The subsequent step for OpenAI’s reasoning fashions is o3, a mannequin previewed on Dec. 20. o3 and its smaller cousin, o3-mini, outperformed o1 in coding, math, science, and ‘conceptual reasoning’ assessments designed to evaluate human-like intelligence and analysis purposes. ‘Reasoning’ features a security function referred to as deliberative alignment, wherein the mannequin makes use of a “chain of thought” to forestall customers from jailbreaking or tricking it into bypassing security measures.
In the meantime, Google’s Gemini 2.0 Flash Considering Experimental mannequin treads related floor to OpenAI o1’s reasoning capabilities.
‘12 Days of OpenAI’ brings new instruments and new generative AI performance
The o3 announcement got here on the shut of OpenAI’s “12 Days of OpenAI” marketing campaign, a vacation season sequence of product updates. These bulletins, from Dec. 5 to Dec 20 (excluding weekends), showcased new options for OpenAI’s generative AI instruments, with some accessible now and others nonetheless in testing.
Day 1: The $200 ChatGPT Professional and o1 updates
On Dec. 5, OpenAI launched a brand new subscription tier for ChatGPT: the Professional plan. For $200 monthly, the Professional subscription brings OpenAI o1, o1-mini, GPT-4o, and Superior Voice to ChatGPT. It additionally permits entry to o1 professional mode, a extra compute-intensive model designed for troublesome issues skilled engineers and researchers face.
On the identical day, OpenAI introduced an up to date, extra detailed system card for the hotly-anticipated o1 mannequin.
Day 2: The Reinforcement Effective-Tuning Analysis Program
With the Reinforcement Effective-Tuning Analysis Program, OpenAI launched a brand new software for builders and machine studying engineers to create personalized fashions for particular duties. It’s anticipated to launch publicly in alpha testing in early 2025.
Day 3: Sora video generator
OpenAI’s photorealistic video generator, introduced early final 12 months, is now accessible for ChatGPT Professional customers. Whereas AI video creation is simpler than ever, fashions like Sora nonetheless battle with complicated, fast-moving topics and may usually be recognized by a too-perfect glossiness. Sora movies will likely be watermarked in line with C2PA requirements to establish them as AI-generated.
SEE: Study the fundamentals of generative AI with a number of the many free programs accessible from Microsoft and LinkedIn, up to date for 2024.
Day 4: Canvas
Canvas, a coding interface launched in beta in October, turned usually accessible in December. The present model of Canvas understands and writes Python and integrates with customized GPTs, permitting builders to connect with their apps. It additionally permits customers to view prompts and outputs side-by-side for simpler reference.
Day 5: Apple on-device AI with ChatGPT
Apple Intelligence obtained its anticipated ChatGPT replace in the course of the 12 days of OpenAI. The on-device Apple Intelligence can now entry ChatGPT servers for extra complicated queries that the onboard chip can’t deal with.
Day 6: Superior Voice with Video
Superior Voice mode, accessible to ChatGPT subscribers, can now converse about pictures in your pc display or by way of your digital camera. The mode brings extra pure speech and versatile responses to the audio model of the chatbot.
Day 7: Tasks
As of Dec. 13, ChatGPT Plus, Professional, and Crew customers can manage their chats into Tasks, or separate cases. Tasks let customers assign particular directions that apply solely inside Mission, and related sources will be saved with it. This function will likely be accessible to Enterprise and Edu customers in January.
Day 8: ChatGPT search upgrades
ChatGPT search obtained a number of tweaks after the December launch, together with a brand new maps interface, speedier response occasions on cell, and extra performance for Superior Voice to carry search up to the mark with the remainder of the paid-tier voice choices. Search is now accessible to customers on the free tier, so long as they log in with an electronic mail deal with.
Day 9: New options, choices, and upgrades for builders
Day 9 was all about builders, with a wide range of bulletins:
- Builders can now entry OpenAI o1 within the API.
- Numerous upgrades for the API had been launched, together with an easier WebRTC integration, a 60% worth discount for GPT-4o audio, and help for GPT-4o mini at one-tenth of earlier audio charges.
- Choice fine-tuning permits for improved customization.
- Go and Java SDKs are actually out in beta.
Day 10: 1-800-CHATGPT
Taking a cue from Google’s traditional Voice Search, OpenAI has opened a telephone and WhatsApp line for its generative AI. Customers can ask natural-language questions, and the chatbot will reply without spending a dime. OpenAI considers this function experimental, noting that its availability and limitations might change.
Day 11: Extra choices for apps
Day 11 introduced a protracted listing of connections from ChatGPT to extra coding apps and instruments, together with VS Code forks, Jetbrains IDEs, extra Terminal apps, and extra. (Initially, it supported iTerm 2, Terminal, TextEdit, VS Code, and Xcode.) Three new app integrations arrived, connecting ChatGPT to Apple Notes, Notion, and Quip. Superior Voice Mode can now work with varied different desktop apps of the consumer’s selecting.
OpenAI notes that ChatGPT gained’t work together with desktop apps with out the consumer’s permission.
Plus, Professional, Crew, Enterprise, and Edu customers can use the brand new app integrations.
Day 12: o3 and o3-mini
OpenAI saved the most important information for final: o1 is not the corporate’s foremost mannequin. As a substitute, o3 – now in early entry for security and safety researchers – improves coding, math, and science efficiency. The corporate additionally pioneered a brand new approach referred to as deliberative alignment, used to maintain o3 on-mission. Security researchers can apply to check o3 here.