Mesolitica’s Post

OpenAI announcements today, fresh from the oven! Got two from my wish list! (realtime API and prompt caching) Realtime API: OpenAI introduced a public beta of its Realtime API, enabling developers to create low-latency, speech-to-speech applications using AI-generated voices. Vision Fine-tuning: Developers can now fine-tune GPT-4 models using images, improving tasks involving visual understanding. Model Distillation: New feature allows developers to fine-tune smaller models like GPT-4o mini using larger models, saving costs. Prompt Caching: This feature reduces API costs by 50%, allowing developers to cache frequently used context between calls

🧠 OpenAI Launches SimpleQA: Improved Fact Checking ✅ SimpleQA by OpenAI tests language model accuracy with 4,300 fact-based questions to improve reliability. #OpenAI #FactChecking #SimpleQA https://lnkd.in/eNvKDHG9

Recent AI breakthroughs from the Chinese company DeepSeek showcase resilience in the AI race, even in the face of limited access to advanced chips. I hope we are entering the next stage of the AI revolution, where success is not solely determined by pouring billions into infrastructure or ingesting massive amounts of unstructured data into training processes. Instead, the focus should shift toward developing more efficient model architectures and improving the quality of the data used for training. The democratization of AI, an emphasis on smaller, fine-tuned models, and innovative approaches to how these models interact with one another offer a more sustainable path toward achieving AGI #ArtificialIntelligence #MachineLearning #SustainableAI #DeepSeek #AIBreakthroughs

Still at super early stage but this seems to be promising and interesting. By leveraging function call features of OpenAI API, we can create an interactive chat with our internal system data without having to worry about the data security issues as we have total control over what data we want to provide to the model

DeepSeek R1 is Opensource and 96.4% cheaper than OpenAI o1 while delivering similar performance. You can also run it on home-grade hardware. OpenAI o1: $60.00 per 1M output tokens DeepSeek R1: $2.19 per 1M output tokens Intelligence truly is too cheap to meter! ⚡

OpenAI Dev Day recap! (Speech to Speech API finally released) Denys Linkov and I went through the OpenAI Dev Day releases yesterday and explain what each of them are and how they impact you. 1. Realtime API: This is the official release of openAI's speech to speech model that allows you to stream speech or text (and later video) to OpenAIs API and receive responses back in realtime. We spent most of the video talking about this as it'll have a big impact on democratizing access to high quality voice agents. 2. Prompt Caching on the API: Useful if you're doing a high volume - it caches larger responses and saves them for a 5-10 minute period but has a reduce cost for use. 3. Finetuning the Vision Model: Being able to fine tune their vision models with your own data set. 4. Model distillation - which allows you to fine tune a smaller model with the outputs from a larger more expensive model. I'll send over the full video after its uploaded :)

OpenAI o1 is here … Today I’ve seen the preview launch of OpenAI's latest release, OpenAI o1, and can’t wait to play with it more. What's New in OpenAI o1 compared to 4o (spoiler alert … it isn’t a better naming convention) … 🚀 Speedier performance … Quicker responses but there’s action descriptions now to show you what it’s doing while processing your command 🧠 Smarter language understanding … More natural interactions, yet to see if this has a significant improvement to the voice conversation experience 🎨 Enhanced creativity … Generates innovative ideas to inspire and assist. I hope this is good because charting and image generation on the previous version was 💩 🔧 Greater customisation … Tailors to our specific needs for a more personalised experience. Not sure how this interacts with the new memory features but I’m hoping it at least more reliably maintains context through longer conversations and thought development 🔒 Improved security … Stronger protection to keep our data safe I believe human & machine co-creation is just getting started. This should be a nice little level-up … and with Elon having just launched a 100k GPU data centre in 122 days for Grok training the competition is hotting up!

We were all very excited about OpenAI realtime when we heard the demos etc. But it's not practical to use in most small business voice support systems. At the costs he shows here it's more expensive than human operators. For simple support needs Vapi is still on top - cheaper to deploy and develop with, cheaper to run, almost as good. I'm sure OpenAI will slash pricing at some point but I'm also equally sure as this YouTuber notes that services that allow for easier voice platform building like Vapi will integrate it. Some platforms already have - I tested a custom voice agent using realtime + chainlit (python library) and that's looking like an exciting combination if realtime pricing comes more into line with what customer support actually costs to provide etc. But for now I'm still generally recommending folks start out looking at what can be put together for them in Vapi first if it's a straightforward task especially.

OpenAI just dropped their Spring update. No GPT5 yet but some cool new updates: - New Model GPT4o launched which is faster and cheaper than GPT4-Turbo - Free accounts now get access to custom GPT's, GPT4o, Browsing, File Uploads, Data analysis - There's now a Desktop app which you can run offline, share your screen with and talk to - Chat GPT Voice is now real-time, can be interrupted, has a new video mode (ie it can see), can detect emotions and can vary it's speaking style to represent different emotions This is a big step forward in terms of widening access with many additional capabilities coming for free users.

Recent Updates from OpenAI dev Day, Real time Voice API looks interesting so far, let's see how actually it turns out to be. #openai #devday