Tim Laverty

Got inspired by a dutch podcast that creates stories from wikipedia pages. I used gpt-4o and their text-to-speech model to build something that sounded human. Of course it's not as great as the original podcast but this tech works way better than expected! Try it yourself 👇 Huggingface Space: https://lnkd.in/eKdT3XBY How did I do it? - Use <break time="1s"/> tag on strategic places to create breaks. - The TTS model is smart enough to understand these tags! - Prompt to use colloquialism english. - Use filler words like "um", "well", ... Bonus: I also generated an intro song using facebooks "MusicGen" and tweaked it a bit with with online tool called "AudioMass" Tools I used: AudioMass: https://audiomass.co/ MusicGen: https://lnkd.in/eAfjHZzA If your dutch speaking; the original podcast is called "Nooit Geweten", I can highly recommend: https://lnkd.in/ex66X6_a #gpt-4o #text-to-speech #tts #huggingface #spaces #podcast

14

2 commentaren

Can’t wait for this one to start ☝️ What a wonderful & practical opportunity to learn how to code with the help of AI tools like Replit Bite sized lessons to gradually build that confidence, and better yet with like minded builders. Come & Join me 🔥😍🔥 Thank you Harold D. and Max H. #CodingWithAI

15

4 commentaren

My article "Green Code: Why the World Needs More Than Just Efficient Algorithms" got featured as the cover story in SDN Magazine (Dutch)! As developers, we have the unique power to create software that doesn’t just serve people, but also serves the planet. 🌱 In this article, I explore how we can make a meaningful, positive impact by writing code that is not only efficient but eco-friendly too. 💡 If you're curious about how technology and sustainability go hand in hand, this is for you! Let's work together to create a greener digital future. Reach out if you want to learn more or get a copy of the magazine! 🌿 #GreenCode #SustainableSoftware #Sustainability

107

6 commentaren

This is the latest episode of Spilling the Source: https://lnkd.in/ekZXrest If you contribute to an OSS library, or are simply enthusiastic about one and would like to come on and talk about it then please get in touch, or reply to this.

2

Dutch startup orq.ai is on a mission to make AI accessible for everyone! 🛠️ Their no-code platform empowers engineers, product teams, and even non-technical folks to build and ship amazing AI-powered features. With support from Google Cloud and the Google for Startups program, Orq.ai is accelerating the adoption of generative AI, helping businesses innovate faster than ever. 📈 Why Orq.ai is making waves: 🌊 Less than an hour to start developing AI apps 🌊 Secure, reliable deployment of AI solutions 🌊 Increased collaboration between teams Learn more here ↓

15

1 commentaar

In the past week we've seen OpenAI release GPT-4o and Google release a new version of Gemini Pro 1.5, together with their new Gemini Flash model. The big news about these announcements is efficiency, speed and cost. Yay! 🐢 The challenge with GPT-4 has been its slowness - whilst it's a great model, it's so slow that I've had to rule out using it in several situations. If you need to make multiple calls to the model (very common) and you have a user sitting waiting on a response, GPT-4 is not viable. Why is it slow? Because it's a large model and larger models use more compute and take longer to respond. 🏎️ Enter GPT-4o, which is much, much faster. Why is it faster? Because it's a better optimised model that uses less compute and is able to respond quicker. And it's cheaper - 50% of the cost of GPT-4-turbo. When you make lots of calls to a model, a 50% cost reduction is a big deal. ⚡️ Now let's look at Google. Gemini 1.5 Pro and its stablemate 1.5 Flash both bring significant optimisations: "1.5 Pro, despite using significantly less training compute, surpasses 1.0 Ultra, a state-of-the-art model, on text capabilities like math, science and reasoning, code, multilinguality and instruction following. Similarly, 1.5 Flash outperforms 1.0 Pro, despite being more lightweight and efficient." 💸 Emphasising the importance of faster, more efficient, cheaper models, Gemini Flash is a new class of model explicitly designed to address these requirements. Because often we need something that responds quickly and costs less. When I say costs less, Gemini Flash is around 1/10th the cost of Gemini Pro. 🔥 But Google's not done with Flash as a smaller, more efficient model. The Gemini Technical Report also talks about a further model, Flash-8B, which is still under development. It's an even smaller model, optimised for uses such as large-scale data labelling and high-throughput agents, where speed and cost are major drivers. Of course Flash-8B is smaller and not as good as 1.5 Pro, but it's probably *good enough* for a whole class of uses. Not everything requires the most sophisticated AI available. 👉 These new models may not have set the world on fire with dramatic new capability leaps, but they do set *my* world on fire with faster, cheaper, more efficient models. They use less energy, they generate less CO2 (remember, we have a climate emergency), they cost less to use and they respond faster. If you need high-end capability then GPT-4o and Gemini Pro 1.5 both push the envelope on capability. This feels like the advancement we need - the option to pay more for better models, or pay less for more optimised and faster models.

5

1 commentaar

No human intervention allowed during deployment. Humans will eventually make mistakes. Deployments should be fully automated and should be executable, over and over again. Ideally in an incremental fashion. But also from scratch if needed in case of a disaster (mind the data). So how do you do that? First use 'infrastructure as code' to provision any shared infrastructure such as networking, security, communication or storage infrastructure. Ideally your infrastructure as code allows composition in its own right. We use Azure Bicep, which has support to load bicep modules from the azure container registry. This allows teams to maintain their own templates for e.g their subnet, while the network is still provisioned as a whole. The same goes for their storage, security and communication needs. Once the infrastructure is in place each composed host can be provisioned independently through its own deployment orchestration pipeline. Teams are free to choose the orchestration technology of their choice. Kubernetes, container apps, app services, service fabric, or whatever else they feel most comfortable with. Finally there is the data. Make a habit from automatically restoring and rebuilding your data on a regular basis Data on the inside: maintain backups and restore, as is. Data on the outside: no need to maintain backups. You should be able to rebuild this data from the original source of truth at all time. Rebuild it from scratch into a new version for every deployment containing schema changes. I personally prefer Event Sourcing for data on the inside. As the storage schema stays the same over time, it is very easy to backup and restore. While the concept of Projections allows you to rebuild the 'data on the outside' over and over again targeting new versions.

27

1 commentaar

OpenAI didn't have a great week, losing out twice: first to Runway's Gen-3 and then to the Moshi voice AI model. Here’s what else you missed in AI this week: 🧠 𝗜𝗺𝗮𝗴𝗲𝘀 𝗳𝗿𝗼𝗺 𝗯𝗿𝗮𝗶𝗻 𝗮𝗰𝘁𝗶𝘃𝗶𝘁𝘆 👉 Radboud University developed AI to reconstruct images from brain activity recordings accurately. 👉 The AI uses fMRI and electrode recordings, significantly improving reconstruction accuracy. ​ 👉 This tech could help stroke victims and recreate dreams, despite dataset limitations. 💬 𝗠𝗼𝘀𝗵𝗶 𝘃𝗼𝗶𝗰𝗲 𝗔𝗜 👉 Kyutai's Moshi, a real-time AI voice assistant, offers 70 emotions and speaking styles. 👉 Moshi features 160ms latency, surpassing OpenAI, and will be open-sourced soon. 👉 Developed by 8 researchers in four months. 🍏 𝗔𝗜 𝗼𝗻 𝗔𝗽𝗽𝗹𝗲’𝘀 𝗩𝗶𝘀𝗶𝗼𝗻 𝗣𝗿𝗼 👉 Apple is adapting its new AI features for the Vision Pro headset. 👉 Features include improved Siri, writing tools, and an OpenAI-powered chatbot. 👉 High price tag might deter users, despite advanced AI capabilities. 🎥 𝗥𝘂𝗻𝘄𝗮𝘆’𝘀 𝗚𝗲𝗻-𝟯 𝗔𝗹𝗽𝗵𝗮 👉 Runway's AI video generator Gen-3 Alpha is now available to all users. 👉 Gen-3 Alpha improves character consistency, camera motion, and scene transitions. 👉 Limited generation time for standard plans may challenge power users despite impressive features. 🏃‍♀️ LLM routing framework 👉 RouteLLM is an open-source framework for cost-effective routing between large language models. 👉 It routes queries between expensive and cheaper LLMs, optimizing cost and performance. 👉 RouteLLM achieved cost savings while maintaining 95% of GPT-4's performance in tests. Did you like this post? Please share ♻️ it with your network! Let me know about your favorite AI news this week in the comments 👇 --------- ➕ Follow Sebastian Illing for the latest AI news. ♻️ Repost to keep your network up to date.

39

19 commentaren

🌟 The Right Architecture for Your SaaS Platform 🌟 When it comes to building a SaaS platform, the architecture you choose can make or break your project. And guess what? The choice isn’t black and white—it depends on various factors. 🤔 Let’s talk about monolithic architecture first. It’s like a cozy studio apartment where everything you need is in one space. It’s perfect for simpler applications, quick launches, and when you want to keep things straightforward and snappy. Plus, if your team is small and you’re looking to hit the market fast, going monolithic can be like a speedboat—swift and efficient. 🏁 Now, onto microservices. Imagine a sprawling mansion with different rooms for everything. This is the go-to for complex, scalable applications that grow with your user base. If you’re planning to frequently update features, need high fault tolerance, and love the idea of using a variety of tech stacks, microservices are your best bet. It’s like having a team of specialists, each handling their own department. 🏠 For SaaS development, your choice should be tailored to the specific needs of your project, the expertise of your team, and the complexity of the application you’re dreaming up. Whether you’re a startup or a seasoned enterprise, the architecture you choose is the foundation of your service. So, choose wisely! 🧐 What’s your experience with monolithic vs. microservices architectures? Drop your thoughts below! 👇 #SaaSDevelopment #SaaSSolutions #SoftwareArchitecture #TechTalks #TechChoices #BuildingTogether

4

That pesky ECDSA signature refuses to go away, and we are struggling with it in a distributed world of trust. Luckily EdDSA is much more scalable, and now used in many distributed signature applications (and where we can split the private key up into secret shares and distribute them over several nodes). But we have a problem ... elliptic curve methods are going to be cracked by quantum computers. So can we take our PQC signature methods and scale them to distributed signatures? Well, no, as Dilithium doesn't support it. But Raccoon can, and also has better side channel resistance (and with a similar performance and key size to Dilithium. So, meet Raccoon: https://lnkd.in/d2QMHSM9

30

2 commentaren

TIL it's "confabulate" not "hallucinate" Thanks Beren Millidge! And this example from Birgitta Boeckeler demonstrates the future of coding is shifting to include more code review time. Suddenly those backlog items to refactor to a modular code base are looking mighty appealing... #ai #github #coding #programming

1

AR Glasses are making their way out into the world. Full AI integration, mixed reality, voice assistants, codec avatars and more...fully bringing home the point we’ve been saying for the past two years that AI and Spatial Computing/RT3D have been intertwined and will become further intertwined. It looks like a truly mainstream product is only a year or two away. https://lnkd.in/gpK9Xtwx

5

Many people have been asking about the importance of long context windows for language models (LLMs). Long context is a way to add more information at prompt time by generally prepending this information before the user prompt. This is a more dynamic way of conditioning models than fine-tuning. However, the problem is inherently quadratic in time and space. At Google , we have found a solution to address this need by splitting the context into many windows, which are dynamically compressed and updated. This approach is currently a hot research problem, and we are constantly working to improve it. Stay tuned for more updates! #llms #largemodels #longcontext #compression #attention #quadratric #approximation

70

6 commentaren

The game of large language models (LLMs) got a major plot twist! Meta's unveiling of Llama, a family of AI models pushing the boundaries of parameter size and processing power, is a game-changer. Why it matters? Llama boasts a staggering 405 billion parameter model, dwarfing previous heavyweights like Google's LaMDA and OpenAI's Jurassic-1 Jumbo. This sheer size unlocks unprecedented capabilities in: -> From scientific papers to philosophical treatises, Llama can digest vast amounts of data, leading to deeper comprehension. -> Imagine writing styles that perfectly mimic different authors or create compelling narratives on the fly. -> From machine translation to chatbot interactions, Llama holds the potential to transform how humans and machines communicate. -> Llama isn't just about brawn. Meta stresses its commitment to open-source access and responsible AI development. This fosters collaboration and helps mitigate potential biases often associated with large datasets. Meta's ambitious project is a harbinger of the future, where AI can not only process information but truly understand it. Imagine personalized education tailored to individual learning styles, or AI assistants that anticipate your needs before you even ask. https://lnkd.in/dYn_cTjB #MetaLlama #ai #LLMs #futureoftechnology #naturallanguageprocessing #opensource

6

STOP USING AI FOR EVERYTHING According to Alex de vries phd by 2027 ai will consume enough energy to power the netherlands for a year. Stop using ai to write terrible code for you, write terrible code yourself like the rest of us. Stop using ai for “marketing” and use google keyword planner and make horrible logos in photoshop like the rest of us. Don’t get me wrong, ai is exciting and has many great use cases but you do NOT need to shove it into every app that you write just for the sake of having it. Ai in its current state is what binary is to python. A neural network is essentially just a regular program with decision based weights in between the input and output. We have a LOT of work to do before ai becomes performant. From one tech bro to another, please, for the planets sake stop using ai for every single thing you do. We have plenty of tools that work just fine without it. #ai #artificialintelligence #sonofanton Source:

1

🚀  Llama 3.1 is a game-changer! I had to blink my eyes today when I ran some tool-calling experiments (in Dutch): the 70B version performed perfectly on tough cases, just like GPT-4o. And Officially Llama 3.1 does not even speak Dutch. I ran it on a double RTX 4090 GPU for 80ct/h (20 euros a day)! ‘Tool calling’ is the glue between human speech and the system executing the spoken request. So far self-hosted LLMs were not strong enough to do this reliably even after fine-tuning, so an external LLM like OpenAI GPT-4 was always needed. No more! Now banks, insurance companies, and governments who are afraid of privacy and security issues with external LLMs can run a high-quality model on 4000 euros of graphics card! The limit is gone! Last but not least, the 8B version (which you can host for 15ct an hour) is also really good at tool-calling. It has minor glitches, but if your use case is not too demanding it is a blazingly fast-responding little beast.

26

6 commentaren

Two months ago, I was at an AI meetup in Amsterdam. But here's the truth: AI can help us understand and improve human behavior. Sounds futuristic? It's happening now. Charlotte Gerritsen showed a solution that recognized behavior in crowds: ↳ detecting when chaos might start, ↳ identifying changes in movement patterns, ↳ capturing facial expressions to predict stress. This technology allows proactive measures: ☑ Calming people down. ☑ Preparing exits. ☑ Ensuring safety at events. Imagine bringing this to the office: ☑ Capturing stress moments. ☑ Noticing when teamwork breaks down. ☑ Identifying toxic competition. With this data, we can: • Implement team-building activities. • Introduce mindfulness breaks. • Foster a healthy work environment. But there's a dark side. Misuse is possible. Abusive monitoring will drive people away. AI itself, as everything else simply is. We are the ones that give it meaning. It's up to us to decide: ☑ Use it for good. ☑ Avoid misuse. Let's use AI to create supportive, safe spaces. It's more than being nice. It saves you bare $. Aetna the insurance company has proven to save 3'000 $ per employee annually. Just by reducing cost due to a mindfulness program they've rolled out. Lead with empathy. And technology will follow. Just as people will. Follow Damian Nomura for a daily dose of Human - AI inspiration

46

40 commentaren

A very eloquent summary of the current status of the EU GDPR and AI legislation. The time, energy, and money invested in restrictive legislation that does not provide any consumer benefit could have been leveraged elsewhere. The amount of data collection in EU surprises me every time, the fact that it comes with a long boilerplate explanation about my privacy rights does not make me feel any better either. The EU should rethink how it legislates not only tech.

4

1 commentaar