Ian Ozsvald

To all those fellow film buffs and data nuttheads out there, Letterboxd just dropped a new "Stats" feature which displays descriptive statistics for each of your film lists, no idea if this is only available to paid subscriptions, but deff give it a try if you like data. The variables they chose to display and how they decided to do it is actually spot-on, and I'm just guessing this will improve over time. While going through the report for one of my lists, I came across this very elegant approach at genre decomposition: Themes & Nanogenres. - Themes: A single short sentence describing an idea or topic. This is excellent, and their descriptors appear to be on point. This is also what Goodreads has been lacking since it's inception. - Nanogenres: Single keywords that descend into the genre hierarchical structure. Granular-enough, very very nice. Knowing that you can export Letterboxd data onto structured formats, there are some interesting data science projects waiting to be done there with this new feature. I'll be 100% doing that whenever I have some time. Append /stats to your list's URL in order to display the report. According to the algorithm, more than half of the films in my WIP top 20 list have "Surreal and thought-provoking visions of life and death" as the main Theme (thought-provoking), Liv Ullmann is my favorite actress (yep, no objection there), and Luis Buñuel, Andrei Tarkovsky, and Ingmar Bergman form my holly triad (no surprise there as well). I keep telling people that the 60's was THE BEST time to be alive for cinema, but they sometimes dismiss me as annoying. Thanks Letterboxd for building such an amazing app! This is the single service where I'm genuinely proud to be a Patron.

2

2 Comments

Yet another indicator about why LLMs are not databases. They are non-deterministic (in other words, they are stochastic), meaning that even with the same prompts and contexts, different host GPUs can change the output of the LLM in different ways. Even with RAG, replicability is not guaranteed. The underlying message here? LLMs value as generators is unquestionable, and it is a good tool for natural language prompts. However, LLMs likely will never be reliable enough to serve as databases, precisely because they are chaos machines: they are extraordinarily dependent upon starting conditions. https://lnkd.in/gBDei_ye

276

35 Comments

Need for diverse sources of #data, like #RWE or real world evidence, was a topic in the #AI panel at Patient-Centered Outcomes Research Institute (PCORI) meeting yesterday. With reports of “model collapse”, we need more data. AI “eating” AI-generated data leads to bad AI. Next week will be speaking at GovAISummit on data lakes and hope to make the case for user-generated data for patient-centered outcomes as part of how we define #quality in #healthcare #datascience.

5

1 Comment

What I think is coming to light is that fine-tuning models involves a trade-off roughly equivalent to attempting to build detours around established routes within the models. Detours often put stress on routes that are less ideal to describe implicit connections, and that very frequently end up causing travellers (i.e., the traversal network) into areas that aren't actually that well connected to most of the graph in order insert to the new functionality for the fine-tuning. RAG, similar, bypasses the dataset entirely, except for final elucidation of concepts. You can think of fine-tuning as adding technical debt - you make the knowledge base more specialized at the risk of losing generalized knowledge. This is another reason why AGI may be further off than we want to admit - specialists become specialists primarily by deliberately eschewing learning things in other areas to devote their attention to one particular problem domain. I'm pretty good at being a systems architect, but I could not tell you squat about plumbing or surgery beyond analogies I can make from what I know.

7

One of the best blogs I've come across on quantization, with outstanding visualizations. Highly recommend a read for anyone looking to understand LLM quantization in more depth! 📐 - https://lnkd.in/e_WTgq6x At Felafax AI (YC S24), we are building Infra to make full precision multi-GPU fine-tuning easier. #AI #Infra #LLM

16

3 Comments

A recent insight from Trustible's latest newsletter edition (https://lnkd.in/eBPVePBR) Anthropic's Claude Sonnet 3.5 can now control computers by interpreting screen images and issuing commands to move the mouse and type. This is an important step in Agentic AI and aligns with similar projects by Google (browser control) and Apple Intelligence (enhanced iPhone AI). As a consumer: this is really cool. But what are the new and existing risks that organizations should be aware of with AI-enabled computers? 🔐 𝐃𝐚𝐭𝐚 𝐏𝐫𝐢𝐯𝐚𝐜𝐲: Screen captures could expose sensitive data. Anthropic says user-submitted data is never used to train AI and Apple mitigates this with mostly on-device processing. But unclear for future tools. 💻 𝐄𝐱𝐜𝐞𝐬𝐬𝐢𝐯𝐞 𝐀𝐠𝐞𝐧𝐜𝐲: AI could gain extensive control over a computer, necessitating action limits (e.g., restricting Claude from changing system settings). ⚠️ 𝐏𝐫𝐨𝐦𝐩𝐭 𝐈𝐧𝐣𝐞𝐜𝐭𝐢𝐨𝐧 𝐕𝐮𝐥𝐧𝐞𝐫𝐚𝐛𝐢𝐥𝐢𝐭𝐲: Bad actors could embed malicious prompts in web content, guiding the AI to unintended actions. 🧑‍🏫️ 𝐄𝐝𝐮𝐜𝐚𝐭𝐢𝐨𝐧 𝐆𝐚𝐩: Current AI Agents require some coding ability and typically go through a technical review process before they are deployed. In contrast, “computer use” AI may be widely accessible, raising risks without adequate user awareness. AI that can control a computer (and other general-purpose AI Agents) are likely to become more common over the next year. 𝐓𝐫𝐮𝐬𝐭𝐢𝐛𝐥𝐞 can help orgs understand the risks of emerging AI capabilities so that you can adopt them responsibly and securely.

14

1 Comment

What's "Model Collapse" in AI? Glad you asked... A fascinating study published in Nature this week presents findings which seem to answer a question I've been pondering for a while... As more and more AI-generated content gets published on the web, it will form a growing proportion of the data sets on which future generations of AI models are trained. That's bound to have some implications, right? Right. The study found that after a few generations, AI models start to forget original sources and deprioritise more marginal ideas and perspectives. As they become increasingly self-referential, they tend towards a bland centre-ground that fails to represent the whole spectrum of thought: "We discover that indiscriminately learning from data produced by other models causes ‘model collapse’—a degenerative process whereby, over time, models forget the true underlying data distribution... In other words, the use of LLMs at scale to publish content on the Internet will pollute the collection of data to train their successors... Being trained on polluted data, they then mis-perceive reality." This will be an interesting challenge for AI model developers to resolve. And it should give those of us who work with content (even more) pause for thought about how much AI-generated content we dump on the web. Read the full paper here: https://lnkd.in/ecAgEnbZ #AI #content #contentmarketing

10

3 Comments

We've just launched a major new feature Ballpark called Data Tables - a powerful way browse, analyse and export your research responses. https://lnkd.in/eCf3Uwxw Why spend time building a table?! • As our customers grew, so did the volume of responses gathered for surveys and usability tests, going from tens of responses, to hundreds and now thousands. The only way to parse this effectively was to export to a spreadsheet and do your work there, which sucked. We added the most used functionality of a spreadsheet to Data Tables. • Customisation of the data is super important - everyone has different ways they want to view information. Data Tables allow you to show/hide columns and export based on the view you've set up. • Performance is supercharged, thousands of responses load within a second Now live on every plan https://ballparkhq.com/ #userresearch

55

5 Comments

Unlike many other MMMs, we publish all of our model results into our platform so that our clients can go out and validate them. We’ll never hide these results. But what happens if the output of our MMM doesn’t match the results from a client’s lift tests? First, an important point – to know whether your model is or isn’t working, you need to be able to validate it. This is hugely important to us, and it’s why we’ve made backtest results available in the Recast app itself so our customers can actually look at them and come to us if something looks off. Because let’s be honest, MMM is an extraordinarily difficult statistical challenge. And while I think Recast does far and away the best across the MMM industry, there are still times that we might miss. If we over/under-predict one time, the model will adjust so that we don’t make the same mistake twice. Now, if backtests are off systematically, that’s an indication that something may be wrong in the configuration of the model – and our customer data science team and our modeling team will go in and take a harder look. This is almost always case by case, but there are a couple of things that we generally look for first: - Is there something the model doesn't know about? I remember a client once had raised prices… but didn’t tell us. Once we added a pricing context variable into the model that problem went away. - Is the model just flat-out wrong? Maybe we aren’t capturing seasonality well, maybe channels are hitting their upper or lower bound… it can be a number of things. This requires that we go into model investigation mode in order to get the model into a good place – this is obviously very important to us so, in the few instances where this happens, it becomes a top priority to get it right!

10

(Gift link) - Nice oped from Martin Casado and Katherine Boyle at a16z VC on the blatant bias for Big Tech in D.C. on all things AI. I've never seen such government favoritism for incumbent tech companies in my 40+ year career. It's frankly obscene, self-destructive behavior. It may be some part exploitation of ignorance on the part of Big Techs (you've been duped), it may be some part faulty ideology, but I don't think there is any question it's also political corruption. It's one of the most extreme examples of government capture we've ever seen. To clarify my position -- I don't think it was necessary to even mention open source in this op-ed as it muddies the water with two issues, even if related (I support some open source, though in LLMs it does have security considerations -- all LLMs do). I'm also not against all regulation in AI, but I am fiercely opposed to regulation that favors dominant monopoly incumbents and/or new ventures with multiple billion dollar valuations that don't even have a product. As I repeat often, the super majority of problems and risks that need regulating in AI are caused by LLM firms and their Big Tech partners. That's where the focus should be, and since they are conflicted, the government shouldn't be relying on them for advice like they are, or internal AI advisers who came from them who are on the payroll of non-profits funded by tech billionaires with a myriad of conflicts. It's ugly. Bottom line: While the U.S. has a serious problem with the Big Tech oligopoly and their oligarchs, we are not quite yet an oligarchy, but we're getting very close to meeting that definition. We better fix it before it's too late. ~~~~~ Quote: Although the public-facing argument for AI regulation is to promote safety, we believe the true purpose is to suppress open-source innovation and deter competitive startups. Many emerging technology companies are well-equipped to help solve national-security challenges, but of the 22 members on the board, none represent startups, or what we call “little tech.” Only two are private companies, and the smallest organization on the board hovers around $1 billion in value. The AI companies selected for the board either are among the world’s largest companies or have received significant funding from those companies, and all are public advocates for stronger regulations on AI models.

1

Can the UK public sector use AI? Yes, but pubic bodies have to follow the Generative AI Framework for HMG: https://lnkd.in/eDXvWrKf The issue is this is a very long document with a huge number of links out to other documents. On the one hand one might argue that effectively all it says is – follow hmg policy for implementing IT. On the other hand – there’s really nowhere else that this diverse and overlapping set of policies have been set out in one place. In the next videos I’ll dive into the Principles, Ethical themes and how from a practical perspective the document raises more questions than it answers. #ai #ethics #publicsectorai

7

The future of Coding / Software development is undergoing a paradigm change. Fast and Furious. And Effective?

2

Positive AI/QA moves from the UK government: They see AI Assurance as a key area for productive and responsible AI use. https://lnkd.in/ekZxfg59 #AI #AIQA #AIAssurance

6

Just some tidbits on what has been happening in #AI in the UK over the last couple of weeks Healx: Cambridge-based Healx has raised $47m (£37m) in a Series C round after its neurofibromatosis treatment received regulatory approval. Founded in 2014, Healx uses AI for rare disease treatment discovery, having raised over $100m, including a $56m Series B in 2019. Expect hiring in engineering and AI/ML roles. Retrograde: Entrepreneur Grace Beverley launched Retrograde, an AI-powered talent management system for content creators. Co-founded with Jake Browne and Gary Meehan, Retrograde offers an AI agent for tasks like inbox management and deal negotiation, taking a 10% service fee. It's an early-stage company with high growth potential. Ocula Technologies: Belfast-based startup Ocula has raised £3.25m to boost e-commerce ROI using AI. Founded in 2021, Ocula works with retailers like Boots and AO. Its AI tool, Ocula Boost, improves product page SEO. They're hiring for roles in London and Belfast, using a Python tech stack on GCP. Aveni Scottish fintech Aveni secured £11m from Lloyds Bank and Nationwide. Founded in 2018, Aveni uses LLMs for financial services, including NLP systems for customer conversation analysis. Hiring in Scotland and England for sales and engineering roles, with a Python, Node, and Typescript stack. Fractile: London-based AI chip startup Fractile emerged with a £12m seed round. Founded in 2022, Fractile aims to challenge Nvidia by improving chip performance for AI inference using in-memory compute technology, claiming to be 100x faster. Hiring for ML compiler and modeling engineers. Hey Savi: UK fashion search platform Hey Savi raised £2.2m in pre-seed funding. Founded by an all-female team, Hey Savi uses AI for personalized fashion shopping. Currently hiring a head of engineering to build out the tech function. Edit Cloud: London-based Edit Cloud raised £2m for post-production software. Founded in 2022 by Simon Green, the startup aims to streamline post-production collaboration with cloud-based workflow tools. Hiring likely as the team grows. Xapien: AI due diligence startup Xapien raised £8m in Series A funding led by YFM Equity Partners. Founded in 2018, Xapien’s platform conducts due diligence checks quickly. Current hiring includes CS and marketing roles. CultureAI: Manchester-based CultureAI raised $10m (£7.7m) to detect employees' risky cyber behavior. Founded in 2018, the startup's software helps IT and cybersecurity teams. Hiring technical and sales roles in London and Manchester. Iconic AI: London-based Iconic AI closed a $4m (£3m) pre-seed round for AI-driven video game development. Founded in 2023 by ex-Snapchat engineer John Lusty, Iconic AI aims to expedite game development. Hiring game-specific animation, design, and engineering roles. #technews

10