• Get The Gist
  • Posts
  • DeepSeek’s Janus-Pro Outshines DALL-E 3 in Image AI

DeepSeek’s Janus-Pro Outshines DALL-E 3 in Image AI

Plus: Chinese AI Model ‘Kimi k1.5’ Outshines GPT-4o

Welcome to Get The Gist, where every weekday we share an easy-to-read summary of the latest and greatest developments in AI—news, innovations, and trends—all delivered in under 5 minutes! ⏱

In today’s edition:

  • DeepSeek’s Janus-Pro Outshines DALL-E 3 in Image AI

  • Chinese AI Model ‘Kimi k1.5’ Outshines GPT-4o

  • Meta AI Assistant Gains "Memory" for Recommendations

  • xAI is Preparing to Release Grok-3

  • And more AI news….

Top Developments

DeepSeek

Image by: DeepSeek

The Gist: DeepSeek’s Janus-Pro model is setting new standards in image generation, outperforming top competitors like OpenAI’s DALL-E 3 on key benchmarks, all while being developed at a fraction of the cost.

Key Details:

  • Janus-Pro-7B excels in image generation and analysis, leading on industry benchmarks like GenEval and DPG-Bench.

  • The model was developed for less than $6 million, compared to OpenAI’s $100 million, causing ripples in the AI industry.

  • Janus-Pro is available for free on Huggingface, offering parameter sizes from 1 billion to 7 billion.

  • Its innovative architecture balances visual encoding and processing, enhancing flexibility and performance.

Moonshot AI

Image by: Kimi.ai

The Gist: Moonshot AI’s Kimi k1.5 has emerged as a serious competitor in the AI race, surpassing GPT-4o and Claude 3.5 Sonnet in key benchmarks with its multimodal reasoning and reinforcement learning capabilities.

Key Details:

  • Kimi k1.5 processes text, images, and code, excelling in tasks like mathematics, coding, and visual reasoning.

  • It scored 96.2 on MATH 500, 77.5 on AIME, and ranked in the 94th percentile on Codeforces.

  • Built using reinforcement learning, Kimi uses a "Chain of Thought" approach, breaking problems into smaller steps for improved reasoning.

  • The model handles long-context tasks (up to 128k tokens) and delivers efficient outputs by reusing prior results.

Meta

Image by: Meta

The Gist: Meta’s AI assistant now remembers user-shared details to offer personalized recommendations, such as meal ideas or weekend activities, based on user activity across Meta’s apps.

Key Details:

  • The feature allows users to share specific details, like dietary preferences, which the AI remembers for tailored suggestions.

  • Currently available on Facebook, Messenger, and WhatsApp in the U.S. and Canada, the assistant’s memory can be updated or deleted by users anytime.

  • Recommendations are influenced by personal data such as location, interests, and activity across Meta’s platforms.

  • For example, based on your profile and app usage, the AI could suggest local events or dining spots that align with your preferences.

Quick Gist

Pika Art AI has launched Pika 2.1, an advanced AI video creation tool featuring enhanced motion control, realistic physics simulation, and customizable effects (Read More).

xAI is preparing to release Grok-3, an advanced AI model shown to outperform competitors in multiple tests, with the launch scheduled for next week (Read More).

OpenAI introduced ChatGPT Gov, a version of its AI service designed for federal agencies, ensuring security and compliance when handling non-public sensitive data (Read More).

Meta has open-sourced the Large Concept Model (LCM), a language model that improves multilingual summarization and hierarchical reasoning through advanced sentence embeddings (Read More).

The Qwen team released Qwen2.5-VL, an advanced vision-language model with capabilities like detailed image recognition, video event understanding, and structured data outputs in multiple formats (Read More).

THAT’S IT FOR TODAY!

That’s it for today, see you tomorrow! 👋

If you have any questions, feedback, or requests, drop us an email at [email protected]. We love hearing from our readers! 😊

P.S. If this email was forwarded to you, you can sign up for free by clicking here!