Peter Sarlin

Matías Otero Johansson Marco Mezger One thing that makes history interesting is that because reality turns out to be complicated, you always end up questioning a lot of your assumptions about what happened. So in the neo-liberal telling of history, the Soviet economy was a total disaster under Brezhenev, and then Gorbachev reformed it, and then it became capitalist, and then we all live happily ever after. https://lnkd.in/gD639bms The trouble is that this is not the way most Russians see things. In the eyes of most Russians the 1990's were an absolute disaster and it turns out that the villain in the story is Mikhail Gorbachev and Boris Yeltsin. Gorbachev took an economy that had some problems and then ran it into the ground. The thing about the Soviet economy was that it was growing at decent rates until 1973, and then something happened. The interesting thing is there is a lot of disagreement as to what happened. My explanation is that 1) The oil shock of 1973 required a reallocation of resources. The West had a market economy, whereas the Soviet economy didn't. 2) You didn't have any mechanism in the Soviet economy for incorporating new technology (Apple Computer)

2

5 kommenttia

There have been approaches like CBRoberta based on BERT etc., there may be cases where it outperforms LLMs, but I don't think there are privacy etc. concerns regarding public communication of CBs for using commercial models (results also indicate that they seem to perform better as they evolve, which will only lead to increased performance). As costs are also likely to decrease, I believe the use of specific NLPs or training models will only work in niche situations. Another issue is likely to be the changing tone of individual central banks and their eagerness to act as they say they will. We have used GPT4 on CBRT speeches once but real life data may not just align with what will happen. https://lnkd.in/dRiQrP-g

5

1 kommentti

Joe Rogan did a very interesting podcast with Marc Andreessen yesterday: https://lnkd.in/d4y3X_v8 - and it's worth watching all the way through, whether you agree with him or not. Among economics, politics, AI, drones, general topics of (de)regulation, Marc Andreessen mentioned something I have been harping on since at least 2018 - that in the face of AI fakes and general fake or doctored content, the only way to get to actual truth is to actually have "supply chain tracking" for information - and just disregard information that can't be tracked to respected sources. And blockchains, used as databases and not for FinTech and general scammyness, are a uniquely suited technology for this - they are globally accessible, distributed, and designed to protect the validity of stored information. Unfortunately, the demand side for this is low, and so far I haven't been able to find any source of funding for this project. I'm 100% convinced a solution based on these ideas will be created and used, either by me or by someone else. https://lnkd.in/dFpfm66N

16

8 kommenttia

Interesting chat with Peter Guest / WIRED on European AI, tech & competitiveness 👇 below a few topics on AI sovereignty, more in the article. "The dominance of American models is driving many in Europe to talk about the concept of “AI sovereignty”: making sure that the core “digital infrastructure” behind the AI boom isn’t entirely controlled by private companies outside of the continent." But I believe focus shouldn't be on just sovereign AI models. ▶️1️⃣▶️ "We have a very good basis in Europe in terms of data, compute, and talent,” he says. “But that’s not enough, you need distribution. And that happens through software products, usually software platforms.” ▶️2️⃣▶️ “Marc Andreessen said in 2011, software's eating the world, software's eating traditional industries … AI is accelerating that,” “Unless we ensure that we have these [...] sizable software product companies in Europe, we're not going to be able to create value in Europe with AI.” ▶️3️⃣▶️ "Only two of the 10 most valuable public European corporations are in tech: German business software provider SAP and Dutch semiconductor equipment maker ASML. Six of the world’s 10 most valuable public corporations are US tech companies. Nvidia and Microsoft, the two largest, are each worth more than 15 times SAP", as covered by Peter. Thank you Peter Guest and WIRED the interview and for an insightful coverage on AI sovereignty! Link in the comments. #AIforPeople

146

16 kommenttia

A few observations on BRICS in my latest piece for Valdai 👇 https://lnkd.in/dgUuNaUW

7

new and #openaccess article in IJRM - International Journal of Research in Marketing: "More transactions but a lower average transaction value: How #mobilepayment #apps influence consumer purchases through time-savings" with Steven Lu https://lnkd.in/eW2jSnts

28

Happy to share that our latest paper, "Assessing the solvency of virtual asset service providers: are current standards sufficient?" is out. Centralized cryptocurrency exchanges, as any corporation, can become insolvent: our latest paper explores whether current standards are enough to assess the solvency of virtual asset service providers (VASPs) and proposes solutions for better oversight. Thanks to my coauthors Esther Segalla Michael Sigmund Burkhard Raunig Felix Zangerl Bernhard Haslhofer! Scuola IMT Alti Studi Lucca Oesterreichische Nationalbank Complexity Science Hub Link to the paper: https://lnkd.in/dpg6QDrz #research #cryptocurrencies #economics #academic #datascience

21

1 kommentti

What are the risks from artificial intelligence? 🤖 This MIT project proposes a comprehensive AI Risk Repository of over 700 AI risks, extracted from 43 frameworks and categorised into a Causal Taxonomy (how, when, why risks occur) and a Domain Taxonomy (7 domains & 23 subdomains). How can you use the AI Risk Repository? - Get a clear view of the AI risk landscape 🌍 - Access new risks and research on a regular basis 🔍 - Use as a reference for research, development, policy and audit 🛠️ - Support AI curricula and policy frameworks 📚 https://airisk.mit.edu #AI #AI safety #Machine learning #ResponsibleAI #DataScience

24

1 kommentti

🍍 In the first week of June, several noteworthy research papers were published that provide valuable insights into the inner workings of large language models (LLMs). These studies delve into various aspects of LLMs, offering a deeper understanding of their cognitive processes. 1️⃣ When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMs: This survey investigates the conditions required for successful self-correction in large language models (LLMs) by critically surveying prior work, and establishes clearer definitions of research questions targeted in self-correction research. It also provides a checklist for designing appropriate experiments to evaluate self-correction. ArXiv: https://lnkd.in/g75pVugb 2️⃣ Iteration Head: A Mechanistic Study of Chain-of-Thought: This paper gives insights to the inner workings and conditions that give rise to Chain-of-Thought (CoT) reasoning capabilities in transformer language models, and explains why transformers gain CoT abilities when trained with next-token prediction in a controlled and interpretable setting. The paper also hypothesizes that specialized "iteration heads" emerge in transformers to enable CoT reasoning. ArXiv: https://lnkd.in/gWVYuHmx 3️⃣ What Makes Language Models Good-enough?: The authors create a good-enough language processing (GELP) evaluation dataset to test the effects of different architectural features on language processing. By doing so, they examine the effects of number of layers and self-attention heads on language models' performance in good-enough language processing. ArXiv: https://lnkd.in/g84-meww 👉 Read further below to get the latest in-depth highlights, methodologies, experiments & results on NeuroPurrfect AI: https://lnkd.in/g526DyVf #AI #MachineLearning #LLM #ArtificialIntelligence #NeuroPurrfectAI

46

2 kommenttia

https://lnkd.in/d6wiYTry If the results of this study by the Anthropic research team are confirmed and can be scaled and generalised, this may be the most important advancement in empirical LLM studies and xAI since the introduction of gpt 3.5 instruct the first big decoder-only model that attained human-level linguistic performance.

1

What is the right pricing model for AI? Should it be a monthly fee or a flat rate per token? Do you pay extra for more knowledge? Three years ago, I was focused on server-less computing for AI and how to allocate inference engines. At the time, consumption based pricing was the future. Maybe it still is? Vikram Sreekanti and I have been thinking about pricing models and we just published our latest thoughts: https://lnkd.in/gUdKPDRF

27

1 kommentti

So OpenAI decided to pause the use of Sky, the demo voice of GPT4o, because it is considered as too close to the tone of Scarlett Johansson from the movie #Her (image and news: https://lnkd.in/eyNvivSc). My three cents: 1) I think it shows once again that digital innovators, especially in the US, are heavily influenced by #PopCulture and #ScienceFiction. We see it with #Musk and Asimov, #Zuckerberg and Ready Player One, and now #Altman and Her. 2) Research is probably needed to unpack the positive aspects and negative aspects of such influence. Sci-Fi can of course foster the imaginary (as established in the #DesignFiction literature for year: https://lnkd.in/egQcRrrK) but can lead to non-desirables futures as well in terms of sustainability, ethics or in this case, potential legal issues with the voice of an actor being potentially mimicked. 3) Her is one of the most influential sci-fi movie of the decade, and has inspired numerous innovations (incl. controversial ones such as AI girlfriends). With Benoît Vanderose, we discuss this in our Pop-Code podcast episode on the movie: https://lnkd.in/eZvuZXCt

30

4 kommenttia

"Fine-Tuning is a nightmare in practice..." (advice from X). I agree. I have built ML and AI models for clients for 20 years. eHow. eBay. Walmart. Blackrock. Even Google (well, Aardvark). It's hard. But the payoff is there --if you can get it right. Best example: eHow was the first $1B IPO since Google. So if you are Fine-Tuning your own models, how can you know if you are on the right path ? Weightwatcher can help. I invented weightwatcher to help my clients who are training and/or fine-tuning their own AI models. And it's open-source. How can it help ? Here are over a dozen examples of how to interpret weighwatcher results for Instruction Fine-Tuned models. Generally speaking, if you get the Fine-Tuning right, your model will follow the predictions of the weightwatcher HTSR theory. And it when it doesn't, that special case can be useful too. If you know what you are looking at. If you have a fine tuned model, and you have the base model, this is all you do: pip install weightwatcher import weightwatcher as ww watcher = ww.WeightWatcher() details = watcher.analyze(model=model, base_model=base_model) Weightwatcher will remove the instruction fine-tuned components from the base model and analyze them for you Want to learn more ? Check out the examples: https://lnkd.in/gS8bS3tM Have questions? Join our Community Discord: https://lnkd.in/gZQF64Bw Or ping me here. And if there are cases you think we should add, please let us know. WeightWatcher is a one-of-a-kind must-have tool for anyone training, deploying, or monitoring Deep Neural Networks (DNNs). #talkToChuck. #theAIguy

59

11 kommenttia

I have been building production machine learning models for clients on Wall Street, Main Street, and in Silicon Valley for over 20 years. I believe that the next generation of AI models will use large scale instruction fine tuning. And to achieve this, you need to know if the fine-tuning is working. To solve this problem, I invented weightwatcher. If you want to get into the space and understand how to fine-tune models at scale, please check out these examples of what you should expect from a good model

64

1 kommentti

Helping ordinary citizens exercise their #data_sovereignty through decentralized trading of personal resources (e.g., data, SW, and compute devices) on the web for a wide spectrum of AI application developments is an effort to reinforce the levee that gives economic middle stability. Yes, that levee #Jaron_Lanier mentioned in his book, "Who Owns the Future". Why decentralized trading? Because the resource-hogging big tech companies have not adequately compensated the users for what they have supplied for many years, and it is destroying the "levee" necessary for avoiding any social unrest in the future. P2P trading seems to be the only viable way to help citizens monetize the resources they "own", unless the government takes practical actions to make sure data owners get proper payment they are entitled to, or the big tech companies voluntarily seek healthy symbiosis with the data contributors. However, the P2P trading of these personal resources should be facilitated in such a way that ensures scalability that is comparable to what the central big tech companies would offer, and most of all, secures transactions and protects privacy. Without such means, users would remain to be exploited by the #siren_servers. The following two papers we published in the special issue of #Journal_of_Web_Engineering introduce how we satisfy those critical requirements. Web 3.0 Chord DHT Resource Clustering https://lnkd.in/gRYgfKGF Self-sovereign and Secure Data Sharing Through Docker Containers for Machine Learning on Remote Node https://lnkd.in/grcDXHBr #데이터자기주권

3

New Working Paper "Do experimental asset market results replicate? High-powered preregistered replications of 17 claims", joint work with Christoph Huber, Felix Holzmeister, Magnus Johannesson, Christian König-Kersting, Anna Dreber and Jürgen Huber. Link to paper: https://lnkd.in/gR7nv3Gt Experimental asset markets provide a controlled approach to studying financial markets. We attempt to replicate 17 key results from four prominent studies, collecting new data from 166 markets with 1,544 participants. Only 3 of the 14 original results reported as statistically significant were successfully replicated, with an average replication effect size of 2.9% of the original estimates. We fail to replicate findings on emotions, self-control, and gender differences in bubble formation but confirm that experience reduces bubbles and cognitive skills explain trading success. Our study demonstrates the importance of replications in enhancing the credibility of scientific claims in this field.

79

2 kommenttia