Cascade Strategy

There is a recurrent thread going on in tech circles that contrasts the philosophical approach of companies that have adopted the lean startup practice of rapid customer development cycles, and pivoting with your learnings, vs companies that are thesis-driven, have a big mission and are relentless in pursuing it, bending reality to their will — using examples of product visionaries like Jobs, Gates and Musk. 1) 2) One very recent example (at the time of writing): The conversation has degraded into a love/hate discussion around companies doing experimentation, as if being rigorous about your hypotheses and attempting to invalidate your riskiest assumptions somehow precludes you from having a vision of the future and how you’ll change it. I’ll try to address some of the confusion around that in another post. This one, I’ll use to structure how a Product Manager should think about that future, build out his product vision in material terms, turn it into a product strategy and then pursue that strategy with tactical actions that enable the real product shape to emerge from both convictions at a macro scale, and experimentation at a micro-scale. What is Strategy? Good books and practitioners of product strategy have tried to outline what strategy is, and how it translates into actionable practice over the past few years. Thinking about strategy is not new, of course, but its application to the internet economy and digital products has really been defined more clearly over the past 10 years, and books such as Good Strategy / Bad Strategy and, in particular, 7 Powers — recommended by some of the best founders and investors in the industry - have helped mold this collective knowledge base through amazing synthesis. Others, like , , , or , write, teach and distill product strategy through their writing, analogies and analysis. Let’s define strategy through some of their insights: Ben Thompson Tren Griffin Web Smith Gibson Biddle Ryan Singer “Good works by focusing energy and resources on one, or a very few, pivotal objectives whose accomplishment will lead to a cascade of favorable outcomes” — Good Strategy / Bad Strategy strategy “How will your product customers, in , ways?” — Gibson Biddle delight hard to copy margin-enhancing The set of conditions creating the potential for persistent differential returns. a route to continuing Power in significant markets.” — 7 Powers “Power: Strategy: The ideas of strategy and power are directly linked to the fact that, once you’ve achieved it, you’ve established a competitive advantage that can be defined by both the economic benefit that it accrues to your company, as well as the barriers that it imposes on competitors and new entrants. In other words, it creates a . Moats are usually applied to product companies when they’ve established sufficient differentiation that it appears insurmountable for a competitor to cross the gap competitively, given the current market landscape. Google’s dominance in is one such example of a moat (everywhere but China and Russia). Anyone trying to make a dent on web searches would have to go head-on with G’s technological dominance, brand power, distribution advantages (such as the 9B$ premium it pays for ). For any competitor to disrupt Google in this space, a significant market transformation would have to occur (one such as the explosion of decentralized internet content that led us to the need for Google in the first place). moat search being the default search engine on iOS It’s important to note that Google turned into Alphabet, which is a dominant conglomerate of internet companies, and has probably overgrown its original mission to “organise the world’s information and make it universally accessible and useful”. Their vision gave way to new visions, and their strategy has evolved alongside them. It has emerged from their thesis, through a culture of experimentation at a massive scale. So while it is clear that the company has changed drastically throughout its lifetime, it is also relatively obvious, in hindsight, that a similarly large and world-changing company would be unlikely to originate from a different strategy. Thesis-driven vs emergent — a false dichotomy So before we move on to the 7 powers lenses on Strategy and Power, and how they can be applied to your product development, a picture is starting to be formed as to how working out your strategy and developing a thesis for how the world (or your company) will be different when you fulfill your larger, more ambitious goals, does not change the fact that you’ll pack a lot of assumptions into your problem, solution and execution spaces, that you can invalidate through rapid and rigorous experimentation at a local level. The outcomes of your experiments — which are ways to de-risk your tactical bets and measure the compatibility of your views and the market’s — coupled with your conviction and relentless pursuit of your thesis, will drive your company forward. (Side notes on emergence) On the importance of uncertainty and emergence, Alex Danco notes, in his newsletter: Two Truths and a Take First, there’s the problem of emergence. If you start from biology, it’s pretty easy to work your way backwards by reducing biology into its component parts — you’ll find chemistry and physics as its underlying components, with nothing unaccounted for. But it’s hard to go the other way. Life is a complicated, adaptive, messy thing. If you start with physics and chemistry, it’s hard to identify what exactly is the thing in here that gives rise to life. We say that life is an “emergent” phenomenon, which lets us dodge the question of “emerges through what, exactly?” This talks to the relative ease of working from an consummation of strategy and boiling it down to its essential tactics and experiments, but also to how hard it is to predict that exact set of steps, and what they would lead us to achieve, had we not have run those experiments (including the many that failed). ex-post The largest players in the industry report in their experiments, which, arguably, could be even worse / more catastrophic if they weren’t able to measure them. 10–30% rates of success Writing it down identifies one key tenet of great strategy has the fact that it must begin by: Good Strategy / Bad Strategy clearly identifying the current obstacle in the way to obtaining strategic power, set out a guiding policy (the guardrails and overall direction of your strategy) and establishing a set of coherent actions, which together, will allow us to overcome the obstacle. This exercise is very often missing at the highest levels of Product leadership. Product people, when they can think of a better version of the world, also may think of a series of tactics (typically designed to fit in a slide deck) that can be used to kickstart the product engine, but there’s usually lack of good directional policy — one that allows individuals to make informed trade-offs — and even of coherent and reinforcing actions, which, being successful would compound to let us overcome the obstacles on our way to market power. Amazon’s product function (in effect guided by their executive shaman’s communication style — ), seems to have recognized both the human inability to clearly visualize and describe the changes we want to produce in the future world once our product succeeds, but also how we fail to outline and critically analyze our guardrails and tactical actions in both timely and effective manners. see his ’97 shareholder letter This has translated to the way the company does product briefs, through their use of 6-pagers as product briefs and the use of the . This gives PMs and Execs the latitude to explain the product in detail and examine its key assumptions and risks, outlining a plan to execute against them. What key hypotheses they have, how they coordinate and compound, how they address the obstacles in their way, and how they’ll eventually lead to a moat. Working Backwards methodology The purpose of these documents is to lay out your product thesis, with a market and customer-centric voice, and then backtrack all the way to how they’ll play out in policy and tactics in order to make you win. This means crystallizing your vision of the future in such a way that it is actionable, but that leaves enough latitude for regular failure, of a non-catastrophic nature, while in pursuit of that vision. That is good product definition. (For more examples of good product definition, definitely check out Basecamp’s and Ryan Singer’s attempt to synthesize product shaping). Shape Up NEXT UP: 7 Powers In the next post, we’ll address the , described in the book, by Hamilton Helmer, which gives us a framework to think about the key types of strategic powers (aka ), their sources, the benefits they accrue to their holders, and the barriers they impose on challengers. Product leaders can use the framing of the 7 powers of strategy to design their policies and tactics, formulating an approach to achieve their goals. 7 strategic powers Moats

Good Strategy, Bad Vision? Product Management Paradoxes — Part I

Being at the forefront of automation, why is Amazon downplaying the impact of AI and robotics? Breakthroughs in machine learning (ML) and robotics are drastically enabling a faster and cheaper means of delivery and manufacturing. We all want our packages instantly at lower costs and companies want higher margins. But what’s the overall impact on our society as we automate more and more tasks in factories and warehouses? Among the world’s largest companies, Amazon, in particular, has faced heightened scrutiny. The e-commerce giant has been the major driving force behind the now widespread reality of warehouse automation and same-day delivery, pushing its peers and suppliers like DHL to test a range of robotic technology just to keep up. Amazon was criticized for pressing workers to physical limits while continuously automating jobs, impacting 125,000 full-time hourly associates and an additional 120,000 seasonal workers across its logistics and warehouse centers in the United States. Against the backdrop of public scrutiny, we see Amazon delaying, rather than encouraging, more candid and transparent discussions of how we as a society transition. During a recent press tour, Scott Anderson, director of Amazon Robotics Fulfillment said that . More specifically, he thinks that it would take at least 10 years to automate the processing of a single order picked by a worker inside a warehouse. Is this fact or a publicity ploy to distract the public from an urgently needed discussion of how we transition and get ready for this inevitable massive trend? fully automated workstations in warehouses are at least 10 years away What Amazon Got Right While we may still be years away from “fully” automating every step of complete orders in warehouses, certain stages of orders are already being automated, and there will be a cascade of automation during the next few years. In other words, the technology is already here to automate enough to be efficient and cost-effective. Despite all the fancy videos flowing around the internet showcasing automated warehouses, the reality is that only . Sophisticated automation technologies such as sorters, AGVs(automatic ground vehicles), and palletizers often require over $50m investment with an estimated payback period greater than five years. over 10% of warehouses in the U.S. are using sophisticated automation technologies In addition, while mobile robots like Kiva and Fetch are currently in use, robotic arms used for piece picking are only just beginning to deploy. Why? As mentioned in , most robotic arms are powerful but blind. They are programmed to perform the same tasks repetitively. Even small changes could pause the production line for recalibration. They can’t adapt to changes in the environment and are not nearly as dexterous as human arms. my previous article Compared to robots, humans are still much more flexible and have a better visual perception as well as gripping capabilities (We still haven’t invented any gripper that is nearly as dexterous as human fingers). Furthermore, humans don’t use low-level languages to describe seemingly intuitive tasks like grasping. While most of us can easily pick and place various objects without even thinking about it, robots need to learn to grasp from scratch as they lack crucial context including physics like friction and center of gravity. (That’s why we can’t simply apply current metrics for human to robotic grasping. I will explain more about this in my next post.) The picking process can be really complex, ranging from zone picking, discrete picking, all the way to batch picking. It also involves several steps including receiving, sorting, palletizing, depalletizing, picking, checking and packing, labeling, and shipping. Many of these steps have been automated but we are still years away from fully automating this process. Amazon is right in highlighting the limitations of current AI and robotics technology compared to the superior cognitive and manipulation ability of humans. However, there are important trends that Amazon hopes you miss. What Amazon Hopes You Missed isn’t needed. Simply automating part of the order assembly is sufficient to automate many jobs. Labor costs account for an average of . Given that whopping number, it’s not hard to understand why budget-conscious warehouses are driven to automate wherever they can along the process. That’s also why the warehouse robotics market is forecasted to exceed USD 4 billion by 2022 with a CAGR of 11.8% according to MarketsandMarkets. The rapidly growing market demands have attracted not only robotics companies like ABB, Fanuc, and KUKA but also agile robot maker, Boston Dynamics, who recently acquired California based startup Kinema. Total automation (level 5) 65% of most warehouse facilities’ operating budgets Robots are also getting smarter. A good example is piece picking, which has been for many years. Now startups like Osaro and PlusOne are deploying machine learning-enabled robot arms that can recognize and manipulate a wide range of objects in warehouses. Even if the robots can only deal with 80% of the work and people still need to handle the rest of edge cases, this 80/20 strategy of automation to human intervention already ensures companies can reduce significant costs on labor and system integration. the holy grail of robotics Additionally, over time, the algorithm will improve and get closer to full autonomy as these collect more data and learn from trial and error. The recent development in cloud robotics, meta-learning, and transfer-learning also accelerates the process of robots learning and adapting to changes in the environment. level 4 autonomous robotic systems As mentioned in my , the market is growing rapidly. According to Tractica, worldwide sales of warehouse automation technology are expected to increase from $8.3 billion in 2018 to $30.8 billion in 2022. Industry giants including robotics makers (e.g. ABB, KUKA, Fanuc), material handling companies (e.g. Muratec, Dematic), retailers (e.g. Amazon, Ocado) are heavily investing in AI and robotics technology by building in house teams as well as working with startups that disrupt the industry with innovative solutions. previous post High-density storage system startups like Attabotics and Commonsense Robotics emerged as retailers demand faster and more efficient delivery solutions. The needs for AI-enabled robotics extend beyond warehouses. As suggests, an estimated $766 billion total wages in the US are spent for predictable physical work, which is more likely to be automated by robots. a recent Mckinsey report The top three markets where the most predictable physical work resides are: Accommodation and food Manufacturing Transportation and warehousing ) (source: Mckinsey report According to the US Bureau of Labor Statistics, the number of warehouse workers grew from 662k in 2009 to over 1.18m in 2019 with a CAGR of 5.96%. If the growth continues, there will be over 2.1m warehouse workers by the year 2029. Even if it really takes Amazon and other retailers 10 years to “fully automate every step of order”, that implies that more than two million jobs being automated in the next decade. Is 10 years a long time? Is it enough time for millions of workers and for the society to get ready for such a drastic shift? What we do know for sure is that millions of people will be impacted if we continue to deny the inevitable future of automation and waste the last chance we have to get prepared. Whether discontinuing its “Amazon picking challenge” or putting out articles downplaying the pace of automation, Amazon seems to be trying everything it can to avoid public scrutiny. There is no doubt that repetitive jobs will be replaced by automation, as they should be. Amazon knows this better than anyone else. Why would the company be planning to spend $700m over the next six years to retrain a third of its U.S workers, if they are not anticipating large job cuts in the near future? Is this too little too late? Can big companies including Amazon take a much more active role in making their plans for automation more transparent and beneficial to all parties? If we do it right, this new wave of automation can result in a much brighter future where companies will enjoy higher productivity and employees will be freed from repetitive and physically demanding jobs and have better working conditions. A recent Harvard study found that than their employers give them credit for. People are not afraid of automation but they are definitely fearing uncertainty. workers are more eager to embrace change and learn new skills Instead of wrongly assuming that workers don’t want to change and therefore soft-pedaling plans for automation, companies should work with their employees to transition to working with robots in jobs that are dangerous or repetitive and that workers would rather not do by themselves. This will help people get higher-skilled jobs with better wages in our future economy. Click here if you want to see more articles like this! . Bastiane Huang is a Product Manager at OSARO, a San Francisco based startup building software-defined robotics. She has worked for Amazon in its Alexa group and with Harvard Business Review as well as the university’s Future of Work Initiative. She writes about ML, robotics, and product management. Follow her here

What Amazon Doesn’t Want You to Know About Its Automation Strategy

The Snowflake Hack and Its Domino Effect

স্নোফ্লেক হ্যাক এবং এর ডমিনো প্রভাব

Der Snowflake-Hack und sein Dominoeffekt

El truco del copo de nieve y su efecto dominó

Le Snowflake Hack et son effet domino

स्नोफ्लेक हैक और इसका डोमिनो प्रभाव

スノーフレークハッキングとそのドミノ効果

눈송이 해킹과 도미노 효과

O truque do floco de neve e seu efeito dominó

Взлом Snowflake и его эффект домино

Kar Tanesi Hilesi ve Domino Etkisi

Hack bông tuyết và hiệu ứng Domino của nó

雪花黑客及其多米诺骨牌效应

Authors:
(1) Ronen Eldan, Microsoft Research (email: roneneldan@microsoft.com);
(2) Mark Russinovich, Microsoft Azure and Both authors contributed equally to this work, (email: mark.russinovich@microsoft.com). Table of Links Abstract and Introduction
Description of our technique
Evaluation methodology
Results
Conclusion, Acknowledgment, and References
Appendix 5 Conclusion The ambitious endeavor of teaching a Large Language Model (LLM) to selectively forget, or ”unlearn”, is a testament to the nuanced complexities inherent in the world of artificial intelligence and machine learning. Widely regarded as a daunting task, any attempt at enabling such a functionality in LLMs stands at the vanguard of innovative solutions, and in this light, our proof of concept arguably underscores progress. Firstly, our research demonstrates that unlearning, though challenging, is not an insurmountable task, as the positive outcomes in our experiments with the Llama2-7b model suggest. Yet, this achievement must be contextualized with prudence. Our current methodology—basing our evaluation on prompts presented to the model and assessing the resultant completions—though effective in certain scenarios, could potentially be blind to more adversarial means of extracting information. It’s conceivable that non-traditional or intricate methods, such as delving into token probability distributions, might inadvertently reveal the model’s latent familiarity with unlearned content. Diving deeper into the potential generality of our technique, a pertinent observation emerges when considering the unique attributes of the Harry Potter series. The books are replete with idiosyncratic expressions and distinctive names—traits that, in hindsight, may have abetted our unlearning strategy. The pronounced presence of Harry Potter themes across the training data of many LLMs further compounds the challenge. Given such widespread representation, even the slightest hint in a prompt might stir a cascade of related completions, underscoring the depth of memory ingrained in the model. A nuance of our methodology involves a reliance on GPT-4’s existing knowledge of the Harry Potter universe. To detect specific anchored terms and devise generic counterparts, the expertise of GPT-4 proved useful. This raises the question whether our technique achieve similar efficacy when stripped of such vast prior knowledge. Preliminary experiments show that entity extraction can still be effective when this knowledge is absent, and we speculate that the lack of familiarity with idiosyncratic expressions can be addressed with simple n-gram frequency analysis but we leave a more thorough study for future work. Extending our approach to other types of content, particularly non-fiction or textbooks, presents its own set of challenges. Unlike the fictional universe of Harry Potter, non-fiction content will not possess the same density of unique terms or phrases. Furthermore, non-fictional texts often embed higher-level constructs such as ideas, concepts, or cultural perspectives. It remains uncertain to what extent our technique can effectively address and unlearn these more abstract elements. This would clearly necessitate adaptations of our technique. In conclusion, while our technique offers a promising start, its applicability across various content types remains to be thoroughly tested. The presented approach offers a foundation, but further research is needed to refine and extend the methodology for broader unlearning tasks in LLMs. Acknowledgement The authors would like to thank Yanan Cai for helping to configure and manage the Azure GPU VMs used for this work. References [BHT+19] Yonatan Bisk, Ari Holtzman, Jesse Thomason, Jacob Andreas, Yoshua Bengio, Joyce Y Chai, Mirella Lapata, Angeliki Lazaridou, Ryan J Maynez, Piyush Narang, et al. Piqa: Reasoning about physical commonsense in natural language. arXiv preprint arXiv:1911.11641, 2019. [CLC+19] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2924–2936, 2019. [GTB+21] Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language model evaluation, September 2021. [JLZ+22] Yiwen Jiang, Shenglong Liu, Tao Zhao, Wei Li, and Xianzhou Gao. Machine unlearning survey. In Fifth International Conference on Mechatronics and Computer Technology Engineering (MCTE 2022), volume 12500, pages 1596–1603. SPIE, 2022. [JYY+22] Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. Knowledge unlearning for mitigating privacy risks in language models. arXiv preprint arXiv:2210.01504, 2022. [MCKS18] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. [NHN+22] Thanh Tam Nguyen, Thanh Trung Huynh, Phi Le Nguyen, Alan Wee-Chung Liew, Hongzhi Yin, and Quoc Viet Hung Nguyen. A survey of machine unlearning. arXiv preprint arXiv:2209.02299, 2022. [Row07] J.K. Rowling. The Harry Potter Series. Bloomsbury, 1997-2007. Comprising: Harry Potter and the Philosopher’s Stone; Harry Potter and the Chamber of Secrets; Harry Potter and the Prisoner of Azkaban; Harry Potter and the Goblet of Fire; Harry Potter and the Order of the Phoenix; Harry Potter and the Half-Blood Prince; Harry Potter and the Deathly Hallows. [SLBBC19] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. arXiv preprint arXiv:1907.10641, 2019. WCY+23] Lingzhi Wang, Tong Chen, Wei Yuan, Xingshan Zeng, Kam-Fai Wong, and Hongzhi Yin. Kga: A general machine unlearning framework based on knowledge gap alignment. arXiv preprint arXiv:2305.06535, 2023. [YBS19] Vikas Yadav, Steven Bethard, and Mihai Surdeanu. Quick and (not so) dirty: Unsupervised selection of justification sentences for multi-hop question answering. arXiv preprint arXiv:1911.07176, 2019. [ZFBH+23] Dawen Zhang, Pamela Finckenberg-Broman, Thong Hoang, Shidong Pan, Zhenchang Xing, Mark Staples, and Xiwei Xu. Right to be forgotten in the era of large language models: Implications, challenges, and solutions. arXiv preprint arXiv:2307.03941, 2023. [ZHB+19] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791–4800, 2019. [ZNIS23] Haibo Zhang, Toru Nakamura, Takamasa Isohara, and Kouichi Sakurai. A review on machine unlearning. SN Computer Science, 4(4):337, 2023. This paper is available on arxiv under CC 4.0 license. Authors: (1) Ronen Eldan, Microsoft Research (email: roneneldan@microsoft.com); (2) Mark Russinovich, Microsoft Azure and Both authors contributed equally to this work, (email: mark.russinovich@microsoft.com). Authors: Authors: (1) Ronen Eldan, Microsoft Research (email: roneneldan@microsoft.com); (2) Mark Russinovich, Microsoft Azure and Both authors contributed equally to this work, (email: mark.russinovich@microsoft.com). Table of Links Abstract and Introduction Description of our technique Evaluation methodology Results Conclusion, Acknowledgment, and References Appendix Abstract and Introduction Abstract and Introduction Description of our technique Description of our technique Evaluation methodology Evaluation methodology Results Results Conclusion, Acknowledgment, and References Conclusion, Acknowledgment, and References Appendix Appendix 5 Conclusion The ambitious endeavor of teaching a Large Language Model (LLM) to selectively forget, or ”unlearn”, is a testament to the nuanced complexities inherent in the world of artificial intelligence and machine learning. Widely regarded as a daunting task, any attempt at enabling such a functionality in LLMs stands at the vanguard of innovative solutions, and in this light, our proof of concept arguably underscores progress. Firstly, our research demonstrates that unlearning, though challenging, is not an insurmountable task, as the positive outcomes in our experiments with the Llama2-7b model suggest. Yet, this achievement must be contextualized with prudence. Our current methodology—basing our evaluation on prompts presented to the model and assessing the resultant completions—though effective in certain scenarios, could potentially be blind to more adversarial means of extracting information. It’s conceivable that non-traditional or intricate methods, such as delving into token probability distributions, might inadvertently reveal the model’s latent familiarity with unlearned content. Diving deeper into the potential generality of our technique, a pertinent observation emerges when considering the unique attributes of the Harry Potter series. The books are replete with idiosyncratic expressions and distinctive names—traits that, in hindsight, may have abetted our unlearning strategy. The pronounced presence of Harry Potter themes across the training data of many LLMs further compounds the challenge. Given such widespread representation, even the slightest hint in a prompt might stir a cascade of related completions, underscoring the depth of memory ingrained in the model. A nuance of our methodology involves a reliance on GPT-4’s existing knowledge of the Harry Potter universe. To detect specific anchored terms and devise generic counterparts, the expertise of GPT-4 proved useful. This raises the question whether our technique achieve similar efficacy when stripped of such vast prior knowledge. Preliminary experiments show that entity extraction can still be effective when this knowledge is absent, and we speculate that the lack of familiarity with idiosyncratic expressions can be addressed with simple n-gram frequency analysis but we leave a more thorough study for future work. Extending our approach to other types of content, particularly non-fiction or textbooks, presents its own set of challenges. Unlike the fictional universe of Harry Potter, non-fiction content will not possess the same density of unique terms or phrases. Furthermore, non-fictional texts often embed higher-level constructs such as ideas, concepts, or cultural perspectives. It remains uncertain to what extent our technique can effectively address and unlearn these more abstract elements. This would clearly necessitate adaptations of our technique. In conclusion, while our technique offers a promising start, its applicability across various content types remains to be thoroughly tested. The presented approach offers a foundation, but further research is needed to refine and extend the methodology for broader unlearning tasks in LLMs. Acknowledgement The authors would like to thank Yanan Cai for helping to configure and manage the Azure GPU VMs used for this work. References [BHT+19] Yonatan Bisk, Ari Holtzman, Jesse Thomason, Jacob Andreas, Yoshua Bengio, Joyce Y Chai, Mirella Lapata, Angeliki Lazaridou, Ryan J Maynez, Piyush Narang, et al. Piqa: Reasoning about physical commonsense in natural language. arXiv preprint arXiv:1911.11641, 2019. [CLC+19] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2924–2936, 2019. [GTB+21] Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language model evaluation, September 2021. [JLZ+22] Yiwen Jiang, Shenglong Liu, Tao Zhao, Wei Li, and Xianzhou Gao. Machine unlearning survey. In Fifth International Conference on Mechatronics and Computer Technology Engineering (MCTE 2022), volume 12500, pages 1596–1603. SPIE, 2022. [JYY+22] Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. Knowledge unlearning for mitigating privacy risks in language models. arXiv preprint arXiv:2210.01504, 2022. [MCKS18] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. [NHN+22] Thanh Tam Nguyen, Thanh Trung Huynh, Phi Le Nguyen, Alan Wee-Chung Liew, Hongzhi Yin, and Quoc Viet Hung Nguyen. A survey of machine unlearning. arXiv preprint arXiv:2209.02299, 2022. [Row07] J.K. Rowling. The Harry Potter Series. Bloomsbury, 1997-2007. Comprising: Harry Potter and the Philosopher’s Stone; Harry Potter and the Chamber of Secrets; Harry Potter and the Prisoner of Azkaban; Harry Potter and the Goblet of Fire; Harry Potter and the Order of the Phoenix; Harry Potter and the Half-Blood Prince; Harry Potter and the Deathly Hallows. [SLBBC19] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. arXiv preprint arXiv:1907.10641, 2019. WCY+23] Lingzhi Wang, Tong Chen, Wei Yuan, Xingshan Zeng, Kam-Fai Wong, and Hongzhi Yin. Kga: A general machine unlearning framework based on knowledge gap alignment. arXiv preprint arXiv:2305.06535, 2023. [YBS19] Vikas Yadav, Steven Bethard, and Mihai Surdeanu. Quick and (not so) dirty: Unsupervised selection of justification sentences for multi-hop question answering. arXiv preprint arXiv:1911.07176, 2019. [ZFBH+23] Dawen Zhang, Pamela Finckenberg-Broman, Thong Hoang, Shidong Pan, Zhenchang Xing, Mark Staples, and Xiwei Xu. Right to be forgotten in the era of large language models: Implications, challenges, and solutions. arXiv preprint arXiv:2307.03941, 2023. [ZHB+19] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791–4800, 2019. [ZNIS23] Haibo Zhang, Toru Nakamura, Takamasa Isohara, and Kouichi Sakurai. A review on machine unlearning. SN Computer Science, 4(4):337, 2023. This paper is available on arxiv under CC 4.0 license. This paper is available on arxiv under CC 4.0 license. available on arxiv

Who’s Harry Potter? Approximate Unlearning in LLMs: Conclusion, Acknowledgment, and References

Real-World Resilient Strategies for Fintech Projects

How Useful are Computational Models to Traders in Their Decision Making?

The System Design Cheat Sheet: Relational Databases (Part 1)

Let's learn about via these 53 free stories. They are ordered by most time reading created on HackerNoon. Visit the to find the most read stories about any technology. Machine Learning Uses /Learn Repo 1. Use Cascade Models to Get Better Speed and Accuracy in Computer Vision Tasks Great way to improve your Computer Vision models metrics 2. The Four Types of Machine Learning | Part 2 In the previous post, we saw the first two types of machine learning. In this post, we will discuss the other two types of machine learning. These are — Semi-su 3. The Convergence of AI + IOT is Imminent & Your Competitor is Already Preparing for It With the emergence of ever-cheaper and robust hardware, 5G connectivity around the corner, and most importantly, a growing list of real world use cases, we can all agree that IOT projects are here to stay. But is that where it ends ? 4. Adversarial Examples In Machine Learning Explained There are easy ways to build adversarial examples that can fool any deep learning model and create security issues no matter how complex the model is. 5. Reimagining Customer Behavior Through Machine Learning Here what you need to know about Machine Learning in business market strategy 6. Machine Learning: Santa’s little helper He’s making a list and checking it… well… once, because his model has a 98% accuracy 7. How AI Is Transforming The Field Of Education Artificial Intelligence (AI) is one of the wonders of the modern world, which isn’t going to cease to amaze the most intelligent of human beings. Like other fields, the field of education is also gaining the maximum amount of benefits out of the promises of the AI-powered technology. 8. Artificial Intelligence on Making Court Decisions In Estonia, artificial intelligence will play a crucial role in the future of human beings. It will play a crucial role in the functioning of judicial institutions, which is a great innovation. According to the authorities in this Baltic country, they are empowered to arbitrate cases of minor offenses autonomously. Unbelievable but true! 9. The GPUs for Deep Learning: NVIDIA vs AWS vs Azure and More Take a deeper dive into what a GPU is, when you should use it or shouldn’t for Deep Learning tasks, and what is the best GPU on-premises and in the cloud in 202 10. Debunking 4 Common Myths About Machine Learning Machine learning is a subset of artificial intelligence that involves the use of algorithms and statistical models to enable computers to improve them. 11. How Document Classification Can Improve Business Processes The process of labeling documents into categories based on the type of the content is known as document classification. It can also be defined as the process of assigning one or more classes or categories to a document (depending on the type of content) to make it easy to sort and manage images, texts, and videos. Document classification can be done using artificial intelligence, machine learning, and python. 12. How To Get Started With Machine Learning With The Right Mindset You got intrigued by the machine learning world and wanted to get started as soon as possible, read all the articles, watched all the videos, but still isn’t sure about where to start, welcome to the club. 13. How Companies Are Actually Using AI in Everyday Practices When thinking of AI (artificial intelligence), mixed emotions often come to mind. For movie buffs, we might immediately see images of Will Smith battling it out with humanoid AI creatures in IRobot or the even more realistic looking depiction of artificial intelligence in the movie aptly called AI. In our human minds, AI is something that could potentially lead to a catastrophic apocalypse as machines take over the world. 14. Every Way Natural Language is Better Than SQL Since the dawn of time, humans have communicated through gestures, drawings, smoke, or speech. Along the way, Structured Query Language (SQL) made its way into human life so we could speak to databases. However, it’s time to revert back to our natural language and rethink how we talk to our data. 15. How Machine Learning Can Help With Fraud Prevention Technology, without a doubt, has eased up a lot of issues for us, including the likes of fraud prevention. But before we start talking about the technological inputs pertaining to the same, it is necessary to understand why fraud prevention is actually required by businesses and why this genre of functionality is wide-spread and extremely popular. Firstly, financial firms are probably the most affected ones as fraudsters are actually interested in siphoning off money more than anything else. Secondly, fraudulent activities aren’t restricted to one vertical and it is a challenge for the firms to develop newer strategies for combating evolving threats. 16. Introduction to Different Machine Leaning Tools Machine Learning is one of the emerging technologies of the present IT industry. This technology has now become the talk of the town and has seen  an abnormally high growth over the few years. 17. Deploy Computer Vision Models with Triton Inference Server There are a lot of Machine Learning courses, and we are pretty good at modeling and improving our accuracy or other metrics. 18. Basic Use Cases of AI, ML, Deep Learning and Internet of Things The world’s most influential companies and technologies are influenced by the efficiency of Artificial intelligence and similar technologies. Whether it is Facebook or Amazon, Google or Microsoft, all firms are harnessing AI techniques and algorithms to introduce high-level performance and streamlined operations. 19. How is AI-powered Software Development influencing Pet Tech Technology is revolutionizing all aspects of our lives. Let's take a look at how AI-powered software development influencing pet tech. 20. An Introduction to AI-Based Visual Inspection for Defect Detection “Why should I care about a cool new technology until it’s solving any of my problems?” – this was the exact conversation I had with the executive of a water purification plant over a warm cup of coffee. 21. Announcing ModelDB 2.0 release Since we wrote ModelDB 1.0, a pioneering model versioning system, we have learned a lot and adapting it to the evolving ecosystem became a challenge. Hence we decided to rebuild from the ground up to support a model versioning system tailored to make ML development and deployment reliable, safe, and reproducible. 22. Most Useful ML Ops Applications Many ML Ops tools allow overseeing the entire machine learning model life cycle. Here are some of the most worthwhile ones to consider. 23. Vladimir Vapnik's New Learning Model Vladimir Vapnik recently gave a talk about a new theory of learning he is working on. 24. Evaluating Regression Models in Machine Learning Model evaluation is very important since we need to understand how well our model is performing. 25. Breaking into Deep Learning: Transforming the World Without Expert Input Deep learning is a subdivision of machine learning in which Artificial Neural Networks (ANNs) learn from a huge influx of data to produce high-quality output. 26. Model Tests are Essential for Building Domain Knowledge Testing protects against regressions. But the most important product of testing is domain knowledge and powerful capabilities. 27. Can Graph Neural Networks Solve Real-World Problems? In this article, we will learn about GNNs and its structure as well as its applications 28. The Future of Artificial Intelligence: To Kill or To Heal? The evil cyber-intelligence from the Matrix and a cyborg killing machine from the Terminator movies - that’s what most people used to imagine when talking about the future of artificial intelligence. 29. What’s the Difference Between AI And Machine Learning? It’s nearly impossible to have a conversation about technology without mentioning artificial intelligence (AI) or machine learning (ML). 30. Here’s How Libratus, a Poker-Playing AI, Bluffed Four Professional Texas Holdem Players The world of technology is changing at a faster rate than we can possibly fathom. Long gone are the days when we were the sole trailblazers in a human-tech relationship when the incentive resided in our hands. 31. What's the Difference Between MLOps and AIOps An overview of the MLOps and AIOps worlds to understand what they mean, how they relate to DevOps, and how they compare in terms of benefits. 32. Machine Learning Applications Across Different Industries 1952 witnessed the world’s first computer that could learn while it was running. It was a game of checkers developed by Arthur Samuel. It has barely been half a century since then and we’re already having a conversation about whether we should commercialize self-driving cars or not. Machine Learning gave birth to some of these great advancements in technology and we’re going to dive deep into everything it can do to make our lives much easier than ever before. 33. How AI and Machine Learning Changes Modern-day Industries It’s important to know exactly how AI and Machine Learning are transforming the industries and their role in our technological advancements. 34. Applications of Artificial Intelligence in Business Artificial Intelligence, the concept of ‘machines with brains’ has been in the spotlight around the last seven decades. A theoretical notion that started as simple rule-based automation in the 1950s' has now grown so much that now the scientists are trying to make human-like robots. The question- What AI might do to us? — has created a lot of controversies in and out of the scientific community. 35. 8 Machine Learning Trends that Impact Business in 2021 and Beyond Let’s discover the latest innovations in machine learning in 2021-2022 and go over various examples of how this technology can benefit you and your business. 36. Success With Machine Learning Projects in Python In this article, we will give you a sense of the applications for machine learning and explain why Python is a perfect choice for getting started. 37. Build A Commission-Free Algo Trading Bot By Machine Learning Quarterly Earnings Reports [Full Guide] Introduction 38. What is Machine learning and Why is it Important? Machine learning is quite an exciting field to study and rightly so. It is all around us in this modern world. From Facebook’s feed to Google Maps for navigation, machine learning finds its application in almost every aspect of our lives. 39. Is Automation Jeopardizing Our Future? Is automation jeopardizing our future? What's its impact on our future and the job market? And how can we cope up? 40. Universal Data Tool Introduction: Weekly Update 2 If you haven’t heard of the Universal Data Tool, it’s an open-source web or desktop program to collaborate, build and edit text, image, video, and audio datasets with labels and annotations. 41. An Introduction to Variational Autoencoders Using Keras In this tutorial, we’ll explore how Variational Autoencoders simply but powerfully extend their predecessors, ordinary Autoencoders. 42. How To Improve Data Quality When With Unsupervised Machine Learning There won’t be any business insights if the data quality is poor. 43. How Machine Learning is Used in Astronomy Is Astronomy data science? 44. Will Machine Learning Algorithms Encroach on Content Marketers? Since its inception, the field of online content marketing has been in a constant state of flux. This tendency is built into it — since it operates in an ever changing online world, all it can ever do is change along with the internet. For better or for worse, the rapid advent of machine learning algorithms in online processes will only lead to more change in the field of content marketing. 45. How To Identify Trees with Deep Learning Idea / inspiration 46. Pecan.ai Raises 11 Million to Bring Machine Learning to Business Analysts Pecan.ai has just come out of stealth, raising an $11M Series A, to enable business analysts to build machine learning models automatically. Dell Capital led the round, joined by S capital and bringing the total funding of the company to $15M. 47. Embeddings in Machine Learning: Everything You Need to Know Here's a deep dive into the history of machine learning embeddings, common uses, and current infrastructure solutions, including the vector database. 48. Quantum Machine Learning Using TensorFlow Quantum INTRODUCTION 49. How Data Analysis Helps Unveil the Truth of Coronavirus These days we are all scared of the new airborne contagious coronavirus (2019-nCoV). Even if it is a tiny cough or low fever, it might underlie a lethargic symptom. However, what is the real truth? 50. Big Tech Companies Need Machine Learning Engineers Machine learning is no longer a sci-fi concept, but an actual application of AI technology we use every day. 51. No Such Thing As The Intelligent Edge But it will be much bigger than the Internet. 52. Why It’s Very Difficult to Create AI-Based Slow Motion Over the last few years a number of open source machine learning
projects have emerged that are capable of raising the frame rate of
source video to 60 frames per second and beyond, producing a smoothed,
'hyper-real' look. 53. Optuna Vs. Hyperopt: Which Hyperparameter Optimization Library You Should Choose Thinking which library should you choose for hyperparameter optimization? Thank you for checking out the 53 most read stories about Machine Learning Uses on HackerNoon. to find the most read stories about any technology. Visit the /Learn Repo

53 Stories To Learn About Machine Learning Uses

CASCADE STRATEGY

cascade.app

-- employees

Since n.d.

COMPANY RANKING

CASCADE STRATEGY

EVERGREEN INDEX #2094

Representing this company? Click here to claim and customize this page!

Cascade Strategy WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!

CASCADE STRATEGY

cascade.app

-- employees

Since n.d.

COMPANY RANKING

CASCADE STRATEGY

EVERGREEN INDEX #2094

RELATED COMPANIESRANK

STORIES ABOUT CASCADE STRATEGY

Representing this company? Click here to claim and customize this page!

Cascade Strategy WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!