David Peterson

Very nice description of why and how you can have AI locally, specifically in Chrome on a PC. I really like the "Nano" in Google Gemini Nano to indicate this is a small model LLM. I don't care how many tokens it can handle, just that it is big enough for the task at hand. #AI #GenAI #LLM https://lnkd.in/e--FgVFs

11

2 Comments

This episode of DOS Won’t Hunt saw Matt Bishop, principal architect at Bitwarden; Artem Kroupenev, vice president of strategy at Augury; Matias Madou, Secure Code Warrior’s CTO and co-founder; and Joel Carusone, senior vice president of data and AI at NinjaOne, come together. They discussed the use of AI in software development, its benefits and risks, and how developers want AI to be implemented in the cycle.

11

"File Uploading is Easy, Until it Isn’t." Discover why moving massive data is easier with the MASV file transfer platform. Learn more from Marc Paquette here: https://lnkd.in/g8sPe6C4 #FileTransfer #DataManagement

4

Execution postponed... again, this time until 2025 :) Why? Simply - sandbox doesn't work. Happy at least google didn't push implementation despite dramatic results for publishers. As a reminder, interesting podcast (link in comment). Long story short: - an extra 1,500 milliseconds to load (yes, in word, one thousand five hundred!) - viewability - decreased to 39% - as a result of the above, a drastic CPM drop #cookieless #sandbox

40

2 Comments

Great to see a new model from Anthropic introduced today. Congrats to the team on shipping another great update! I've been working on a project to benchmark the marketing knowledge of large language models, which I've just used to test it. 🥁🥁🥁 DRUM ROLL.... ...I can confirm that Claude 3.5 Sonnet OUT PERFORMS all other models in marketing knowledge! 🏆🏆🏆 Here are the scores for the top 5 models: 🧠 83.8% Anthropic Claude 3.5 Sonnet 🧠 83.0% Anthropic Claude 3 Opus 🧠 81.9% OpenAI GPT-4o 🧠 81.0% OpenAI GPT-4 Turbo 🧠 80.7% Google Gemini-1.5 Pro When it comes to the marketing industry, Anthropic's LLMs are clearly leading the way! My project currently has 1,600+ multiple-choice marketing questions, and is still a work in progress, but if you want to learn more, check it out on GitHub here: https://lnkd.in/dHH4J8RS #GenAI #Marketing #Research

9

3 Comments

Ready to revolutionize your approach to platform engineering? Join us at SKILup Day: SREs and the Rise of Platform Engineering on April 30, 2024, to explore the evolution of SREs in platform engineering. https://bit.ly/4az7z8R Discover how SREs are reshaping DevOps workflows at scale, playing a crucial role in driving efficiency and resilience. At SKILup Day: SREs and the Rise of Platform Engineering, we’ll discuss: ➡️ Resilience Redefined: SREs leverage platform engineering to fortify systems against disruptions, ensuring uninterrupted service delivery even amidst unpredictable challenges. ➡️ Scalability at the Core: Through Platform Engineering, SREs architect solutions that effortlessly scale with evolving demands, enabling businesses to adapt and thrive in rapidly changing environments. ➡️ Efficiency Amplified: By embracing automation and data-driven decision-making, SRE-led platform engineering maximizes operational efficiency, empowering teams to focus on innovation and strategic initiatives. Don’t miss out! Register now to gain valuable insights and actionable strategies on how to empower your journey to SRE and platform engineering. https://bit.ly/4az7z8R #SKILupDay #SRE #PlatformEngineering #DevOps

4

Interesting piece on how vector databases (an esoteric technology) have finally found use in the generative AI era ---- Vector databases have been around for decades, but are now emerging as something of an industry standard for AI businesses to use alongside a technique called retrieval-augmented generation, or RAG. When combined, businesses can link their private data with large-language models like OpenAI’s GPT-4, allowing the AI to perform data analysis, summarization and other tasks on their data. Without them, AI models are limited to what they have learned from their initial training on public data online, up to a certain point in time, and are more prone to factual errors called “hallucinations.” Vector databases are different from traditional databases with columns and rows because they are designed to store a massive amount of data as “vectors,” or numerical representations of the raw data New York-based startup Pinecone was an early entrant in the vector database AI space. Valued at $750 million after raising $100 million from Andreessen Horowitz and other investors last April, Pinecone has been building its vector database since 2019—a fortunate few years before OpenAI’s ChatGPT kicked off the AI gold rush in late 2022. ----- This one of those things you may learn about in computer science, along with red-black trees, and never ever use! #ai #chatgpt #llms #databases

8

3 Comments

Expertise matters more now than it perhaps ever did. This is my key takeaway from recent deep-dives into LLMs, AI models, RAG systems, embeddings, modern knowledge systems, etc. Here's the irony: these brute-force trained LLMs don't truly understand what they're saying or why. If they were human, they'd have severe impostor syndrome (because they'd truly be imposters). While they can generate relevant and often correct content, they lack fundamental understanding. They can provide convincing explanations for their outputs, but they neither possess core truths nor mechanisms to consistently adhere to them. The utility of LLMs depends entirely on your needs: ✍️ Need a smart draft of your writing? Fire away. 👨‍💻 Want prototype code for a proof of concept? Sure – but you wouldn't use that code to control a car or rocket. 🎨 Looking for frontend design inspiration for prototypes and soliciting feedback? LLMs can help (as seen on my prototypes at https://frigg.cloud, https://lnkd.in/ebmu3Tcm, https://lnkd.in/esj5wtRP, and https://lnkd.in/eb3bFGm3 ) – though they won't produce best-in-class products. Consider my domain of expertise: building integrations and apps on marketplaces. Let's use Slack as an example. I can ask an LLM to suggest integration designs and provide API documentation as context. However, expertise is crucial for: - Understanding Slack's strategic direction toward "Agent" models - Knowing user expectations (like matching slash commands with right-click actions) - Identifying truly valuable features versus nice-to-haves that can wait for user demand - (... and more) Expertise remains critical, even as we approach systems with genuine "understanding." While AI will develop specialized expertise – from the mundane (photo categorization) to the profound (revolutionizing government inefficiencies) – human expertise becomes more vital, not less. In fact, as LLM-based AI proliferates content, applications, and ideas at unprecedented rates, we need experts more than ever to identify subtle but crucial flaws in AI systems' outputs. This perspective assumes, like me, that you're aiming for excellence. If you're not, that's fine – but understanding the limitations and proper applications of AI tools is key to using them effectively.

1

"What I love about Opstream, was it created an orchestration workflow for all of those different journeys to take place and brings everybody involved in the process - the vendors, the requesters, the approvers, to the same source of truth. So we're all on the same page with what's going on." - @Sean Warren Some platforms talk about transformation. Opstream delivers it. With self-service procurement workflows, Pushpay eliminated bottlenecks and saw immediate value. You don't have to take our word for it. PushPay’s results speak for themselves. Check it out here: https://lnkd.in/eyyJyz_a

9

Scientists have successfully mapped the mind of a large language model by extracting millions of features from the middle layer of Claude 3.0 Sonnet. This provides a preliminary conceptual map of its internal states during the midpoint of its computation, marking the first detailed glimpse into the workings of a contemporary, production-level large language model. https://lnkd.in/eigaV3mr #ai #innovation #genai

4

My new standard attack is evolving to: "Why are you anthropomorphizing LLMs? Would you anthropomorphize 7zip or JPEG?" I'm reposting this article to help reframe the discussion. LLMs are, in fact, super lossy compressors. That's it. That's all they are. They don't care about your feelings because super lossy compressors don't have feelings and they don't have care. Stochastic parrot, or my new improved knockoff, stochastic pigeon, do not get to the heart of the problem. A compression / decompression algorithm is not a sentient being. End of story. Argument over. #WrittenByMe https://lnkd.in/gGaqyRMg

4

11 Comments

When Uri Levine was building Waze he didn’t describe it as an AI-powered crowd-sourced navigation system. Instead, he simply said, 'I’m going to help you avoid traffic jams.' Why? Because solving problems your customers truly care about is the key to success. What does it take to build not one, but TWO companies valued at over $1 billion? Discover the lessons from his journey on this week’s Corporate Competitor Podcast! https://lnkd.in/eJXC2KGw

9

Epoch AI has released a new AI benchmarking service that's notable for a few reasons. First, they do great independent research. Second, the tool they've provided is really user-friendly and useful. Third, they're not using saturated benchmarks like MMLU, and from what they are using—GPQA Diamond and Math Level 5—you can see that models continue to rapidly improve and we haven't hit a plateau. That conclusion was based on using saturated benchmarks. Data is here: https://lnkd.in/gTmg6HkV

1

The future of AI may be as close as R2-D2 and C-3PO with the rise of small language models. #SLMFuture #embodiedintelligence #pocketai #aiagents #cambrianexplosion #humanaiconnection

2

Anthropic continues to deliver new models that customers enjoy using in a variety of ways. Today, Anthropic released an upgraded Claude 3.5 Sonnet model with awesome new capabilities, and it’s available now in Amazon Bedrock. The updated Claude 3.5 Sonnet shows wide-ranging improvements on industry benchmarks, with strong gains in agentic coding and tool use tasks. Anthropic also introduced a new capability in public beta: computer use. In Amazon Bedrock, developers can direct Claude to use computers the way people do—by looking at a screen, moving a cursor, clicking buttons, and typing text. Claude 3.5 Sonnet is the first frontier AI model to reliably use computers in this way—albeit experimentally—and Anthropic expects the capability to improve over time. Such a cool feature that I know our customers will find useful. https://lnkd.in/gwBQf8zH

2,508

46 Comments

Improving output quality from LLMs is a major task. I love this concept, the GenRM method is a bit like pair programming or co-authoring. It brings in a level of checks and balances similar to what you'd get when two people collaborate. In pair programming, both developers watch for errors, and in co-authoring, writers critique each other. GenRM mimics this by generating different outputs and using reasoning to verify them, catching mistakes just like a partner would. It’s all about having that extra layer to boost accuracy, reducing the chances of errors slipping through. #LLMQuality #AICollaboration #GenRM #AI #Artificialintelligence

4

3 Comments