Ryan Barnes

I tested 11 CSV editors to write a listicle. (For those of you who don't know- I hate numbers and Excel) So, this was quite challenging at first. I was down the rabbit hole. Didn't know what to do, or where to start from. I'd never manually tried software before #writing about them. I'd look at their website, customer reviews, and demos on YouTube. (Now, I realize how stupid it was. The insights you get from trying each software yourself is pure GOLD. It was time-consuming but once I tried 4-5 software, I knew what exactly to do.) I used ChatGPT to understand why people use CSV editors. After trial and error, I came up with a few use cases. Next, I tried to understand what features should a CSV editor have. After listing down 7-8 features, I began downloading multiple software. Logged in and played with it. Just hovered here and there. I uploaded a massive CSV file to each platform to understand how powerful it was. Then, I tried out all the other features my readers would look for in a CSV editor. I got insights no YouTube demo or customer review could give. I read a few competitive articles and found out how different mine was. It was packed with the exact features that mattered. None of the generic copy-pasted crap. My takeaway: Take an extra day and try out the platforms you're writing about. This was the first time I did and this is how I'll write all my #listicles.

37

17 Comments

LLMs CANNOT generalize outside their domain distribution. You could argue that they can generalize within the distribution of the domain where they have seen 70% of the data. Even with 70%, they fail terribly. Here is a simple thought experiment. Suppose LLMs were expected to behave like the mathematician Girolamo Cardano in 1501. Could they have independently come up with the imaginary number “i,” the square root of -1? They would likely react with rejection. Logic does not allow you to conjure a number like “i,” because the square of no number can yield a negative. This requires intuition or imagination, a conviction so strong that one accepts the existence of such a number and is willing to discover its properties, leading to the development of complex numbers. This leap laid the groundwork for the discovery of quantum mechanics. This is not simply a case of out of domain generalization; it represents a mysterious leap into the improbable. It is unlikely that this would emerge as a behavior of even a 1,000 trillion parameter model, as it remains constrained by the distribution coefficients of its domain. We need to rethink AI models from ground-up if we need to mimic this unique human trait. #AI #Humans #Math

8

1 Comment

After decades of experience in building foundational AI models, the simplest rule of thumb I've observed that works empirically as a framework for hybrid modeling AI platforms, when applied in the following strategic order, is: 1) Start by extracting simple local features, prioritizing multi-scale extraction early on. 2) Then, focus on broad-range (dilated) interactions, ensuring high resolution by incorporating skip or residual connections. 3) Next, recalibrate features by modeling inter-dependencies (squeeze and excite). 4) Follow this by capturing long-range dependencies through some form of state management (e.g., multi-head attention). 5) Enhance expressivity through non-linear transformations (such as non-linear feedforward networks). 6) Finally, condense the representation using global pooling or dense layers. By applying these layers in multiple blocks through stacking, you will generally achieve effective numerical conditioning rapidly, minimizing the need for extensive ablation studies. This approach has consistently proven to be reliable. I stand by this rule all the time and works for me in general. #AI #ArtificialIntelligence #Engineering

4

It's no secret AI/ML X Biology is hot right now, with millions of dollars in funding going to companies with ML guided discovery strategies. This field is the next wave of startups, similar to CAR-T and CRISPR before them. Candidates looking to enter this field often ask about wetlab skills that pair well with ML guided strategies. Here are some skills requested by leading AI/ML X Bio companies: - Display Technologies: Yeast, Phage, mRNA, and Ribosome display are in high demand, with Phage being the most popular. Yeast is also gaining traction because of it's eukaryotic Expression benefits. - HTS Screening: With the evaluation of molecules derived by AI/ML, extensive screening is inevitable. - Automation: Developing assays and workflows for automated liquid handlers is crucial in this field. - CFPS / IVTT: Ideal for expressing multiple proteins simultaneously, despite challenges with folding and disulfide bonds. But some experts can help you solve these issues. ( We know them 😊) - SPR & BLI: Candidates skilled in SPR are highly sought after by companies. - Molecular Dynamics Data: Individuals who can integrate molecular dynamics data into ML algorithms for structure and binding prediction are in high demand. Upskilling in any of the mentioned skills can significantly benefit your career in AI/ML X Biology. While there may be other skills to consider, these are the most requested ones based on current trends in the industry. Unashamed Segue CoFactor is currently recruiting for display experts on both coasts. We are also in need of research associates with SPR. Please reach out for more information. #ML #biology

37

1 Comment

Ted Chiang and Terence Tao: art, math and AI There is much I dislike about Ted Chiang's manifesto in The New Yorker two weeks ago (https://lnkd.in/gdXhCQ2K). But his narrow understanding of #AI, and specifically #genAI and #LLMs, is a topic for another time. 👉 A more positive take relies on a striking parallel between Chiang's (deeply flawed, in my view) main point that art is characterized by the myriad decisions required at all scales, and Terence Tao's observation (https://lnkd.in/gAjs2N7K) that getting to the solution of a really hard math problem requires a lot of decisions about where to go next -with so many dead ends that avoiding them is a key survival skill. The big difference between the two is that Tao can see the potential of AI to support the decision-making, even if the critical forks in the road are defined and decided by the talented human. Chiang thinks it removes meaning. I think the two should talk. 💎 This parallel is not as far-fetched as one might think. The Library of Babel, the quintessential Jose Luis Borges short story from his 1941 collection "The Garden of Forking Paths" (a perfect title for our purpose here!), bridges the two worlds. In the Library of Babel, all books already exist (up to a certain length), buried in an ocean of gibberish. So the decisions Chiang is talking about are akin to routing decisions in the Library: which book, which shelf, which section, which octogonal chamber, which floor, ... at all scales. Similarly, if a solution to a mathematical problem "pre-exists", the right sequence of decisions will lead to it. In both cases, the expert (or artist) is the scout, guided by experience, context, intuition, emotions (even in math). I know I am being naughty, here is a depiction of the Library of Babel generated by chatGPT4o.

1

1 Comment

Experimentation platforms need to have a symbiotic relationship between platform engineering and data science: I’ve been engineering experimentation platforms from scratch for the last 8 years. Even after building random allocation, data pipelines, and the scorecard, the work never stops there. The science around experimentation keeps growing. Think about how experimentation has grown so much from differences in means of randomized data. 1. There was an evolution of what treatment effect we are modeling. The average treatment effect, to the conditional treatment effect, to the time dynamic treatment effect. This evolution allowed us to analyze data segments, and look for patterns in the treatment effect over time, such as novelty effects or diminishing returns. 2. There was an evolution in inference. Bayesian testing is more and more popular, and so is always valid inference like the SSRM. 3. Then there was also an evolution in the type of experiment we run. From AB testing, to quasi experimentation, to multiarm bandits, to encouragement experiments. It’s hard to keep up with all these advancements in science. And it’s almost impossible to have an experimentation platform keep up in implementation. To have a truly stellar experimentation platform, you have to find ways to work smarter. Platform engineers should not be the only ones responsible for implementing these changes in experimentation. Data scientists want to be part of the discussion. Invite them and involve them! But how? It’s not so easy though. A production experimentation stack can be very complex, and the barrier to entry is frequently overwhelming to a data scientist. I believe we have to compromise. First, experimentation platform teams can offer a python library to perform metrics computation. Second, the team can offer a separate, stand alone library to perform causal inference on data that came from the first library. These libraries need to form the analytics stack for the experimentation platform. The beauty of this is that the libraries can be used for prod, but can also be used in local notebook environments. If you remove the overhead of the production environment, and enable the DS to use the metrics and inference libraries in local notebooks, they will be able to easily extend the libraries with new innovations. And when the libraries are extended, those new ideas can be merged upstream back into the platform. Our [slightly old] paper describes how we think about democratization, how to modularize for engineers and data scientists, how to reduce the needs for distributed computing so that DS can operate in local environments, and also how to scale it back for the production environment. I’ve been carrying that way of working for the past 6 years across 3 companies. I’d love to hear if this resonates with your team. https://lnkd.in/gsDBEmtX

91

1 Comment

I guess some people confuse "AI Research" as "searching" the web for what types of APIs and solutions are available! This is not a DIY search! 😂 If you're in AI research, you're developing foundational models. Roughly 70% of what you do is math and the remaining 30% is numerical conditioning needed to make empirical outcomes align with theoretical expectations. Foundational research isn’t about tweaking pre-built solutions; it’s about advancing the science itself, often working with novel architectures, experimental training paradigms, and unexplored datasets. Then, there’s "Applied ML/AI Engineering," which some swing as AI research but is fundamentally different. In applied AI/ML, you’re taking existing models, either proprietary or open source, and engineer solutions through RAG, fine-tuning, and custom frameworks. The work here involves adapting general-purpose models to fit specific tasks, scaling systems, and solving for performance, latency, or usability. Applied AI Engineering is challenging, but it’s built on the shoulders of foundational AI research, not in place of it. This is not research. The right title is "AI/ML Engineering". And, of course, data science is entirely distinct. It’s about extracting insights from data, often using statistical techniques and traditional machine learning algorithms to find patterns, make predictions, or inform decisions. Data scientists focus on structuring, analyzing, and visualizing data, building models to support strategic goals, not advancing AI capabilities themselves. Here, the focus is primarily on answering specific business questions or providing actionable insights, not on exploring the theoretical limits of artificial intelligence. 90% of the time I see titles that does not reflect the work people do! This is primarily because of orgs not normalizing their rubrics. #HR #PeopleOps #AI #ML #Org #startups

7

12 Comments

🎉 Excited to share that our paper, "Refining Text-to-Image Generation: Towards Accurate Training-Free Glyph-Enhanced Image Generation" has been accepted at WACV 2025! 🔗 https://lnkd.in/gv3736ym This paper tackles a critical issue in text-to-image generation: generating precise, readable, and complex text within images. ✨ Key Contributions: - LenCom-Eval Benchmark:   - 📊 Introduced LenCom-Eval, a new benchmark to rigorously test models on generating images with lengthy, complex, and rare text sequences.   - 📂 Included challenging datasets, like MARIO-Hard and Aug-MARIO-Hard, which assess models’ abilities with rare word combinations and text with intentional variations, pushing the limits of accuracy and clarity in text generation. - SA-OcrPaint Framework for Enhanced Text Quality:   - 🧊 Simulated Annealing (SA): Reduces text overlap by optimizing the layout of text bounding boxes, making complex layouts clearer and more readable.   - 🔠 OCR-Aware Recursive In-painting: Corrects spelling and formatting errors by using OCR to detect and fix inaccuracies, enhancing text fidelity without retraining. 👏🏽 Sanyam Lakhanpal, Shivang Chopra, Vinija Jain, Aman Chadha, Man Luo

64

3 Comments

In-Context Learning at Extreme Scales: My Thoughts As large language models (LLMs) continue to expand their context capacities, the domain of in-context learning (ICL) at extreme scales has piqued my interest. A recent paper by researchers at Carnegie Mellon University and Tel Aviv University provides a detailed exploration into this area, unveiling both impressive capabilities and notable challenges. 🔍 Key Insights: Performance Improvement: The study demonstrates that ICL can improve significantly by integrating hundreds to thousands of examples into the context window—far beyond traditional few-shot settings. This is particularly impactful in complex tasks like 77-way banking intent classification, where long-context ICL matches or surpasses traditional finetuning. Input Sensitivity: Long-context ICL shows reduced sensitivity to input shuffling compared to shorter contexts. However, organizing examples by label dramatically reduces performance, emphasizing the need for diverse context configurations. Source of Gains: Performance gains primarily arise from accessing more relevant examples, rather than aggregating knowledge across many examples during encoding. 📉 Challenges & Drawbacks: Sensitivity to Example Order: Despite reduced sensitivity, the performance still varies with input order, indicating potential instability in practical scenarios. Performance Saturation: The models often do not fully leverage their extended context capacities, showing performance saturation before reaching the maximum context length. Computational Cost: The increased computational cost associated with using large context windows could limit practical deployment, particularly in resource-constrained environments. Efficiency Issues: There is a noted inefficiency in how models utilize the encoded information within large datasets, suggesting that current architectures may need refinement to better use available data. 🚀 Future Directions: The paper suggests several exciting avenues for future research, including optimizing the trade-offs between computational costs and performance, exploring model adaptation strategies, and refining ICL approaches for specific applications or datasets. As we push the boundaries of what LLMs can achieve, effectively leveraging their full context could unlock new capabilities. This area is ripe for further exploration and could lead to major advancements in how we deploy AI. For those at the cutting edge of language model development, I highly recommend this paper for its thorough analysis and forward-looking perspective. Check out the full work at https://lnkd.in/d7AmGh4y. What are your thoughts on the future of in-context learning at such scales and the challenges discussed? #AI #MachineLearning #DeepLearning #LanguageModels #InContextLearning

17

Giant Self-Licking Ice Cream Cone Great paper in Nature Magazine (Open Access! "AI models collapse when trained on recursively generated data" https://lnkd.in/gZpAZUhu) on what Jathan Sadowski called "Habsburg AI": data inbreeding leads to unseemly results (https://lnkd.in/gXMWhkJV: "a system that is so heavily trained on the outputs of other generative AI's that it becomes an inbred mutant, likely with exaggerated, grotesque features. It joins the lineage of Potemkin AI."). Or "Model Autophagy Disorder", MAD (https://lnkd.in/gyNE24KQ). Or "Giant Self-Licking Ice Cream Cone". https://lnkd.in/gXFY_Pkb

7

Communication is everything! Something might not be working but the knowledge and solutions are in your peers if you can tap into it. Knowing when you don’t know is key to success.

1

Tej Pandit's work is set to be presented at #neurips workshop on Scalable Continual Learning for Lifelong Foundation Models. In his Masters thesis, he has used structural plasticity to address the learning capacity, though this approach has upper bounds. Now, he proposes CASPER , a simple approach that combines the power of hypernetworks to generate the weights for multiple neural networks to solve the continual learning problem. With auto-generative replay, the hypernetwork generates samples of previous networks. This method mitigates catastrophic forgetting without relying on an ever-expanding memory buffer. Learn more: https://lnkd.in/gqTBZNnD

98

2 Comments

As part of my PhD research, I did long interviews with many engineers from SaaS product companies to understand their onboarding journey, esp. how they ramped up on the product. In many cases, I realized that seemingly small things had an outsized positive impact on their onboarding journey (they seemed to have learned better, faster and more). As I get ready to work on a paper to formally report the results, I thought I will share some of the takeaways which may be useful for new campus hires who are going to join a company in a few weeks. Three seemingly small actions that impacted onboarding significantly: * 𝐀𝐬𝐤𝐢𝐧𝐠 𝐪𝐮𝐞𝐬𝐭𝐢𝐨𝐧𝐬: Mostly, new campus hires do not ask questions. Some of them said that is because they are used to not asking questions in class, and onboarding feels like a classroom! However, in many instances, the new hire started asking questions at some point, and their experience changed significantly, for better. There were external situations which caused them to start asking questions (those are stories for a different day, very fascinating!), but they benefitted a lot when they started asking questions. * 𝐓𝐚𝐤𝐢𝐧𝐠 𝐧𝐨𝐭𝐞𝐬 𝐩𝐮𝐫𝐩𝐨𝐬𝐞𝐟𝐮𝐥𝐥𝐲: Not many people reported taking notes. I found there were a few who took notes with specific goals - one of them wanted to capture what they didn’t understand, another took down everything so they could revise within 24 hours, still others had a mental model of the product where they filled the gap of understanding through these notes. People who had specific goals when taking notes seemed to have had much better experience of onboarding than those who didn't take notes or took them without clear purpose. * 𝐓𝐞𝐬𝐭𝐢𝐧𝐠 𝐭𝐡𝐞 𝐩𝐫𝐨𝐝𝐮𝐜𝐭: Some people were assigned the task to read and execute test cases, others ha mentors telling them to do so. Either way, those who engaged in this activity reported much better experience with onboarding (they understood the broad picture of the product better) than those who did not. If you are joining the workforce soon as a new campus hire engineer, do keep these in mind and try them out during your onboarding journey. Do stay proactive and ask for help (for ex, if your mentor is not responsive and so your questions are unanswered), you will be surprised at how supportive onboarding teams generally are! If you have specific questions on how to do well in your first job out of college, feel free to connect and discuss. #newcampushires #firstjob #careeradvice

37

9 Comments

Most things in drug discovery are difficult to generalize: every step of the process depends on its own set of goals, experiments and models. However what is true is that the impact of AI will depend on its relevance for every problem, and that this relevance will depend on the quality and availability of the data used as input. AlphaFold for instance is poised to transform hit discovery, but its existance relies on the very well documented PDB database. Many other steps in the drug discovery process don’t have an easily accessible database to train models on. AI is set to transform research for new medicines, but the winners will be the data producers, who will have the capabilities to fine tune and adapt the models on their own specialized data. This is what we’re working towards at Orakl Oncology and it’s a lot of fun to figure this out!

27

1 Comment

The increasing low-quality & hostile review s in the machine learning community are really concerning.

21

Physics helps get stuff done. Its application has put rockets in space, semiconductors in phones, and eclipses on calendars. For some philosophers, this is all physics offers. It is a mere instrument, albeit of great power, giving us control over tangible things. It is a set of gears and widgets (wavefunctions, strings, even electrons) to crank out predictions. In contrast to instrumentalists, scientific realists argue that the success of theories shows that they map onto the structure of the world, symbols in equations carry the imprint of real entities. This is an old debate in the philosophy of science. While we touch on some arguments for either position, this episode focuses on the phenomenology of physics researchers. What do physicists believe? Céline Henne is a philosopher at the University of Bologna. Alongside Hannah Tomczyk and Christopher Sperber she has fielded the most comprehensive survey of the attitudes of physicists towards the reality of the objects of their study. From looking at the answers to dozens of questions from several hundred physicists, they have distinguished several camps of belief. It's an elegantly designed survey, simply reading the questions forces a consideration of one's own position. It makes us think about what some of our deepest theories are actually saying. Take the survey here: https://lnkd.in/gh4TjFCy and listen on YT, Spotify, Apple or at multiverses.xyz https://lnkd.in/gJdkU7Tc

10

1 Comment

Super excited to be releasing AlphaFold 3 today, developed by Isomorphic Labs and Google DeepMind: our next generation AI model for predicting the biomolecular structures and interactions of proteins, DNA, RNA, small molecules, and more: https://lnkd.in/gY7deAqk The AlphaFold 3 paper is published in Nature today, bringing together more training data from PDB with new neural net architectures and a diffusion module that generates the 3D coordinates of each atom. https://lnkd.in/gSZ5k2Dj When looking at the accuracy of this model, for interactions between proteins and other molecule types we see at least a 50% improvement compared to existing methods, and for some important categories we have doubled the prediction accuracy. We’ve been using these bleeding-edge models day-to-day at @IsomorphicLabs for drug design on our internal and partnership projects. There’s so much scope for advancing rational structure based drug design! https://lnkd.in/g7df5p4Q And @GoogleDeepMind have also developed AF Server which makes a lot of these capabilities accessible for free for non-commercial research https://lnkd.in/gJGC-xK3 Very proud of the teams across Isomorphic Labs and Google DeepMind for all the amazing research and work that has gone into this. Read more: AF3 blog: bit.ly/44yfaCw AF3 for Drug Design: bit.ly/4a5o3EM Nature paper: bit.ly/3yaLLSL AFServer:bit.ly/3JWY1Zy

2,765

44 Comments

😯🤯 It's super fascinating how LLMs resemble humans in so many aspects. Ever heard of the "lost-in-the-middle" issue with LLMs? 📏 As long contexts become more common, it's really important for folks working in the space to be aware of this issue with LLMs. 💡 A really popular paper from November 2023, titled "Lost in the Middle: How Language Models Use Long Contexts," empirically demonstrates that LLM performance varies significantly depending on the position of relevant information within the context. 📖 Some Insights: ⛳ The paper analyzes LLM performance on tasks requiring identification of relevant information within input contexts, including multi-document question answering and key-value retrieval. ⛳The performance significantly varies based on the position of relevant information within the context. ⛳LLMs tend to perform best when relevant information is at the beginning or end of the context, while performance degrades when information is in the middle. ⛳ A distinctive U-shaped performance curve is observed. Performance peaks when relevant information is at the start or end of the context and declines significantly when it's in the middle. 🤯 The U shaped performance curve is so similar to human attention spans during a typical presentation. 🤯I found the human attention chart from renown professor and university researcher J.W. Niemantsverdriet's research which suggests that the attention of a typical human audience doesn’t maintain a consistent level throughout presentations. It peaks in the beginning and end (which we can all agree on 😁 ) It's truly intriguing how LLMs often resemble humans in many ways. Typically, they reflect human traits because they're trained on human-generated data. However, this "lost-in-the-middle" resemblance in this case is just 🤯 🤯 . 🚨 I post #genai content daily, follow along for the latest updates! #genai #llms #contextlength

82

2 Comments