Wendy Chang

Leopold Aschenbrenner — formerly of OpenAI's Superalignment team, now founder of an investment firm focused on artificial general intelligence (AGI) — has posted a 50,000-word, five-chapter, 165-page paper, "Situational Awareness: The Decade Ahead". which is being quoted by Vinod Khosla and everyone for a crystal ball for the future of AI. and including which industries or professions are gonna be destroyed. 1. "DONT BET AGAINST AI - If there's one lesson we've learned from the past decade of AI, it's that you should never bet against deep learning.We're literally running out of benchmarks." 2. "strikingly plausible that by 2027, models will be able to do the work of an AI researcher/engineer." 3. "There is a potentially important source of variance for all of this: we're running out of internet data. That could mean that, very soon, the naive approach to pretraining larger language models on more scraped data could start hitting serious bottlenecks" (this is stuff that Wootzapp is working on and synthetic data that I'm working on) https://lnkd.in/gPtb2nu7

33

3 Kommentare

What’s next for Langauge Models #OpenAi By Colin Jarvas, the Chief Architect. Super interesting talk by Colin at #TOA Berlin. 😎He talked about : 1. Some good and bad examples where the chatgpt was used. 2. What are places it’s been extensively used and are successful: 👩‍🚒Usecase : - Customer Service - Autonomous Assistant - Content creation 👩‍💼User experience : - Balancing control and automation - Multi-modality - Translations 3. What is coming next ? - Increasing intelligence of text model - Increasing multi-Modality And more. Parting Advice for all the people who are using OpenAI’s platform is there in the video. #ai #openai #whatsnext #chatgpt #AIEngineers

2

Hey all 👋 I've published a new blog post: How We Integrate AI in epilot GmbH. In this post, I've shared our learnings, challenges, next steps, and the mindset we have around integrating AI at epilot GmbH. I’d love for you to give it a read and share your feedback!

31

Who can find job on LinkedIn Today: From Developer to C-Level Last week, we finished our 12th LinkedInPUMP group. With this group, we have over 150 participants and an incredibly high NPS. Here are the key conclusions from 12th Group: 1/ Developers and Analysts are in Trend Again! It was amazing to see all the analysts and developers in LinkedInPUMP group getting flooded with messages from recruiters. It’s not just 1, 2, or 5 messages each — it’s 10-20! Now I can say for sure: if you want international job offers in development and analytics, come to me! We hacked the system! 2/ C-Level can get a job only through Networking. Unfortunately, no one in Europe is eager to hire not local COO, CEO, CFO (and also Directors). The only option is proactive networking, using connections with former colleagues, including “Russian-speaking companies.” We found that many C-level professionals do not have clear positioning and personal branding when entering the international market. So by the end of the training, all our C-level participants repackage themselves and find “their niche.” We have a lot of experience in this. 3/ You can promote your Services and work in a company at the same time. What to do if you work in a company and want to earn extra money (by selling services, consulting)? It’s easy - we found a way to combine both, and the system gives very fast results. If you have the desire! 4/ Not only IT specialists can build an International Career. Here is a list of professions that have good chances for European job offers (if experience allows): Project Managers, Product Managers, all types of M&A, all types of marketing, HR (with an advantage for Talent Acquisition and Recruiters), Finance (controlling, treasury), Logistics (a bit harder, but there are options). It’s great that we built a training system where everyone achieves results! In 9 days, we start the LAST LinkedInPUMP group this year! You won’t find a better opportunity — in November, people are still active, so you can grab the last job openings before the New Year holidays. Hurry if you want to get a job offer by spring! Sign up here! https://lnkd.in/efEE2JAf

18

1 Kommentar

First morning of GenAI Summit highlights: From Lukas Biewald talk on Evaluation - while there is not a common quality benchmark to AI quality, it is critical to define your own and build it right after your PoC and way before moving to production - when working with AI, keeping records of your experiments, both failed and successful, is probably the most valuable part of your IP. That is counterintuitive to most SW development teams, used to only "commit to their repo" the "working code". From Tengyu Ma talk state of RAG - the 2023 discussion between long-context vs fine-tuning vs RAG is finally over, and RAG is the clear way to go - in the near future Embedding models and LLMs will handle most of the work (like multimodal embeddings generation), and we'll be able to remove many of the manual tricks that we need today to make RAG accurate - privacy = on-premise 👀

15

3 Kommentare

Biggest meetup in Berlin so far – when people are still exchanging ideas at 11PM, you know it was worth it! 🎉 My team and our CTO, Björn Niederfranke, took the stage to share our journey with AI in Software Engineering, myth-busting some big misconceptions, together with Ulrich Homann, Corporate VP at Microsoft. Here’s what we debunked: 1️⃣ "AI will replace Software Engineers." AI can automate many low-complexity tasks, freeing engineers to focus on more complex challenges. But from our experiments, no LLM has managed to write deployable code in place of a human – especially in complex, legacy environments. 2️⃣ "Fewer engineers are needed with AI." In reality we’re drowning in requirements. AI’s role will be to support engineers in handling more, not in reducing teams. 3️⃣ "Just get a copilot, and you're done." After a year of experimenting, our estimate is around a 5% efficiency gains by next year. Like any tool, results vary based on the skill of the person using it. 🍓 Lastly, we shared our latest project: using Graph RAG on our codebase to overcome the context window limit, we are building a ChatGPT-like experience for understanding and optimizing our software. Exciting times! Big shout out to my colleagues for organizing such a great event 🙌 Gregor Schumacher; Kevin Rampelmann; Marcel Gocke; Regina Gerber; Anna Werner; Sujay Joshy; Axel Scherer 👉 Interested in this space or considering a role at Riverty? Let’s connect! #AI #SoftwareEngineering #TechMeetup #Berlin #Riverty #GitHubCopilot

67

5 Kommentare

System 1 and System 2 Thinking for LLMs - but no #AGI One week after #OpenAI's #o1 release is a good point to reflect on the paradigm change implied by o1's working principle. Let's break it down. Also so called chain-of-thought is no rocket science. So far models had no chance to take time to answer. Every answer was basically the first best shot. Very similar to our human intuition (as labelled by Daniel Kahneman as system 1 thinking), which is great in many cases and just wrong in other cases. Us humans have the great luck to be able to reflect and resonate. Therefore, shiftting response processing from system 1 to system 2. 2+2 might work nicely with intuition. 38*3 is arguably tough without further processing. Same holds true for LLMs. o1 is now specifically trained to reflect on own responses and therefore can take some time to answer. We are paying for better response quality with higher processing time. Finally, we can trade time for quality also with LLMs. Especially valuable for analytical tasks like coding and math. Not so much for creative tasks, where a clear right and wrong just do not exist. This is also what we see reflected in o1 benchmarks. #norocketscience #o1 #gpt #genai

17

2 Kommentare

Recently, I had an intriguing task: I saw my wife enter our daughter's play rehearsal schedule in the calendar manually. I decided to see if I could use LLMs to take events from a detailed rehearsal schedule in PDF format and extract calendar events for my daughter’s rehearsals. This was more than a simple extraction - it had to be personalized - it required understanding nuanced details like recognizing that “full cast” events should include her (LLMs initially missed that), roles she was playing, conflicts etc. I tested several AI tools, including GPT-4o, Gemini, and Claude 3.5 Sonnet. While the initial data extraction had its hiccups (it took multiple passes to get all the events) - I was a bit surprised by that, the real magic came in what these tools could do once the data was extracted: • I got close to 40 events for 2 weeks, then I asked the events on the same day to be grouped together. It was very impressive - it not only grouped but embedded the event details that were grouped into the description. • Renaming the events to improve clarity and context was just a simple prompt. • Finally automatically generating an .ics file with ready-to-import calendar events. This process highlighted how AI’s strength often isn’t just in extracting data but in post-processing it into actionable, user-friendly outputs. In this experiment Claude 3.5 Sonnet performed the best. The key insight? AI tools require thoughtful guidance and iteration, especially when personalizing outcomes - this needs domain expertise. As a tech leader, this reinforced the importance of adaptability—not just evaluating tools for accuracy but leveraging their transformative potential. What are some of your favorite use cases for AI in streamlining everyday workflows? #AI #Productivity #Leadership #TechInnovation

93

14 Kommentare

Very excited that we shared our blockchain solution, LETUS, at SIGMOD-2024. LETUS introduce a re-designed architecture with delta-encoding and version-based indexing schema, also with block tiering and state pruning proposed. LETUS is also universal for different blockchains, adaptable for Ethereum, BNB Smart Chain, etc. Evaluations shows that, LETUS improves the throughput by up to 15.8 times for the ERC-20 smart contract transfer workload. And reduces the space usage by up to 80.3%, reduces disk bandwidth usage by up to 95%. During the 2023 Asian Games in Hangzhou, AntChain (augmented by LETUS) successfully supported the digital flame ignition event, where over 100 million torchbearers from more than 100 countries lit the digital torches in real-time online. Final version paper link: https://lnkd.in/g5h736Zb PDF link: https://lnkd.in/gyNqy4r4

1

🚀 We've open-sourced TimesFM – a decoder-only foundation model for time-series forecasting 📈 Key Highlights: ✅ Large-Scale Training Data: Trained on a corpus of over 100 billion real-world time points, including Google Trends and Wikipedia pageviews. ✅ Zero-Shot Forecasting: Offers high-quality forecasts even on unseen datasets, without requiring fine-tuning or additional training. ✅ Outstanding Benchmark Performance: Achieved competitive results across a comprehensive suite of benchmarks. 🤗 Explore the model on Hugging Face and GitHub. I'm encouraged by the early feedback and look forward to seeing how #TimesFM will be used in real-world applications for forecasting and anomaly detection. 🙏 A special thanks to the talented research team behind this achievement: Abhimanyu Das Rajat Sen Yichen Zhou Weihao Kong

44

1 Kommentar

My ICSE paper with Christian Bird and Tom Zimmermann is now live on the ACM website. Check it out to learn how you can best leverage Objectives and Key Results (OKRs) in your org!

44

AI memory is fundamental for conversational AI and AI agents. It is critical for building personalized experiences and improving reasoning. Check out Letta’s course to understand how to use AI memory to enhance the customization and quality of your AI app.

15

Combining Fuzzy Search and Prefix search in the same query AND keeping the latency down is harder than you might think! It will be exciting to see applications with high query rates using this. The latest advances in embedding in multiple resolutions and multi-phase ranking will enable new use cases in many organizations:

26

Today, on when AI and hard-coded rulesets collide. Engineer: "Our AI sometimes generates invalid domain names. That seems like a problem." PM: "Definitely a hallucination. Let's hardcode so that those don't get shown to the user." Me: "Hey AI, can you give me some good unclaimed URL ideas for my business?" The AI: Here are some ideas to get you started: - [invalid URL removed] - [invalid URL removed] - [invalid URL removed] - [invalid URL removed] - [invalid URL removed] There's always a corner case you missed. Always.

6

It is the last day at Berlin Buzzwords and the show is definitely living up to its hype. Following yesterday’s post, I am happy to share a link to my session that explains how the limits of search can be shattered using domain specific computing. Interested in learning about the architecture of domain specific computing? Or how it achieves hyper candidate generation, better ranking efficiency and fast menory access? Tune in to this talk to learn more. https://lnkd.in/dD4GKZdK #vectorsearch #keywordsratch #search #elasticsearch #bbuzz

9

It is the last day at Berlin Buzzwords and the show is definitely living up to its hype. Following yesterday’s post, I am happy to share a link to my session that explains how the limits of search can be shattered using domain specific computing. Interested in learning about the architecture of domain specific computing? Or how it achieves hyper candidate generation, better ranking efficiency and fast menory access? Tune in to this talk to learn more. https://lnkd.in/dD4GKZdK

48

Nature is healing: https://lnkd.in/dztKzJw4

3

Finally - my review of Stephan 🚀 Schmidt's book. It is a short book with many good thoughts on the whole spectrum of a CTO's job, all general enough to immediately think about how those apply to one's situation - basically like a meditation.

7

1 Kommentar