Iker Huerga

Thoughts on this? >> NVIDIA AI Research Unveils ‘Star Attention’: A Novel AI Algorithm for Efficient LLM Long-Context Inference >> Comment below! >>> lqventures.com #AI #IoT #mhealth #industry40 #healthtech

1

Study published in Nature doesn't conclude that LLMs are not ready for clinical decision-making https://lnkd.in/ekRZGyjf The study highlights the limitations of some open-source large language models (LLMs) in clinical decision-making. It’s not surprising they struggle, as they underperform humans. Exciting times ahead: proprietary models like GPT-4, which outperform humans on medical information understanding benchmarks, should be tested next! Chart with MedQA MSLE accuracy. The top evaluated model Meditron significantly underperforms human experts whereas as GPT-4 overperforms https://lnkd.in/edM8dB5H

6

Nothing new here. This result is obvious to statisticians. Statistics is the engine of the AI-mobile. So LLM problems directly arise from classic statistical problems like selection bias (drift, lack of "fairness", non-random sampling) and spurious correlation (confounding, reverse causation, leakage).

16

☕ Freshly Brewed Quantum News from The Daily Qubit x The Quantum Insider ⚆ Independent researchers used quantum machine learning to classify ten types of cancer and identify biomarkers, showing QML's advantage in handling complex biological data. ⚆ EDF is collaborating with Pasqal to apply quantum computing in energy forecasting, material simulation, and smart electric vehicle charging. ⚆ PhD student Luke Mortimer ported the classic game DOOM to quantum computers, dubbed "Quandoom." Although current quantum hardware can't meet the steep 72,376 qubit requirement, the game runs on simulators at 10-20 FPS ⚆ Scientists from Sorbonne Université, NAVAL GROUP, and QC Ware Corp. developed a quantum convolutional neural network using Hamming-weight circuits, improving image classification with faster processing and fewer parameters. ⚆ On the most recent episode of The New Quantum Era, host Sebastian Hassinger joins Martin Schulz, Professor at Technical University of Munich and board member of the Leibniz Supercomputing Centre Center, for a discussion on integrating quantum and classical computing for hybrid infrastructures and resource scheduling. ⚆ QubiCSV, is an open-source platform developed by scientists at Berkeley Lab and the University of Massachusetts Amherst, which helps manage and visualize quantum calibration data with real-time collaboration tools. ⚆ Quantum Computing Inc. extended its agreement with Los Alamos National Laboratory to continue developing its Dirac-3 machine for solving complex optimization problems ⚆ Archer Materials Limited developed a scalable carbon film with 400+ nanosecond electron spin lifetimes, which may improve quantum device manufacturing and support quantum technology development through standard semiconductor techniques ⚆ Scientists from the Federal University of Rio de Janeiro and the Technology Innovation Institute in Abu Dhabi introduce a hybrid quantum-classical framework for simulating matrix functions, reducing runtime and noise sensitivity on early quantum hardware ⚆ Arqit Quantum raised $13.6 million through a stock sale and plans to use the proceeds for corporate development. Don't miss a single qubit. Read more at https://lnkd.in/gcXNfGFM #quantumcomputing #quantumtechnology #innovation #quantumscience #quantumphysics

12

Recursion's lead drug for cerebral cavernous malformation just hit the primary endpoint in a Phase 2 trial, proving safe and tolerable for patients compared to a placebo. But investors closely watching the AI drug developer may not be satisfied with the efficacy showings. Leading up to the announcement, executives had positioned the trial as a modest milestone, with some observers stating that Recursion was attempting to temper expectations for what was an unusual, likely-to-fail drug candidate. What do you think of these trial results? Do they alter your opinion on AI-influenced drug development? #biotech #drugdevelopment #AI #artificialintelligence #machinelearning #clinicaltrial #lifesciences

31

1 Comment

Pop Quiz--How well do you know your Nobel Prize Physicists? A Google Books Ngrams project showed that while fame is becoming more and more fleeting for some, for others, it's non existent. That's the mathematicians and physicists who don't seem to be appreciated by the public. Let's hope for more than just a passing moment of recognition for this year's Nobel Prize winners: John Hopfield (with his many contributions in chemistry, physics, and AI) and Geoff Hinton (who has been more publicly active in warning about the potential dire consequences of AI)! And here's a chance to recognize a few more scientists: How many other NP Physicists can you recognize? (radiologists should at least recognize the famous x-ray of the wife's hand and ring! and of course, not shown in the quiz, Paul Lauterbur, Sir Peter Mansfield, Sir Godfrey Hounsfield, II Rabi, and other MR/CT winners). #ISMRM https://lnkd.in/ey4H5FH9 Geoffrey Rubin, MD, MBA, FACR, Robert Edelman, Neil M. Rofsky, MD, MHA, Carl E Ravin, Derek K Jones, MBE, ISMRM, Garry Gold, Margaret Hall-Craggs, Vikas Gulani, Scott Reeder, MD, PhD, Dan Sodickson, Mark Griswold, Mitsue Miyazaki

13

3 Comments

🔮 What underwrites that power, allowing your computer or phone to run any program you (or anyone else) might conceive, without needing to know it in advance? Or what if I told you there’s no process in this universe—no matter how wild or elaborate—that your mind can dream up yet a computer can’t, in principle, execute? It’s the principle of universality, laid down by Alan Turing, which holds that a single, universal machine can in principle execute any computable process that exists—even those not yet imagined. Alan Turing’s great insight was that a single, universal machine can, in theory, handle any computation you can dream up. Just think: an infinite tape, a read/write head, and a few states, all woven into a mechanism so fundamental that it could emulate any conceivable process in the universe. In that sense, the question of Artificial General Intelligence and the limits of our machines feels almost beside the point. If the universe itself mathematically supports universal computation and building a universal computer with in it —and that certainly appears to be the case—then any process that can exist in reality can, in principle, be computed. Now picture quantum computing, where David Deutsch’s universal quantum Turing machine widens that horizon even further to multiverses. Entanglement and superposition bring a cosmic dimension to universality, so much so that each idea you conceive could be mirrored in some unfolding quantum reality. Deutsch’s Many-Worlds perspective (that Google mentioned in their latest Willow paper) suggests that when you run a quantum algorithm, you might be enlisting entire parallel universes to help carry the computational load. Ultimately, the spark isn’t just in Turing’s or Deutsch’s contributions; it’s in recognizing that these breakthroughs bring you closer to the heart of creation itself. When you hold a phone or sit at a computer, you’re not just wielding a tool—you’re tapping into a universal engine that can reflect and shape ideas as limitless as your own capacity to envision them. And in that realization lies a humbling sense of wonder about what might come next. Here is a fascinating conversation I listened to over the weekend that delves in to some of these deeper insights that makes quantum computing possible How do you link computation to consciousness? What is computational universality?

11

Derek Lowe continues his examination of AI-based drug discovery by comparing the fraction of AI-discovered drugs that passed Phase 1 and 2 trials with those of conventionally discovered drugs, finding little difference. Lowe’s previous articles demonstrated that AI only impacted on a small percentage of the drug development process, just one part of #drugdiscovery. The main part of the drug development process is three phases of trials, which take years to carry out and evaluate. Therefore, the issue is whether AI helps discover drugs that have a higher chance of passing the trials than do conventionally developed drugs. Lowe analyzed a paper written by a “prominent consulting firm.” The paper matches the AI-focused companies (114!) with clinical projects that are reported for them (39). Those clinical projects included AI-discovered targets, small molecules, biologics or vaccines discovered or optimized via AI techniques, or drugs that have been repurposed through AI techniques. For phase 1 trials, he looked at the compounds and targets (listed in article), and what he found “is that in almost every case, the targets were already known to be implicated in the disease under investigation. In some of these examples, in fact there are several drugs already in the clinic targeting the same proteins, or even therapies that are already on the market working through the same mechanisms (C. diff toxin B, e.g.).” “There are some really interesting things on the list, but I do not see how any of them can be classified as "target discovered by AI". Nevertheless, the paper says “of the 24 therapies that have reported Phase I, 21 were successful. The industry standard success rate for Phase I is 66% for all indications and 76% for lead indications, so while the AI-based examples here might be getting through Phase I at a higher rate, both the sample size and outright differences are too small, in my opinion, to make that claim.” “Meanwhile, four out of ten have succeeded in Phase II, which for what it's worth (an even smaller sample!) is exactly the same as for non-AI compounds. And frankly, this is where I'd expect the numbers to be the same, because Phase II success is all about picking the right targets, and (as mentioned above) these were all already targets that people go interested in the old-fashioned way, not because the AI picked them out of the ether and said to go for them. Why shouldn't they fail at the same rate as everybody else's stuff? “So let's watch these number.” “For now, I am not convinced that issuing press releases about your compounds that talk about their discovery through #AI techniques is sufficient to expect greater things from them. I would like for that to be true, but I'm not ready to say that it is yet.” #technology #innovation #hype #science #universities #research #artificialintelligence https://lnkd.in/ebcuEDgN

65

13 Comments

AI in Drug Discovery: Disproving the Hype Skeptics This New Yorker article offers compelling evidence that AI is far from just hype, especially in drug research: ➡️Tangible Breakthroughs • Rapid Results: AI has dramatically accelerated the drug discovery process, screening millions of compounds in days instead of years. • Novel Discoveries: Machine learning algorithms have designed entirely new molecules with potential therapeutic properties. ➡️Real-World Impact • COVID-19 Response: AI tools played a crucial role in rapidly identifying potential treatments and vaccine candidates during the pandemic. • Repurposing Existing Drugs: AI has successfully found new applications for approved medications, saving time and resources. ➡️Beyond Speculation • Practical Applications: Unlike speculative AI hype, drug discovery showcases AI's concrete, life-saving potential. • Industry Adoption: Major pharmaceutical companies are increasingly integrating AI into their R&D processes, validating its effectiveness. This progress in drug research demonstrates that AI is not just buzzwords but a powerful tool delivering tangible results in one of the most critical fields of science. What's your take on how AI is reshaping the pharmaceutical industry? How do you think this impacts public perception of AI's potential?

19

10 Comments

What will happen with all the data at biopharma companies? Will there be incentive to share or will data access differentiate new startups? Researchers at the intersection of AI and biotech discussed this question and more at a panel at #LSINW24. Read my story below to hear from panelists Danielle Greenawalt, Lynda Stuart MD PhD, Sean McClain and Jonathan Cohen, speaking at Life Science Innovation Northwest 2024. #AI #drugdevelopment #biotech

36

Does AI already have a spending problem? Pundits from Goldman Sachs to Gartner are focused on AI ROI. Gartner, in its annual hype cycle, described businesses shifting focus "from excitement around foundation models to use cases that drive ROI." Goldman Sachs analyst Jim Covello, the head of global equity research, is more cautious on the potential to achieve an AI ROI in the first place. “AI technology is exceptionally expensive,” he said. To me, such questions are like saying "math is hard." It can be. Depends on what you are calling "math." But yeah, powerful GPUs aren't cheap, but they happen to be really good at parallel math. The other aspect to all of this is that spending more on data/tech/domain/business talent could help most organizations get a clearer ROI for whatever their tech investment is. Just throwing deep learning at random problems isn't likely to end up with much of an ROI.

7

Very good (16 min) summary on the state of play of the GenAI industry boom. 20/15 Visioneers #genai #llm https://lnkd.in/e-FDtWjR

1

A new framework to detect hallucinations, or false information, in text generated by large language models (LLMs). Hallucinations occur when LLMs generate plausible-sounding but factually incorrect content. The proposed framework aims to identify such hallucinations by cross-checking the generated text against external knowledge sources. 1️⃣ LLMs can generate fluent and coherent text, but may include hallucinated content that is not factually accurate. 2️⃣The framework uses a two-step process: 1) Identify factual statements in the generated text, and 2) Verify those statements against external knowledge sources. 3️⃣For verification, the framework leverages existing question-answering models and knowledge bases to cross-check the factual claims. 4️⃣If a statement cannot be verified or contradicts the external sources, it is flagged as a potential hallucination. 5️⃣ Experiments on various datasets showed the framework could effectively detect hallucinations with high precision and recall. https://lnkd.in/eWD_9Mkw #ai #ArtificialIntelligence #LLM #LLMHallucination

11

5 Comments

Deep Genomics is the latest drug company to announce that AI hasn’t helped them to develop drugs despite testing more than 200 million molecules proposed by AI. Today, it “has zero drugs in clinical trials and many of its plans have blown up. The company halted its Wilson disease program, ditched dozens of its machine learning models, appointed a new chief executive and is pursuing a different approach to using AI. It’s also open to a sale.” “AI was supposed to revolutionize drug discovery. It hasn’t. Not yet, anyway. Machine learning promised to speed up a lengthy and fraught process, and achieve breakthroughs beyond the capabilities of the human mind. But there are no drugs solely designed by AI in the market today, and companies that have used AI to assist with development have suffered setbacks.” Deep Genomics founder Brendan Frey says: “AI has really let us all down in the last decade when it comes to #drugdiscovery. We’ve just seen failure after failure.” The blunt assessment is shocking coming from him because “he has been studying machine learning since the 1990s, and he is the co-author of more than 200 papers and a co-founder of the Vector Institute, a nerve centre for AI research in Canada.” Other companies are reaching similar conclusions. “An eczema drug developed by U.K. biotech BenevolentAI failed to reduce itch and inflammation in trials last year, while Exscientia, another British firm, shuttered trials of its cancer treatment. Recursion Pharmaceuticals in the U.S. said that it would find 100 clinical candidates in its first 10 years. A decade later, it has five in trials.” “If you rewind the clock 10 years, there was a lot of hubris, a lot of #hype and a lot of big claims that were made that have perhaps not played out in reality,” said Daniel Cohen, co-founder of Valence Discovery, which was spun out of Mila, Quebec’s AI institute, and acquired by Recursion last year. “But I think that’s necessary in creating a new way of doing things.” Even if AI-based drugs succeed, it’s not clear “just how much time AI can save during the development process. “I’m not doubting that it can get us there a little bit quicker,” said the CEO of a Toronto investment bank focused on healthcare. “But in the 10-year journey of drug development, you may be saving five months.” In conclusion, the article says: “The setbacks in #drug development could hold lessons for other industries hoping that #AI will unlock massive #productivity gains. Molecular biology is immensely more complicated than building a chatbot, of course, but new #technology is anything but straightforward. There are unwelcome surprises, inflated expectations and dashed hopes.” Very consistent with other assessments of AI for drug development, many of which I have posted. #innovation #startups #artificialintelligence #biotech #science #research https://lnkd.in/gsBQf2ub

391

97 Comments

Good to join Daphne Koller and Derek Lowe to discuss our perspectives on hype vs. reality for AI in drug development. The key challenge from my perspective is a mindset shift. Researchers must think about biology as an engineering problem and a data problem. Casey Ross and Matthew Herper highlighted our discussion in STAT. Read more at https://lnkd.in/gvHQZ_m8 

65

1 Comment

That generative AI content would eventually lead to model collapse has been widely discussed over the last year. In a recent study from Oxford University by Ilia Shumailov and Zakhar Shumaylov they illustrated this on sample models consisting of variational autoencoders (VAEs) and Gaussian mixture models (GMMs). Perhaps a salient question is how AI-generated content can be identified to prevent future models from being "contaminated." Without some kind of standard, this becomes a value judgment, as any data that used an algorithm or data selection could be considered AI-generated, depending on the definition of AI. Maybe a simple filter for fake news or true informational content is worth investigating; a straightforward fact-checking system could prevent any hype around model collapse. Or indeed, checking whether the data originated from a large multi-billion parameter generative model. AI models collapse when trained on recursively generated data https://lnkd.in/eAykANGW The Curse of Recursion: Training on Generated Data Makes Models Forget https://lnkd.in/eGUfAMWJ AI produces gibberish when trained on too much AI-generated data https://lnkd.in/eBKNENry

20

1 Comment

There are a growing number of reports indicating that the genAI progress may be slowing as the costs to train new models increases. More specifically, The Verge reported that the next iteration of Gemini was not meeting the dev team's hopes. The Information notes that OpenAI, too, may be encountering slowing progress, and that its latest model would represent a smaller upgrade than the huge gulf between GPT-3 and GPT-4 (which launched all the way back on March 14, 2023). And then there are rumors swirling around Anthropic, too, which recently launched a newer version of its Sonnet 3.5 model, but quietly revised its plans to debut Opus 3.5. Or maybe scrapped the plans to release it altogether. https://lnkd.in/gSGgVr5a https://lnkd.in/gEjegwyA

8