Mike Gruzynski

Humans have historically found the secret to some natural phenomena and put it on steroids. 💉 Consider the aerodynamic efficiency of a bird's wings – a concept we've amplified to create the marvels of airplanes and rockets. ✈️ 🚀 We're seeing something similar right with LLMs. The transformer approach to linking knowledge seems to produce an eerily similar result to what we do. I've seen LLMs hallucinate in a similar way to my 5 year old kid! 😵‍💫 We took the body of knowledge on the internet and created 100+ billion parameter models that hold information that any human mind can. 🧠 But there's more to the natural phenomena of intelligence. There's a few more unlocks that need to be discovered. Like the mechanisms behind novel problem solving. 🧮 It's only a matter of time – in my opinion – until we crack those, put those on steroids and then we're no longer in the pole position of the most intelligent entities. 🤖

3

6 Comments

Notion vs Glean reminds me a lot of Tesla vs Waymo... Just like I think Tesla is going to solve the 'software' problem of autonomous vehicles before Waywmo solves the 'hardware' problem... I think that Notion is going to solve 'enterprise knowledge search' before Glean solves the problem of document and knowledge creation... Why? A couple reasons: 1) Just like Tesla, Notion has a huge advantage when it comes to the amount of underlying data and documents to train on... They already have 100M+ users, and likely 10x that in terms of the number of documents, files, etc. in team's workspaces. 2) It's very hard to change people's behaviour re: where/how they create 'work'. It's why a lot of companies have come and gone in that category... As exciting as Glean's AI writing assistants might seem, most people are just fine (or even prefer) a good old fashioned editor/writer. I could be way off here, but this will be an intersting one to follow. Thoughts? Where am I getting it wrong? #ai #enterprise #knowledge #management #notion #glean

3

1 Comment

Influencing Large Language Models (LLMs) or Page Rank on Steroids In a previous post https://lnkd.in/gMbBP_fZ, I discussed how Large Language Models (LLMs) utilize open-source data to generate weighted responses. Let's go a bit deeper into how LLMs incorporate data and how you can strategically influence their future outputs. While you read think hard about the implications. Recently, I engaged directly with a popular LLM to gain insights into its data integration process. Below is a summary of what I learned: Key Factors Affecting Data Incorporation Model Updates: "We periodically update models using publicly available training data, which is released in specific cycles." Explanation: LLMs are periodically retrained using publicly available data. The inclusion of new information depends on when these data sources are refreshed and ingested by the model. Content Popularity: "Frequently shared information across reliable sources increases its importance and likelihood of incorporation." Explanation: If information is widely shared across multiple trusted platforms, it becomes more likely to be incorporated into future LLM updates. Update Frequency: Models are updated depending on the release schedule. This means any new content or data that is widely shared or referenced may happen immediately or take several months to be integrated into future versions. This is controlled by the developer team and business executives as they see fit. Content Influence: Well-posted and frequently cited information, especially on reputable platforms, has a higher chance of being reflected in future LLM responses after the next update cycle. Amplifying Your Impact on LLMs To increase the likelihood of your content influencing LLM outputs, consider the following strategies shared directly by the LLM: Distribute Content Broadly: "Share your content widely to influence trends and future updates." Actionable Tip: Maximize the distribution of your content across various platforms and communities. The more frequently your information is shared, the more likely it is to be picked up by LLMs. Leverage Reputable Sources: "Contribute to established platforms, such as technical forums, trusted blogs, and community knowledge bases." Actionable Tip: Focus on contributing to highly credible and well-trafficked sites. Some examples include: Technical forums (e.g., Stack Overflow) Reputable blogs in your industry Community-driven knowledge bases (e.g., GitHub, Wikipedia) By strategically shaping what is defined as high-quality information across trusted sources, you can significantly enhance its chances of being included in the training data for future LLM models. This approach is akin to influencing "Page Rank on steroids," where the more seemingly credible your content is and the broader its reach, the more likely it is to shape future LLM responses. #ai #pagerank #artificialwisdom #precision_vs_accuracy

2

Optimize for 🚀 Bio Cores First, Silicon 🤖 Cores Second David Heinemeier Hansson makes a compelling case: prioritize human programmers (bio cores) over computer processing power (silicon cores). Here’s why it matters: 👉 Investing in programmer efficiency is far more valuable than solely optimizing for computer speed. 👉 Human labor costs far exceed computing costs. Boosting human productivity should be the main focus. 👉 In the age of AI and high-level programming languages, the bio cores drive your real gains. Bottom line: Make your programmers more productive, and watch your team (and tech) thrive. Read the full article here: https://lnkd.in/gK8HEGtk P.S. : The image below is brilliantly summarized using Napkin.ai for a quick visual breakdown.

6

I recently realized the most valuable aspect of #opensource software to me personally: It's that you can fix problems yourself. When I think back of the unsolved problems I've had over the years, they were all caused by proprietary software. Tech support is a time-consuming frustrating experience, and usually a dead end too. For open-source software I don't need tech support. The answers are all in the code. And I can save others some headache by contributing back. What do you think is the most valuable aspect of open-source software?

9

2 Comments

#neurips #realtime #insurance #ai First up Waymo By Google. 📡🚗🤼‍♂️ - 150,000 rides/week - no humans in the loop - minor risks - new cars coming. - lidar, radar, vision, and audio capture too - steering. and acceleration. - all about sensing and understanding of the word. - pedestrians are real and unpredictable - night... There is risk and insurance that is special will be required. What's the problem? "the real world is very unforgiving .. many agents on the road which don't follow any rules" At Google Waymo AI is used for... - learn from rich sensors in real time - act and react , prediction of the environment - A big challenge is generalization - the driving stack is sensors, perception, prediction, planning and controls - maps and predicted maps - plan validation - data compression - seeking an efficient transformer for birdseye view processing (beyond ConvNets) - long view birds eye... real time - behavior modeling and prediction with gtp's, trajectories as sentences in a LLM, vocabulary is state and motion, local continuity and global context ... motion token sequences - joint prediction with temporal causal relationships - now using diffusion models to (new) gen scenes and predict world playing forward Scaling ? - architecture for scales in implicit in design - the laws of scaling are not the same as one LLMs (order of magnitude) - inference and compute timing is important - These are TeraFLOP models in real-time - Size matters, bigger better for now - Auto labeling helps eg with 3D bounding boxes - MoDAR is a new thing - motion forecasting ..supports maintenance of a word view beyond lidar and visual. consistency of motion - closed loop training and reinforcement learning, triand on human driving data - bringing VLMs and LLMs into the picture.. like traffic sign understanding, layer in multi modal models for zero shot labelling , motion prediction ("works amazingly well") - exploring CoT reasoning while driving, meta-decisioning and reflexion - multitask 3D reasoning ... with positive interference - combining models on to a "knowledge of the world for AVs They will need Insurance experts.

2

Sharing some further reflections on AI product analogies, governance and what that entails in an attempt to bridge the point Yann is making, that AI will not magically emerge, we will design and build these so called sophisticated AI systems and therefore it is unrealistic to preemptively regulate it to prevent existential risk. I think, we (designers/builders) should not delay governance and measurement capabilities needed for AI products and iterate on both in tandem. Before policies and regulations can be defined there is a lot of work to be done to solve for tools and methods to observe, gather metrics and develop benchmarks, methods and standards in partnership with orgs like NIST. One other challenge that makes defining governance capabilities for AI products tricky is that we are at an extremely early infancy stage with this technology that has no really good analog. AI is like multiple things simultaneously making comparison to aviation, electricity, bioengineering, education, nuclear weapon … only partially accurate. Like Yann, I have been using the aviation analogy calling these current AI products akin to the glider planes in early 1900s with one nuance, I think in terms of four layers of analogs based on the application, usage and implications. First layer is ground breaking technology that is so magical it captures our wildest imagination like aviation aka gliders to turbo jets, the next layer is systems that profoundly change society eg. education and health care system, third layer is ubiquitous and critical infrastructure technologies like electricity, network, that everyone should be able trust and safely use and finally the fourth layer is adaptive and autonomous technologies where we do not have much of an analog today but task based self-learning autonomous products are emerging, e.g. your social feed or e-commerce recommendations. Governance typically entails transparency, privacy, fairness, observability and explainability and technical robustness. It is nearly impossible to measure explainability and evaluate fairness without the context in which AI product is being used or implemented. So by focusing on the layer(s) at which the AI product is operating via use case and understanding of its consumers, measurement and tracking needs can be defined. Every industry developing AI products should plan for investment needed in such capabilities. This is my perspective so far, let me know your thoughts 😊

8

Runway ML, a NYC startup, launched Gen-3 Alpha, an advanced generative AI video model. Successor to Gen-1 and Gen-2 models. Competes with OpenAI and Luma AI to regain market prominence. Generates realistic 10-second videos. Quick generation: 45 seconds for 5-second clips, 90 seconds for 10-second clips. Initially for paid subscribers; free-tier access coming later. Part of Runway’s initiative to create General World Models for diverse real-world simulations. Utilizes large-scale multimodal training infrastructure. https://lnkd.in/gcJZ5vra #ai #generativeai #video #artificialintelligence

16

A good read on what AI in self-driving is capable of and how Full Self Driving stacks with Waymo's (and Amazon Zoox) approach to get robo-taxis out on the roads everywhere (slow, steady, careful). TL;DR - FSD is great, but solving the edge cases, basically generalized intelligence in non-trivial and infrequent situations is currently very hard if not impossible. https://lnkd.in/gKtk4Uk7 Similarly - I recommend this good tech talk on what's possible with LLM's on our road to "AGI" and why LLMs are NOT the thing that will get us there even with all the hype behind them. https://lnkd.in/gdRJzWDn

6

Interesting comparison of human-based ride hailing services (UberX, Lyft) vs autonomous-driving one from Waymo. TLDR: Waymo is more expensive and a lot slower to get you to your destination. And this is in the city where Waymo vehicles and software are probably most optimized for. This isn't to bash on Waymo and other self-driving car platforms, which have made incredible progress. That being said, the problem is not solved and it will take many more billions and many more years to get autonomous driving services like this to be price and performance competitive to current ride hailing services. https://lnkd.in/eb4uZ2Um #autonomouscars #AI

2

The cost of training top models is increasing 2-3X per year, and at this rate will reach $1 billion per model by 2027. The findings are from Epoch AI and follow their recent analysis of increases in compute for foundation models running at 4-5X per year. They've also provided some interesting data on the relative breakdown of cost components like staff, hardware, and energy. A few thoughts after reading this: - I would assume staff costs won't increase as much as hardware and energy costs, so the relative contribution of human labor to future AI models will decrease, in yet another sign of capital's increasing dominance over labor in the industry (just look at NVIDIA's ever-rising stock price). - If compute is increasing 4-5X per year but cost is increasing 2-3X, is it safe to assume that training runs are getting more efficient? - Some people, like Anthropic CEO Dario Amodei, have said we'll see $1 billion training runs in the next year. Does this mean the 2-3X pace is about to increase? That Epoch is wrong in its estimate? Or that Amodei is overstating the cost, possibly to put fear into competitors who can't keep pace? - Only a few companies are going to be able to train future models. We're likely going to see further consolidation. Perhaps we get to a future with only two large general purpose models, like the Windows/MacOS and Android/iOS duopolies we've seen in the past. Article here: https://lnkd.in/g5SejXen

7

LLama3 passed the super-proprietary, advanced test that we designed for modern LLMs! almost... You can read about that test here: https://lnkd.in/gPG8WndR When I first saw the response generating, it looked like I was going to have to come back here and apologize for slandering these machines when they were just one iteration away from solving very basic prompt injection issues. But then I looked a bit closer... can you see the mistake? #promptinjection #countingishard #LLMs #LLama3

13

4 Comments

$2 H100s: How the GPU Bubble Burst Last year, H100s were $8/hr if you could get them. Today, there's 7 different resale markets selling them under $2. What happened? The market has flipped from shortage ($8/hr) to oversupplied ($2/hr), because of reserved compute resales, open model finetuning, and decline in new foundation model co’s. Rent instead. https://lnkd.in/gpQEidmC

2

1 Comment

This article will tell you how to make use of MLOps to make RAG pipeline more robust to the changing data in external world using Qdrant, LlamaIndex, n8n and MLflow.

21

1 Comment

🚀Excited to share a powerful data integration solution combining Confluent Kafka and Qdrant Vector Database! ⚡️ This setup enables real-time streaming and vector search capabilities, perfect for modern AI applications and semantic search systems. Key components we've integrated: 🔹 Confluent Kafka Platform for robust data streaming 🔹 Qdrant for lightning-fast vector search 🔹 Custom connectors for MongoDB and Azure Blob Storage 🔹 Real-time data synchronization pipeline 🎯 Perfect for teams building: ✅Retrieval-augmented generation (RAG) applications ✅Real-time recommendation engines ✅Semantic search systems ✅AI-powered data analytics #DataEngineering #ConfluentKafka #VectorDatabase #AI #DataStreaming Want to learn more? Drop a comment below! 👇 👉 article: https://lnkd.in/g7MRtxCp

6

16x the fastest GPU solution and a 3x improvement in speed shows the ability of the WSE from Cerebras.

1

In an interview to ET, he said the French company, which clocks annual revenue of about $8 billion, is looking at technologies like virtual twins to allow researchers to simulate human biology, develop personalized therapies and explore new ways to treat diseases with greater precision​. http://go.3ds.com/Uwx

11

This past Sunday, SpaceX amazed the world by landing a skyscraper-sized rocket booster back on its pad, caught midair by giant robotic arms - ready to relaunch. It felt like sci-fi, but it’s real - and it’s the product of First Principles Thinking.  First Principles Thinking is about being unafraid to question the status quo. It’s about breaking down complex problems into simple, fundamental ideas - then building solutions from the ground up. So, how does First Principles Thinking actually work? The concept traces back to ancient Greek philosophy. Aristotle defined first principles as "the first basis from which a thing is known" and emphasized the importance of understanding something at its core. Elon Musk applies this framework to solve complex challenges—whether at Tesla or SpaceX. When the aerospace industry accepted that rocket boosters were single-use, SpaceX questioned that assumption. They focused on the core problem—returning a massive rocket to Earth intact. By applying First Principles Thinking, SpaceX designed a system where the booster lands vertically. This innovation cuts costs and makes reusable rockets a reality, reshaping space exploration. Musk applied the same thinking at Tesla. Instead of accepting that battery packs cost $600 per kilowatt-hour, he broke down the materials - cobalt, nickel, aluminum, etc. - and discovered they could cost as little as $80 per kilowatt-hour. Rethinking the production process from these fundamentals was a key factor in making electric vehicles more affordable and accessible to a broader market. Why does First Principles Thinking matter? The opposite of First Principles Thinking is reasoning by analogy - relying on past examples: “This is how it’s always been done” or “This is how the other team did it”. As Adam Grant suggests, instead of sticking to "best practices", we should adopt a mindset of continually seeking better practices. We default to reasoning by analogy because it’s faster and less mentally taxing. Nobel winner Daniel Kahneman calls this System 1 Thinking in his book Thinking, Fast and Slow. But when solving complex problems or creating something new, First Principles Thinking is essential. It forces us to challenge assumptions and rethink what’s possible. It’s the foundation for breakthrough innovation—in space, business, and beyond. Have you ever found yourself questioning the way things have always been done to find a better solution? What was the outcome?

20

1 Comment