Tuana Çelik

Really excited by the release of GPT-4o mini. As well as being 60% cheaper than GPT-3.5 and way more performant, the longer output token limit will come in handy for our podcast generation tool. Perhaps most interesting will be the opportunity to fine tune; fine-tuned open weight models have been shown to outperform GPT-4 for specific use cases, so I am really looking forward to seeing what fine-tuned GPT-4o mini can do! AI generated podcasts: https://lnkd.in/ev_ceij6 Finetuned models beat OpenAI’s GPT-4: https://lnkd.in/eFccUKir

8

The original INRIA 3DGS code keeps getting improvements from the team, to make it more convenient - including Taming3DGS optimizations, which can cut training times in half for the same results 😀

34

1 commentaar

In PyData Eindhoven definitely the building of trusted Generative AI applications is the main theme. Not only the building of Retrieval Augmented Generation pipelines is being discussed, but also, how to make them performant and traceable. A lot of the discussions made me remember some talks by Dr. Chris Hillman (Who you should definitely follow), and this article by our colleague Chetan Hirapara (Who you should also definitely follow). https://lnkd.in/g5JYrhHX

32

2 commentaren

Sonnet 3.5 has been makingn waves. Personally I have found that perhaps it lacks breadth of knowledge ( not able to comprehensively answer many questions as well as openai) but its definitely faster , mostly gives better code and provides projects and artifacts features which are reason enough for switching to it for techies doing using it for programming tasks.

1

Quantum Internet is Happening Imagine an internet that doesn't just connect devices, but entangles quantum states across cities, setting the stage for a whole new era of secure, high-speed data transfer. A research team led by QuTech has just achieved a groundbreaking connection between Delft and The Hague, using 25 km of existing optical fibre [https://lnkd.in/e7hubyWD]. It's a step towards making the "Quantum Internet" a reality, and the possibilities are staggering. Reinventing cybersecurity This new internet could transform cybersecurity by creating unbreakable communication protocols. Unlike traditional encryption, quantum encryption uses entanglement, meaning that any attempt to intercept data would disrupt the quantum state and immediately signal an intrusion. This could enable real-time, ultra-secure communication channels for critical sectors, effectively setting a new standard in data security. Quantum-powered collaboration With an operational quantum network, organisations could one day connect quantum computers across continents for combined processing power exponentially greater than today's fastest supercomputers. Imagine globally distributed quantum computing power making complex simulations from climate modelling to drug discovery - more efficient and accessible than ever before. This quantum networks won't just make us faster, they could make us fundamentally better at solving complex problems. Real-time applications Long-distance quantum links require exceptionally precise stabilisation, with accuracy down to the wavelength of photons. QuTech's recent successes reflect innovations in protocols and control capable of maintaining stable entanglement over metropolitan distances. This stability opens the door to quantum timing applications, potentially transforming fields such as autonomous systems and even AI model training, where real-time distributed processing can be essential. Tomorrow's infrastructure The team's modular architecture means this setup can be scaled to connect multiple quantum nodes, enabling the addition of new quantum processors and adapting to future technologies. With support from EU research agencies, this project could form the basis of a robust quantum infrastructure across Europe and beyond. And the network's scalability makes it a leading contender for global deployment, elevating it beyond a laboratory experiment as a blueprint for the future of internet infrastructure. So, a new "Internet" is on the horizon, and its potential applications are both vast and revolutionary. We're at the dawn of a new era where quantum and classical networks will integrate, transforming the way we communicate, compute and collaborate on a global scale. #AI #Cybersecurity #DataSecurity #DigitalInfrastructure #DistributedComputing #FutureOfTech #Innovation #InternetOfTheFuture #Networking #Photonics #QuantumComputing #QuantumEncryption #QuantumInternet #QuantumNetworks #TechInnovation

2

One of the best blogs I've come across on quantization, with outstanding visualizations. Highly recommend a read for anyone looking to understand LLM quantization in more depth! 📐 - https://lnkd.in/e_WTgq6x At Felafax AI (YC S24), we are building Infra to make full precision multi-GPU fine-tuning easier. #AI #Infra #LLM

16

3 commentaren

🚀 Calling all AI founders in the Netherlands! 🇳🇱 We need your insights! 🤔 As announced at the AI salon, Techleap embarked on a research journey to explore the conditions needed to build successful AI companies in the Netherlands. With a lack of unicorns in our ecosystem, it's time to make a change! Why? Because Techleap identified the following:  1. Addressing the challenges faced by Dutch founders in staying abreast of rapid international AI advancements and fostering a competitive edge in the global market  2. Lack of founder participation in shaping crucial AI regulations, especially in anticipation of compliance with the upcoming EU AI Act  3. Strategically mapping out the landscape of AI developments to drive growth and innovation within the Dutch tech ecosystem, ultimately aiming to elevate the emergence of unicorns. By filling out our survey, you can help shape steer the Dutch ecosystem. Your input is crucial as we navigate through the ever-evolving international AI landscape. 🌍 https://lnkd.in/g_y2CZBC Thank you! Myrthe Hooijman #AIFounders #TechEcosystem #NetherlandsAI #UnicornsInTheMaking 🚀💡🦄

67

8 commentaren

How much Big Data is there in the real world? A new dataset from the Redshift team lets us figure it out empirically; I dig into the data in this blog post. TL;DR: Most databases don't have big data, when people do have big data they still query their small data, and when people do query their big data they only read a small part of it. Only a tiny of fraction of users ever interact with Big Data, so maybe we don't need Big Data tools as much as we thought?

156

5 commentaren

🚀 Groundbreaking advancements in the field of AI and language models! 🚀 𝐁𝐢𝐭𝐮𝐧𝐞: 𝐁𝐢𝐝𝐢𝐫𝐞𝐜𝐭𝐢𝐨𝐧𝐚𝐥 𝐈𝐧𝐬𝐭𝐫𝐮𝐜𝐭𝐢𝐨𝐧-𝐓𝐮𝐧𝐢𝐧𝐠 - a novel approach that dramatically enhances instruction-tuning of pretrained decoder-only large language models (LLMs). Our team, comprising Dawid J. Kopiczko (Vrije University of Amsterdam Amsterdam), Tijmen Blankevoort (Meta), and Yuki M. Asano (University of Amsterdam), developed Bitune to integrate both causal and bidirectional attention, achieving remarkable improvements in various downstream tasks. 🔍 𝐊𝐞𝐲 𝐈𝐧𝐬𝐢𝐠𝐡𝐭𝐬: Bitune leverages two sets of parameters: one for causal and another for bidirectional attention. These features are combined using a weighted average with trainable coefficients to generate new tokens. Our method introduces only a minimal set of new parameters through parameter-efficient fine-tuning techniques (PEFT), making it highly efficient. 📊 𝐌𝐚𝐭𝐡𝐞𝐦𝐚𝐭𝐢𝐜𝐚𝐥𝐥𝐲 𝐈𝐧𝐭𝐞𝐫𝐞𝐬𝐭𝐢𝐧𝐠 𝐅𝐢𝐧𝐝𝐢𝐧𝐠𝐬: During the prefilling phase, Bitune processes the instruction using both causal and bidirectional attention. This leads to an enhanced representation of the query. This method preserves the autoregressive nature of response generation while significantly improving performance on tasks such as commonsense reasoning, arithmetic, and language understanding. 📈 𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 𝐁𝐨𝐨𝐬𝐭𝐬: Consistent zero-shot improvements across models like Gemma-2B, Llama2-7B, and Phi-2. Significant gains in commonsense reasoning and arithmetic tasks, demonstrating the versatility and robustness of Bitune. 🏆 𝐖𝐡𝐲 𝐁𝐢𝐭𝐮𝐧𝐞 (𝐁𝐋𝐒𝐓𝐌) 𝐢𝐬 𝐒𝐮𝐩𝐞𝐫𝐢𝐨𝐫 𝐭𝐨 𝐄𝐧𝐜𝐨𝐝𝐞𝐫-𝐎𝐧𝐥𝐲 𝐌𝐨𝐝𝐞𝐥𝐬: While encoder-only models like 𝐁𝐄𝐑𝐓 use bidirectional attention, they lack the autoregressive capabilities crucial for generative tasks. Bitune combines the strengths of bidirectional attention for richer context understanding with the generative prowess of decoder-only models, resulting in superior performance. This is a leap forward in making LLMs more effective and efficient in real-world applications, from general-purpose assistants to medical diagnostics and beyond. Excited to see the impact of Bitune in the AI community!

16

1 commentaar

🚀 Looking to streamline fine-tuning for models like Llama 3, Mistral, and Mixtral? Our latest blog shows how Anyscale simplifies LLMs fine-tuning with Ray: 🔑 Key takeaways: • Scalable fine-tuning with LLM-Forge • Model Deployment with RayLLM Read more: https://lnkd.in/gQwD4A2U

34

Wake up babe, new finetuning method just dropped 👀 Introducing DoRA! A new alternative to LoRA that claims to give you more finetuning, with less data. DoRA, introduced in June 2024, builds on LoRA to make finetuning more efficient with fewer parameters. That means greater model performance with less data, (or get the same model performance with less data + compute). This can enable finetuning for use cases with poorer quality data, and makes finetuning more sclable and efficient. More context for folks interested below 👇 LoRA (and related methods like qLoRA) ‘finetune’ models by adding millions of parameters, containing the proprietary data you want to finetune the model around, to a ‘large’ model with billions of parameters. By doing so, you’ve adjusted a large enough segment of the models behaviour based on your own data, thereby ‘fine-tuning’ it for your use case. ‘Rank’ refers to the dimensionality of the data being added. ‘Low Rank’ matrices would have fewer columns/dimensions, vs. high ‘rank’ matrices would be rich data with a lot of columns. You can think of 'rank' as an approximation for 'how rich is that data being added'. DoRA lets you achieve the same amount of finetuning but with lower ranked data. According to the paper, "DoRA quality superior to LoRA especially in lower ranks: The difference in quality of DoRA of rank 8 and LoRA of rank 8 appears to be more significant than when training ranks of 32 or 64 for example." Where does this come from? A seemingly subtle but important change to how the new neural networks are built. In LoRA, weight vectors are approximated using fixed low-rank matrices added to the model's existing weights. And for each weight vector, both the MAGNITUDE and DIRECTION are adjusted together. In contrast, DoRA decouples the magnitude and direction of weight vectors. This allows for independent, dynamic adjustments during fine-tuning. This separation enables more precise and efficient fine-tuning. DoRA is able to dynamically adapting to task-specific requirements, often leading to better performance and resource utilization compared to LoRA's fixed-rank approach. 📚 https://lnkd.in/gNYBNF7Y https://lnkd.in/gSyFm27J

22

2 commentaren

One of the best indicators of a data science team's knowledge and maturity around LLM usage is the level of attention they are currently giving to upfront data preparation. 📣 Tomorrow, we're excited to be teaming up with Jerry Liu and the team from LlamaIndex to host a webinar on best practice data preparation for RAG workflows. We'll be showing how their LlamaParse solution for intelligent parsing and document transformation can be combined with Deasie's metadata creation workflow to build accurate and scalable knowledge retrieval over complex documents. When: Thursday 18th July, 9am PT / 12 ET Where: https://lu.ma/ufx77ve8 Who: Relevant for both AI engineers as well as company leaders looking to build their data foundation for GenAI

46

In this great video, you will find 18 nerds that do kickass voodoo stuff with regards to LLMs. OpenAI as of 2023 has 375 employees. Google in contrast employs 190.000 people and they are not able in my opinion to compete. Perplexity and Groq.com are other examples of outsmarting large players. Conclusion: This is the time of disruptive startups in the AI space - innovators dilemma on full display! On the other hand, I am constantly amazed how smart Microsoft is playing this game. https://lnkd.in/dYCDSZ_A

2

A great read on the performance and financials of LLM fine-tuning. Really hoping the outcome is a tool that helps overcome the biggest elephant in the room, Mr. Cost! #llm #finetuning #finance #computaioncosts

2

1 commentaar

Replacing an expensive generalist LLM like GPT-4 with a cheaper more specialized model like a fine-tuned Mistral 7B is a tried and true approach to really drive your GenAI expenses down. We've got a handy guide on how to start your fine LLM cost optimization journey with OpenPipe for fine tuning and OctoAI for model serving https://lnkd.in/e8udh8de

84

You have to love it when you can get a negative review that essentially says "Vendor made a change that makes their work less exploitable" when the software remains under an OSI approved license. People don't inherently understand open source companies. We didn't make KumoMTA open source because we're a bunch of anti-capitalist hippies, we did it because we considered it to be the best option for our customers. To keep that decision in place, our customers need to support us, otherwise the next option becomes either closing up shop or moving to a model that doesn't make supporting our efforts optional. Fortunately for us we work in a very niche space where our customers recognize the value we bring them and the importance of us continuing to do so, and steadily growing numbers of them are becoming sponsors and ensuring the continuous future of KumoMTA. Because of this unique combination of factors we haven't had to look at open core models or source available licenses to succeed.

6

1 commentaar

The fun never stops! Welcome to Cillers next hackathon at Magello Group on April 27, in collaboration with the world's leading software tech companies: Couchbase, Kong Inc., Curity, Polytope, Redpanda Data and Databricks. The challenge: Build A Super Capable, Scalable, Reliable And Secure Software System In Just One Day! A Cillers Hackathon is an amazing forum for developers to engage hands-on with game-changing software system technologies and collaborate with peers from Sweden's leading companies. It’s a great way to stay up to date with the newest tech adopted by Apple, SpaceX, Uber, Cisco, LinkedIn, VISA, Walmart, Electronic Arts and Akamai to build their systems faster. ⏩ Raving Reviews From Hackathon Participants: ⭐️ Developer at Klarna: "Had an exhilarating experience at the hackathon - felt like stumbling into the Y Combinator of the Scandinavia!" ⭐️ Developer at H&M: "Thanks for spearheading such an exceptional event, and fostering an environment conducive to technical exploration and innovation." ⭐️ Developer at SEB: "Incredibly inspiring! The Cillers hackathon is a great opportunity to try out the latest tech, alongside a community of talented and creative minds." ⭐️ Developer at Scania Group: “A lot of fun joining this one! Met a bunch of nice people, and got to work with some cool new technology." ⭐️ Developer at Treebula: "It was a fantastic journey blending innovation, teamwork, and cutting-edge technology!" You are free to build what interests you the most, but if you like, we'll have many suggestions for you. All are welcome to join, from beginner to expert, individuals or teams, anyone eager to learn. We'll have tech experts at the event for inspiration and support. This hackathon is about sharing knowledge, sparking inspiration and fostering a supportive environment for all. Read more and register here: https://lnkd.in/djahEgC2 Welcome to a full day of inspiration, good energy, and valuable learnings. We hope to see you at the hackathon!

43

🌍 Live from Berlin are thoughts on eBPF, Linux Kernel, Tracing, Auto-Discovery, Anomaly Detection, Edge Delta... Awesome technical talk by our very own Ozan Sazak as he walks through writing and implementing eBPF within an application at WeAreDevelopers conference. 💯 What makes it so powerful? Engineers cannot predict the future, and within #observability it's very difficult to find the right logic and query to ensure you have coverage for tomorrow's production outage. eBPF gives a reliable and standardized stream of data that allows you to have a lot more coverage than simply relying on logs and metrics. If you're curious about what we're doing in this space, come talk to us.

62

3 commentaren