Top stories today:

1. OpenAI releases o1 as its new AI model with “reasoning” abilities
2. AirPods Pro 2 gets FDA approval to serve as hearing aids
3. Adobe revenue +11% to $5.41B in Q3 FY24, guidance misses estimates
4. Global new unicorn count flat in Aug., with 8 +$1B-valued startups
5. Meta continues to dominate global VR headset market with 80% share

0. Data and calendar

1. OpenAI releases o1 (codenamed Strawberry) as its new AI model with “reasoning” abilities

• o1-preview & o1-mini are available for ChatGPT Plus & Team users.

• Enterprise and Edu users will get access early next week.

• TBD: o1-mini will be available for free but no release date disclosed.

• “Completely new optimization algorithm and a new training dataset specifically tailored for it.”

Writing code and solving multistep problems is easier with o1

• 83% of problems were solved by o1 in an Olympiad exam vs. 13% by GPT-4o.

• A “completely new optimization algorithm” is used to train the model.

o1 improves over GPT-4o on a wide range of benchmarks

• However, o1 can be slower — it takes 10 secs+ to answer some questions.

o1 costs 3-4x more than GPT-4o to devs

• $15 per 1M input tokens on o1-preview vs. $5 on GPT-4o.

• $60 per 1M output tokens on o1-preview vs. $15 on GPT-4o.

o1 is more manipulative than GPT-4o

• o1 carries a “medium” rating for chemical, biological, radiological and nuclear weapon risk, according to OpenAI.

Our view: this appears to be the 1st large jump in model type and quality since GPT-4 was launched in March 2023

• o1 is likely still a Transformers model, possibly with clever chain-of-thought prompting and fine-tuning.

• OpenAI is suggesting that o1 is something completely novel, but that could be marketing speech.

• We’ll wait to see independent benchmarks and whether other LLM builders launch similar models.

• Relatedly, ChatGPT crossed 11M paying users, making $225M+ monthly or $2.7B annually.

3. AirPods Pro 2 gets FDA approval to serve as hearing aids