Rahul Chakkara

It took me a while to understand why LLMs are still really bad at displaying text on an image... Anthropic has released the first ever detailed look inside a modern, production-grade large language model. It helped me understand it a bit more, and why it still can't create an amazing image that includes my business name 'INSIGHT7' Might help, might put you to sleep... #insights #ecommerce #supplychain #analytics #datascience https://lnkd.in/gAQwaa5s

8

I have always argued that the obsession with Artificial General Intelligence (AGI) is misguided. If you have several narrow AI models that are very good at coordinating work, it will do pretty much the same for a large number of applications. This is exactly what OpenAI's 'Swarm Framework' is doing. Read more about this perspective: https://lnkd.in/eDSceJWG #aiinnovation #openai #aialignment #agi #artificialgeneralintelligence

1

‘Sensational’ Proof Delivers New Insights Into Prime Numbers https://lnkd.in/eCC6aPJ6 Two particular points I'd like to highlight: (1) the authors' use of a "gambit", which presumably has much wider applicability; (2) is the question of existence of zeros with real part 3/4 set on a path similar to the question of existence of odd perfect numbers?

2

Paul Adamson is a colleague from the UK who has started a promising podcast about the impact of #AI in IT. I was honoured to be the third guest on his show. We discussed about how we, in Spain, build up our own toolchain, bottom up, to be used alongside ChatGPT and Github Copilot to implement AI Augmented Software Engineering. Furthermore we discussed the wider impact of AI on Software Engineering and the Industry. Recommended. If not for this episode, then at least for the other two. #GenAI #leadership #mentoring #globalagiledevops #GetTheFutureYouWant #softwarengineering

23

Really proud of our CEO Sarah Cardell, who has been named one of Time Magazine’s 100 Most Influential People in AI. Under Sarah’s leadership, the CMA has invested in the technical skills and specialist expertise needed to really get to grips with these fast-moving and complex markets. We acted early to build our understanding and evidence base, through in-depth research and broad stakeholder engagement. And we’re taking an impartial, evidence-led, and timely approach to reviewing a small number of deals which could be problematic in competition terms for UK consumers and businesses. As with all the CMA’s work, the goal is the best possible outcomes for people, businesses and the UK economy. It’s inspiring to work with Sarah and other hugely talented and committed colleagues, like Karen Croxson, Joel Bamford and Will Hayter, helping ensure AI delivers on its promise of transformative benefits for everyone. Competition and Markets Authority #AI #foundationmodels #artificialintelligence #competition #innovation https://lnkd.in/eBzaGEF7

33

Very intriguing to observe the surge in work by commercial organizations utilizing scientific research techniques, with a significant portion of the latest projects being shared on arxiv to expedite the publication process. One recent study stands out for its focus on evaluating AI outcomes, including the AI's capacity to summarize extensive bodies of work and introducing a metric for positional bias. The study's approach of utilizing cost analysis to compare results is commendable. By opting for a model with slightly lower correlation but substantial cost savings, they underscore the current belief that AI solutions excel in enhancing efficiency rather than replacing human expertise. #AI #ArtificialIntelligence #CostAnalysis

8

Making sense of the UK government’s AI framework – Part 2 The Generative AI Framework for HMG raises all the questions you need to ask when implementing AI in any organisation and particular a public sector body. However, it does not provide answers meaning that we are likely to find different bodies answering question in different ways. I’ll look at specific issues in further videos, but to give you one example that covers technology and governance: ·      “Put in place evaluation and auditing structures, tracking data provenance, design decisions, training scenarios and processes.” Source: Generative AI Framework for HMG > Ethics > Transparency and explainability > Practical recommendations. https://lnkd.in/eDXvWrKf What are these structures? Are they different from existing ones e.g. doesn’t the Service Manual and Central Digital and Data Office (CDDO) have sufficient structures for evaluating design decisions?

3

1 Comment

In 2025 we will see a leap in voice-enabled AI. Most language models only support text. The biggest obstacle to voice AI is achieving human conversational speed. LLMs need to respond in around 200 milliseconds for this to happen. Amazon, Apple, Facebook, Google and other specialised providers are in a race to develop these models. They will drive innovation in industries such as Education, Financial Services, Health, Public Services and Retail. The benefits will include improved accessibility, more natural interactions and hands-free use. The early adopters will be able to provide services that are more personalised and responsive. Voice AI is the new lever to transform customer experience.

7

1 Comment

https://lnkd.in/eWK3ZVsM RAGs have become the standard architecture for providing LLMs with context in order to avoid hallucinations. However even RAGs can suffer from hallucination, as is often the case when the retrieval fails to retrieve sufficient context or even retrieves irrelevant context that is then weaved into the LLM’s response.

4

Alot of academics have also known this for decades. Research from Emily M. Bender and Melanie Mitchell show the risks of focusing on language 'form' without meaning. Cognitive science innovations in RRG Linguistics and brain science, Patom Theory, combine to link 'semantics to syntax' for implementation on machines. The human brain's ability as a baby to acquire their 'mother tongue' shows the blueprint to capture this law of nature, that is language. Let's use it for the trustworthy *and* last interface we will ever need. #patomtheory #RRGLinguistics

4

"LLM Adoption in the Enterprise - A Guide to Building Meaningful Products with Generative AI", very interesting report authored by Isabelle Nguyen, and published in March 2024 by O'Reilly. 👇 This report discusses barriers to #AI adoption among thought #leaders, outlining over five chapters how #AI can drive unprecedented #innovation. It addresses misconceptions, explores common applications, examines the product mindset required for #AI deployment, details the development cycle with a use case, and explains transitioning to production. Link: https://lnkd.in/dCqWfu5V #ailaw #machinelearning #deeplearning #artificialera #web3 #web4 #web5 #fintech #digitalera #digitaltransformation #generativeAI #trustworthyAI #explainableAI #aiethics #artificialintelligence #explainableAI #responsibleAI #foundationmodels #GPAI #GAI #LLMs #NLPs #GenAI #frontierAI #narrowAI #AGI #ethicalAI #adversarialAI

44

𝐓𝐡𝐞 𝐌𝐢𝐧𝐢𝐬 And 𝐓𝐡𝐞 𝐌𝐚𝐱𝐢𝐬 In an age where bigger often means better, the trend of adopting smaller, more efficient language models—Mini GPTs and the Small Language Models—stands out as a remarkable shift. The prevailing wisdom might suggest that higher parameters lead to higher generalisation, but the current landscape of AI research and application is proving otherwise. 𝐓𝐡𝐞 𝐑𝐢𝐬𝐞 𝐨𝐟 𝐌𝐢𝐧𝐢 𝐆𝐏𝐓𝐬 These Minis, typically ranging between 10 to 50 billion parameters, are not just a compromise—they're a strategic choice. 𝐒𝐨, 𝐓𝐡𝐞 𝐁𝐞𝐧𝐞𝐟𝐢𝐭 ? 𝐄𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲 Mini GPTs require less computational power, making them more accessible for a wider range of applications. 𝘛𝘩𝘦𝘺 𝘤𝘢𝘯 𝘣𝘦 𝘥𝘦𝘱𝘭𝘰𝘺𝘦𝘥 𝘰𝘯 𝘦𝘥𝘨𝘦 𝘥𝘦𝘷𝘪𝘤𝘦𝘴, 𝘦𝘯𝘢𝘣𝘭𝘪𝘯𝘨 𝘳𝘦𝘢𝘭-𝘵𝘪𝘮𝘦 𝘱𝘳𝘰𝘤𝘦𝘴𝘴𝘪𝘯𝘨 𝘢𝘯𝘥 𝘳𝘦𝘥𝘶𝘤𝘪𝘯𝘨 𝘵𝘩𝘦 𝘳𝘦𝘭𝘪𝘢𝘯𝘤𝘦 𝘰𝘯 𝘤𝘦𝘯𝘵𝘳𝘢𝘭𝘪𝘻𝘦𝘥 𝘤𝘭𝘰𝘶𝘥 𝘪𝘯𝘧𝘳𝘢𝘴𝘵𝘳𝘶𝘤𝘵𝘶𝘳𝘦. 𝐂𝐨𝐬𝐭-𝐄𝐟𝐟𝐞𝐜𝐭𝐢𝐯𝐞𝐧𝐞𝐬𝐬 Training and running large models is expensive. Some early estimates - a Mini GPT can reduce costs by up to 80%, making advanced AI capabilities more affordable for smaller enterprises and startups. 𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 𝐎𝐩𝐭𝐢𝐦𝐢𝐳𝐚𝐭𝐢𝐨𝐧 Despite their size, these smaller models can be fine-tuned to excel in specific tasks. 𝘙𝘦𝘴𝘦𝘢𝘳𝘤𝘩 𝘧𝘳𝘰𝘮 𝘚𝘵𝘢𝘯𝘧𝘰𝘳𝘥 𝘜𝘯𝘪𝘷𝘦𝘳𝘴𝘪𝘵𝘺 𝘥𝘦𝘮𝘰𝘯𝘴𝘵𝘳𝘢𝘵𝘦𝘴 𝘵𝘩𝘢𝘵 𝘢 𝘸𝘦𝘭𝘭-𝘵𝘳𝘢𝘪𝘯𝘦𝘥 𝘔𝘪𝘯𝘪 𝘎𝘗𝘛 𝘤𝘢𝘯 𝘰𝘶𝘵𝘱𝘦𝘳𝘧𝘰𝘳𝘮 𝘪𝘵𝘴 𝘭𝘢𝘳𝘨𝘦𝘳 𝘤𝘰𝘶𝘯𝘵𝘦𝘳𝘱𝘢𝘳𝘵𝘴 𝘪𝘯 𝘥𝘰𝘮𝘢𝘪𝘯-𝘴𝘱𝘦𝘤𝘪𝘧𝘪𝘤 𝘢𝘱𝘱𝘭𝘪𝘤𝘢𝘵𝘪𝘰𝘯𝘴, 𝘢𝘤𝘩𝘪𝘦𝘷𝘪𝘯𝘨 𝘶𝘱 𝘵𝘰 90% 𝘰𝘧 𝘵𝘩𝘦 𝘢𝘤𝘤𝘶𝘳𝘢𝘤𝘺 𝘸𝘪𝘵𝘩 𝘰𝘯𝘭𝘺 𝘢 𝘧𝘳𝘢𝘤𝘵𝘪𝘰𝘯 𝘰𝘧 𝘵𝘩𝘦 𝘳𝘦𝘴𝘰𝘶𝘳𝘤𝘦𝘴. 𝐍𝐨𝐰 𝐬𝐨𝐦𝐞 𝐧𝐮𝐦𝐛𝐞𝐫𝐬 According to a report from OpenAI the deployment of Mini GPTs has grown by 65% over the past year, indicating a robust adoption curve. - Businesses leveraging these smaller models have reported a 30% increase in operational efficiency due to faster processing times and reduced infrastructure costs. - In the healthcare sector, Mini GPTs have been instrumental in accelerating research and development, with some models reducing drug discovery times by up to 40%. The Minis are not just a fleeting trend; they are a testament to the evolving understanding of how AI can be harnessed most effectively. #minigpt #minillm #llminnovation

3

The value of any model is its ability to increase understanding of a phenomenon. The #GartnerHypeCycle is only partially helpful at explaining the evolution of #GenAI. But this conceptual approach from Drew Breunig helps navigate what he terms 'the ocean of use cases. He splits them into 3 categories: 👑 Gods - the human replacement use cases (AGI). Will impact us in 'hype and hand-wringing' and will 'affect funding, regulation, and politics' 👷‍♀️ Interns - focused on the grunt work, and are used and supervised by experts ⚙ Cogs - reliable components, complete one task, and are enabled by data Check out the article - worth a careful read https://lnkd.in/eRSUnSUq #genai #usecases #digitalmedia

15

3 Comments

Encouraging article from Tim Smith at Sifted regarding why AI is thriving in Brexit Britain. Key takeaways include: 🙌 A great talent pool and strong research universities 🏦 Access to a world-leading financial services sector 📜 Less strict regulatory stance on AI compared to the EU 💡 Encouraging innovation while keeping on top of potential risks 🌍 Prepared to champion the work of overseas companies publicly 🗣 Shared business language of English removes potential friction This article came off the back of last week's news that Wayve raised $1B from Softbank, Microsoft and Nvidia, making it the new record round for a European AI company! It will be interesting to see if this flourishing AI ecosystem can buck the trend of British tech companies reaching a certain size/maturity before transitioning to US ownership/equity markets... #executiveleadership #talent #ai

16

The point I was (poorly) making here, is as follows: AI is ushering in a monumental shift in human evolution, transforming our lives in unforeseeable ways. Popular films such as Terminator, Matrix, and I, Robot have depicted scenarios where AI becomes sentient and perceives humans as existential threats. We all know how this plays out... However, another significant concern is the concentration of AI control in the hands of a few, with major companies like Amazon and Nvidia owning much of the critical hardware and infrastructure required to operate AI. A proposed solution to this centralisation risk is the use of blockchain through Decentralised Physical Infrastructure (DePIn). DePIn describes a network of hardware - storage, sensors, processors, etc... with control and monetisation distributed across a blockchain - democratising the control of AI's core components. Although this makes it more difficult to shut down AI if needed (i.e. if it goes rogue), it reduces the risk of a few powerful individuals making unilateral decisions about humanity’s fate. What's worse? The human element or AI turning on us? And do you think AI is something to be celebrated or feared? Or is it both? Only time will tell...

17

4 Comments

🚨 The Entropy Barrier: The Hidden Cost of AI Supremacy 🚨 A chilling reality is emerging in the world of Large Language Models (LLMs). Recent research has unveiled a stark truth: L ≈ (3.1×10^8 / C)^0.050 This isn't just another equation. It's the financial doomsday clock for AI innovation. 💰 Every 20x increase in compute yields just a 1 nat decrease in loss. 🏦 The cost? Potentially billions in infrastructure and energy. 🔓 The reward? Inching ever closer to the entropy of natural language itself. But here's the terrifying part: As we approach this fundamental limit, the costs skyrocket exponentially. Only the tech giants and nation-states may be able to foot the bill. Are we witnessing the birth of an AI oligopoly? The implications are staggering: • Small players could be priced out of the market • Innovation might stagnate in the hands of a few • The democratization of AI? A distant dream As we race towards the singularity, one question looms large: Who will control the future of artificial intelligence? The clock is ticking. The entropy barrier awaits. And the cost of admission is rising by the second. Are you prepared for the coming AI apocalypse? #AIApocalypse #EntropyBarrier #FutureOfAI #TechOligarchy

12

7 Comments

🚨 The Entropy Barrier: The Hidden Cost of AI Supremacy 🚨 A chilling reality is emerging in the world of Large Language Models (LLMs). Recent research has unveiled a stark truth: L ≈ (3.1×10^8 / C)^0.050 This isn't just another equation. It's the financial doomsday clock for AI innovation. 💰 Every 20x increase in compute yields just a 1 nat decrease in loss. 🏦 The cost? Potentially billions in infrastructure and energy. 🔓 The reward? Inching ever closer to the entropy of natural language itself. But here's the terrifying part: As we approach this fundamental limit, the costs skyrocket exponentially. Only the tech giants and nation-states may be able to foot the bill. Are we witnessing the birth of an AI oligopoly? The implications are staggering: • Small players could be priced out of the market • Innovation might stagnate in the hands of a few • The democratization of AI? A distant dream As we race towards the singularity, one question looms large: Who will control the future of artificial intelligence? The clock is ticking. The entropy barrier awaits. And the cost of admission is rising by the second. Are you prepared for the coming AI apocalypse? #AIApocalypse #EntropyBarrier #FutureOfAI #TechOligarchy

3

2 Comments

UK FPTP (first past the post) electoral system vulnerabilities. Electoral Calculus is a well-regarded polling / MRP outfit in the UK. The image is their latest prediction of the most likely outcomes of the upcoming GE. It has the Tories on 23% of the vote, with a central prediction of 66 seats out of 650 (so about 10%) but with a massive range of uncertainty - anywhere between 37 and 225 of the seats. The fact that it's potentially so volatile, in terms of seat outcome, is apparently based on uncertainty as to the precise impact of Reform UK (Farage)'s 12% on sapping the Conservative vote. While in some ways I wouldn't exactly weep at a Conservative wipeout, the fact that the difference in seat outcome is *so much* more volatile than the likely overall % changes does seem to illustrate the vulnerability of FPTP as a voting system. Whichever party it is, I don't think one party winning 519 seats out of 650 on around half of the vote is a good outcome in a Parliamentary system, especially not the 'winner takes all' version which we have in the UK. This sort of system weaknesshas long been ignored when the two main parties still have representation in the 100s of seats, as it's only affected the 'lesser' parties and the two big ones seem to long for complete power without the tedium of coalition. I wrote a piece a while back noting that we've seen in microcosm already where things could end up, in the form of 2015 general election in Scotland, in which the SNP won over 90% (56 out of 59) Scottish seats in the House of Commons on just under 50% of the vote. Fortunately the existence of the Scottish Parliament and its proportional representation system kept the other parties alive and able to recover. But there won't be that safeguard in this case. And yes, of course the big parties are themselves coalitions in a certain sense, but privateish debates within Labour or Conservative factions are no substitute for public debates fronted by different parties, whatever one may think of the quality sometimes. Links: EC's main table - https://lnkd.in/etWPkX_T EC's April 2024 voter migration analysis - https://lnkd.in/ey44AZRX EC's June 2024 commentary (before today's announcement by Farage) - https://lnkd.in/eUaydqD6 2015 UK GE in Scotland - https://lnkd.in/eAUTzr27

15

19 Comments