• unwind ai
  • Posts
  • Finetune GPT-4o & Gemini for FREE

Finetune GPT-4o & Gemini for FREE

PLUS: Qwen2 beats GPT-4o in Math, Google Gemini 78% Cheaper

Today’s top AI Highlights:

  1. Opensource Qwen 2 Math models beat GPT-4o, Claude 3.5 Sonet, and Llama 3.1 405B

  2. AI agent that automates business processes without any coding

  3. Google Gemini 1.5 Flash gets a major price cut

  4. Fine-tune GPT-4o mini with 2M training tokens free daily

  5. Google DeepMind introduces the first AI agent that plays amateur-level table tennis

& so much more!

Read time: 3 mins

Latest Developments

The Alibaba Cloud team has just released Qwen2-Math, a series of math-specific large language models built upon their existing Qwen2 LLMs. These models demonstrate excellent mathematical problem-solving capabilities, outperforming much larger state-of-the-art open and closed-source models like GPT-4o, Gemini 1.5 Pro, and Llama 3.1 405B! Qwen2-Math series is opensourced and includes both base and instruction-tuned models.

Key Highlights:

  1. Different sizes - Qwen2-Math has base and instruction-tuned models in 1.5B, 7B, and 72B parameter sizes.

  2. Dominating the Benchmarks - Qwen2-Math-72B-Instruct surpasses the performance of GPT-4o, Claude-3.5-Sonnet, Gemini-1.5-Pro, and Llama-3.1-405B on benchmarks like GSM8K, Math, MMLU-STEM, OlympiadBench, and CollegeMath - a serious contender in the mathematical reasoning arena.

  3. Complex Problem Solving - Qwen2-Math can tackle a wide range of mathematical tasks, from basic arithmetic to challenging competition-level problems. Think IMO Shortlist and even the International Zhautykov Olympiad!

  4. Availability - The suite of models is available on Hugging Face, GitHub, and Modelscope.

Bardeen, an AI-powered automation tool, has just raised $3 million in strategic funding from Dropbox and HubSpot. This platform helps businesses build reliable AI agents and automate repetitive tasks, like data extraction, lead generation, and contact updates, all using natural language. Bardeen simplifies workflow automation without requiring any coding.

Key Highlights:

  1. Visual Workflow Planning - Bardeen shows you exactly how it will execute an automation before it runs. This allows you to review and modify the steps, ensuring accuracy and preventing unexpected errors.

  2. Easy Automation Creation - You can create automations using simple language instructions or by recording your actions. This makes Bardeen accessible to both technical and non-technical users.

  3. Robust Integrations & Security - Bardeen integrates seamlessly with popular apps like Google Sheets, HubSpot, and more. Plus, with its SOC 2 Type 2 compliance, it meets the highest security standards for businesses.

Google released Gemini 1.5 Flash in May in Google AI Studio and via API, which closely matches 1.5 Pro’s performance across all benchmarks, but for much less latency and cost. Now, Google has significantly reduced the cost of 1.5 Flash API, expanded language support for Gemini 1.5 Pro and Flash to over 100 languages, and made Gemini 1.5 Flash tuning available to all developers. Plus, they've revamped the developer documentation for an improved user experience.

Key Highlights:

  1. Cheaper Flash - The cost for Gemini 1.5 Flash input tokens has been reduced by 78% (from $0.35 to $0.075 per million tokens), and output tokens by 71% (from $1.05 to $0.3 per million tokens), making it a highly cost-effective option for high-volume applications.

  2. Tuning Power - You can now fine-tune Gemini 1.5 Flash models with your own data for improved performance. Tuning is free, and the inference cost is the same as the base model.

  3. Multilingual Models - Both Gemini 1.5 Pro and Flash now support queries in over 100 languages and scale your AI apps worldwide.

Quick Bites

  1. Google DeepMind has developed the first AI agent that plays table tennis at an amateur human level. It uses a combination of low-level skills and long-term strategic planning. The robot trains in a simulated environment and refines its skills by playing against humans, adapting to various opponents and continuously improving its performance.

  1. Anthropic is expanding its model safety bug bounty program to focus on finding and mitigating universal jailbreak attacks. It offers rewards of up to $15,000 for identifying vulnerabilities in high-risk domains like CBRN and cybersecurity.

  2. Hugging Face has acquired XetHub, a platform for developing large-scale AI models, to scale up its systems and manage an increasing number of AI models. The acquisition will help Hugging Face host hundreds of millions of models on its platform.

  3. After being accused of a laundry list of safety concerns, Democrat lawmakers have written to OpenAI CEO Sam Altman about how OpenAI handles whistleblowers and safety reviews as former employees complained that internal criticism is often stifled. They want to understand whether federal intervention may be necessary.

  4. OpenAI has expanded GPT-4o mini fine-tuning access to all developers on usage tiers 1-5. You’ll get 2M training tokens daily for free through September 23.

Tools of the Trade

  1. Flowith: Your AI assistant for deep research and work, providing an intuitive canvas interaction for long-text generation and multi-threaded conversations. It features an advanced agent system that integrates with numerous external tools and supports cutting-edge AI models.

  1. Wordware: An IDE that enables anyone to build complex AI Agents and applications using natural language. Domain experts and engineers can now iterate 20x faster with prebuilt tools, API deployment, tracing, and more.

  2. Automations: provides scripts that let you use AI to generate terminal commands and Git commit messages directly from your terminal, eliminating the need to memorize them. 

  3. Awesome LLM Apps: Build awesome LLM apps using RAG for interacting with data sources like GitHub, Gmail, PDFs, and YouTube videos through simple texts. These apps will let you retrieve information, engage in chat, and extract insights directly from content on these platforms.

Hot Takes

  1. True Story!
    Day 1 - OAI loses 3 leaders. Everyone declares them over. Proclaim LLMs are a commodity
    Day 2 - OAI posts a strawberry photo, the same accounts declare they have achieved AGI
    The difference 🍓🍓🍓 make! 🫡 ~
    Bindu Reddy

  2. Sama, if you want to stop bleeding talent to Anthropic, you need to release GPT-5 now and prove OpenAI is the future. ~
    Flowers

Meme of the Day

That’s all for today! See you tomorrow with more such AI-filled content.

Real-time AI Updates 🚨

⚡️ Follow me on Twitter @Saboo_Shubham for lightning-fast AI updates and never miss what’s trending!

PS: We curate this AI newsletter every day for FREE, your support is what keeps me going. If you find value in what you read, share it with at least one (or 20) of your friends!

Reply

or to participate.