Ramin Assadollahi

I had the luxury to wake up in the sunny alps this morning! Yesterday, I enjoyed some great #AI in the #Alps real talk and discussions: 1️⃣ On AI certifications: Why sampling the application domain (a.k.a. constructing the test set) is the crux of the matter and which methods can be used to approach that systematically on a meta level. 2️⃣ On AI-based simulation: That the next generation surrogate models are LLM-based, much more powerful and probably won't use any a-priori physics based constraints. 3️⃣ On leadership and change management: Why German organisations tend to make slow ('ich muss da erst nochmal mitnehmen') and timid ('ich muss da erst noch jemanden abholen') decisions. + some great life advice in the evening 😉 Thanks for organizing this event and making it memorable! Robert Weber Peter Seeberg Barbara März Johannes Brandstetter Bernhard Nessler Klaas Nebuhr Dr. Andreas Nauerz Michael Fischer Kirsten Prill Bastiaan O. Werner Faulhaber Fabian Kreutmayr + many more... #ai #alps #collaboration #realtalk

77

7 Kommentare

People are so focused on the Input Context of LLMs but often overlook that there are limitations in the output. Long input contexts still present two problems: • The model does not give equal attention to all parts of the text (this is quite commonly discussed). • The output is not as long as the input context. Therefore it might not be granular enough. For many tasks, such as "Extract ALL the topics in this long text" chunking the text yields much more granular results. Obviously, the caveat is that the chunking strategy becomes a very important piece. For advanced cases, beyond simple summarization, chunking is still the better strategy. #LLMs #Chunking #LongInputContext

13

1 Kommentar

📊 AI's Future: Insights from Leopold Aschenbrenner's "Situational Awareness" Former OpenAI researcher Leopold Aschenbrenner's thought-provoking essay offers a glimpse into AI's potential trajectory. Key takeaways: 📈 Trust the trendlines: Deep learning's consistent progress defies skepticism. 🚀 Rapid advancements: AI continues to surpass expectations across various benchmarks. 🧠 AI researchers by 2027?: Plausible scenario of AI performing complex cognitive tasks. 🤝 From chatbots to coworkers: Anticipate more agent-like AI interactions. 📉 The data dilemma: Potential bottleneck as we exhaust available internet data for training. While speculative, Aschenbrenner's analysis provides valuable insights into high-level AI discussions in Silicon Valley. As we navigate this rapidly evolving landscape, it's crucial to stay informed and adaptable. #ArtificialIntelligence #AIFuture #DeepLearning #TechTrends

3

2 Kommentare

𝐎𝐩𝐞𝐧𝐀𝐈'𝐬 𝐆𝐏𝐓-𝟒𝐨 𝐀𝐈𝐦𝐬 𝐭𝐨 𝐛𝐞 𝐲𝐨𝐮𝐫 𝐛𝐞𝐬𝐭 𝐟𝐫𝐢𝐞𝐧𝐝, 𝐧𝐨𝐭 𝐣𝐮𝐬𝐭 𝐲𝐨𝐮𝐫 𝐚𝐬𝐬𝐢𝐬𝐭𝐚𝐧𝐭. ChatGPT-4o was announced yesterday. I feel that some people were expecting a more improved GPT version, featuring a larger model, better answers, faster responses, and other technical enhancements. However, after watching "𝘐𝘯𝘵𝘳𝘰𝘥𝘶𝘤𝘪𝘯𝘨 𝘎𝘗𝘛-4𝘰", I realized that these are not the main goals of the next ChatGPT version. It's about the humanization of AI. The live demos in the video showcase various interactions with ChatGPT, which now can: - Use audio, video, images, and text all at the same time, - Have a faster response time for voice inputs to allow real-time communication, and - Be available as a desktop app to be omnipresent. ChatGPT is positioning itself not merely as an "assistant" but more as a friend—maybe even your best friend—with whom you talk all the time and share your thoughts, moods, and more. Also, this leads me to think why ChatGPT might not try to build or rival a search engine as some rumors stated. As we will probably talk to ChatGPT more than to any other human, and as such, it can become the go-to for everything, everywhere, and be all things at once. A lexicon, a teacher, a friend ... a life companion? For those who watched the memorable movie "Her" with Joaquin Phoenix, this vision of ChatGPT is probably far closer to where it's heading than a replacement for the Google Search.

16

3 Kommentare

I’ll be at ICLR’24 in Vienna next week and look forward to connecting with many of you, old and new friends! My team (https://lnkd.in/gmejMnn7) will be presenting the following papers: - Chain-of-Table: Evolving Tables in the Reasoning Chain for Table Understanding: https://lnkd.in/gFJ9epQE - SPADE: Semi-supervised anomaly detection under distribution mismatch (TMLR featured paper): https://lnkd.in/gRCdFCAk - TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting: https://lnkd.in/gs8-M5eG This represents just a small subset of our recent papers (https://lnkd.in/gHMzZi9W) on pushing SoTA to make LLMs work for organizations’ real-world problems. Given how the foundation model quality is starting to saturate and the gap to open-source is reducing, we're particularly investing in not just improving the core foundation model, but also in developing tooling on top of LLMs, in the spirit of compound systems (https://lnkd.in/gwdWd45S). These projects involve things like better adaptation methods (e.g. distillation, adaptation methods for embeddings and grounding), improving LLM tool usage, improving LLM usability with automatic prompt optimization, and developing LLMs for high-demand use cases such as Text2SQL. We also have several projects working on improving the core foundation model, such as tackling the lost-in-the-middle problem of long-context LLMs. If this sounds interesting, feel free to drop by the Google both on Tue/Wed/Thu 9.30am, or email me directly at tomaspf@google.com to set up time during the conference or over Google hangout after. #iclr24 #deeplearning #LLM #googlecloud #machinelearning

84

Apple targets Google staff to build artificial intelligence team - SWI swissinfo.ch https://buff.ly/4biSyYK - 🔍 Apple has been hiring AI experts from Google to expand its global AI and Machine Learning team - 🏢 Apple has a secretive European laboratory in Zurich, Switzerland - 📉 Apple's shares have slipped, putting pressure on the company to announce game-changing AI features - 💬 Apple is focusing on deploying generative AI on mobile devices - 🧠 Apple has been developing AI products, such as Siri, based on neural networks - 🤝 Apple's top AI team includes former key figures from Google - 🧑‍💻 Apple has acquired numerous AI start-ups in the past decade - 📱 The next big thing for Apple could be AI smartphones with more powerful chips - 🤐 Apple has been tight-lipped about its AI plans - 💡 Apple's foray into generative AI features may be revealed at its Worldwide Developers Conference in June

2

1 Kommentar

The new episode of Spilling The Source is out. this time I talk to Marco Vinciguerra about Scrapegraph AI. A site scraper that uses LLM models for processing. https://lnkd.in/eJ2m6APz

5

2 Kommentare

#LLM features are no longer a mystery - they are being decoded. The recent paper from #Anthropic attempts to "map" the mind of their #LLM #Claude3. We will be able better understand the way models "think". Also, these features can be used to influence the model behavior to the desired one. A broad range of safety-oriented features are observed: including deception, sycophancy, bias, and dangerous content. Read more in the paper below: https://lnkd.in/e-s8rkyc #machinelearning

31

#AI is not hype when it comes to #lowcode and #nocode! #AI floats all boats - and by boats, I mean #lowcode, #nocode or any model / structured application development platform. AI can help structure spaghetti code for sure, but that's another article for another day. This is why I have never been more excited to be attending #Pegaworld next week. After founding my #RPA company in 2005, that Pegasystems acquired in 2016, I got to spend over 7 years with some of the great minds under the Alan Trefler leadership. And in that time, I became more passionate than ever, that the concepts within #lowcode would be THE future of all enterprise software application development. And I've continued to be passionate, and even more vocal ever since my tenure with Pega ended. And you all thought I was retiring ;) So, with the advent in more recent times of #GenerativeAI, it has become clear to me that whilst others are still figuring out the AI hype, or where to apply it, these #lowcode technologies were born to infused with AI from the get go. Just waiting for their time. I simply couldn't retire in the midst of this revolution combining #GenAI with the power of #Lowcode. Check out my discussion on the "Low Down on Low Code" #podcast below; See you all in #Vegas my Pegasystems friends, I am looking forward to seeing you and also everything, that's new! HFS Research Joel M. Phil Fersht Rob Koplowitz Dave Marcus John R Rymer John Bratincevic Leon Trefler Don Schuerman Damon Lockwood Eric Musser John Higgins Ken Parmelee Brent Leary Analysis.tech Kieran Gilmurray Pascal BORNET Rob Walker Kerim Akgonul https://lnkd.in/eN4AEcT3

17

7 Kommentare

Recently, tech social media blew up with a discussion of whether we need AI control systems (or whatever you want to call them*). Again. The sensationalist takes could be approximately summarized as: “We’d better do something urgently so AGI doesn’t outsmart us!” versus “AI is still pretty dumb, chill out.” I’d like to offer a perpendicular perspective in the linked blog post I wrote for you. (In which we both chill out and build something.) In short: "Nevermind AGI, GenAI, or anything AI. We're not smart enough to design, build, and steer complex technology responsibly and technology is getting more complex quickly. Let's build tools that make us more capable." Or, more bluntly: "Have you noticed that the people in charge are never the human-in-the-loop? There's a loop they need to be in... and they're not there." *For the semantics fan club, I put a glossary of 57 AI-control-related terms for you at the bottom of the blog post. If you're tempted to argue about word meanings, please enjoy those first. 57 of them should surely provide all the nuance you can guzzle in one go. And color-coded too! <3

524

82 Kommentare

Cassie Kozyrkov created two very good articles related to AI on Medium. Thank you Cassie. 1️⃣ The first article ("Glossary: Controlling AI Systems") describes a long list of AI related terms. What is also interesting, and useful is that she also paints a picture showing where the focus is of each topic (e.g., engineering vs policy). 2️⃣ The second article ("Why We Urgently Need AI Control Systems") highlights the need for making AI technologies more accessible and useful to the domain experts who need them - this includes the need for having experts and leaders involved in testing and validation. 📃 Articles https://lnkd.in/e9JmFtzc https://lnkd.in/eWk6M2Ze 📖 Glossary https://lnkd.in/e-dzeVWk

1

A very nice article expalining the Small Language Models. https://lnkd.in/dgTuspiS The Microsoft Phi-3 is one of these.

3

New OpenAI features (GPT-4 Omni) in a nutshell - Deutsch 👇 Impressive news from the OpenAI Spring Update: 👉 Native Clients for ChatGPT 👉 New Model called "GPT-4o" (o is the letter O like omni:-)) 👉 Improved modalities for voice (you can interrupt the model, no lags) and vision (integrated recognition) 👉 ChatGPT with GPT-4o free for everybody - so everybody can also use the GPT Store (pro users get 5x bigger capacity limit) 👉 API access to GPT-4o 👉 Real time translations - the Demo was impressive! Nothing about improved agents support - but I am sure not too far away ;-) -- Beeindruckende Neuigkeiten aus dem OpenAI Spring Update: 👉 Native Clients für ChatGPT 👉 Neues Modell namens "GPT-4o" (o ist der Buchstabe O wie omni :-)) 👉 Verbesserte Modalitäten für Sprache (man kann das Modell unterbrechen, keine Verzögerungen) und Sehen (integrierte Erkennung) 👉 ChatGPT mit GPT-4o kostenlos für alle - damit kann auch jeder den GPT Store nutzen (Pro-User bekommen 5x größere Kapazitätsgrenze) 👉 API-Zugang zu GPT-4o 👉 Übersetzungen in Echtzeit - die Demo war beeindruckend! Nichts über verbesserte Agentenunterstützung - aber ich bin mir sicher, dass es nicht allzu weit her ist ;-)

8

What Karl says. There's going to be two amazing talks about Hopsworks based #ai and #ml systems: * Build a personalized Bitcoin (BTC) virtual assistant in Python with Hopsworks and LLM function calling by Javier de la Rúa Martínez (https://lnkd.in/eQfWpFRU ) * Build TikTok's Personalized Real-Time Recommendation System in Python with Hopsworks by Jim Dowling (https://lnkd.in/eXEHePjG) If you are in the #berlin area next week - go and take a look!

7

When integrating LLMs with a private knowledge platform for your enterprise, there's one vital question you need to ask yourself: What model do we want to use? And when it comes to choosing a model, you have two options: A. Building your own model or enhancing an existing model;  B. Using a pre-trained model: a commercial (or not) off-the-shelf model. Each of the options has its pros and cons. This infographic shows the main pros and cons of the two options and comparison of what it would mean for your organisation to adopt one strategy over the other. What challenges are you facing integrating LLMs with your private knowledge platform?

169

17 Kommentare

OpenAI are holding a product announcement at 10:00 PST / 18:00 BST / 19:00 CET today: https://www.openai.com/ Is GPT4 getting upgraded today? Is it getting sibling products and services? CEO Sam Altman has been talking up a storm in recent weeks about the future of AI ('Universal Basic Compute' access for all global citizens? A cognitive version of 'let them eat cake' I fear). At the same time he has stated that today will not be GPT5, nor search, but will feel 'magical'. Hmmm. The only way to find out will be to listen in or come back here afterwards to discuss what they are releasing. My hope is that they build on the practical benefits of custom GPTs and add Agents, and improve the voice interface. Having well integrated search functions to allow for clear referencing would also be neat, especially if it is done on the custom knowledge that is uploaded. Maybe we'll hear more about the rumoured integration into iOS 18. Whatever it is, you can be sure that here at Curistica we'll be first to explore the strengths, weaknesses, risks, and real world use cases for you! What are you hoping for?

23

9 Kommentare

In a strange way, Europe appears to be the dark horse. The EU seems absolutely determined to ban everything, including AI and energy. On the other hand, we have this incredible AI company Mistral AI - the leading open-source AI company right now! Keep up the good work, European creators. We are actively looking for European founders who can leverage the incredible ecosystem we are currently building. We are specifically focusing on vertical AI models and applications in the areas of #energy, #transportation, #industry, #materials, #carbon, #climate, and the #built environment. #france #europe #AI

21

💢Despite the hype around #ai, particularly #LLMs and #RAGs, these advanced features haven't significantly impacted decision-making in dynamic and regulated environments. Do you know why? Because #llms #rag and generative #ai are not the right choices of technologies for decision-making in dynamic and regulated environments - no matter what! So what is the right choice? ✨The Decision Intelligence is all about the right choices of #ai algorithms and techniques for decision-making in organizations. And what is the "Decision" in #DecisionIntelligence? 💫A "decision" is a hierarchical and multi-step model representing a holistic view of business decisions. - It has all the components and sub-components of the business decisions - It depicts the relationship between those components and sub-components - It specifies the required input information - It enables the implementation and encapsulation of any business logic with different shapes and forms, such as #rules #optimization #machinelearning, and so on for those components and sub-components. 🚀The "Decision" in "Decision Intelligence" is a neuro-symbolic AI model that enables the building of composite AI solutions by combining the best technologies such as #businessrules #machinelearning #optimization #orchstration #data. What about Intelligence? ⛳Intelligence has multiple layers. It essentially starts with "Knowledge Core" and "Intelligence Core," and then it can be extended to "Expert Intelligence," "Cognitive Intelligence," and "Autonomous Intelligence." Each layer of intelligence is a prerequisite to the next one, making it so powerful and effective for decision-making in organizations, especially in regulated and dynamic environments such as #banking #insurance #government #financialservices #health #transportation #logistics and so on. 💣If you are a data and analytics leader (#CDAO), a data product leader, or a #DataScientist with a passion for AI in decision-making, the "Decision Intelligence" and "Decision Intelligence Platform" are your friends; unfortunately, not #LLMs and #RAG.

7

1 Kommentar