Asankhaya Sharma

Can we trust machines to judge machines when it comes to LLMs? There's an ongoing battle between manual and automated evaluation of LLMs. Manual evaluation is predominantly used by many companies, especially large ones with sufficient budgets. This approach involves internal QA teams or external consultants who manually create test cases and evaluate the outputs in spreadsheets. This process is time-consuming, but can provide insights into: - model's performance - specific failure modes such as hallucinations - tone discrepancies - biases - copyright violations - privacy issues Smaller companies without such budgets often resort to using expensive engineering time to manually review outputs, which can be inefficient. The primary advantage of manual evaluation is the accuracy and reliability of human judgment. But it's not scalable and introduces human biases, which can affect the consistency of the evaluations. Also manual processes cannot keep up with the rapid development cycles and the volume of outputs generated by modern LLMs. Automated evaluation aims to reduce the reliance on human labor by employing metrics and models to assess LLM performance. Techniques used in automated evaluation include: - Metrics-Based Evaluation: It utilizes established metrics such as BLEU score, ROUGE score, and perplexity. These metrics help assess aspects like fluency, accuracy, and relevance. But these metrics often do not correlate well with human judgments and can be inadequate for comprehensive evaluation. - Heuristic-Based Functions: These are specific functions to test particular capabilities or behaviors of the LLM e.g. toxicity detection or PII (Personally Identifiable Information) detection. These approaches can be tailored to specific use cases but may lack generalizability. - Model-Based Evaluation: It uses another trusted language model to evaluate the outputs of the target LLM. This can automate the evaluation process but requires careful calibration to ensure reliability. Automating this process can be a huge unlock. But developers attempting to create their own automated evaluation systems report very low confidence in the results. Lot of opportunity on this front.

3

1 Comment

Super proud of Mike and his team. Many people asked me over the last year why long context and reasoning are so important. My answer is simple: long context is, in other words, a much better memory that allows LLM to utilize its ability to comprehend incoming text fully which as we all know is a million miles better than RAG. A very clear use case is with processing huge documentation like corporate wiki. Now, the deal with reasoning over long context is also simple: if models can get an ability to process long context but do that poorly then what's the value is in doing? And that's why a proper benchmark like BABILong is so desperately needed. It helps you to open your eyes on how inefficent the majority of the modern methods for processing long context by LLMs are. It helps you to understand what the real goal is. Let's make LLMs understand long context better together!

18

2 Comments

1/ As part of the #Ursa announcement, we have made the #Kafka API a first-class citizen in our data streaming platform. While the #Pulsar API is excellent as a unified API for combining queuing and streaming, the Kafka API will remain popular for data streaming. Elevating the Kafka protocol to a first-class citizen and embracing it as an open protocol is a continuation of our technology vision of enabling data sharing across teams, departments, and organizations. I have elaborated on this in our blog post: https://lnkd.in/gB9tp2gJ 2/ Our vision and openness do not stop at just supporting the Kafka API. We extend it by embracing the #KafkaConnect API. As Jay Kreps mentioned in his tweets, Kafka Connect has thousands of connectors, and hundreds of ISVs develop for it. It is integrated into dozens of commercial services, all of which contribute to it. https://lnkd.in/grJXjvUE 3/ While we have built the Pulsar I/O framework for #ApachePulsar, we have not attempted to replicate Kafka Connect and create a fragmented ecosystem. Instead, we leverage the Kafka ecosystem as much as possible. For example, we have a Kafka Connect adapter that allows our users and customers to run any Kafka Connectors on #ApachePulsar. In fact, all the Pulsar #Debezium connectors are built in this way. 4/ We have taken the Kafka Connect adapter work one step further by supporting Kafka Connect natively in our product via a Unified Connector Runtime. This eliminates the need to adapt Kafka Connectors as Pulsar I/O connectors. We can natively run any Kafka Connector on StreamNative Cloud via this runtime. We believe in Kafka Connect as an API for integration. As a vendor, we can build a differentiated cloud product that delivers user experience, performance, reliability, and many other added values through innovations in runtime and platform levels in a significantly better way. 5/ The Kafka Connect & Unified Connector Runtime is currently in early access and will be released soon. If you are interested in working with us as a design partner, we would love to receive your feedback early on. You can reach out to Neng Lu, Kundan Vyas, or myself. If you are an ISV, we would love to partner with you on your Kafka Connect connectors. Let’s talk! 6/ Finally, we believe that vendor innovation can occur at the runtime and platform levels to deliver a faster, cheaper, and better product without introducing restrictive licensing. For example, the secret behind supporting the Unified Connector Runtime is a flexible runtime powered by Pulsar Functions and the Function Mesh. Therefore, "To Bento or not to be Benthos" is no longer a question, as a Benthos or Bento pipeline can also be run as a function on StreamNative. If you would like to see us support Benthos or Bento on StreamNative, show your support by liking or sharing the post, and we will make it a native feature. 

32

We are excited to back Morphing Machines. As Dr. Ranjani, CTO and co-founder says: “For the first time in India and globally, we have a processor that is truly runtime reconfigurable, catering to emerging workloads.” While Nvidia’s growth and high valuation are driven by robust demand for “infra-compute” in high-performance GPUs, spurred by Gen AI, there is a distinct need for innovative solutions tailored to a new wave of reactive applications. Enter Morphing Machines. It addresses this demand with its unique product designed for next-gen telecom, autonomous vehicles, collaborative robots, avionics, and online data analytics. This product is specifically engineered to handle heterogeneous workloads and diverse tasks with varying performance demands simultaneously, ensuring a high level of scalability. As we move from an institutionally-backed research team at IISC to productization, Golden Sparrow is collaborating with Speciale Invest and others to support Deepak Shapeti, who joined as CEO three years ago after earning his Masters at Stanford. He brings not only technical expertise but also the business acumen necessary to guide the product to commercial success. Venkat Raju Afreed Faizan Anand Rao Ritik Singh Michael Marmor

68

5 Comments

How many r’s in strawberry? Came across a bunch of AI threads in the past day discussing this topic. The debate varied far and wide from is AI going to replace jobs, using the correct “technical” prompt, comparisons of LLMs to deep technical discussions on how the LLMs actually work. People debating these threads include people from various walks of life, executives, users, engineers and scientists. Many opinions and facts. My fascination about anything technology boils down to two questions. Reliable use cases (“where” and “how” can I really use this) and Risk of failure (“when” will this not work?). These principles apply to any “tool” that our kind (species) has ever come about (invented, discovered etc.) including what we are calling “AI”. The answer to “how” we use anything lies somewhere within what I term the “where-when” spectrum. Coming back to AI, the marker we choose in the “where-when” spectrum decides everything from financial gain (or obliteration), risk management to ethics. This marker is bound to progress with time unless it becomes cost prohibitive. For now, we need safeguards (systemic / manual interventions) to ensure we stay within boundaries that keep the equation positive. The responsibility to ensure AI remains a value added proposition is in each of our hands. Let’s be responsible and put safeguards in place so we are not bitten (and bitten bad)! I plan to share more information about this. If you have ideas or questions please leave a comment. If you are in my network you know where to reach me to chat up! #AI #ResponsibleAI #Innovation #Risk

12

Most of the enterprise search solutions are backed by the Lucene ecosystem (Elasticsearch, Solr). Given the recent boom in AI technologies, there seems to be a shift in new-gen search implementations. New search frameworks are leveraging embeddings, vector databases and other related technologies. Shifting to these new suits of technologies would be a costly affair for well-established enterprises that have already invested heavily in their existing ecosystem. While Lucene has expertise in text-based search, embedding models unlock the semantic search capabilities. This unlocks the potential of free-flowing search and more relevant results. However, introducing more tech pieces means more complexity in the system. Fortunately, recent Lucene releases now support vector search with improved vector dimensions. Similarly, Vespa also introduced vector search a couple of years ago with improved dimension support in recent times. So organisations using the Lucene/Vespa ecosystem and looking to improve search with semantic capabilities might achieve that without moving out of their existing ecosystem or introducing more complexity in the system. #search #lucene #vespa #ai #embeddings #openai #semanticsearch

61

Catching up on some of the great presentations from #PEPR24 presentations, and came across Damien Desfontaines’ talk on synthetic data. I obviously disagree with lots of it, but I think Damien generally makes his points in good faith. So I’m going to (try to) do the same in response. The part of his talk that stood out to me as most obviously incorrect and uncharitable was the discussion of the incentives for improved privacy measurements in synthetic data. I’ve included his graphic below as a summary of his argument that there’s no commercial incentive to improve empirical privacy evaluations of synthetic data for context. Sharing three thoughts: 1) If there’s no commercial incentive for better privacy evaluations with synthetic data, differential privacy companies shouldn’t exist. The complexity and form factor of differential privacy imposes enormous incremental friction to configure and use relative to synthetic data. If there’s no commercial case for better synthetic data privacy metrics where none of that friction exists, then how could there possibly be commercial demand for more friction and better privacy metrics — the asserted value of differential privacy? As far as I know, Tumult Labs isn’t a charity, so someone there must believe there’s a market for stronger privacy guarantees — we agree with that, even if Damien doesn’t 😂 2) That’s the theoretical argument, but I’m a synthetic data guy, so I’ll make an empirical argument, too: prospects and customers care deeply about the rigor of privacy tests for synthetic data and it is a buying criteria in *every* deal in our funnel. For *privacy* use cases (more on this in the next paragraph), buyers are not just asking for better metrics; they’re demanding it. My data on this is MUCH better than Damien’s — since he doesn’t actually sell synthetic data to customers — and he’s completely wrong. 3) Despite their website copy, most synthetic data companies are delivering to non-privacy use cases *at least as often* as they’re delivering to privacy-enhancing analysis use cases. Synthetic data is a really useful technology for data quality work, test data management, and anonymization. And in practice, most synthetic data platforms are most heavily used for data quality work (imputing NULLs, class rebalancing, dataset augmentation, etc.). Privacy is — at best — a secondary criteria for these use cases, so the basic privacy metrics are sufficient. At Subsalt, we work exclusively on anonymization use cases and we win deals on the basis of our privacy work. We’re living proof that there’s a business case for more rigorous privacy work in synthetic data. I don’t expect this to solve the eternal (and cursed) debate about empirical vs. theoretical privacy metrics, but I do think it’s worth clarifying that if you want to know about the structure of a market — in this case the market for synthetic data — you should listen to someone that knows at least something about that market.

13

5 Comments

To start, I have to concede that the quote from our website (yes, Subsalt is the anonymous company in Damien Desfontaines’ post; I don’t think he even tried that hard to make it anonymous 😂) overstates the real risks of re-identification in synthetic data. The copy will be changed, because it actually is a good example of sloppy privacy messaging. Mea culpa. But despite being a pretty good dunk, it’s not outcome determinative on the issue we’re actually discussing. The relevant question is: is there market demand for more rigorous privacy metrics in the synthetic data space. The answer is unambiguously “yes.” Customers are not satisfied and they justifiably don’t believe the most commonly used metrics add up to “anonymization.” So why haven’t most synthetic data vendors adjusted their “privacy scores” to accommodate this dissatisfaction? The most commonly referenced companies in synthetic data go to market with freemium versions of their product that can be adopted and used by end-users without a multi-stakeholder purchasing process. The users that actually adopt these tools tend to use them disproportionately for *non-privacy* use cases. And more stringent privacy tests are not neutral to the UX for these users: they’re a source of friction and complexity. So the versions of these synthetic data products you can try off-the-shelf stick to basic privacy metrics that don’t introduce friction. The dirty little secret in synthetic data is that VERY few synthetic data workloads in production are premised on sufficient data anonymization through synthesis. That’s because synthetic data deals where anonymization is the primary use case have traditionally died in the risk, privacy, legal, and compliance departments *because those stakeholders want more rigorous privacy guarantees* that conflict with the needs of the non-privacy users of synthetic data platforms. If every large organization on earth had deployed a synthetic data platform for privacy use cases in spite of this weakness, you’d be able to conclude that there’s a lack of demand for more rigorous privacy work in the space. But all of the synthetic data vendors (combined) have a relatively tiny penetration into the market. The biggest source of today’s “non-consumption” (e.g. the fully uncaptured portion of the addressable market) is the misalignment between the standard synthetic data form-factor and privacy measurements and buyer demand for anonymization. Despite the sloppy copy on our website, that is the latent demand we hear from prospects and customers every day and we take the problem very seriously. I’d be happy to share more about our privacy metrics and approach, because I share the deep concerns and commitment to providing rigorous privacy preserving tools so organizations can do more with sensitive and regulated data — even if I don’t share the skepticism of all empirical privacy metrics.

9

2 Comments

LinkedIn team (Juan Pablo Bottaro Karthik Ramgopal) shared a helpful deep dive into their deployment of an LLM-powered job search assistant. Highly recommend this read! Few highlights:  💡 Relatively easy to setup RAG-based pipeline. Achieved 80% of the basic experience in 1 month, then 4 more months to reach 95% mark. 💡 Evaluating the quality of application outputs is hard. I hear this from many enterprises developing LLM applications. (check out Okareo by Matthew Wyman Boris Selitser) 💡 Balancing latency with quality is important. While certain techniques like Chain of Thought could improve accuracy and reduce hallucinations, it will increase response times, which can affect user experience. https://lnkd.in/gtGGwf5P #largelanguagemodels #RAG #EBR #LLM #ml #ai #hallucinations #evaluations #

4

Mathematical Explanation of How DLSFH Addresses Perturbative Nonrenormalizability @SSRN The Dodecahedron Linear String Field Hypothesis (DLSFH) leverages string theory's approach to circumvent the issue of perturbative non-renormalizability in gravity. https://lnkd.in/dRKX-Mtx

2

Measuring intelligence is like gauging the spiciness of a ghost pepper with a ruler, it's a tricky business. Claims of AI being "10,000 times smarter" than humans usually refer to specific abilities, things as processing massive datasets or performing complex calculations. It's akin to comparing a calculator to the human brain. The calculator excels at math but lacks the nuanced "smartness" we associate with humans. Intelligence is multi-faceted and resists quantification. AI might be a whiz at crunching numbers, but we're still holding the castle when it comes to understanding the nuances of human emotions and context. And that's not because we invented the castle in the first place! So, the next time you hear AI touted as superior to humans, take it with a grain of salt. It's like declaring a car 10,000 times faster than a human, true in a narrow sense, but it doesn't capture the full picture. The caveat? AI is evolving. While today it might be a specialized tool, a decade from now, it could be as versatile and adaptable as the human mind. It's a possibility worth pondering in light of unprecedented technological advancement. The world would never be the same. https://lnkd.in/g4CRhkBX

7

1 Comment

Interesting insight from Santosh Rao on Marketplace by APM My research talks about metrics Tech/SaaS Leaders should measure depending on your scale https://lnkd.in/gGfBWgn5

4

When the Universe speaks, you're smart if you listen. I've received a lot of PMs in the past few weeks, and more than half asked for practical guidance re causal analytics, both for GTM and more broadly, as well as an explanation of terms, likely issues to be thinking about, and all the kind of real life stuff that is the difference between Concept and Reality. Two people went further, asking me to show them how to operationalize a lot of the technical principles that Dale W. Harrison outlines in his newsletter, which is sponsored by ProofAnalytics.ai. That's a great idea. So....there are a lot of places we could start, but let's begin with what the Business cares about most, whether from GTM or other functions: Effectiveness Materiality Efficiency Sustainability 1. #Effectiveness is defined as "the quality of results derived from the money, time, and effort invested to attain those results." For example, given the fact that Marketing is a nonlinear multiplier of Sales effectiveness, Marketing is considered effective when it multiplies the effectiveness and efficiency of Sales beyond what Sales could have attained by itself. Analytically, Effectiveness is about the causal effects your investments had on a desired outcome, considering time lag and external headwinds and tailwinds. 2. #Materiality identifies the significance of the issues or factors to a company's business. Materiality is a sort of "1b" derivative of effectiveness. This is not about effectiveness, per se, but rather how crucial X is to the success of the business.  Materiality is a relative idea commonly expressed as a stack rank. 3. #Efficiency is a ratio between the value of X and what the resources needed to produce it. Before you can calibrate the efficiency of something, you first have to determine how effective / valuable it is. Once something is understood to be effective and significant to the forward momentum of the business, the goal should be to match spent to results more and more advantageously to the company. This is not just about cutting the cost; it is about spending the right amount of money to extract the full value of something that's clearly effective. 4. For your GTM efforts to be #Sustainable means that they have to generate enough cashflow into the company to justify the cash needed to pay for all of it. This is important, as you can be very effective and be unsustainable, just as a company can be growing and still go out of business because it ran out of cash. ROI, CAC, LTV, etc are not effectiveness metrics. Your C-suite knows this, so when you emphasize efficiency metrics to imply your effectiveness, there's a big hole in your argument. Tomorrow, I'll be on the Hard Skill Exchange debate on "GTM Productivity v. Performance" with Mark Ogne of Symplexity.AI, Mariana Cogan of Hexagon Manufacturing Intelligence, Christian Idiodi of Silicon Valley Product Group, and Jen Anderson of Contento. Looking forward to sharing and learning!

79

18 Comments

Episode 9 is out ! Data Science Dialogues Full Pod on : https://lnkd.in/ggs9YhF5 Ft. Saurabh Sharma and Amber Guglani Powered by : Data Science Pathshala In this pod we talked about : 📊 Data Science Journey 🤔 Perception Among Colleagues 🌟 What Sparked Your Interest in Data Science? 🚗 Gen AI's Impact on Automobile Automation 💼 Assessing Gen AI's Workforce Efficiency 🧠 Challenges Presented by Gen AI 💰 Investing in Gen AI for Startups 🔮 The Future Landscape of Gen AI in 10 Years 🛡️ Regulating Gen AI for Public Safety 🔄 Addressing Scalability with Gen AI 📚 Learning Pathways for Gen AI

13

Connecting Founders and Funders At ODIGITAL Startup Drivers, we are committed to democratizing access for everyone to invest in early-stage, technology-driven startups. We know this is the prime asset class of the financially savvy and ultra-rich for creating financial abundance and wealth. We believe this is also the best asset class for outperforming galloping inflation and local currency weakness. We understand that the more of our citizens who become equity owners of these high-growth ventures, the more people who are uplifted to higher realms of financial abundance and wealth in our communities. To achieve our mission, we are dedicated to connecting visionary and resourceful technology startup founders with funders who wish to support their growth and scaling initiatives. We have partnered with Tekedia Capital to give everyone exclusive access to co-invest a minimum of $250 in their investment deals. We recently closed a funding round on May 31st with 10 of the world's best investment deals, many of which came to us at very attractive discounts. These were all high-potential ventures poised to deliver exponential returns on investment at future exits. These startups are from diverse sectors, such as Artificial Intelligence (AI), Fintech, Cybersecurity, Proptech, Manufacturingtech, and others. These businesses operate in different markets such as the USA, Canada, Nigeria, Benin Republic, Congo, with 3 YCombinator companies included. We believe the best time to invest in technology startups was 20 years ago, and the next best time is now. Nvidia is the dominant supplier of artificial intelligence (AI) hardware and software, and it has seen its stock price soar by over 3,500% in the past 5 years, with many shareholders becoming billionaires along the ride to a current market cap of $3.34 trillion. Chip-maker Nvidia became the world’s most valuable semiconductor company after its share price climbed to an all-time high on June 18th, 2024. It's share price has nearly doubled since the start of this year, and it is now more valuable than Microsoft and Apple. The company's meteoric rise has been fueled by its dominance of the chips needed for artificial intelligence (AI). In the accompanying video, NVIDIA founder and CEO Jensen Huang discusses his leadership style of torturing his team members into greatness. If you are a tech startup founder looking for funding or an investor who wants to co-invest with us in our upcoming investment cycle in Sept./Oct. 2024, hurry and connect with me via any of the options listed below for more information. * Send me a direct message * Email - invest@startupdrivers.com * Click - https://lnkd.in/dHd8K9ZN - to join our WhatsApp group * Visit - https://startupdrivers.com Join Us. Let’s Win Together. Kindly repost for the benefit of your network, and follow me for more posts like this if they resonate with you.

2

Great to see this demo of EnterpriseGPT from an Upekkha batchmate Dilip Ittyera . EnterpriseGPT addresses the most important concern with GenAI - Hallucinations. Recently an Airline was used by its customer because the AI came up its own refund policies which they couldn't offer. EnterpriseGPT solves this by promising 3 things Accuracy - You always get the right answer Reliability - You always get the same answer for the same question Explainability - You always get the source and references for your answer What this means is that your AI is not going to cook-up important information like - Company policies & Terms of service. It's a great alternative for scenarios where you run the risk of heavy law suits when you share the wrong information. If you're thinking of adding AI into your support or client facing workflows, get in touch with Dilip

6

An interesting article caught my attention, India has less than 2,000 senior engineers who can build core AI products. In the race to equip India's workforce with AI skills, there's an evident paradox. On one hand, we witness technology giants like Tata Consultancy Services, Infosys, Accenture, and Microsoft undertaking massive upskilling initiatives, training hundreds of thousands of employees. Yet, on the other hand, we confront a stark reality—a severe shortage of seasoned engineers capable of building core #AI products. This dissonance begs the question: are we investing in skill acquisition without ensuring its effective utilization? The numbers might be impressive, but they mask a deeper issue—the gap between training and practical application. While the focus on mass skilling is commendable, it's imperative to ask whether we're fostering an ecosystem conducive to #innovation and #entrepreneurship. India's future prosperity hinges not merely on acquiring technical competencies but on cultivating a mindset of relentless innovation and problem-solving. We must transition from a culture of routine employment to one that champions #creativity and #disruption. This requires a fundamental shift in our approach—a departure from the traditional 9-to-5 mindset towards a culture that celebrates risk-taking, experimentation, and learning from failure. Absolutely, the need for this paradigm shift is most critical at the top level. Here, bureaucratic tendencies often hinder execution and decision-making. What's required is a shift towards a results-oriented, mission-driven approach, where leaders prioritize #action and drive change effectively; even if require to make some hard decisions. At this pivotal juncture, India cannot afford to rest on its laurels. The AI revolution demands not just skilled workers but visionary #leaders who can harness technology's potential to drive meaningful change. It's time to bridge the gap between skill acquisition and real-world application, empowering a new generation of Innovation pioneers to chart the course for India's technological prowess on the global stage. Let's seize this opportunity to shape a future where innovation knows no bounds. Let's build together, lets contribute our best of efforts. https://lnkd.in/gHfjGQEt #AI #Innovation #Entrepreneurship #Startup #WorkForce #IndiaAI

33

1 Comment

As LLM usage and development has exploded, evaluation frameworks and platforms have quickly become a critical component of the development pipeline. As you might expect, one scalable, effective approach to evaluating LLMs is also by leveraging other LLMs. One such LLM evaluation model is the newly announced PROMETHEUS 2—an open-source language model that sets a new standard in evaluating other language models. Evaluation LLMs like PROMETHEUS 2 are generally assessed based on their ability to prove a high correlation between human ratings and those of state of the art LLMs such as GPT-4. Based on this latest paper, it appears that PROMETHEUS 2 excels at this task, and can be tailored to specific evaluation criteria beyond general attributes like helpfulness and harmlessness. Some key highlights of PROMETHEUS 2 include: - High correlation and agreement with human evaluators and proprietary LLM judges on multiple benchmarks. - Unified evaluation capabilities for direct assessment and pairwise ranking. - Developed using an innovative weight merging technique, combining models trained on different formats for robust performance. You can find this paper here: https://lnkd.in/dG5PwEFT What are some other open source tools that you used to evaluate LLMs for quality, performance and safety/responsibility? Woojin Kim Khan Siddiqui, MD MD Rajiv Shah HOPPR #AI #MachineLearning #OpenSource #LLMs #AIResearch #Prometheus2 #Evaluation #AICommunity

28

2 Comments

Interesting moves on the AI front - https://lnkd.in/gh6CG9js

11