EU Economy and Finance’s Post

More support for my claim that LLM's are slowing down in progress: https://lnkd.in/gysPHbHs In the medium and long term (and perhaps near term), exponential increases in costs, etc is a negative trend, not a positive one. Non-linear costs in compute, power, data, money required to be invested is unsustainable. I see others getting the wrong meaning out of the data in the above article. This data really means is that there has been an increased focus on LLM's. Even the compute doesn't seem like it can keep up with doubling every 6 months, but that seems unlikely. Power requirements doubling seems possible, but also unlikely, to be sustainable. Costs 2.4x'ing every year is not sustainable. Probably the biggest issue is data. There's only so much data out there and new data is going to get harder to gather and clean as people try to block AI from gathering it. There is no way 3x data every year is sustainable.

What is RAG? RAG is a technique for augmenting LLM knowledge with additional data. LLMs can reason about wide-ranging topics, but their knowledge is limited to the public data up to a specific point in time that they were trained on. If you want to build AI applications that can reason about private data or data introduced after a model's cutoff date, you need to augment the knowledge of the model with the specific information it needs. The process of bringing the appropriate information and inserting it into the model prompt is known as Retrieval Augmented Generation (RAG).

Incredibly excited today to have published this massive new exploration of what AI can do for government – and what the government should do for AI to help our public services flourish. Any government serious about delivery must give this its full attention. I learned so much working on this joint project between the Tony Blair Institute for Global Change and Faculty: what AI can and can’t do; how we make it safer and fairer; the positive change it unlocks for how we engage with citizens, run the back-office of government and make policies. Most of all, it has filled me with a real optimism: if the government chooses to, we have the means to all but eliminate backlogs; make the civil service an incredible place to work on complex problems with cutting-edge tools; put strategic decisions on a new footing with rapid complex simulations and real-time insights. All of that while saving the Treasury up £40 billion a year in the next five years. Give it a read and let me know what you think. Massive thanks to co-authors Tom Westgarth and David Railton and the wider team of contributors and reviewers Rachel I., Laura Britton, Ursule Kajokaite, Benedict Macon-Cooney, Jeegar Kakkad, June Shin McCarthy, Roger Williams, John Gibson, Nijma Khan, Paul Maltby, Sona Hathi, Filip Wolski, Matt Clifford, Alex Chalmers, Mike Keoghan, Roger Taylor and Sir Patrick Vallance among others. https://bit.ly/4au1FoG

How do LangChain and LlamaIndex differ? 🤔 Here's a quick breakdown: LangChain is your go-to for multi-purpose LLM-powered applications, offering an array of tools for diverse needs. Its strength lies in flexibility and advanced AI capabilities. On the other hand, LlamaIndex shines in search and retrieval applications. With a focus on lightning-fast data retrieval and crisp response generation, it's your solution for streamlined search experiences. Choosing the right one depends on your specific needs: ➡️ Opt for LangChain if you require a versatile framework with extensive AI functionalities. ➡️ Go for LlamaIndex if your priority is rapid data retrieval and precise response generation in search applications. Choose wisely based on your project's needs: https://lnkd.in/gQnacsPd

𝐁𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤𝐢𝐧𝐠 𝐋𝐋𝐌𝐬 𝐉𝐮𝐬𝐭 𝐆𝐨𝐭 𝐄𝐚𝐬𝐢𝐞𝐫 The Stanford AI Index 2024 report highlights a major challenge in the LLM industry: a lack of standardized testing. This makes it difficult for organizations to choose an AI provider that aligns with their needs and values. At QuantPi, we believe in transparency empowered by data you can trust. And that's why we developed the LLM Test Suite, a comprehensive and efficient way to benchmark that goes beyond subjective assessments. 𝐐𝐮𝐚𝐧𝐭𝐏𝐢'𝐬 𝐋𝐋𝐌 𝐓𝐞𝐬𝐭 𝐒𝐮𝐢𝐭𝐞 𝐚𝐥𝐥𝐨𝐰𝐬 𝐲𝐨𝐮 𝐭𝐨: 👉 Test diverse LLM models for your specific application. No more relying on cherry-picked results! 👉 Compare apples to apples. Our standardized tests ensure a fair and holistic evaluation. 👉 Make informed decisions. Choose the LLM that best meets your needs and ethical standards. Curious as to what can be tested? We performed a subset of our LLM assessment across different trustworthy dimensions such as performance, robustness and bias with the SQuAD2.0 validation dataset, where SQuAD stands for Stanford Question Answering Dataset for Microsoft's Phi-2 and Google's Gemma 7-b. Dive into the results here: https://lnkd.in/g5kvxyTi #ResponsibleAI #StanfordAIIndex

Improving Public Consultations with AI - Use Case from the UK Government's Incubator for Artificial Intelligence � � The Consultation Analyser from the UK Government's Incubator for Artificial Intelligence (i.AI) is an AI-based tool designed to automate the analysis of responses to public consultations, with the aim of improving the efficiency and fairness of the policy-making process. It was developed in partnership with No10's data science team. � The Analyser identifies patterns and themes within public feedback and presents these findings to policymakers via dashboards. This approach allows for more streamlined analysis, potentially reducing the �80 million annual cost associated with public consultations. � The Consultation Analyser uses topic modelling to extract themes from responses and ensures data confidentiality by running on internal servers. This method is intended to increase the objectivity of the analysis, freeing up human resources for deeper insight into the patterns identified.

𝐓𝐡𝐞 𝐫𝐞𝐜𝐞𝐧𝐭 𝐬𝐮𝐜𝐜𝐞𝐬𝐬 𝐨𝐟 𝐳𝐞𝐫𝐨-𝐬𝐡𝐨𝐭 𝐭𝐢𝐦𝐞-𝐬𝐞𝐫𝐢𝐞𝐬 𝐦𝐨𝐝𝐞𝐥𝐬 𝐰𝐚𝐬 𝐞𝐱𝐩𝐞𝐜𝐭𝐞𝐝. ➡️  Two new papers show that 𝐒𝐜𝐚𝐥𝐢𝐧𝐠 𝐋𝐚𝐰𝐬 of Larger Models also apply in time-series. ➡️ 𝐒𝐜𝐚𝐥𝐢𝐧𝐠 𝐋𝐚𝐰𝐬 are empirical rules that describe the relationship between an LM’s parameter size, tokens(dataset size), training time, and performance. ➡️ First introduced in [Kaplan et al.], and were later re-examined in Deepmind’s Chinchilla paper. ➡️ This explains why larger DL forecasting models perform better on non-toy datasets. Scaling-laws for Large Time-series Models: https://lnkd.in/dVqZ4_7e Scaling Law for Time Series Forecasting: https://lnkd.in/gBei--C9 For latest developments in AI research, feel free to subscribe to my newsletter: AI Horizon Forecast Newsletter: https://lnkd.in/gBei--C9

In lower compute cost we trust

RAG is not nearly dead, nor will it ever be. Having a large context window does not imply making a good retrieval. 𝐁𝐮𝐭 𝐭𝐡𝐞𝐫𝐞'𝐬 𝐦𝐨𝐫𝐞: ◉ The more irrelevant information, the more noise in the output, therefore worse results, keep it clean. ◉ RAG is much more complex than extracting information and inserting it into the context, before reaching this point, rankings and evaluations are made, and pipelines will become more complex as agentic systems evolve. ◉ Even assuming it's a good idea to fill the context with data you don't need (it's not), using third-party APIs the input token is billed, and whatever the case, even if you are hosting your own, the GPU is going to suffer as the context gets larger, so, again, keep it clean. ◉ The main benefit of RAG is that it enables an LLM to access real-time data from current internal and external knowledge sources, which makes it more versatile and responsive to new information. ◉ In the vast majority of cases, you want to filter by parameters/metadata that allow you a concrete and well-spun retrieval, filling your context with unnecessary information can lead to misleading results. #RAG #LLM #context #AI #vectordb #innovation #genAI #generativeai

Insightful panel discussion on "Governance in the age of Generative AI" by Digital trade and Data Governance Hub at The George Washington University. The topics includes - Varied sources of data used in training/finetuning LLMs - Synthetic data for generative AI - Impact of data openness - A view on opensource and closed LLMs - Role of governments on governance of LLMs - Discussion on shared data governance #datagovernance #artificialintelligence #conference #aigovernance #responsibleai #data