digital system able to outperforming humans in economically significant tasks.

Adding to the intrigue is a brand new anonymous chatbot model appearing on the open-source LMSYS Arena, reportedly exhibiting advanced reasoning beyond GPT-4o. This recent development follows nearly 17 months since OpenAI released GPT-4, with competitors like Anthropic, Google, and Meta gaining ground. The AI community watches with bated breath as the corporate attempts to keep up its lead within the AI race. Read more.

Andrej Karpathy, a distinguished AI researcher and former OpenAI scientist, has scrutinized the present method of coaching language models, Reinforcement Learning from Human Feedback (RLHF). Karpathy likens this system to teaching a parrot to parrot human speech, critiquing it for counting on human input to direct AI behaviour as an alternative of nurturing true intelligence.

He draws a stark contrast between RLHF and the “real RL” used to coach DeepMind’s AlphaGo, which involved the AI learning through self-play, optimizing its performance by simulating 1000’s of games, and developing superhuman capabilities without direct human intervention.

Karpathy contends that RLHF is basically a “vibe check,” aligning AI behavior with human preferences quite than equipping it with the tools for independent thought. He warns that this reliance on human guidance may limit the AI’s potential to evolve into a very autonomous problem solver. Read more.

Prompt Poet, acquired by Google from Character.ai, is redefining prompt engineering for big language models (LLMs). This tool simplifies the creation of sophisticated prompts through a low-code interface, making it accessible to each technical and non-technical users. It integrates YAML and Jinja2 for flexible prompt templates and efficient context management, allowing users to include external data seamlessly. In practice, Prompt Poet streamlines prompt design by managing complex instructions and dynamic data, reducing manual input errors and enhancing model responses. Tested with various scenarios, akin to every day planning assistants, Prompt Poet demonstrates significant improvements in context-aware output. This development not only showcases a step forward in AI prompt engineering but in addition hints at Google’s shift towards more intuitive and accessible AI tools. Read more.

Apple is poised to introduce its generative AI model, Apple Intelligence, with analysts predicting a premium tier carrying a $20 price tag. Slated for a late-year launch, Apple Intelligence goals to reinforce Siri’s capabilities and introduce generative features like automatic email and image creation.

While the rollout will initially exclude some regions, a subscription model could also be employed for premium features. Analysts suggest that Apple might incorporate these features into the prevailing Apple One subscription, which costs $19.95 per thirty days and bundles various services. Apple’s services division, which generated $24.2 billion within the last quarter, could leverage this recent AI system to further monetize its offerings. The strategy aligns with industry trends, where firms like OpenAI and Microsoft also charge for advanced AI services. Apple’s move is seen as a way to deepen user engagement and generate additional revenue. Read more.