#203 - Gemini Image Gen, Ascend 910C, Gemma 3, Gemini Robotics

665 snips

Mar 17, 2025

This week in AI, there's buzz about OpenAI's new 'deep research' feature raising cybersecurity alarms. Google is making headlines with a massive investment in Anthropic, enhancing its tech partnerships. Meanwhile, Huawei unveils its Ascend 910C chip, challenging Nvidia's dominance amid trade tensions. Also on the table is Waymo's expansion of its robotaxi service and new advancements in image generation technologies. Plus, tensions between the US and China continue to grow as AI execs face travel restrictions.

Ask episode

AI Snips

Chapters

Books

Transcript

Episode notes

INSIGHT

OpenAI's Agent Strategy

OpenAI is unbundling its AI agent tools, allowing businesses to build their own.
This strategy mirrors open-sourcing and lets OpenAI learn from user implementations.

INSIGHT

Gemini's Native Images

Google's Gemini 2.0 Flash now has native image output within its chat flow.
This allows impressive multi-turn image editing, preserving context and consistency.

ANECDOTE

Waymo Expands Again

Waymo expands robotaxi services to more Bay Area cities, including Palo Alto and Los Altos.
Andrey is excited about potential commute integration, showcasing Waymo's steady expansion.

Get the Snipd Podcast app to discover more snips from this episode

Get the app

Deep Research System Card

Documentation for OpenAI's Deep Research Capability

OpenAI Team

The Deep Research System Card provides details on how OpenAI's Deep Research capability was developed, its capabilities, and associated risks. It highlights the model's ability to search, interpret, and analyze vast amounts of online data, while also addressing potential issues like hallucinations and privacy concerns.

Superintelligence and AI Startups

The Rise of Autonomous Systems

Hypothetical Author

This hypothetical book delves into the concept of superintelligence and its role in shaping the future of AI startups. It discusses how companies like Reflection AI are pushing the boundaries of autonomous systems, aiming to revolutionize industries through advanced AI technologies.

Sesame Releases Its Base AI Model

Behind the Viral Virtual Assistant Maya

Sesame Team

Sesame's CSM-1B model is a 1 billion parameter AI model used for generating human-like speech. It employs residual vector quantization (RVQ) and is built on Meta's Llama framework. The model is open-sourced under the Apache 2.0 license but lacks substantial safeguards against misuse.

Claude 3.7 Sonnet System Card

Technical Overview of the Hybrid Reasoning Model

Anthropic Team

The Claude 3.7 Sonnet System Card provides an in-depth look at the hybrid reasoning capabilities of the model, including its ability to switch between quick responses and extended thinking. It also covers safety evaluations and responsible scaling policies.

Open-Sora 2.0: Training a Commercial-Level Video Generation Model in $200k

None

Xinying Guo, Tom Young, Chenhui Shen, Zangwei Zheng, Xiangyu Peng

Open-Sora 2.0 is a commercially viable video generation model that demonstrates cost-effective techniques for high-quality video synthesis. It was trained for approximately $200,000, significantly cheaper than comparable models. The model uses a hierarchical data filtering system, a deeply compressed video autoencoder, and a diffusion transformer architecture to achieve high-quality video generation.

Reka AI Open Sourced Reka Flash 3: A 21B General-Purpose Reasoning Model that was Trained from Scratch

Not applicable

Not specified

Reka Flash 3 is a general-purpose reasoning model developed by Reka AI, featuring 21 billion parameters. It is designed for tasks like general chat, coding support, instruction following, and function calling. The model is optimized for low-latency applications and can handle contexts of up to 32,000 tokens. It is available under the Apache 2.0 license.