OpenAI's New Model, the Reflection Drama, & Model Merges

TLDR;

📰 News on: 🍓 OpenAI’s Upcoming Model,

🫣 the Reflection Drama,

🤔 Who’s Lying, Mistral or Alibaba, Glean?,

🥵 SB-1047 bill,

🫢 Hume’s new emotion model.

[TREND OF THE WEEK] Arcee’s SuperNova Model Merge

OTHER PREMIUM CONTENT

Other New Premium Content

OpenAI’s Strawberry. It’s coming soon, but what do we know so far?

How Many GPUs Do I Need? Learn to estimate how many GPUs you’ll need to serve your customers. It’s actually straightforward! (Only Full Premium)

The AI Superhuman Forecaster, a wake-up call for the world? (All Premium members)

NEWSREEL

OpenAI’s New Model, Out in 2 Weeks

It seems that the wait is finally over. According to The Information (paywalled, but your boy is here to give you all the details), OpenAI intends to release its new model, Strawberry, in the next two weeks through ChatGPT.

According to the source, the model will be given ‘time to think,’ a paradigm shift we’ve talked about repeatedly. In other words, taking up to 20 seconds, Strawberry will ‘explore the space of possible solutions’, which means the model has time to correct its first impressions, explore new possible solution paths, and eventually settle for the best solution.

This is similar to humans' System 2 thinking mode when solving a complex task, instead of questions that can be responded to innately, like “2+2” (called System 1 thinking).

Hence, as a rule of thumb, you’ll want to use Strawberry in cases where the model requires complex thinking, like math or coding.

Interestingly, some beta testers don’t seem that impressed. The model struggles to acknowledge when the task is simple and does not require search, meaning that in some cases, you should directly use GPT-4o and use Strawberry only when needed. Some have questioned whether the price increment will be justified.

TheWhiteBox’s take:

Of course, these new features, active search, and monkey sampling, mean that the model will be much more expensive to run, as the average number of generated tokens per user request will be much higher.

Read my Notion article (available to Premium members only) for a detailed analysis of the expected improvements on Strawberry.

Therefore, although I expect the input prompt price per million tokens to remain the same, the costs explode during generation as the model performs an active search, thereby making the output tokens much more expensive for them… and for you.

In summary, expect having to pay more to access Strawberry, be that through the web subscription or higher $/million output tokens in the API. But will it be worth the presumably much higher price, with some speculating up to $2000/month?

In the meantime, Bloomberg reports that OpenAI is hoping for $6 billion at a $150 billion valuation for its next round, an increase, information also echoed by The Information and higher than the originally reported value (a $120 billion valuation).

CONTROVERSY

Reflection 70B Creator Accused of Fraud

Matt Schumer, the person behind what could be considered the biggest news of the past week, a model called Reflection…

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