80% o3 discount, Apple's AI Update, Exa, & More

TLDR;

Important note: I’ve switched my sending email to [email protected] to fully transition to the new brand (beforehand, you were receiving my emails with my previous domain, @thetechoasis.com.

This week, we cover model releases from Apple and OpenAI, DeepSeek’s irresistible performance-to-cost ratio, and Claude’s safety testing backfiring on them.

We also discuss Meta’s acquisition of Scale AI and OpenAI’s insane revenue growth, as well as Salesforce’s dirty move against competitors. Finally, we take a look at a new hot AI tool, Exa, and o3’s dramatic price reduction.

Enjoy!

EDGE AI

Apple’s New AI Models

As part of their WWDC event a couple of days ago, Apple presented updated research on their new foundation models that will soon power Apple Intelligence: an on-device, three-billion-parameter model and a server-based one, which is located in Apple servers with extreme security compliance.

The models show comparable results to several popular models but don’t expect state-of-the-art performance.

Using vibe evals, the on-device model is comparable to Alibaba’s similar-sized Qwen models and Google’s Gemma. In contrast, the server-level model is a mid-tier model in terms of what the frontier of non-reasoning models stands for today.

An interesting aspect of the release is that it will enable app developers to fine-tune the model, thereby integrating Apple Intelligence more effectively into their apps.

The secret here is that they are relying on LoRA adapters, a fine-tuning method where a small subset of weights is added to the model that is trained during the new skill being learned, while the actual model remains frozen (unchanged).

Thus, at runtime, for the given task, the adapter weights are loaded into the model, transforming its behavior to suit that particular task. This way, we avoid having to train the entire model, which is expensive, while also allowing us to have one single foundation model and hundreds of adapters; it’s like having several models for the size of one, as each adapter is several orders of magnitude smaller than the base model.

Think of this as follows: Imagine a handyman and their toolbox. For some tasks, a specific screwdriver from the toolbox may be required. For others, a larger one might be needed. However, in other cases, it may require a completely different tool, such as a hammer, a utility knife, a tape measure, or an adjustable wrench.

Here, the handyman (the foundation model) is the same every time, but uses different tools (adapters )depending on the task.

The point of LoRAs is that, instead of having a handyman specialized in screwdriving, another in measuring, and another in hammering, we can use the same base handyman and simply adapt a small part, in that case, the used tool.

Thus, Apple will allow app developers to train their own tools on the foundation model, and I hope that one day, they will also enable bespoke adapters for users that tailor the model to their preferences.

TheWhiteBox’s takeaway:

Nothing extraordinary, but it’s progress, which is to say a lot of Apple recently. Now is the…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.