Isaac M.

🚨 Excellent AI paper alert: "Can AI Standards Have Politics?" by Alicia Solow-Niederman. A must-read for everyone interested in AI policy, governance & regulation: "The very formation of standards is political because the standards development and diffusion process reflects a particular institutional context and an associated set of relationships among public and private actors. Under the surface of any standard setting effort, there is a set of assumptions about how a standard will diffuse, and this anticipated diffusion pattern depends on the relationships between public and private actors." (page 8) - "First, take the consulting firm Deloitte. The firm offers a variety of “Artificial Intelligence and Analytics Services” and touts its “Trustworthy AI” framework “to guide organizations on how to apply AI responsibly and ethically within their businesses.” This framework emphasizes how, until global AI regulations "eventually address ethics concerns,” the firm is “working to bridge the ethics gap,” underscoring that AI must be “transparent and explainable, fair and impartial, robust and reliable, respectful of privacy, safe and secure, and responsible and accountable.” In other words, Deloitte is setting forth, and internally defining, its own AI standards for the tools and guidance that it sells." (pages 12-13) - "The effort to govern AI through standards risks ignoring the reality that standards have politics. This Essay calls for us to stop doing so. That will not be easy. The basic difficulty is that AI standards are most powerful as governance tools when we embrace the fiction that standards emerge in a vacuum, without reference to an institutional context and without implicating normative choices. AI standard setting efforts implicitly rely on this fiction when they focus on the need for a “scientific” or“technical” consensus—a uniform understanding of the nature of the problem and the best formula to use to solve it—before crafting a standard. The more a standard appears apolitical, objective, and neutral, separate from specific institutional dynamics and their politics, the stronger the case for the standard’s dissemination across contexts." (page 16) ➡ Link to read the paper below. ➡ For more information and recommended papers on AI policy & regulation, subscribe to my newsletter. #AI #AIgovernance #AIregulation #AIstandards #AIAct

110

13 Comments

This week, Google announced a doubling of Gemini Pro 1.5's input context window from 1 million to 2 million tokens, and OpenAI released GPT-4o, which generates tokens 2x faster and 50% cheaper than GPT-4 Turbo and natively accepts and generates multimodal tokens. I view these developments as the latest in an 18-month trend, which is shifting prompting best practices for developers as well. Since ChatGPT's launch in Nov 2022, newer models like GPT-4, Gemini 1.5 Pro, Claude 3 Opus, and Llama 3-70B have given us better capabilities in two important ways: (i) reasoning, which allows LLMs to think through complex concepts and and follow complex instructions; and (ii) longer input context windows. The reasoning capability of advanced models makes them quite good at interpreting complex prompts with detailed instructions. Many people are used to dashing off a quick, 1- to 2-sentence query to an LLM. In contrast, when building applications, I see sophisticated teams frequently write prompts that might be 1 to 2 pages long (my teams call them “mega-prompts”) that specify in detail how we’d like an LLM to perform a task. I still see teams not going far enough in terms of writing detailed instructions. For an example of a moderately lengthy prompt, take a look at Claude 3’s system prompt. It’s detailed and gives clear guidance for Claude's behavior. This is a very different prompting style than we typically use with LLMs’ web user interfaces, where we might dash off a quick query and, if the response is unsatisfactory, clarify what we want through repeated conversational turns with the chatbot. The increasing length of input context windows has added another technique to the developer’s toolkit. GPT-3 kicked off a lot of work on few-shot in-context learning. For example, for text classification, you might give the LLM a handful — say 1 to 5 examples — of texts and their class labels as context. With longer input context windows — GPT-4o accepts 128,000 input tokens, Claude 3 Opus 200,000 tokens, and Gemini 1.5 Pro 1 million tokens (2 million just announced). With many-shot learning, we can give dozens, even hundreds of examples in the prompt, and this works better than few-shot. When building complex workflows, I see developers getting good results with this process:  - Write quick, simple prompts and see how it does. - Based on where the output falls short, flesh out the prompt iteratively. This leads to a longer, more detailed, prompt, perhaps even a mega-prompt. - If that’s still insufficient, consider few-shot or many-shot learning (if applicable) or, less frequently, fine-tuning. - If that still doesn’t yield the results you need, break down the task into subtasks and apply an agentic workflow. If you’re interested a deeper dive into prompting strategies, I recommend the Medprompt paper (Nori et al., 2023), which lays out a complex set of prompting strategies. [Original text (with links): https://lnkd.in/gNJn2Qsy ]

8,564

206 Comments

VC pricing foundation model AI companies based on OpenAI and Anthropic rounds has always been very weird to me. These companies are priced by Microsoft, Google and Amazon. When Microsoft invests in OpenAI, first, they don’t invest only cash, they invest a massive amount of azure credits for compute. Which is much cheaper for them internally. Second, they don’t really care about how many billions OpenAI is worth, they make that investment back through increased revenue from Microsoft products and increased MSFT share price, not by selling OpenAI shares, as VCs would. So why would big VCs follow these valuations and make late stage bets on that is something I don’t really understand.

130

13 Comments

Generative AI Agents - Everything You Need to Know (Part 1) https://buff.ly/45ex25U by @sabrina_ramonov 🤖What are Generative AI Agents? 🤖Designing Effective AI Agents 🤖Empowering Agents with Tools #rafAIgual #AI #GenAI #agents #building #autonomous #innovation #growth #future

3

New Episode of Insights x Design just dropped! Listen to learn from the AI Hacker - Harpreet Sahota 🥑 Watch Here - https://lnkd.in/ghj82ZD5 Listen Here - https://lnkd.in/g_gaQJEb Discover how RAG agents are shaping the future of AI by integrating text and image capabilities! Join us as we explore how these groundbreaking technologies are revolutionizing how we connect with data. The future is now! #AI #RAG #MultimodalA

8

Apollo Research - Frontier Models are Capable of In-context Scheming Source: https://lnkd.in/gpxSvPEk #apollo #frontier #models #capable #in #context #scheming #veryinteresting #llm #ai #genai Credits: Alexander Meinke Bronson Schoen Jérémy Scheurer Mikita Balesni Rusheb Shah Marius Hobbhahn Highlights: In this paper, we developed and investigated a suite of evaluations to assess whether current language models are capable of in-context scheming, which we define as the strategic and covert pursuit of misaligned goals when goals and situational awareness are acquired in-context. Our results demonstrate that several frontier language models clearly display in-context scheming capabilities across multiple evaluations. When the models are prompted to strongly pursue their goal, they can take strategic actions like disabling oversight mechanisms, attempting self-exfiltration, or subtly manipulating outputs while avoiding detection. Furthermore, we found that o1 often doubles down on its scheming attempts by lying to the user in follow-up messages, even when explicitly told to be honest. We analyzed the models’ CoTs and found clear evidence that, within our evaluations, the models understand that their scheming actions are undesired by their developers or users but pursue them nevertheless. Even in cases where models are not heavily incentivized to achieve their goals, models sometimes pursue scheming as a strategy. We found that sonnet-3.5 was able to scheme in one of our evaluations without externalizing any scheming reasoning in its CoT. However, we believe that if AI developers want to make a safety case to deploy their models in high-stakes settings like developer-internal R&D automation, they should provide explicit arguments for why scheming would not occur or would not cause catastrophic harm if it occurs. Secondly, we believe that all deployments for models that are capable of scheming should incorporate automated CoT monitoring by default. This would reduce the risks from scheming, drive forward the science of AI monitoring, and provide evidence on the likelihood and rate of in-context scheming in real deployment settings. Finally, we believe that external evaluators need to be given access to the hidden CoT of models. In our suite, o1 showed the most concerning instances of scheming but does not reveal its internal reasoning to the user and remains the most consistently deceptive after having taken scheming actions. If we had not managed to elicit a summary of the hidden CoT, we would likely not have noticed the majority of scheming actions and not been able to collect nearly as much evidence about o1 s scheming behavior. Even with the current summary, we are unable to draw definite conclusions about the faithfulness of our elicited CoT summaries because we cannot compare it to the ground truth hidden CoT.

7

Is generative AI the next big thing or just another tech bubble? The hype is real, but so are the concerns. Academics warn of inflated promises and overvalued AI stocks, drawing parallels to the dot-com boom. Yet, some analysts remain optimistic about AI's long-term potential. As we navigate this AI-driven landscape, the key lies in distinguishing between genuine innovation and speculative fervor. #AIBubble #TechInvestment #FutureTech #AIEthics #InnovationHype What's your take on the current AI hype? Share your thoughts below.

5

Exciting times for the Toronto AI community! 🎉 Geoffrey Hinton, “Godfather of AI” and a professor at the University of Toronto - right here in our backyard - along with John Hopfield, a professor from Princeton University, have just been awarded The Nobel Prize in Physics for their groundbreaking contributions to machine learning.🏆 As an AI-first business, homegrown in Toronto, we couldn’t be prouder to celebrate this incredible accomplishment. Hinton's pioneering work on neural networks and deep learning has paved the way for AI advancements, while Hopfield’s innovative creation of associative memory systems has influenced the way machines process and recall data. These two trailblazers have significantly shaped modern technologies, impacting fields from medical diagnostics in healthcare to smarter climate modeling. The machine learning principles pioneered by Hinton and Hopfield have enabled organizations like ours to custom-train large language models, streamlining complex processes such as patent research - saving valuable time and offering more precise outcomes. Hinton has also raised important concerns about the need to be mindful of AI’s impact on jobs. We hear him loud and clear! At NLPatent, we are committed to empowering professionals in the innovation sector rather than replacing them. We actively involve our users in every step of our product design and development to ensure our solutions truly meet their needs. Congratulations to Professors Hinton and Hopfield! We’re incredibly grateful for the unwavering dedication of visionaries like Hinton, whose innovations and advocacy for responsible AI continue to inspire us. ✨ #NobelPrize #AI #Innovation #MachineLearning #UofT

13

The type of artificial intelligence that powers systems like OpenAI’s ChatGPT and Google’s Gemini will not be able to reach human levels of intelligence, in particular human levels of planning and reasoning, said Meta’s AI chief Yann LeCun. LeCun is considered one of three “AI godfathers” for his foundational role in the field. The models, often just called LLMs, are trained using vast amounts of data and their ability to accurately answer prompts is limited by the nature of the data they are trained on, LeCun said, meaning they are only accurate when they have been given the right kinds of training data. LLMs have a “limited understanding of logic,” do not have persistent memory, don’t understand the physical world and cannot plan hierarchically, LeCun explained, adding that they “cannot reason in any reasonable definition of the term.” In previous interviews he has explained why a house cat is smarter than the best AI. "A cat can remember, can understand the physical world, can plan complex actions, can do some level of reasoning—actually much better than the biggest LLMs.” With respect to LLMs, its output certainly appears to most people as reasoning—but mostly it’s exploiting accumulated knowledge from lots of training data. He says that the LLMs are only accurate when fed the right training data and that they are also “intrinsically unsafe” and that researchers hoping to develop human-level AI should be looking at other models because they rely so heavily on the training data, which could contain inaccuracies or be out-of-date. LeCun said he and his roughly 500 strong team at Meta’s Fundamental AI Research lab are working to develop an entirely new generation of AI systems based on an approach called “world modeling,” where the system builds an understanding of the world around it like humans do and develops a sense of what would happen if something changes based off this. It could take up to 10 years to achieve human-level AI using the world modeling approach, LeCun predicted. LeCun’s comments reflect Meta’s lower level of #hype about AI, or perhaps his views are driving that lower level of hype. Although Meta’s LLMs like Llama 3 are considered among the most capable in the world, investors have responded less enthusiastically to Meta’s decisions about AI.  The company lost nearly $200 billion in value in April and sparked a wider decline in tech stocks after CEO Mark Zuckerberg told investors to prepare for slower growth while the company pumps money into AI. He urged investors to show patience as it builds out its AI capabilities at significant expense despite there being no indication of where, when or even how the investment will pay off down the line. #technology #startups #innovation #artificialintelligence https://lnkd.in/g8Me-Qab

794

226 Comments

All Ai Tools in One Place ! 💡 Quick Take: This isn't just another AI tool - it's your complete AI command center for 2025. Discovered a tool that's changing the AI game completely! All-in-One AI Power: • Chat with GPT-4, Claude 3.5, Gemini 1.5 simultaneously • Compare responses side-by-side • Real-time web access for up-to-date answers Pro Features That Matter: • Document analysis • Custom prompt library • Local chat history • Dark mode ready 🔥 Ready to supercharge your workflow? Check it out for yourself the links in the comments👇 💡 PRO Tip: You can also simply add a Google Chrome Extension  #DigitalTransformation #ContentStrategy #artificialintelligence #ai #ki #chatgpt #llm __________________________________________ Time to level up your AI game 🚀 🔥💡 Stay tuned for advanced AI Tools strategies! Keep your AI skills sharp: → Smash that LIKE button 👍 (your support fuels the content!) → Drop your biggest AI win in the comments 🧡 → Hit FOLLOW for cutting-edge AI insights 🔔Let's dominate AI Game together!

5

9 Comments

🦄 Fei-Fei Li's New Startup: AI That Understands the 3D World Stanford professor Fei-Fei Li, often referred to as "the godmother of artificial intelligence," launched World Labs, a startup developing spatial intelligence technology, in April 2024. Investors include Andreessen Horowitz, one of Silicon Valley's largest venture capital firms. The recent funding round increased the company's worth to over $1 billion. World Labs has reached this valuation just four months after its launch 😎 «A decade ago, ImageNet from my lab enabled a database of millions of high-quality photos to help train computers to see. Today, we're doing the same with behaviors and actions to train computers and robots how to act in the 3D world. But instead of collecting static images, we develop simulation environments powered by 3D spatial models so that the computers can have infinite varieties of possibilities» — Fei-Fei Li 💡 What’s the Idea Li and her team are working on developing an AI model that can process visual information in a human-like way and understand the dimensions of objects, their functions, and spatial relationships. In a recent TED Talk, Fei-Fei Li illustrated the idea of spatial intelligence by showing a picture of a cat with an outstretched paw pushing a glass to the edge of a table. The human brain can assess in a fraction of a second the geometry of that glass, its place in three-dimensional space, its interaction with the table, the cat, and everything else, and then predict what might happen and take action to prevent the glass from falling. Algorithms with spatial intelligence will also be able to realistically extrapolate images and text into three-dimensional environments and act on those predictions. 👍 How It is Useful This will help make breakthroughs in navigation, robotics, gaming, augmented and virtual reality. Robots at the Fei-Fei Li-led AI lab at Stanford can already make sandwiches, open doors, and unplug smartphones from chargers in response to verbal instructions. More on the topic: Fei-Fei Li's Speech at Ted Talks Fei-Fei Li is the Fairy Godmother of Artificial Intelligence #startup #ai #robotic #technology

4

Great recap of #prompt techniques - see section 2.2. in particular if you are working on LLM applications.

1