DeepSeek vs GPT, Mistral Update, Gemini Gets Memory

PLUS HOT AI Tools & Tutorials

Nov 22, 2024

Hello and welcome to our weekly roundup!

In today’s issue: another king of the hill among reasoning models (it even beat OpenAI's o1), a whole bunch of reasons to try Mistral's AI, and an update that promises to make Gemini really useful.

Let’s get started.

Oh, and if you want to learn how to create videos with AI, this is the place to go:
Quick Video Creation with AI
Stepan Ikaev
·
Nov 20
Read full story

This Creators’ AI Edition:

Featured Materials 🎟️
News of the week 🌍
Useful tools ⚒️
Weekly Guides 📕
AI Meme of the Week 🤡
AI Tweet of the Week 🐦
(Bonus) Materials 🎁

Your advertisement could be featured here!

Sponsor a spot in our newsletter to connect with subscribers interested in AI, technology, and startups. For inquiries, click the button below.

Featured Material 🎟️

DeepSeek’s Newest Model Surpasses OpenAI’s o1

The week's big news is the release of the new DeepSeek model. A few days ago, a startup announced on X that it was launching a preview of its first reasoning model, DeepSeek-R1. This platform was developed similarly to OpenAI's o1 and is its direct competitor. As the first tests have shown, DeepSeek outperforms the AI industry leader in many key parameters. Let's break down what's going on here.

DeepSeek is a Chinese AI startup that has been gaining attention recently. Founded in 2023 and backed by quantitative hedge fund High-Flyer Capital Management, it aims to create artificial general intelligence (AGI) and is positioning itself as a competitor to OpenAI and Google.

Reasoning models are different from standard LLMs because of their ability to “fact-check” their answers. They spend more time thinking about the answer and how to respond to the prompt, which avoids many of the “hallucinations” found in popular AI. It's not a panacea (mistakes still happen) but makes AI more useful.

Like o1, released by OpenAI in September, DeepSeek-R1 plans ahead when faced with complex problems, solving them individually to guarantee an accurate answer. In doing so, it can take up to 10 seconds to “formulate”. The thought process is fully transparent, allowing users to follow the individual steps required to generate an answer.

Keep your mailbox updated with key knowledge & news from the AI industry

DeepSeek-R1 now outperforms o1 on two key tests, AIME and MATH. The former uses other AI models to evaluate LLM performance, and the latter is a series of complex text-based tasks. In addition, the model has shown that it correctly answered several “tricky” questions that stumbled existing models such as GPT-4o and Anthropic PBCs Claude.

However, don't be in a hurry to remove ChatGPT from your bookmarks. Strictly testing results are only part of the picture. X users who have tested DeepSeek on the strength of DeepSeek have already found some problems with reasoning. In particular, the new model does not solve some logical problems like playing tic-tac-toe and also has limitations in terms of topics for conversations. For example, DeepSeek refuses to answer political questions related to Chinese politics.

When asked about incidents such as the Tiananmen Square massacre, Chinese President Xi Jinping's relationship with Donald Trump, and China's potential invasion of Taiwan, he has consistently replied that he is “not sure how to approach these kinds of questions.”

DeepSeek-R1 can already be accessed via the DeepSeek Chat app. It's free but is limited to sending 50 messages per day. According to the company's posts on X, DeepSeek plans to release open-source versions of its R1 series models and related APIs.

News Of The Week 🌍

Mistral Unveils New AI Models, Agents, and Chat Features

French startup Mistral has rolled out a whole bunch of updates that compare its AI capabilities to current ChatGPT models. The Le Chat can now search the web with quotes (just like Search for ChatGPT does). It has also gained support for the canvas tool, allowing users to modify, transform, or edit the generated one. This should be especially useful for web page layouts and data visualization.

In addition, Le Chat now handles large PDFs and images for analysis and summarization, including files containing graphs and equations. The platform includes Black Forest Labs' Flux Pro model for image generation. And Le Chat can now host common automated workflows for tasks such as scanning expense reports and processing invoices; Mistral calls these AI “agents.”

New models make the above changes possible. In particular, thanks to Pixtral Large, the second platform in the Pixtral family that can process text and images simultaneously. It has 124B parameters and outperforms key players (including Claude, Gemini, and GPT) in some multimodal tests. Another is Mistral Large 24.11, the new model that brings “notable improvements” in understanding long contexts.

One of the few European AI startups continues to prove that it didn't raise $640M this summer by accident. Previously, I favored other models, virtually ignoring Mistral. But now it's starting to feel like I made a big mistake.

OpenAI's GPT-4o Enhances Creative Writing

It's not as big as Mistral, but it's still a nice update from OpenAI. On Wednesday, the company unveiled updates to its GPT-4o language model designed to improve creative writing capabilities. The developer says the AI now produces more natural and interesting results. Changes have also been made to how GPT-4o processes uploaded files. The chatbot will now act faster.

In addition, OpenAI has brought ChatGPT's Advanced Voice Mode to the web. The feature, which makes chatting with ChatGPT more natural, will become available to paid ChatGPT customers this week. To access it, you must be a Plus, Enterprise, Teams, or Edu subscriber. The web launch follows the debut of Voice Mode in ChatGPT's iOS and Android apps in September.

It's not that significant news, but it provides a good reminder that OpenAI's approach is different from how Apple and Google operate with their operating system updates. As soon as a startup makes progress, we immediately get a corresponding update.

Sharing is caring! Refer someone who recently started a learning journey in AI. Make them more productive and earn rewards!

Refer a friend

Apple Is Reportedly Building Siri Powered by LLMs

According to Bloomberg sources, Apple is developing a new version of its voice assistant, Siri, powered by extended large language models (LLMs). This Siri is part of Apple's attempt to catch up with BB competitors who have already introduced features similar to those of Google's Gemini Live. The new assistant is expected to completely replace the Siri interface as early as spring 2026a.

Based on Bloomberg's description, the new Siri will be very similar to Advanced Voice Mode, but with the same access to personal info and apps that Siri has today.

Google Gemini Can Remember Things Now

Starting this week, Google One AI Premium Plan users can teach Gemini about their interests and preferences. The model, available as part of the Gemini Advanced subscription, now remembers all conversations and uses that data to generate the next responses (something similar has been in ChatGPT for a while. To test it, ask the AI to describe or roast you). Users can also view and edit their data via the Saved Info page.

The update is expected to be particularly useful for writers, developers, and students who regularly use AI for single-category tasks.

This may be the most significant update for Gemini since its release. One of ChatGPT's biggest advantages is its ability to use data from past conversations when generating new responses. This often makes life easier and cuts down on prompt engineering time. Who knows, maybe now Gemini will become more competitive.

Useful Tools ⚒️

Buddler - SEO Growth Hacking Solutions

HumanLayer - an API and SDK that enables AI Agents to contact humans

Locofy Pro – Build frontend 10x faster with design-to-code AI

Momen – Build your MVP and scale it; no code is needed.

ReadPartner - Instant AI Summarization & News Automation

ReadPartner AI is an abstracting tool that helps you quickly digest information from websites, YouTube videos, and documents. It supports multilingualism, allows for summaries in different languages, and offers automated AI news digests delivered via email or messaging apps. You can try it through its website and Chrome extension (I recommend the second option). A great tool for creators, students, and anyone who works with large amounts of information.