Daniel Miner’s Post

𝐎𝐩𝐞𝐧𝐀𝐈'𝐬 𝐒𝐭𝐫𝐚𝐰𝐛𝐞𝐫𝐫𝐲 𝐦𝐨𝐝𝐞𝐥 𝐢𝐬 𝐟𝐢𝐧𝐚𝐥𝐥𝐲 𝐨𝐮𝐭. 𝐒𝐨𝐦𝐞 𝐢𝐧𝐢𝐭𝐢𝐚𝐥 𝐝𝐚𝐭𝐚. 𝐓𝐡𝐞 𝐠𝐨𝐨𝐝, 𝐚𝐧𝐝 𝐭𝐡𝐞 𝐛𝐚𝐝 OpenAI has finally unveiled its new o1 models, aka "Strawberry". A lot had been speculated about its capabilities, and it being the first AGI model. Let's see some basic ideas about how it works, benchmarks, researchers findings and some examples of how people is using it: 𝑯𝒐𝒘 𝒊𝒕 𝒘𝒐𝒓𝒌𝒔 𝒄𝒐𝒎𝒑𝒂𝒓𝒆𝒅 𝒕𝒐 𝒑𝒓𝒆𝒗𝒊𝒐𝒖𝒔 𝒎𝒐𝒅𝒆𝒍𝒔: o1 implements Reinforcement Learning to "think" before it provides answers via a private chain of thought. This means that the model's "ability" to reason is not only limited by the pretraining phase, it can also happen during inference, therefore opening another potential scaling way forward (bad news: sustainability). 𝑩𝒆𝒏𝒄𝒉𝒎𝒂𝒓𝒌𝒔: o1 exceeds an expert human in PhD-science questions. This is truly remarkable, and opens the questions of how are we going to validate some model outputs going forward, if we humans don't simply have the capacity to validate the facts. 𝑹𝒆𝒂𝒍𝒊𝒕𝒚 𝒄𝒉𝒆𝒄𝒌: while o1 is a big step forward in reasoning capabilities, it is far from AGI. It can still fail in simple reasoning tasks such as tic-tac-toe, as researchers from OpenAI acknowledge. It actually performs worse than GPT-4o for some use cases. 𝑾𝒂𝒓𝒏𝒊𝒏𝒈: with this new reasoning capabilities, more challenges arise. o1 tries to find creative ways of solving the task it has been assigned. In some cases, researchers have found that this might result in security issues, such as the model hacking docker containers. Benchmarks and cherry-picked use cases aside, I am really interested in the possibilities that combining these models with tools such as Cursor or Replit might bring. As an example, you can see in the following link a video of a 3D snake game being created in under a minute using o1 + Replit: https://lnkd.in/eRdeeaue Additional reading: Ethan Mollick's blog post after using the preview model for a month: https://lnkd.in/e7AMj_d5 #openAI #o1 #reasoning #generativeAI

Swiss Team NTT DATA DACH at the #dataandaileadershipday and #cloudleadershipday taking place at Zürich Sihlcity. Our colleague Daniel Miner did a fantastic speech on the benefits of GenAI in the software development lifecycle. Thanks to Daniel Graf to once again orchestrate our presence at the show. #GenAi #futuretech

💪𝗖𝗼𝗻𝗾𝘂𝗲𝗿 𝘁𝗵𝗲 𝗺𝗶𝗱𝘄𝗲𝗲𝗸 𝗵𝘂𝗺𝗽! 𝗪𝗲 𝗮𝗿𝗲 𝗶𝗻𝗷𝗲𝗰𝘁𝗶𝗻𝗴 𝘀𝗼𝗺𝗲 𝗺𝗶𝗱𝘄𝗲𝗲𝗸 𝗺𝗼𝘁𝗶𝘃𝗮𝘁𝗶𝗼𝗻 𝘄𝗶𝘁𝗵 𝗮 𝘀𝗽𝗲𝗮𝗸𝗲𝗿 𝗮𝗻𝗻𝗼𝘂𝗻𝗰𝗲𝗺𝗲𝗻𝘁! Expert speaker from NTT DATA, Daniel Miner is spilling the secrets behind Gen AI models. Discover how these cutting-edge technologies can turbocharge your coding efficiency, automate documentation, and accelerate rapid prototyping. Understand how GenAI revolutionises debugging, code review, and user interface design, propelling productivity and innovation throughout the software development lifecycle. From conception to deployment and maintenance, see for yourselves the transformative power of GenAI! 🎙️https://bit.ly/4bJMau9 🚀 Don't let the midweek slump hold you back! Join us on June 19th and elevate your prospects to new heights! https://bit.ly/4c7rqMY Daniel Graf #midweekmotivation #GenAI #futuretech #softwaredevelopment #Innovation #dataandaileadershipday #techconference

NTT is conducting research aimed at realizing general-purpose AIs under the concept of the AI Constellation. This approach leverages the collective intelligence of small, highly specialized, and unique AIs to solve the increasingly large-scale and complex problems faced by society and businesses. https://lnkd.in/g7vnJr-n #NTT #AI #AIConstellation

As always, and despite the progress of what was presented yesterday by OpenAI, it is good to look for a critical counterpoint. Some examples shared by Gary Marcus on X, show that #GPT-4o continues to hallucinate, making absurd errors when faced with simple tasks and, of course, still far from reasoning. Some captures from examples in the thread below. With #GenAI models getting as powerful as we are seeing, it is even more important to hone our critical spirit in the face of the answers they give us, and, above all, to understand their limitations. 🔗 https://lnkd.in/dmiD2mWt

For me, data meshes do for data what APIs do for applications: driving scalability and accessibility for developers and ultimately business users. Read this fantastic blog post of my colleagues Dr. Najmeh Foroozani and Esmaeil Nikumanesh on how data meshes could and should become a part of your data strategy: https://lnkd.in/epAauSbh #datamesh #datastrategy NTT DATA DACH

🚀 Wie können Unternehmen in nur drei Monaten eigene KI-Anwendungen realisieren? Unser neuer Fachbeitrag "In acht Schritten zur eigenen GenAI" von Daniel Miner, Head of AI NTT DATA DACH, gibt Antworten und Einblicke in den Weg zur erfolgreichen Implementierung von KI-Anwendungen im Unternehmen mit praktischen Erfolgsbeispielen von #ChatGPT und weiterer Anwendungen. 💡Auf einen Blick die wichtigsten Schritte: 1️⃣ Technologie verstehen: Informieren auch Sie sich über die Möglichkeiten und Grenzen von ChatGPT und anderen KI-Modellen, damit Sie sich realistische Ziele setzen können. 2️⃣ Erwartungen managen: Setzen Sie sich klare und messbare Ziele, um zu praktischen Ergebnissen zu gelangen. 3️⃣ Menschen im Fokus: Die Mitarbeitenden mit ihren Fragen und Anliegen wahrnehmen und aktiv in den Prozess einbinden. 4️⃣ Daten qualifizieren: Gewährleisten Sie, dass Ihre Trainingsdaten qualitativ hochwertig und ausreichend sind, um optimale Ergebnisse zu erzielen. 🌐 Bereit für Ihre eigene GenAI? Hier geht es zum vollständigen Artikel 👉https://lnkd.in/eWbJcQax #KünstlicheIntelligenz #GenAI #Unternehmensentwicklung #KI