Bal Balaji

Thrilled to hear #OpenAI: We're rolling out new features and improvements that developers have been asking for: 1. Our new model GPT-4 Turbo supports 128K context and has fresher knowledge than GPT-4. Its input and output tokens are respectively 3× and 2× less expensive than GPT-4. It’s available now to all developers in preview. 2. Assistants API and new tools

1

🧠 The myth: More data makes an #AI model smarter and therefore more accurate 💪 The truth: It’s not just about data—it’s about #parameters With the around OpenAI’s #GPT4o it might seem that feeding more data into these models makes them smarter, the real game-changer is the number of parameters they have like Meta #Llama 3.1 with 450b parameters 💡 More parameters mean these models can process and understand data in more sophisticated ways, leading to increased accuracy and the appearance of being “smarter.” More parameters = higher trustworthiness, even with more focused data like the proprietary #data in your organization In the world of conversational user interfaces becoming the norm, more parameters will be crucial for a seamless experience. While yet too slow and too expensive for real fast interactions between humans and machines, with #quantumcomputing we will move in a new area of responsible #UI #design #AI #MachineLearning #LLM #conversationaldesign #hmi #Innovation

6

1 Kommentar

How do we know that we are in an AI bubble? At COMPUTEX every vendor, be it fans or pc cases or cables, is advertising with AI because maybe their products end up in a system that is running AI? I'm also very skeptical about edge AI for LLMs, do these current "AI" chips really speed up a LLM or are they just good for image processing (Microsoft Recall)? As I understand, the bottleneck for fast LLM inference is the memory bandwidth, and that's why we are using GPUs because their VRAM interface is still superior. I guess this might change, but I don't think the current generation of AI hardware is there yet. https://lnkd.in/dYcMqHTH

1

1 Kommentar

💡 The 2024 🇨🇭 Open Source Study confirms that 97% of Swiss corporations are using #opensource software …. and 40% are already using #opensource #ai tools and models. Here my Top 5 takeaways from the study ⬇ 1. There are already 700’000+ open source #ai models on Hugging Face 2. #TensorFlow and #PyTorch are the top used #ai open source tools. 3. #Llama and #Mistral are by far the most used open source #ai models. 4. #opensource is used where nobody can really see it! None of the #cloud environments we’re using in private and professional live would work without Open Source! 5. #linux, #docker and #kubernetes are key technologies to run today’s Cloud environments. A lot of interesting articles about the economics of #opensource and real world examples are complementing the study. Thanks to CH Open, swissICT and Berner Fachhochschule BFH for running an providing this interesting study! Here you can find the details (in German): https://oss-studie.ch/ Matthias Stürmer Alain Gut Christian Keller Richard Zobrist Katharina Göppinger Lise Crettaz Philipp Spaeti

33

4 Kommentare

https://lnkd.in/ew6aHRF7 One of my hobbies is collecting Anthuriums. This GPT predicts parent plants that match your desired Anthurium plant. Hoping this leads to some innovative hybrids! #anthurium #hybrid #gpt

1

2 Kommentare

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

#ObsessivelyEnthusiasticInvestor #UnreasonablyEnthusiasticInvestor #IrrationallyEnthusiasticInvestor at REFASHIOND Ventures: The Industrial Transformation Fund. We look for the people we believe are doing the best work in Industrial Transformation & Supply Chain Technology; Attempting the impossible. Then we go all-in on them. With obsessive and unreasonable enthusiasm. Join us. #IndustrialTransformation #SupplyChain #Innovation #Technology #Startups #VentureCapital #DemandTheImpossible #UnreasonableEnthusiasm #ObsessiveEnthusiasm #IrrationalEnthusiasm 

6

IMVC 2024 just released Dr. Amir Alush great talk about going from messy visual data to high quality datasets: https://lnkd.in/dGBDUUHQ . Try it yourself on your own data at https://lnkd.in/ds4vwrpP!

24

𝗪𝗲 𝗻𝗲𝗲𝗱 𝗮𝗻 𝗮𝘃𝗮𝗹𝗮𝗻𝗰𝗵𝗲! An avalanche of responses and data, that is, that gives a better picture of the 𝗰𝘂𝗿𝗿𝗲𝗻𝘁 𝘀𝘁𝗮𝘁𝗲 𝗼𝗳 𝗚𝗲𝗻𝗔𝗜 𝘂𝘀𝗮𝗴𝗲 𝗮𝗰𝗿𝗼𝘀𝘀 𝘁𝗵𝗲 𝗲𝗻𝘁𝗶𝗿𝗲 𝗦𝗗𝗟𝗖. This survey specifically focuses on human-machine teaming—how we collaborate with AI tools in our workflows today. Why participate? 𝗞𝗲𝗲𝗽 𝗛𝘂𝗺𝗮𝗻𝘀 𝗙𝗶𝗿𝘀𝘁: Share your experiences on how GAI is reshaping workflows and the evolving human role in software creation and maintenance. 𝗦𝗵𝗮𝗽𝗲 𝘁𝗵𝗲 𝗳𝘂𝘁𝘂𝗿𝗲: Your experiences can influence leading practices and future innovations. 𝗕𝗲 𝗽𝗮𝗿𝘁 𝗼𝗳 𝗴𝗿𝗼𝘂𝗻𝗱𝗯𝗿𝗲𝗮𝗸𝗶𝗻𝗴 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵:Your insights will contribute to a comprehensive study on GenAI's impact on the dynamics of human-machine collaboration. 𝗤𝘂𝗶𝗰𝗸 𝗮𝗻𝗱 𝗲𝗮𝘀𝘆: The anonymous, not-for-profit survey takes just 10-15 minutes of your time. 𝗔𝗺𝗽𝗹𝗶𝗳𝘆 𝘆𝗼𝘂𝗿 𝘃𝗼𝗶𝗰𝗲:  If you're part of a software value stream, this is your chance to share unvarnished thoughts on GenAI's real-world applications. Help us spread the word! After completing the survey, please share it with your network. The more diverse our respondents, the more valuable our insights will be for the entire tech community. https://lnkd.in/e_PKhDVH Remember, your participation is completely anonymous, and the research findings will be made public for the benefit of the entire industry. #ArchAITecture #GenAI #SDLC #SoftwareEngineering #SoftwareDevelopment #Research MITRE, Ben Amaba, Phillip Laplante, David Linthicum, Patrick Debois, Nathen Harvey, Alan Shimel, Mike Smoyer, Wesley Reisz, Hasan Yasar Special thanks to Krzysztof Kowalik for the photo via Unsplash

13

2 Kommentare

Cool article from amazing engineers Rithvik Panchapakesan and Danmei Xu about search, embeddings and quantization! These are the folks powering Snowflake's semantic search! #snowflake #vectorsearch #ai

230

3 Kommentare

Gurke & Blind, the famous #consulting company from Stuttgart has just released this #chart, based on #empirical data, that clearly shows we're aiming towards a new #Hype Pit. To my fellow XR do-ers : time to brace and continue the good work, no matter what 😎 . The new #Hype Peak is probably underway. #virtualreality #socialnetworks #augmentedreality

9

6 Kommentare

Open source is already under challenge. Conversations about "banning open source" in AI raise genuine concerns. This piece I wrote for The New Stack explains why. https://lnkd.in/eu5W86Dc #opensource #AI #opensourceai

67

13 Kommentare

Hugging Face recently released fineweb dataset for LLM pretraining. The 🍷 FineWeb dataset consists of more than 15T tokens (44TB) of cleaned and deduplicated english web data from CommonCrawl. They published a report detailing the steps and decisions taken in creating such a dataset. Key takeaways: - Starting point: All 96 snapshots of CommonCrawl web data published by non-profit CommonCrawl https://commoncrawl.org/ - Library used for processing data at this scale: Datatrove https://lnkd.in/gcSY2mwJ - Quality of dataset measured by training small models (1.82B, Llama architecture with 2048 seq length, GPT2 tokenizer) on random samples of dataset (~28B tokens) and evaluating them on a set of "early-signal" benchmark tasks. (from lighteval). - Longer training of these models on 350B tokens was also done - These models evaluated on CommonSense QA, HellaSwag, OpenBook QA, PIQA, SIQA, WinoGrande, ARC, MMLU - Steps followed: 1. URL filtering 2. Text Extraction 3. Language Filtering 4. Gopher Filtering 5. MinHash deduplication (independent dedup per dump) 6. C4 Filters 7. Custom Filters (new heuristic filters) 8. PII Removal - FineWeb-Edu https://lnkd.in/g33tUNcz is a further processed and filtered dataset created from Fineweb by using a classifier to identify educational content. - An educational quality classifier was developed using annotations generated by LLama-3-70B-Instruct to generate educational quality score from 0 - 5. - Model used was Snowflake-arctic-embed with a classification head and single regression output on top of it, trained on 450K llama-3 annotations for 20 epochs with a learning rate of 3e-4 with frozen embedding and encoder layers. - It was then used to select documents with a score greater than the threshold of 3 Checkout the full report here: https://lnkd.in/g3d-pKK8

5

1 Kommentar

🆕 Benchmarks 201 with Clémentine Fourrier of the Hugging Face Open LLM Leaderboard! Why Chat Arenas are **OVERRATED**: - they bias sycophancy/assertiveness over factuality - annotators are not representative of humanity - models are argmaxxing public arena data - not reproducible - can't give fine grained test of model capabilities Why you should STOP using LLMs as Judges: - Position bias (prefer first answers) - Self bias (prefer their own answers) - Length bias (prefer long answers) - Discrete bias (can't do continuous ranges) - Reproducibility (GPT4 closed source - prefer Seungone Kim's Prometheus or Beijing Academy of Artificial Intelligence(BAAI) JudgeLM) One of our best ones 🔥

29

1 Kommentar

Another quick note on evaluations and benchmarks for agentic workflows. As I mentioned in the article I shared yesterday (link in comments), agentic systems are still not very reliable. However, increasingly robust benchmark proposals are emerging in various verticals: 𝑻𝒐𝒐𝒍 𝒖𝒔𝒂𝒈𝒆: Here, WorkBench stands out, which tests an LLM’s ability to use 26 software tools to operate on five simulated workplace databases: email, calendar, web analytics, projects, and customer relationship management. Tools include deleting emails, looking up calendar events, creating graphs, and looking up tasks in a to-do list. 𝑺𝒐𝒄𝒊𝒂𝒍: SocialDial is a benchmark that tests potential social violations which, for example, may depend on cultural context. Anecdotally, in China, reserving places with many occurrences of the number 4 is considered bad luck. Similarly, there is CharacterEval for role performance. 𝑷𝒍𝒂𝒏𝒏𝒊𝒏𝒈: Google and colleagues propose a planning benchmark called Natural Plan that includes thousands of prompts as benchmarks in categories such as travel planning, virtual calendar schedules, and physical location schedules, where the metric to optimize is to make more people fit in a place at one specific time. These tests have unambiguous responses, providing a good set of tests for an agentic workflow. 🔗 WorkBench: https://lnkd.in/dUyynJTx 🔗 Natural Plan: https://lnkd.in/dJNVwkXZ 🔗 In addition to those mentioned, this is a good repository with references to benchmarks for agentic workflows: https://lnkd.in/d5V3JNeB ---- #AI #AgenticSystems #Automation #Innovation #LLM #Agents #multiagents #AI

4

3 Kommentare

New top LLM? From Anthropic ‘Introducing Claude 3.5 Sonnet—our most intelligent model yet. This is the first release in our 3.5 model family. Claude 3.5 Sonnet now outperforms competitor models on key evaluations, at twice the speed of Claude 3 Opus and one-fifth the cost. Try it for free: https://claude.ai.’ TRY IT FOR FREE! #GenerativeAIforRealEstatePeople

13

1 Kommentar

Fantastic development in OpenAI: In the future, humans will need to supervise AI systems much smarter than them. We study an analogy: small models supervising large models. Read the Superalignment team's first paper showing progress on a new approach, weak-to-strong generalization: https://lnkd.in/dHCYHvat There is lots of low-hanging fruit and many promising directions for future work. We think this is an exciting opportunity for the ML research community. To kickstart more research, we're launching Superalignment Fast Grants in partnership with Eric Schmidt. We will award $10M in grants for technical research on aligning superhuman AI systems, including weak-to-strong generalization, interpretability, scalable oversight, and more. Apply here by Feb 18: https://lnkd.in/dfNYEbwM.

3