Posts in: AI

The weird way an #LLM does the maths - pretty smart-ish!

Never has a technology been so widely adopted yet so mysterious in its inner workings. #Anthropic had to put some serious effort into a method, that gives a glimpse into how their model #Claude „thinks“.

So Anthropic’s deep dive reveals that while the model ends up with the correct answer, its internal process is anything but conventional. It uses quirky, approximate methods that differ starkly from the neat explanation it later provides.

When asked to add 36 and 59, Claude first approximates by summing “40ish and 60ish” and “57ish and 36ish” to arrive at a rough 92ish, then separately focuses on the last digits (6 and 9) to ensure the final number ends in 5. Combining these steps, it correctly computes 95. Pretty smartish. 😅

When asked how it arrived at this result, it gives you a common approach that you can find everywhere online, rather than what it actually did.

So, there’s another reason why it’s best to be careful and check what they say, rather than just believing them.

Full article in technology review

Even more detailed information on more examples in the research paper

This playbook is an inspiring resource with insights into creating impactful #DataScience and #AI #4good projects. It captures DataKind’s strategies and offers valuable lessons to improve the journey. Claiming to be continually updated with new insights and user feedback, it has the potential to become a go-to guide for those diving into #Data4Good initiatives.

🚀 Meet Europe’s new AI Factories: 🇦🇹 AI:AT | 🇧🇬 BRAIN++ | 🇫🇷 AI2F | 🇩🇪 JAIF | 🇵🇱 PIAST | 🇸🇮 SLAIF

With an investment of €485 million, these centres are providing SMEs and start-ups with access to advanced supercomputing, boosting Europe’s AI capacity.

AI Gigafactories, coming soon, will expand Europe’s AI power. Highly relevant! 🌍✨

See full press release.

🚀 Exciting times in #AI development! Showing the potential of #OpenSource! While #DeepSeek as such is open source, it is not yet possible to fully reproduce the model. Now, the #OpenR1 project on #HuggingFace is on a mission to fill the gaps in the DeepSeek-R1 pipeline and make this advanced AI approach accessible to all, allowing anyone to reproduce and build upon it. Work in progress, worth keeping an eye on! 🌐🤝

🔗 Check out the project here

#LLMs: “The worst-case scenario would probably be a dictatorship that writes its propaganda into models on a grand scale”

The phenomenon of “artificial intelligence” is fuelled by the fascination that emanates from a machine that is difficult to distinguish from humans when communicating. Even if human communication is only simulated, people react to this communication as they are used to as social beings.

So here is the thing: Through the training the models had, vertain values and believes are subtlely infused. Using methods from psychology, researchers have found that responses differ according to the gender with which the model is addressed. Political colouring can also be determined with these psychometric tests.

Scientific studies confirm that people are influenced by them. And this already happens through the writing assistants, who complete or change sentences, for example. They are thus part of the writing process and can subtly change the opinions of the writers.

Interesting: large language models LLMs can, for example, help supporters of conspiracy theories to reduce their belief in these theories. But this also harbours the danger that the opposite effect can occur if the models are systematically polluted.

The article in the magazine #ct delves deeper into this dicussion (German, paywall): www.heise.de/select/ct…

🚀 The rise of large language models (#LLMs) and machine intelligence applications comes with significant concerns about the substantial resource requirements of these technologies. The demand for expensive hardware and high energy consumption has sparked creative solutions—one of which is #quantization.

• Quantization reduces the size of large language models (LLMs) by lowering parameter precision (e.g., from 32-bit to 8-bit or 4-bit).
• Benefits:
  • Smaller storage requirements enable deployment on devices like smartphones.
  • Improved energy efficiency and lower computational costs.
• Techniques:
  • Post-training quantization (applied after model training).
  • Quantization-aware training (incorporates quantization during training for better accuracy).
• Challenges: Potential accuracy loss and increased implementation complexity.
• Advances: Research on 1-bit models and methods to avoid matrix multiplications to enhance efficiency.

Quantization is a encouraging approach to making LLMs more accessible and sustainable.

For an insightful explanation of this approach (German, paywall): www.heise.de/hintergru…

🚀 Eyes Wide Open, AIs Wide Open: Staying in Control in the Age of AI 🤖

This talk by Yann Lechelle was given at the recent Open Community Experience (OCX) 2024 in Mainz, Germany, and shed light on Europe’s position in the global AI landscape.

Video link: www.youtube.com/watch

My key takeaways:

• Strategic independence: Emphasized the need for Europe to strengthen its technological sovereignty in AI and reduce its reliance on non-European vendors.
• Open Source & Open Science: Emphasized the central role of open source projects and open science in fostering innovation and ensuring transparency in AI development.
• Unified European Approach: Called for coherent policies and investments to strengthen Europe’s AI capabilities, ensuring that they are consistent with universal values and ethical standards.

tl;dr: Europe must strengthen its AI sovereignty through open source collaboration and unified strategies to remain competitive and uphold its core values.

I spent some free time the last few days exploring no-code and low-code workflow automation.

A significant part of my learning involved hands-on experience with #n8n, an open source workflow automation tool. I successfully completed the beginner/level 1 and intermediate/level 2 online courses.

Why n8n?

• Open source: Gives flexibility and transparency.
• Self-hosted 😎 option: Provides steeper learning curve and control over data.

What’s in it for me?

You get a lot! It’s fun, and it can make some of your more mundane tasks, like cross-posting, easier.

And: The future of AI applications lies in AI-enhanced or even AI-agent workflow support. Users will be looking for AI-driven support that goes beyond simple prompts and aims for intelligent, adaptive systems that can make context-aware decisions.

While n8n is user-friendly, it does require some technical understanding. Therefore, the future may lean toward no-code or “assisted low-code” applications, making advanced automation accessible to a broader user base.

For a deeper dive into AI-powered workflows with #n8n, check out this post: blog.n8n.io/ai-agenti…

Another interesting approach to AI-powered workflow applications can be found here: www.entaingine.com

#AI #WorkflowAutomation #NoCode #LowCode #n8n #OpenSource #SelfHosted #AIWorkflows

Introducing #EuroLLM, the latest European addition to the 8B class of LLMs, with a strong emphasis on multilingual capabilities. 🗣️🌍

💡 Why it stands out: • Top performer: Scored the best Borda score (1.0) across multilingual benchmarks. 🏆 • Translation excellence: Shines in tasks like FLORES (88.87) and WMT24 (83.61). 📜➡️🌐 • Versatility: Supports 24+ languages, including Arabic, Chinese, and Russian, powered by 9 billion parameters and a cutting-edge architecture. 🤖✨

Developed by leading European institutions and supported by the EuroHPC infrastructure, EuroLLM showcases the strength of collaborative innovation. 🛠️🇪🇺

For more details and download: huggingface.co/blog/euro…

🚀 Why it matters: In areas of international cooperation, where local language support is key, these models can serve as a powerful asset, rivaling global leaders in the LLM space. 🌐🌟

But let’s keep things in perspective: EuroLLM is in the 8B class, which is impressive but still a step below the likes of GPT-4 and other “Champions League” models. While the size of the models alone is not the deciding factor, there’s still a long way to go - this progress is inspiring and gives us hope for the future. 🌟

#AIInnovation #LLM #EuroLLM #MultilingualAI #TechInEurope #Collaboration #MachineLearning.

In today’s fast-moving landscape, where technologies and customer demands evolve at lightning speed ⚡, traditional businesses face a critical challenge: keeping up. 📉 Their structures - designed for stability and efficiency - often make it difficult to pivot quickly enough.

A solution? Corporate Venturing. 🌟

By partnering with startups, launching internal ventures, or establishing accelerators and incubators, companies unlock a powerful edge: ✅ Access to cutting-edge technologies that reshape industries ✅ New business models to tap into emerging markets ✅ Speed and agility to respond to change faster than competitors ✅ A boost to entrepreneurial culture within the organization

Corporate venturing allows companies to combine their resources, expertise, and networks with the bold innovation of startups—a winning formula for long-term growth and resilience. 🌱

Of course, every opportunity comes with things to manage: 👉 Aligning corporate processes with startup agility 👉 Balancing risk with strategic goals 👉 Integrating new ideas into the core business

💡So will the future belong to those who combine the stability of established companies with the creativity and speed of startups?

Have you seen this approach drive success? Let’s hear your insights! 💬✨