Sign in Subscribe
3 min read

Claude 3.7 scored 8.93% on Humanity's Last Exam, plays Pokémon on Twitch, and Deep Research is available for ChatGPT Plus users.

Model Releases and Updates

Claude 3.7 Thinking Scores 8.93% on Humanity's Last Exam (HLE): Claude 3.7 Thinking, a non-endgame model primarily used for coding, achieved a notable score of 8.93% on Humanity's Last Exam (HLE) despite not having internet access.

OpenAI's Orion Model Details: OpenAI's latest model, Orion, was initially intended to be GPT-5 but faced challenges in achieving significant performance gains over GPT-4. The model was trained with approximately 10x the compute of GPT-4.

GPT 4.5 Spotted in Android Beta: OpenAI's GPT 4.5 has been spotted in an Android beta, suggesting an imminent launch. This indicates significant progress in the development and deployment of advanced AI models.

Claude Sonnet 3.7 Training Details: Anthropic's Claude Sonnet 3.7 model is confirmed not to be a 10^26 FLOP model, costing a few tens of millions of dollars. Future models are expected to be significantly larger.

Claude 3.7 Plays Pokémon Red on Twitch: Claude 3.7, an AI model, is playing Pokémon Red on Twitch, showcasing its ability to interact with a complex environment and make decisions based on reasoning.

Gemma 3 27b Model Release: The post announces the release of the Gemma 3 27b model, part of the Gemini API models list. This is a significant development in the AI community, as it indicates the availability of a new AI model for use and testing.

TinyR1-32B-Preview Model Release: Qihoo360 has released a new AI model called TinyR1-32B-Preview, which surpasses the performance of the official R1 distill 32B model. The new model is a "super" distilled model created by merging three domain-specific 32B reasoning models in Math, Coding, and Science.

API and Service Updates

DeepSeek API Platform Off-Peak Discounts: DeepSeek has introduced off-peak discounts on their API Platform for DeepSeek-V3 and DeepSeek-R1 services, offering significant reductions in pricing for input and output tokens during specific hours.

DeepGEMM Library Release: DeepSeek has released DeepGEMM, a library designed for efficient FP8 General Matrix Multiplications (GEMMs) with fine-grained scaling, as proposed in DeepSeek-V3. This library is optimized for both inference and training.

New Features and Tools

OpenAI's Deep Research Feature: OpenAI has released a new feature called 'Deep Research' for ChatGPT Plus users. This feature allows users to conduct in-depth research on specific topics, generating detailed reports with cited sources.

Advanced Voice Feature for ChatGPT Free Users: OpenAI has released an advanced voice feature powered by GPT-4o mini for free users of ChatGPT, aiming to provide a natural conversation pace and tone similar to the GPT-4o version while being more cost-effective.

Alibaba's Free AI Video Generation Model: Alibaba has made its AI video generation model available for free global use, potentially disrupting the market for other AI video generation companies.

Microsoft's Free Copilot Voice and Think Deeper Tools: Microsoft has made Copilot Voice and Think Deeper available for free with unlimited use. These tools leverage OpenAI's O1 reasoning model, enhancing AI capabilities for users.