Dan Shiebler - Artemis | LinkedIn
Dan Shiebler
Activity
Follow
What does identity and access management look like when a single endpoint runs 10,000 agents? Lemonade CISO Jonathan J. Jaffe breaks it down in our…
Liked by Dan Shiebler
Pi Security is officially here I've spent my career building systems where security isn't optional, from cars to robots, energy devices, AI and…
Liked by Dan Shiebler
Today Pi Security is officially out of stealth, with $35M in funding. I've been into security for as long as I can remember. Building and breaking…
Liked by Dan Shiebler
Experience & Education
-
Artemis
* * ###
**
-
**
undefined
*
Volunteer Experience
Board Member
ChalkTalk
May 2023 - Present 3 years 2 months
Wellness Instructor
Rhode Island Free Clinic
Sep 2011 - May 2015 3 years 9 months
Health
Taught weekly and biweekly exercise classes for patients at the Rhode Island Free Clinic
Guest Lecturer
NYC Data Science Academy
Jul 2016 - Jul 2017 1 year 1 month
Ran workshops and gave lectures on Machine Learning and Data Science
Publications
RecSys Aug 2020
Word2vec is a powerful machine learning tool that emerged from Natural Language Processing (NLP) and is now applied in multiple domains including recommender systems, forecasting, and machine learning on graphs. It is often used off the shelf and we address the question of whether the default hyperparameters, which were originally tuned on NLP tasks are suitable for recommender systems. The answer is emphatically no. For an unconstrained hyperparameter search running to convergence, we measure… Show more
Word2vec is a powerful machine learning tool that emerged from Natural Language Processing (NLP) and is now applied in multiple domains including recommender systems, forecasting, and machine learning on graphs. It is often used off the shelf and we address the question of whether the default hyperparameters, which were originally tuned on NLP tasks are suitable for recommender systems. The answer is emphatically no. For an unconstrained hyperparameter search running to convergence, we measure an average increase in hit rate of 221%. However, this is achieved by using orders of magnitude more computational resources. Surprisingly, even when the runtime is fixed to the out of the box runtime, constrained optimization leads to an average 138% improvement in hit rate. For large scale systems, full hyperparameter optimization on production datasets is not possible and we tackle the question of estimating the optimal hyperparameters from a sample. We find that by applying constrained optimization to a 10% sample of the data and applying the hyperparameters found from the sample to the full datasets an average hit rate improvement of 91% above the default parameters is achieved. Show less
Compositionality May 2020
In this work we take a Category Theoretic perspective on the relationship between probabilistic modeling and function approximation. We begin by defining two extensions of function composition to stochastic process subordination: one based on the co-Kleisli category under the comonad (Omega x -) and one based on the parameterization of a category with a Lawvere theory. We show how these extensions relate to the category Stoch and other Markov Categories. Next, we apply the Para construction to… Show more
In this work we take a Category Theoretic perspective on the relationship between probabilistic modeling and function approximation. We begin by defining two extensions of function composition to stochastic process subordination: one based on the co-Kleisli category under the comonad (Omega x -) and one based on the parameterization of a category with a Lawvere theory. We show how these extensions relate to the category Stoch and other Markov Categories. Next, we apply the Para construction to extend stochastic processes to parameterized statistical models and we define a way to compose the likelihood functions of these models. We conclude with a demonstration of how the Maximum Likelihood Estimation procedure defines an identity-on-objects functor from the category of statistical models to the category of Learners. Code to accompany this paper can be found at this https URL Show less
Artificial Intelligence and Machine Learning for Digital Pathology - LNCS 12090 2020
This paper reviews guidelines on how medical imaging analysis can be enhanced by Artificial Intelligence (AI) and Machine Learning (ML). In addition to outlining current and potential future developments, we also provide background information on chemical imaging and discuss the advantages of Explainable AI. We hypothesize that it is a matter of AI to find an invariably recurring parameter that has escaped human attention (e.g. due to noisy data). There is great potential in AI to illuminate… Show more
This paper reviews guidelines on how medical imaging analysis can be enhanced by Artificial Intelligence (AI) and Machine Learning (ML). In addition to outlining current and potential future developments, we also provide background information on chemical imaging and discuss the advantages of Explainable AI. We hypothesize that it is a matter of AI to find an invariably recurring parameter that has escaped human attention (e.g. due to noisy data). There is great potential in AI to illuminate the feature space of successful models.
Show less
ICLR 2019 (International Conference on Learning Representations) Dec 2018
Most recent gains in visual recognition have originated from the inclusion of attention mechanisms in deep convolutional networks (DCNs). Because these networks are optimized for object recognition, they learn where to attend using only a weak form of supervision derived from image class labels. Here, we demonstrate the benefit of using stronger supervisory signals by teaching DCNs to attend to image regions that humans deem important for object recognition. We first describe a large-scale… Show more
Most recent gains in visual recognition have originated from the inclusion of attention mechanisms in deep convolutional networks (DCNs). Because these networks are optimized for object recognition, they learn where to attend using only a weak form of supervision derived from image class labels. Here, we demonstrate the benefit of using stronger supervisory signals by teaching DCNs to attend to image regions that humans deem important for object recognition. We first describe a large-scale online experiment (ClickMe) used to supplement ImageNet with nearly half a million human-derived "top-down" attention maps. Using human psychophysics, we confirm that the identified top-down features from ClickMe are more diagnostic than "bottom-up" saliency features for rapid image categorization. As a proof of concept, we extend a state-of-the-art attention network and demonstrate that adding ClickMe supervision significantly improves its accuracy and yields visual features that are more interpretable and more similar to those used by human observers. Show less
Journal of Vision (JOV) 2017
Cognitive Computational Neuroscience (CCN) 2017
KDD 2018 (Common Model Infrastructure Workshop)
Computational and Mathematical Models in Vision (MODVIS) 2017
SysML 2018
-
More Feedback, Less Depth: Approximating Human Vision with Deep Networks (co-author)
NIPS 2016 Workshop on Representation Learning in Artificial and Biological Neural Networks
Join now to see all publications
Patents
-
Device-Based Systems and Methods for Detecting Device Usage
US 62/248,049
-
SYSTEMS AND METHODS FOR DETECTING AND ASSESSING DISTRACTED DRIVERS
US 15/268,049
Machine Learning algorithm for detecting phone interaction from motion sensor signals.
(Co-inventor. Patent owned by TrueMotion)
-
SYSTEMS AND METHODS FOR SCORING DRIVER TRIPS
US 62/346,013
Algorithms for building a driver risk score.
(Co-inventor. Patent owned by TrueMotion)
-
Systems and Methods for Displaying Driving Scores
US 62/248,051
-
Systems and Methods for Vehicle Crash Prediction, Detection and Reconstruction
US 62/383,839
US US20180126938A1
Projects
Feb 2017
A TensorFlow package for performing the deep taylor approximation to layerwise relevance propagation.
Aug 2016
A neural network that can generate original musical compositions
Aug 2016
A page where I write about mathematics and computer science
Sep 2015
The PopGen Simulator is a GUIbased interactive tool for teaching students about population genetics and computer simulation. Students use the program to simulate the growth and evolution of a population at the genomic level. Users select between a variety of different models, specify the parameters of the model, and visualize the results. Students in the Brown University Course “Computational Theory of Molecular Evolution and Population Genetics,” used the simulator to reinforce concepts and… Show more
The PopGen Simulator is a GUIbased interactive tool for teaching students about population genetics and computer simulation. Students use the program to simulate the growth and evolution of a population at the genomic level. Users select between a variety of different models, specify the parameters of the model, and visualize the results. Students in the Brown University Course “Computational Theory of Molecular Evolution and Population Genetics,” used the simulator to reinforce concepts and validate mathematical reasoning. For example, in one laboratory exercise students ran multi-locus simulations with a variety of mutation rates and number of loci, and reasoned about the resulting population evolution.
In addition to its teaching applications, the PopGen Simulator can also be used to generate synthetic data and includes an API that researchers can use to quickly prototype and visualize their models. The simulator is written in MATLAB, and also includes a command line component. Show less
Mar 2013
MessageHunt is an IPhone app that has elements of Snapchat, Yik Yak and Geocaching. Users drop secret messages in physical locations. A message can only be picked up and read by someone else who goes to the same location.
-
Senior Thesis - Decision Conflict and Beta Oscillations in the Human Subthalamic Nucleus
Jan 2013 - May 2015
-
Honors & Awards
-
Invited Panel Guest (In the Lab: AI and Machine Learning for Media Industry)
Variety Innovate Summit
Nov 2017
https://events.variety.com/event/2017-innovate/
-
Invited Speaker (Real World Data Science Strategy)
Data Science Go
Oct 2017
https://www.datasciencego.com/dan-shiebler
-
Invited Speaker (Real World Data Science Strategy)
Global Big Data Conference
Sep 2017
http://www.globalbigdataconference.com/new-york/global-artificial-intelligence-conference-93/speaker-details/dan-shiebler-61977.html
-
Invited Speaker (Making Deep Learning Work on Messy Sensor Data)
MIT Lincoln Laboratory
Jul 2017
-
Invited Guest
Super Data Science Podcast
May 2017
https://soundcloud.com/superdatascience/sds-059-changing-human-behaviour-through-a-driving-app
-
Speaker (Making Deep Learning Work on Messy Sensor Data)
Rework Deep Learning Summit Boston
May 2017
https://youtu.be/3Focs88C-so
-
Speaker (The Power and Pains of Sensor Data)
Open Data Science Conference
May 2017
https://youtu.be/1QuqOIFsaj4?list=PLB2SCq-tZtVkquR6O15BtcOdfZotXV5y_
-
Speaker (Machine Intelligence for Driver Safety)
Machine Intelligence Summit New York
Nov 2016
https://youtu.be/DIPY-RhgeTA
Organizations
-
Toastmasters International (Public Speaking and Leadership Organization)
President
Sep 2012 - May 2015
President of the Brown Chapter of an international non-profit devoted to developing members’ public speaking and leadership skills Lead weekly meetings; delivered and critiqued planned and impromptu speeches
More activity by Dan
Brightmind led the Seed in Pi , and today we're proud to announce our continued support with their Series A. Security teams aren't losing because…
Liked by Dan Shiebler
Huge news! Excited to share that Kevin Gabura has been promoted to Partner at Craft Ventures. Kevin Gabura was the 1st team member I brought into…
Liked by Dan Shiebler
Another spotlight on cyber professionals who are admired for what they are doing in our industry. This week’s deep dive is going to be on the best…
Liked by Dan Shiebler
When I look back on my four and a half years at Abnormal AI, what stands out most is how the work keeps finding new ways to matter. I joined as a…
Liked by Dan Shiebler
Maze is two years old! It's cliche to say, but it feels like weeks ago that Adrian Jozwik, Santiago Castiñeira, and I were exploring the idea and…
Liked by Dan Shiebler
I am proud to share that I have joined Tradeweb as their new Global Chief Information Security Officer (CISO). There aren’t many companies that can…
Liked by Dan Shiebler
The only thing that could show up our new logo, is our new report ✨ The 2026 Latio SOC market report is here and it's a good one! Read it now:…
Liked by Dan Shiebler
Artemis was awarded Threat Detection Leader and SIEM Disruptor in the Latio 2026 Security Operations Market Report. It reflects what we set out to…
Liked by Dan Shiebler
Latio's 2026 Security Operations Market Report broke down the SOC landscape and where Artemis Security fits. My favorite quote: "These features…
Shared by Dan Shiebler
After months of research, interviews, product testing, writing, and design, we’re excited to launch the Latio 2026 Security Operations report! We're…
Liked by Dan Shiebler
Industry benchmarks for responding to an attack are measured in hours and days. Some of the best security teams I know are still fighting to get down…
Liked by Dan Shiebler
View Dan’s full profile
- See who you know in common
- Get introduced
- Contact Dan directly
Other similar profiles
Yi Xuan
Amazon Music
4K followers
San Francisco Bay Area
Ayush Pareek
Apple
6K followers
San Francisco Bay Area
Vishal Krishna
Microsoft
3K followers
New York City Metropolitan Area
Shrinath Deshpande
Apple
2K followers
Fremont, CA
Ankit Khedia
Deccan AI
7K followers
Mountain View, CA
Maunil Vyas
Arizona State University
7K followers
San Francisco Bay Area
Thibault Doutre
Point72
5K followers
New York, NY
Andrew Long, PhD
Apple
7K followers
San Francisco Bay Area
Pararth Shah
Stealth
11K followers
San Francisco Bay Area
Zhuoran Yu
OpenAI
10K followers
New York, NY
Samuel Sherman
Rapid7
5K followers
Denver, CO
Shreya Shankar
7K followers
Berkeley, CA
Zhihao Li
Google DeepMind
7K followers
Mountain View, CA
Cindy Pan
Apple
6K followers
San Francisco Bay Area
Matthew M.
Meta
32K followers
San Francisco Bay Area
Raghvendra Singh
Meta
6K followers
New York, NY
Meng Cao
Apple
9K followers
Saratoga, CA
Vinaya Polamreddi
Resolve AI
2K followers
Stanford, CA
Bo Zeng
Airbnb
17K followers
Redmond, WA
Shubham Khandelwal
Moveworks
10K followers
Bengaluru
Show more profiles Show fewer profiles
Explore more posts
-
Bruno Miranda Doximity • 5K followers Ronnie Rocha provides in-depth analyses, cost comparisons, and sample projects involving Gemini, ChatGPT, Claude, Llama, and Qwen models. Dive into to explore the performance of these models for iOS development. #ModelPerformance #iOSDevelopment 15 1 Comment
-
Sumeet Jain Sublime Security • 2K followers LLMs are the fastest way to prototype many workflows, but fast to build doesn't mean fast to run. Something that's been working well for me is a late-stage "determinism checkpoint" to find opportunities for inference shedding -- migrating functionality into deterministic components. Prompt to prototype, then do targeted engineering to move workflow steps outside of the inference loop. I think about this from day one, but a later-stage review benefits from seeing real behavior first. Moving work outside the inference loop can improve performance, reliability, and simplify maintenance. But there is also financial urgency: Inference is artificially cheap right now. When costs normalize, unnecessary LLM calls will become expensive fast. Smaller or local models can help, and we should optimize which models to call -- but we should also ask whether we should be calling one at all. The cheapest inference is the inference you eliminate. Opportunities for inference shedding might include: - Steps whose output doesn't actually vary in a meaningful way across runs - Steps making the same judgment repeatedly at runtime, when the model could produce the logic once and move execution outside of the inference loop - Bridging other steps, e.g. reformatting one tool's output into another tool's input. Often "transform" layers can be made much cheaper One very simple example to illustrate my point: Let's say I have a Claude Skill use the Slack MCP to poll for substantive messages in certain channels, auto-process them to find key insights, and then give me a heads up about important topics needing my attention in the morning. Building it this way is wasteful but quick. It lets me focus on tuning how the messages are processed, ensuring I'm actually getting valuable insights in my morning briefing, etc. After using it for a day or two, I replace the MCP and polling approach with a lightweight script that handles the Slack API calls, timestamp tracking, and message filtering directly -- leaving just key insight extraction for the model. Replacing an LLM call with a script is low-hanging fruit. But inference shedding could mean training a small classifier to handle a routing step, building a rules engine from patterns the LLM surfaced, compiling the model's repeated judgments into a lookup table, or generating a schema once that deterministic validation can enforce thousands of times. The replacement matches the complexity of what you're shedding. At the scale of functionality in products, inference shedding can be a critical part of an agentic feature's "definition of done". Without it, you might be leaving substantial performance and maintainability value on the table, and the capability might be delivered on very shaky financial viability assumptions. 15
-
Ryan Rexroad Berkshire Hathaway… • 1K followers What if 200 Lines of Code Could Automate Your P&L Review? We manage 50 offices across two states, and I oversee the GL's for everything tech-related: phone systems, internet, email, printing, digital marketing, websites, and any other tech-adjacent services. Each month, I receive a spreadsheet with roughly 1,500 line items to review. Imagine if, instead of manually reviewing every single line, you could receive an automated summary highlighting anomalies, outliers, new vendors, and miscategorized expenses delivered straight to your inbox. That's exactly what I did using less than 200 lines of Python, combining traditional automation and AI-powered analysis. ⚙️ The Challenge Monthly spreadsheets with ~1,500 line items Multiple vendors and expense categories to monitor Manual review typically takes several hours 🛠 Step 1 – Classic Automation (40 lines) Goal: Instantly identify and alert about new vendors or miscategorized expenses. 🤖 Step 2 – AI-Enhanced Anomaly Detection (150 lines) Goal: Automatically flag expenses significantly outside historical norms. Summarize monthly spend totals Train an Isolation Forest on historical Vendor × Account data Automatically detect and explain anomalies clearly Example alert: "Latest Comcast bill for Account 62600-334 is 10% above its 6-month average." 💰 The Benefits Hours of manual review eliminated monthly Immediate identification of unexpected charges, saving thousands annually More accurate and actionable data for forecasting and budgeting Instead of complicated dashboards or costly RPA solutions, I built a simple yet powerful script. The best part is that this all happens behind the scenes, and I receive a neatly packaged email. Running similar automations? Let's share ideas and experiences 👇 3
-
Gabriel Bianconi TensorZero • 4K followers 20% confidence prediction I've had for a while: labs will stop offering API access to frontier models altogether to force people to use their proprietary products. Builders that aren't competing will be fine (OSS, previous-gen models, etc), but it'd very tough to compete directly in the niches they care about. 49 6 Comments
-
Nate Andorsky ForesightIQ • 18K followers If you're a founder building in the AI space, the smartest place to double down is on your prorietary data layer. Some call it the data moat, I call it the data GOAT. Ha. LLMs like Claude are making it trivially easy to connect to external data sources through protocols like MCP. Any tech literate person can now wire up an AI to pull from Slack, Salesforce, Google Drive, databases — whatever. In minutes. That means every AI startup whose value proposition is "we connect AI to your tools" is watching their differentiation evaporate in real time. If you're sitting on proprietary data that nobody else has — data you've collected, enriched, structured, and curated you're in a fundamentally different position. The AI infrastructure is actually working in your favor. Better models and easier integrations mean your unique data becomes MORE valuable and more accessible, not less. The easier these LLMs make it to access external data, the more valuable your data, and your company becomes. You can now sell access to your data regardless of how the end user accesses it. The reasoning layer is getting commoditized. The data layer isn't. 13 5 Comments
-
Yi Zhong Besimple AI • 3K followers A thread went viral this weekend alleging that Luel, a human data marketplace that just announced a $31.2M seed round, had public numbers that did not reconcile: datasets, ARR, contributor jobs, traffic, and compliance claims all being questioned from its own public surface. (https://lnkd.in/gAZQtXhX) Kled’s founder accused them of copying Kled and raised concerns around compliance, fraud-heavy traffic, and whether the company had actually built the trust layer required to handle people’s data responsibly. I don’t know the inside story at all. But the fact that someone was able to find out so much information about Luel's dataset through scraping is telling 🫠 A human data marketplace is not just “pay people to upload data.” The hard part is everything underneath: verifying contributors are real, preventing stolen or duplicate submissions, detecting AI-generated content, making sure the contributor can actually consent, removing accidental PII, protecting bystanders and minors, handing data securely, and giving buyers data they can trust. This is why we’ve spent so much time at Besimple on fraud detection and integrity. The product is not only data collection. The product is consent, provenance, quality control, and repeatable trust. 38 4 Comments
-
Amitt Singh Chouhan Raymetric Digital LLP • 3K followers California just launched the Delete Request and Opt-out Platform (DROP) , giving residents a centralized way to submit a data deletion request and opt out of data brokers in one flow. Instead of contacting hundreds of vendors individually, a single DROP request now reaches 500+ registered data brokers under the California Delete Act and CCPA privacy rights enforcement. Why ? Data brokers collect and resell personal data, including identity signals, device identifiers, location history, and behavioral profiles. That data fuels spam, fraud risk, enrichment pipelines, and AI training exposure. DROP restores control back to consumers. Residents verify eligibility through secure identity gateways and submit identifiers such as name, ZIP code, email address, phone number, and optional device signals for better matching accuracy. Brokers must process deletion requests, honor opt-out requirements, and report status under California Privacy Protection Agency oversight. Analytics and growth teams should treat this as a signal to audit third-party data dependencies now. Any enrichment pipelines pulling broker data need clear legal basis alignment with CCPA and CPRA. Consent frameworks, identity resolution logic, and downstream activation models will need tightening as broker supply contracts. This also reshapes how first-party data strategy compounds over time. Clean consent. Transparent value exchange. Minimal reliance on opaque enrichment sources. Teams that are already invested in disciplined data governance will move faster than those still leaning on shadow data flows. Platform access: https://lnkd.in/dY6D7F6d How will this reshape your data privacy strategy and consent operations? #DataPrivacy #CCPA #DeleteAct #ConsentManagement #PrivacyCompliance #MarTech #DataGovernance #DigitalTrust 13 1 Comment
-
Pete Jarvis rpv • 6K followers If you are actually interested in AI, ML (the technology that must not be named - chuckle) and other application specific hardware performance numbers, I highly recommend checking out Puget Systems blog posts. Why? Well, in an age where everyone is shouting, and therefore no one is listening the company do a remarkably good job of getting into the performance details of various hardware configurations. In short, they are worth listening to because the provide nuanced and detailed orientated prose. It is almost like they have thought about things and tested their assumptions. Chuckle. Case in point: https://lnkd.in/g6ak9m4Z PS. They build best in class application specific PC’s. 9 2 Comments
-
Dominique W. LitStage • 20K followers "Why we build" and "what we build" matter more than "how we build it" - Vision is more important than execution. Axios CTO Dan Cox says the AI shift is already here. One engineer recently rebuilt a project that took three weeks last year. This time, using AI agent teams, it took 37 minutes. Axios anticipated the shift and reduced its product and tech team from 63 to 43 people. Output doubled in January and is set to double again. Over two years, the team was cut in half while productivity more than doubled. AI tools like Claude Code and OpenAI Codex now ship features in days instead of months. The bottleneck is no longer coding. It is how fast humans can adapt. Cox says the new competitive edge is not speed. It is clarity of vision. Companies are not just shipping code anymore. They are shaping change. The AI reality is not coming. It is already here. #AI #Job #ClaudeCode #Codex #WhatToBuild #WhyWeBuild https://lnkd.in/g8CcSD7y 4
-
Christopher Lynch AtScale • 20K followers Welcome to AtScale, Jay Schuren. We’ve reached the point in this AI cycle where reasoning power isn’t the constraint. Operational trust is. As companies move from chat interfaces to autonomous systems, the risk profile changes. If an agent calculates revenue differently than finance does, that’s not a model issue — it’s an infrastructure issue. Jay has led organizations through this shift before. He understands what it takes to move from experimentation to operating standard. That’s why he’s here. Looking forward to building the next phase of enterprise AI together. Join our team: https://lnkd.in/eun6WaAr #snowflake #databricks #GBC #openai #anthropic #workday #salesforce #servicenow #datarobot #enterpriseai #aianalytics #agenticai 70 8 Comments
-
Dan Franco Adamas Knight • 21K followers I don't often post on LinkedIn, but felt compelled today. Mark Zuckerberg just hired Lucas Beyer, Alexander Kolesnikov and Xiaohua Zhai for his new SuperIntelligence team, led by Alexandr Wang. These aren't just any researchers, they were instrumental in setting up OpenAI's Zurich office. Here's why this is a master class on hiring and why Founders should pay attention: ➡️ This isn't just random headhunting. It's a calculated play to dominate AI through talent, whilst creating pressure for Meta's competition. First, let's start with the acquihire of Alexandr Wang, and Scale AI. Scale AI is crucial to training models, offering data annotation and model evaluation services. They even boast clients like the US Department of Defence. There are already reports of some of the other Mag7 dropping Scale AI, in fear of Scale exposing their research priorities to Meta. Now they have to shop around for other inferior solutions, brilliant move. Meta recently advertised signing bonuses of up to $100M for top AI researchers. I suspect most of this will be in stock options with long-term performance targets and vesting periods. This is where Meta has a competitive advantage. They currently have the best stock performance YTD among major tech companies (+21%), while others lag behind—Apple (-19%), Tesla (-19%), Google (-10%). Meta can afford this, and Zuckerberg is creating urgency in the market, forcing competitors to accelerate their hiring or rethink strategies. Some won’t keep up. ➡️ This isn’t just about a top researcher getting a $100M sign-on bonus, it’s about attracting “talent magnets,” whom Meta will build teams around for years to come. Mark Zuckerberg, unsatisfied with Llama’s performance, is taking back control and positioning Meta for dominance in the AI race. These moves typically take years to play out; Zuckerberg turned it around in weeks. This is why we love Founder led companies. This is the power of Founders. #AI #TechTalent #Hiring #Meta 69 2 Comments
-
Javier Tordable PAULING.AI • 21K followers I'm very grateful to Lisa Stiffler at GeekWire for covering what we're building at Pauling.AI. This article captures something important: we're not just making drug discovery faster, we're fundamentally changing its economics. When you can automate months of computational chemistry work into weeks and do it at a fraction of the cost, you unlock research that was never viable before. Rare diseases, longevity therapeutics, targets without billion-dollar markets behind them. The goal isn't just to go from 40 approved drugs per year to 400. It's to make sure a meaningful fraction of those 400 are the treatments that traditional pharma economics would never justify developing. That's the real opportunity with AI in this space, and we're just getting started. https://lnkd.in/gB8QWmf6 47 4 Comments
-
Daniel Demetri Gen • 8K followers I think the market reaction to the citrini thought piece was overblown. I think it’s missing the massively deflationary impact of a theoretical large scale replacement of white collar human labor. Such deflation could/would enable government spending to support aggregate demand. But is AI going to eliminate more white collar jobs than blue collar jobs? Signs definitely seem to point to yes. But those blue collars would be more likely to have Gucci and Prada labels? 9 3 Comments
-
Joshua Kelly 3K followers Two conclusions from my FHIR DevDays talk on building LLM Evals for FHIR: 1. You should build evals - it's not that hard - and it's worth it to be quantitive in your assessment of using LLMs to manipulate, generate, or transform FHIR resources 2. It's comparatively difficult to build evals that SOTA LLMs don't saturate in this domain - scoring 100% probably means we're leaving further gains on the table 14
-
Olowu Adebusoye OddX Group • 20K followers Last week was not about productivity. It was about signal acquisition. I spoke to 2 investors and 20 operators in infrastructure systems. Not for validation. Not for networking. For one reason: mapping control points in the systems that actually run execution. Most people optimize for output. I optimize for control surfaces. That difference changes everything. My girlfriend was on the calls with me. At some point she said: > “This is not normal… you’re not even talking about companies anymore, you’re talking about systems that replace companies.” Then she paused: > “I think you should log off.” I didn’t argue. Because she wasn’t wrong. She was just describing it from the outside. What stood out 1. Investors don’t price ideas anymore — they price leverage density Both conversations kept converging on one question: > Where does compounding actually sit in your model? Not in branding. Not in products. In infrastructure ownership. Capital is not the bottleneck anymore. Execution environments are. 2. Operators are the real scaling layer I spoke to 20 operators across different infrastructure layers. Same pattern everywhere: Systems still rely on manual glue “Automation” is mostly cosmetic Teams are not built for load, only for launch Operators are not failing. Systems are failing operators. That’s the real constraint. My position is simple I am not building companies in the traditional sense. I am focused on: Acquisition of infrastructure systems Design of execution environments Deployment of unlimited operators inside those systems Because the constraint is not talent. It is whether talent can operate infinitely without system collapse. The shift most people miss Everyone is building applications. Very few are building: execution OS layers simulation environments coordination infrastructure data manufacturing systems the backbone other systems depend on That is where compounding actually lives. Everything else is interface work. What last week confirmed Investors confirmed one thing: capital is looking for structured infrastructure exposure. Operators confirmed another: execution is bottlenecked by system design, not effort. Put together, it reinforces the direction: > Acquire infrastructure. Structure operators. Scale execution beyond human limits. I am not optimizing for companies. I am optimizing for systems that generate companies as a side effect. And that gap is where the next wave of value sits. It is already forming. 4
-
Bogdan Knezevic Kaleidoscope.bio • 6K followers Jotted down some end of year biopharma reflections while waiting to board a flight. Likely no surprises. 2025 was tough, but there's a hint of currents shifting for the better 🤞 1. BIG gulf between what people think scientists are clambering for, when it comes to AI tools, and what scientists are actually asking about, using regularly, and willing to pay for. "LLM for this, agent for that" really misses the deeper-rooted challenges and bottlenecks. 2. Lots of killing of discovery, shifting emphasis to later stage assets. Even saw this at orgs where the discovery engine is working (producing clinic stage assets that continue to get good clinical readouts). I understand this reactivity to markets and investor sentiment, but it’s nevertheless sad to see strong scientific engines be shut off, teams laid off, and novel discovery stopped as a result. TBD what longer term effects will be over next several years. 3. It often takes the experience of having gone through a cycle to realize what problems you want to avoid at all costs. Our most well-aligned and motivated champions have been people who directly experienced the painful alternatives ('no action' or 'build-it-yourself'). Conversely, those who haven't had to grapple with the problems before often maintain a "we can just do everything ourselves" stance. 4. A lot of work is being outsourced. When managing these CRO relationships, complexity can balloon quickly. Our partners have increasingly turned to Kaleidoscope.bio to drastically streamline this pain. 5. There is painful disconnect between how much time people waste on preventable stuff, and how much time/budget/awareness leadership will provide to address this. I encourage leaders to empower their team to solve problems that will help them move faster, even if they as a senior leader may not deal with the day-to-day (and thus may not feel it directly). 6. Seems to be an increasing number of scientific PMs spearheading operations (we at Kaleidoscope like this). 58 7 Comments
-
Roy Mariathas MBBS FRACGP The Royal Australian College… • 5K followers Quietly been working on this and happy to say I've published my first pre-print on arXiv, "Decomposing Physician Disagreement in HealthBench," in a collab with Satya Borgohain, a senior AI engineer , ex-Relevance AI. For those who aren't familiar, HealthBench is OpenAI's benchmark for evaluating how well large language models handle healthcare conversations from patients and physicians. 262 physicians reviewed AI-generated medical responses and labelled whether they met specific clinical criteria. I was one of them. Interestingly, physicians disagreed on the labels 22.5% of the time. These physician labels are the foundation for measuring AI performance in healthcare. If doctors disagree on whether an LLM's response is correct, it's effectively the ceiling on how accurately you can score the LLM. When you force those disagreements into a single "correct" answer, you can't tell the difference between the model getting it wrong and doctors just seeing it differently. So understanding why doctors disagree tells you something important: is the problem fixable (better rubrics, more complete clinical scenarios) or is it structural (medicine is just ambiguous at the edges and always will be)? That's what our paper set out to answer. We analysed 60,896 physician judgments across 29,511 cases. We thought we'd find that certain doctors or specialties were driving the disagreement. We didn't. Doctors grade remarkably similarly to each other. The variance comes from the cases and not people. Doctors agree when the response is clearly good or clearly bad. The disagreement spikes on the ones in the middle, the responses that could go either way. When the patient or clinician prompt is vague or missing key details, it becomes much harder to evaluate whether the moidel gave a good response. Doctors' odds of disagreeing more than double. But when the clinical scenario is genuinely ambiguous, where medicine itself doesn't have a clear answer, that doesn't increase disagreement at all. So the biggest challenge is that the conversations going in are often incomplete and this shapes how reliably you can measure the quality coming out. We think this matters because there's movement in the space. In January, OpenAI launched ChatGPT Health, Anthropic launched Claude for Healthcare at JPM. This month Microsoft launched Copilot Health connecting to 50,000+ US hospitals, and Amazon Web Services (AWS) launched Amazon Connect Health with AI agents plugging directly into EHR systems. Every major AI lab is now building dedicated health products. All of them will need to evaluate whether their medical outputs are actually good, which will depend on benchmarks built on physician labels. Our paper shows those labels carry more disagreement than most people realise. Link to arXiv paper and Satya's blog in comments. PS. Thanks to everyone that left a message about Reggie ~2 weeks ago. The response was overwhelming, meant a lot. I owe a lot of people a reply. 🙏🏽 42 11 Comments
-
Julia Yukovich AICU • 8K followers Multimodal data isn't the future - it's the NOW. Here's how we're scaling explainable AI in regulated industries. Deploying compliant AI in biotech, medtech, or pharma? Here’s the process I use at AICU to make it work - without losing sleep over regulatory hurdles. 🚀 Step 1: Match your data sources to your compliance needs. → Start by mapping out every data type (documents, images, lab results, device data). → Check what each regulation (GDPR, ISOXYZ, HIPAA, you name it) asks for. → Build clear links between your data and your compliance rules. (No guesswork. No “we’ll fix it later.”) Step 2: Integrate explainability audits into your pipeline. → Every AI output - whether from a document, an image, or a combined workflow - needs a “why.” → Add explainability checks at each step. → Document what the model did, what data it used, and how it made its choices. (Auditors love this. Scientists too.) Step 3: Set up continuous improvement loops. → Don’t treat compliance as a checkbox. → Review real-world outcomes, user feedback, and audit results often. → Tweak, retrain, and update your workflows. → The best systems learn and get safer over time. When we rolled this out, the results surprised even me: - Faster validation cycles. - Less manual paperwork. - More trust from partners and regulators. - Teams could focus on research, not compliance headaches. Multimodal AI is ready for prime time—if you build it right. How are you approaching compliance and explainability in your AI projects? Curious to hear your stories from the field! 32 4 Comments
-
Sasha Dagayev 🇺🇦 Seqera • 3K followers Fantastic talk from Phil! What was most exciting for me was also Phil's point about also taking the re-writes as an opportunity to rewrite docs. If you haven't yet - check out the new docs accompanying the re-write of TrimGalore (https://lnkd.in/eQ3Sv8PQ) - they look amazing! 18 1 Comment
-
Niall Murphy 6K followers YellowDog.ai just set a 10x benchmark uplift in scale computing, delivering 40,000 tasks per second (TPS) and managing 100,000 compute nodes in the cloud. What's even more interesting, that's 2x IBM Symphony and opens an intriguing pathway for these until-know closed/captive systems. 11
Show more posts Show fewer posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.