Diarmuid Curran

Forget the AI doomsayers — could AI's impact be more modest than you think? 🤔 A new paper, "The Simple Macroeconomics of AI" by Daron Acemoglu (found here: https://lnkd.in/eVn4czWF), uses a task-based model to assess AI's impact on the economy, focusing on automation and task complementarities. In his analysis, Acemoglu proposes that if the microeconomic effects of AI are primarily driven by: 💰 Cost savings 🚀 Productivity enhancements at the task level Then the macroeconomic consequences can be elegantly estimated using Hulten's theorem (which asserts that GDP and aggregate productivity gains can be calculated by considering the proportion of tasks affected and the average cost savings per task). Using estimates on AI exposure and task-level productivity improvements, Acemoglu's paper suggests that while the macroeconomic effects of AI are non-negligible, they are relatively modest. His analysis points to a potential increase in total factor productivity (TFP) of no more than 0.66% over a decade. 📈 And even these estimates may be inflated. ⚠️ Early evidence is based on easily learnable tasks, whereas future effects will also stem from tasks that are more challenging, involving context-dependent factors and possibly lacking objective performance measures. Consequently, the predicted TFP gains over the next ten years are even more restrained, estimated at less than 0.53%. Should we be listening to Daron Acemoglu? Who is this guy? Let's examine his credentials. 🕵️ He is writing about AI, but he got his PhD from the so-called "London School of Economics"... notice it's not called the London School of Artificial Intelligence. Suspicious! 🤨 And he's allegedly an "economist," but he's a professor at MIT - which, last I checked, stands for Massachusetts Institute of Technology, not Massachusetts Institute of Economics. Hmm, something doesn't add up here. 🤔 Plus, how can this guy possibly assess the impact on the entire U.S. economy when his alma maters are named after tiny little London and Massachusetts? One's just a city and the other a measly state. Hardly seems qualified to prognosticate on the whole GDP if you ask me! 🗺️ And his supposedly most prestigious honor is the John Bates Clark medal, given to the top economist under age 40. But guess what... Acemoglu is 55 years old! 🥳 While Acemoglu's paper offers intriguing insights, should we take his findings with a grain of salt 🧂 and continue unceasingly to extol the virtues of visionaries like Sam Altman and the impending AI-driven utopia they promise. Vive la révolution! 🎉🤖✨ Or is Acemoglu onto something with his more modest predictions for AI's economic impact? 🤖💰 Share your thoughts in the comments below. What's your take on the hype vs. reality of AI's potential to boost productivity and growth? 🧠💭 #ArtificialIntelligence #Macroeconomics #EconomicsOfAI

2

GPT-4o is getting further integrated into Windows, and OpenAI is reportedly also bringing it or something similar to Apple devices, with an announcement possible at Apple’s upcoming WWDC. From data I found online, this would give OpenAI’s models a footprint on up to 52.4% of devices, and Google’s models most of the rest on Android devices. (Though not all devices will be able to run updated operating systems with the models, however.) If natural language, including voice, becomes the dominant means of interacting with devices, then Google’s chokehold on search might not matter on non-Google devices. If you ask your Windows, Mac, or iOS device to answer a question, it will either just answer it, or search with Bing. Now, Google has a massive search partnership with Apple that could prevent that on iOS devices, but that partnership may also be affected by antitrust actions. So it will be very interesting to see what Apple announces at WWDC.

6

1 Comment

For anyone interested in running LLMs locally, https://www.ollama.com is worth looking into. You can run Llama 3.1 7b on 8 GB of RAM locally on your computer. This means no data is shared with anyone online and all processing happens directly on your computer. For mac they have a great user interface called https://lnkd.in/eWbXx5jq that I use. I'm not saying this is "better" than ChatGPT or Gemini, but definitely cheaper, faster and better when working with sensitive data. Let me know if you like it, because I do. p.s. currently downloading llama 3.1 70b to run on my Mac M1 Extreme with 64GB RAM. Let's see.. (I wish I had faster internet now)

30

3 Comments

The Evolution of AI: From Rule-Based Systems to Deep Learning: I've just published a new article on Medium that explores the fascinating journey of Artificial Intelligence, from its early days of rule-based systems to the advanced deep learning techniques we see today. In this article, I delve into: -The historical milestones of AI development -The shift from rule-based systems to machine learning -The rise of neural networks and deep learning -The impact of AI on various industries -Future directions and ethical considerations Whether you're an AI enthusiast, a data scientist, or just curious about the tech world, I hope you'll find this article insightful and thought-provoking. #AI #MachineLearning #DeepLearning #ArtificialIntelligence #TechInnovation

11

2 Comments

My thoughts on Meta's ROBYN MMM model. For context, Robyn is an open source MMM product that Meta has developed in the last few years. After working on a no. of these models, here are my takeaways: 1. Its not perfect, but its much better than its previous versions A new feature is calibration frame which accommodates incrementality tests and spend, should users want to add any tests. 2. It is time consuming - There are a no. of PIPs and integrations that are needed for it to initially run, and can take more than couple of hours to just set up these integrations with #nevergrad - Even though most of the code is written, there is heavy lifting and decision making that is needed from data scientists. A no. of models are output and needs a seasoned analyst/statistician to know which models to pick and test - It does not have any EDA. You have to do it separately 3. Modeling issues - Like most MMMs, this one is also biased by spend. It heavily takes into account the spend and calibrates contribution proportionately - The optimizer reccomendations are below average. - You need to understand how to use hyper parameters and tweak them based on each model. Needs fundamentals around these 3. Documentation on Robyn's github is pretty extensive There is a facebook group of active MMM users who are helpful, vs the Github page. All in all, I think there are opportunities to improve these outputs, but for a model with less than 10 media variables, I think it does a decent job. What has been your experience? #mmm #modeling #robyn #roi Naavics

66

24 Comments

Name That Bubble AI is in a bubble right now. I can prove it with one slide. In just 6 weeks, the market traded NVIDIA up and down $2 TRILLION DOLLARS. That’s more than the entire value of META. This, with zero intel on their 2Q performance. That is a bubble. If you want the FULL case for the #AI bubble (and why it matters to YOU now) click this: https://lnkd.in/e4prwQ8k

14

1 Comment

Nice. Greatly useful to improve diffusion models generalization with multi-modal data!

9

1 Comment

🧠 Time to Ditch the AI Metaphors: Let's Get Real 🧠 We're giving "AI" way too much credit. It's not learning, training, or intelligent. It's just a complex math problem on steroids playing probability Tetris with words. 🔍 "Neural networks"? Just a tangled mess of nonlinear equations, not digital brain twins. 📚 Machine "learning"? No cute classroom scenes. It's systematically minimizing objective functions, not gaining wisdom. 🏋️‍♂️ "Training" a model? It's crunching numbers, not a Rocky montage. 🎲 "Knowing"? It's as knowledgeable as a toaster is about toast. Just probability, no pondering. Anthropomorphizing AI isn't just annoying - it's dangerous. It inflates the hype and potential threats. Worse, it leads to nonsense like equating AI "learning" to human artists creating. That's like calling plagiarism "collaborative creativity." When AI spits out "George Washington" as the first president, it's not channeling history - it's playing the odds. And if it said “George Jefferson?” Not a "hallucination"? Not a bad acid trip. Just a glitch in the matrix. 🚨 We need a vocabulary overhaul. Stop anthropomorphizing. Start being accurate. Our future depends on it. Ready to rethink our AI lingo? Drop your thoughts below! 👇 #AIReality #DataEthics #AnthropomorphismFail

4

Anthropic Unveils the Black Box of Large Language Models (LLMs) Anthropic has released a groundbreaking paper that, for the first time, inspects the inner workings of a large language model (LLM). Until now, these models acted as black boxes, producing outputs without clear understanding of the underlying reasons. Using a technique called "dictionary learning," Anthropic's team managed to isolate millions of abstract features that represent similar concepts across different contexts and languages, even with image inputs. This breakthrough allows mapping the internal states of the model and understanding the proximity of related concepts. For example, features close to the concept of "inner conflict" encompass relationship breakups, conflicting allegiances, and logical inconsistencies. Remarkably, it is now possible to artificially manipulate these features and alter the model's behavior. This opens up prospects for ensuring the safety of LLMs by activating or deactivating certain features. #ArtificialIntelligence #NLP #LargeLanguageModels #Anthropic #DataScience #MachineLearning #Innovation #TechNews #Research #LanguageAI #Explainability #AIGovernance Full Paper : https://lnkd.in/dZFEgZjW

18

1 Comment

🚀 Testing ChatGPT o1 Preview already, and it’s far superior at solving code issues than 4.0! 🧑‍💻 It’s going to shave DAYS off a project I’m building. ⏳💡 Just this morning, it accurately solved 4 code fixes for me, which had been taking hours yesterday even with debugging help from 4.0. 😅👨‍💻 So, what does this mean? 🤔 People might complain that o1 is a bit slower 🐢, but they’re missing the point. What I’ve already experienced is that while 4.0 sometimes hallucinates decent answers, o1 is actually applying problem-solving techniques 🧠 — an advanced reasoning process that’s running behind the scenes. This is yet another example of why OpenAI is going to win. 🌍🏆 By focusing on the user and delivering meaningful results, they’re proving that prioritizing the customer experience is the key to success in business. ChatGPT is way ahead in trying to understand human intent. 💼✨ 💬 What do you think about this shift in AI problem-solving? Have you tried o1 yet? Share your thoughts in the comments — I’d love to hear your experiences! 👇 #ArtificialIntelligence #AI #ChatGPT #OpenAI #TechInnovation #ProblemSolving #AIRevolution #MachineLearning #AICommunity #CodeFix #TechLeadership

3

AI generated video. LVMs don't understand physics or physiological mechanics. No doubt these will get better with more training and extortionate cost of optimisations. Now imagine you could build a model of video generation by merging together smaller pieces of logic, like Newtonian laws, physiological mechanics, etc... And in time, keep improving the model by updating it with new pieces of information, like the subtle effects of bar height/number on technique. This is the type of tech we're working on at Jiva -- how to merge models, building block by building block. Altogether an entirely different way of achievement actual intelligence. credit: https://lnkd.in/ezv2EgfV

166

52 Comments

A bit on copyright. There have been two different stories in the news relating to copyright and AI this week. The first, about a petition now signed by over 20,000 musicians and writers that argues: “The unlicensed use of creative works for training generative AI is a major, unjust threat to the livelihoods of the people behind those works, and must not be permitted.” https://lnkd.in/e7479hhS I happy to say I've signed this in a personal/musician capacity in solidarity with the other (far more successful!) musicians and writers who have signed. I'm sure we all wear many different hats at different times. The key for me is 'unlicensed'. Clearly I've no problem with AI models being trained, but the creators of the training material should be paid if it was copyrighted, and paid appropriately, including for models that have already been trained. The other story is titled "Microsoft boss urges rethink of copyright laws for AI". It's in The Times, behind a paywall, so I won't link. I agree with the core of the argument that we need copyright laws to updated to keep up with technology, but I can't be the only one that thinks these statements is are absurd: "Speaking after Microsoft’s launch of virtual employees at an event in London, he compared the situation to that of using information from textbooks to formulate new ideas. “If I read a set of textbooks and I create new knowledge, is that fair use?” he said." “What’s copyright?” Nadella asked. “If everything is just copyright then I shouldn’t be reading textbooks and learning because that would be copyright infringement.” I think he's trying to create a link in people's minds between humans reading and AI models being trained, claiming they are both 'just' learning. Something that I think only works because we ended up with the word 'learning' being used in AI (as in Machine Learning) for a process where a model is iteratively optimised to predict an outcome. Maybe The Times has taken this out of context. I'd be interested if anyone knows any more on this. But honestly, if the article is correct, this really, really isn't helping with the debate.

16

4 Comments

Oh look, another "I know Python" developer who thinks printing "Hello World" makes him the next Guido van Rossum! 🐍 Let's be honest - we're all that person who bought the deluxe gym membership but only uses the treadmill. You know Python has list comprehensions, decorators, and generators, but you're still writing nested for loops like it's 2005. Here's a wake-up call: You're the coding equivalent of someone who pays $100/month to walk on a machine when you could just... walk outside. Those advanced Python features? They're like the weight room you keep avoiding because "basic cardio is fine." Stop being the person who brags about their Python skills while secretly Googling "how to reverse a string." Either level up or admit you're just here for the free pizza at the meetups. PS: Yes, I wrote this while procrastinating on learning asyncio. Fight me. #python #coding #developerlife

1

there goes the weekend! experimentation is hard and takes time but it's worth it. I have paid for a subscription for Gemini awaiting for its talents.. it is slowly emerging as the copilot's in the works suite proving to be very important and extremely useful. more exploration required. this advance is not surprising to me because competition is real and the richest companies in the world that are able to attract both customers and capital are vying for leadership here. in the case of Google it's defense and it is pulling out all the stops to compete with openai and Microsoft. I cannot guess at this time who will win this battle and make the most advanced useful artificial intelligence based on generative technology that will change the productivity levels of small medium and large companies. but it's going to be one of the foundation companies for certain... or it'll be someone waiting on the wings as a fast follower. I don't think that group can emerge without massive capitalization so we'll at least have 12 or 18 months warning if there's a new entrant. #ai #gemini #google #iapredoction thanks for the Post Nigel... I know you're biased but I love the bias 😎💥

1

The problem with current LLMs, or any AI that requires ever growing amount of data is that they are driven by the past (data). What's worse, is that we can't trust the outputs... how so then, can they create any predictably rosy future for humanity? How can AGI be near, if this is the direction we are heading? Civilizations started without these large data, but rather by discovering the rules & building blocks of the world. Great ideas grew upon understanding and clever usage of these to meet needs, demands, desires, goals... if LLMs can't understand nor possess the above mentioned, how can we be sure the outputs are correct? Not only the issues of nature, but there are the issues of civilizations, psychology, economics and all those knowledges that come with humanity itself. Issues like "Biases", but whose is correct? "Alignment", but to whose alignment? There are different cultures and schools of thinking and that is also why the world is so interesting to explore and why we are able to advance as a civilization. Are the LLMs and AIs going to unite everyone into one culture with similar biases and alignment? 🤔 #ai #artificialintelligence #agi #llm

4

🚀 The “Last Mile” of AI: A Game-Changer 🚀 Let me break it down… We’ve all heard about Amazon’s “last mile” challenges. Despite having cutting-edge logistics, it’s that final stretch that poses a major hurdle. In generative AI, I call this last mile the custom UX. It’s what turns complex AI into user-friendly tools. 🌟 Runway is a great example of nailing this last mile. They’ve built a UX that helps users create exactly what they envision, without needing to be prompt engineering experts. 🎥 As generative video evolves, directing becomes a crucial last mile challenge. Think about it: How many of us can expertly prompt a “dolly push pan to slow zoom”? (Is that even a thing? 🤔) This is where last mile UX shines. TikTok mastered it, Instagram gave us filters. The goal? Democratize quality through intuitive interfaces. 💡 Want to glimpse the future? Talk to a great UX designer! 📢 Also shout out - if you want to understand the future in a manner that both entertains and informs, follow Gavin Purcell and AI For Humans. I always look forward to these.

3

Hugging Face has published a new so called leaderboard - benchmarking large language models in healthcare settings https://lnkd.in/eJ9vC7WU. #huggingface #largelanguagemodel #healthcare #artificialintelligence #leaderboard #benchmarks

17

5 Comments

Getting AI into production has us once again revisiting another older (relatively speaking) data tool: graph databases. Specifically, some companies are evaluating adding an extra step beyond just traditional RAG. Databases and developers I’ve spoken with have started to see early signs of a two-stage process that wraps standard RAG around graph traversal to improve the accuracy and relevancy of search results. Rather than just retrieve basic information from a vector database, RAG is used as a first step to retrieve nodes of a knowledge graph. That provides a wider, and potentially richer, starting point for traversal. It adds a layer of complexity to the process, but for some companies that have stricter requirements around accuracy (say, medical environments) it stands to potentially be another tool in the future AI data stack. Graphs, too, offer another benefit: explainability. Because relationships on a graph are well-defined, retrieving nodes takes away some of the opaqueness of the relationships in a vector retrieval. Here’s how Emil Eifrem at Neo4j explained it: “Vector spaces are completely opaque—it’s a bunch of numbers, and human beings can’t parse that. Graph spaces meanwhile are completely explicit. I show an apple and a tennis ball, and vector similarity search will say they’re similar but not tell us why. It’s some dimensions in some kind of latent space out there. In graph space, though, an apple and an orange we see are related explicitly because they’re both fruit.” Still, there’s a question of whether machine learning powered search techniques end up “skipping” these steps in the future—and if vector embeddings are all that’s needed. Here’s what Bob van Luijt from Weaviate told me: The vector datatype is everywhere now. So it doesn't matter what type of database you have: time series, document DB, or graph; you can attach embeddings to everything. What embeddings do is find stuff, so in the context of a graph database, that's finding nodes in the graph. Where it becomes interesting is when a database is AI-native. E.g., Weaviate stores graph connections but is by no means comparable with a purpose-built graph DB. However, in machine learning, we have graph representation learning. How is that expressed? Vector embeddings.” How we get to that point of a direct knowledge traversal clearly has a lot of pathways. But it does seem like from all this experimentation that we could see a lot more activity around graph databases in the coming months. #ai #bigdata #graphdatabases https://lnkd.in/gX9ciTU5

57