Joseph B. Elad

LLMs DON’T JUST “KNOW” THINGS. THEY DO THINGS. THAT’S THE POINT. A recent conference paper argued that LLMs cannot produce real knowledge — that they’re just “sophisticated mirrors” reshuffling word patterns, two steps removed from reality. I respectfully but firmly disagree. Not because the philosophical point is wrong, but because the argument is fighting a 2022 version of AI while the 2026 version has already moved on. 1. LLMs are no longer trapped inside text. The claim that LLMs only “absorb what we’ve written” ignores that today’s systems have web search, code execution, tool use, multimodal vision, and real-time data access. They verify claims against live sources and run calculations against actual datasets. Calling them “a summary of a summary” hasn’t been accurate for two years. 2. “Probabilistic” doesn’t mean “unreliable.” Google DeepMind’s AlphaEvolve — using Gemini LLMs paired with evolutionary algorithms — has produced new discoveries in mathematics and computer science. Tested on 50 open math problems, it rediscovered state-of-the-art solutions 75% of the time and found improved solutions 20% of the time. It discovered a method to multiply 4×4 complex-valued matrices using 48 scalar multiplications — the first improvement over Strassen’s algorithm in 56 years. Probabilistic systems producing provably correct, novel results. That’s not hallucination. That’s discovery. 3. AI is already generating clinical-grade knowledge. AlphaFold revolutionized protein folding and earned its creators the 2024 Nobel Prize in Chemistry. AI-designed therapeutics are now in human trials, with positive Phase IIa results for the first fully AI-designed drug. AI-developed drugs show 80–90% Phase I success rates versus ~40% for traditional methods. These aren’t word patterns. These are molecules going into human bodies and working. 4. The real question isn’t “Can LLMs know things?” — it’s “Can we build systems that produce reliable knowledge?” The answer is yes — when you pair LLMs with verification layers, domain-specific fine-tuning, and human expertise. The debate shouldn’t be philosophical. It should be practical. The real risk isn’t trusting LLMs too much. It’s dismissing them too early — and losing years of competitive advantage while debating whether a machine can “really” understand. At Applied AI Asia, we build AI systems for healthcare, agriculture, and education across Asia-Pacific. Every day we see LLMs helping doctors catch diseases earlier, farmers make better decisions, and students learn faster. Not because AI “knows” — but because it augments human knowledge with measurably better outcomes. The organizations that will thrive aren’t asking “Can AI know things?” They’re asking “How do we build AI systems that make our people smarter?” That’s the real question for 2026. Benito Teehankee Dominic Ligot #nemoclaw #googlemind #aiducate #appliedAI #appliedagenticAI

2

2 Comments

𝗪𝗵𝗮𝘁 𝗮𝗻 𝗲𝘅𝗰𝗶𝘁𝗶𝗻𝗴 𝘄𝗲𝗲𝗸 𝗳𝗼𝗿 𝗵𝗲𝗮𝗹𝘁𝗵𝗰𝗮𝗿𝗲 𝗮𝗻𝗱 𝗔𝗜. The announcements from Anthropic and OpenAI are reshaping how we think about enterprise healthcare solutions. One pattern stands out: 𝗟𝗟𝗠𝘀 + 𝗧𝗼𝗼𝗹𝘀 + 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 = 𝗣𝗼𝘄𝗲𝗿𝗳𝘂𝗹 𝗦𝘁𝗿𝗮𝘁𝗲𝗴𝘆 When you combine LLMs with tools (MCP Servers) that access enterprise and reference data, plus structured knowledge assets (Claude Skills) representing industry expertise—something fundamental shifts. Software solutions become 𝘦𝘱𝘩𝘦𝘮𝘦𝘳𝘢𝘭. If you're managing the right data and curating the right knowledge, you can reimagine and recreate solutions as needed. As I wrote previously: "Measure continuously, cut as often as you need." ━━━━━━━━━━━━━━━━━━━━ 𝗧𝗼𝗱𝗮𝘆 𝗜'𝗺 𝗮𝗻𝗻𝗼𝘂𝗻𝗰𝗶𝗻𝗴 𝗛𝗲𝗮𝗹𝘁𝗵𝗦𝗶𝗺 𝗔𝗴𝗲𝗻𝘁 𝗶𝘀 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲 𝗼𝗻 𝗚𝗶𝘁𝗛𝘂𝗯. This is a new version of HealthSim—and a proof point of the architecture above. I reused the same tools and knowledge artifacts from my earlier HealthSim Workspace project (an additive distribution for Claude Desktop). This version? A completely standalone agent built on the Anthropic Agent SDK. 𝗪𝗵𝗮𝘁 𝗶𝘁 𝗽𝗿𝗼𝘃𝗶𝗱𝗲𝘀: → 𝗥𝗶𝗰𝗵 𝗿𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗱𝗮𝘁𝗮 via MCP tools (NPPES providers, CDC health indicators, Social Vulnerability Index) → 𝗘𝗺𝗯𝗲𝗱𝗱𝗲𝗱 𝗵𝗲𝗮𝗹𝘁𝗵𝗰𝗮𝗿𝗲 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 published as Skills (similar to Claude for Healthcare) → 𝗦𝘆𝗻𝘁𝗵𝗲𝘁𝗶𝗰 𝗱𝗮𝘁𝗮 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 engine for cohorts and patient journeys across: • Clinical/EMR (PatientSim) • Health plan claims (MemberSim) • Pharmacy benefits (RxMemberSim) • Clinical trials (TrialSim) • Population health (PopulationSim) • Provider networks (NetworkSim) ━━━━━━━━━━━━━━━━━━━━ 𝗪𝗵𝗮𝘁'𝘀 𝗻𝗲𝘅𝘁? I have a roadmap in planning, but I'm genuinely interested in hearing what 𝘺𝘰𝘶 would find valuable. My vision: a cloud-based 𝗛𝗲𝗮𝗹𝘁𝗵𝗦𝗶𝗺 𝗦𝗲𝗿𝘃𝗶𝗰𝗲 for those who want the capabilities without managing local LLM tooling. 🔗 𝗚𝗶𝘁𝗛𝘂𝗯: https://lnkd.in/g5RB9uU6 What healthcare AI use cases are you most excited about? #HealthcareAI #LLM #SyntheticData #Anthropic #OpenAI #HealthIT #MCP

16

2 Comments

🔴 𝐖𝐞 𝐭𝐚𝐥𝐤 𝐚𝐛𝐨𝐮𝐭 𝐀𝐈 𝐞𝐭𝐡𝐢𝐜𝐬, 𝐡𝐚𝐥𝐥𝐮𝐜𝐢𝐧𝐚𝐭𝐢𝐨𝐧𝐬, 𝐚𝐧𝐝 𝐜𝐡𝐚𝐭𝐛𝐨𝐭 𝐫𝐞𝐠𝐮𝐥𝐚𝐭𝐢𝐨𝐧. 𝐁𝐮𝐭 𝐰𝐡𝐨 𝐢𝐬 𝐭𝐚𝐥𝐤𝐢𝐧𝐠 𝐚𝐛𝐨𝐮𝐭 "𝐊𝐢𝐥𝐥𝐛𝐨𝐭𝐬"?⁣ ⁣ In my post mid-June, I raised an uncomfortable question about while the world obsesses over AI misuse and regulating AI-generated images or ChatBots, where is the global debate on autonomous AI weaponry used in actual wars or so called KillBots?⁣ A few days ago, The National News published a chilling investigation about Israel’s use of AI systems like 𝐋𝐚𝐯𝐞𝐧𝐝𝐞𝐫, 𝐆𝐨𝐬𝐩𝐞𝐥, and 𝐖𝐡𝐞𝐫𝐞’𝐬 𝐃𝐚𝐝𝐝𝐲?tools that assign "𝐬𝐮𝐬𝐩𝐢𝐜𝐢𝐨𝐧 𝐬𝐜𝐨𝐫𝐞𝐬" and, with minimal human oversight, trigger deadly strikes in Gaza. The reported 20-second “human confirmation loop” seems barely more than a fig leaf.⁣ ⁣ Thousands of civilians like 34-year-old Mohsen Obeid’s entire family are collateral to algorithms optimized for efficiency, not compassion. Facial recognition, gait tracking, social media activity, VPN usage and 𝐚𝐥𝐥 𝐝𝐚𝐭𝐚 𝐩𝐨𝐢𝐧𝐭𝐬 𝐭𝐡𝐚𝐭 𝐜𝐚𝐧 𝐥𝐞𝐚𝐝 𝐭𝐨 𝐚 𝐝𝐞𝐚𝐭𝐡 𝐬𝐞𝐧𝐭𝐞𝐧𝐜𝐞. No trial. No transparency.⁣ ⁣ ⚠️ This is not science fiction. This is now.⁣ ⁣ 🔍 If AI can now decide who lives and who dies based on data-fed suspicion, we must urgently rethink 𝐰𝐡𝐚𝐭 "𝐡𝐮𝐦𝐚𝐧 𝐢𝐧 𝐭𝐡𝐞 𝐥𝐨𝐨𝐩" 𝐫𝐞𝐚𝐥𝐥𝐲 𝐦𝐞𝐚𝐧𝐬. The Geneva Conventions, ethics boards, and global regulators are lagging dangerously behind.⁣ ⁣ Let us not just regulate 𝐡𝐨𝐰 AI speaks but 𝐰𝐡𝐨 it targets.⁣ ⁣ 🛑 If we fail to regulate lethal autonomous systems that I call Killbots, we risk unleashing a new era of unaccountable warfare. 𝐈𝐭'𝐬 𝐭𝐢𝐦𝐞 𝐰𝐞 𝐬𝐭𝐨𝐩 𝐥𝐨𝐨𝐤𝐢𝐧𝐠 𝐚𝐰𝐚𝐲.⁣ ⁣ #AIethics #Killbots #LavenderAI #AutonomousWeapons #HumanRights #ResponsibleAI #Gaza #WarandTechnology #EthicalAI #HumanInTheLoop #GenevaConventions #TechForGood https://lnkd.in/eYiewzRy

18

4 Comments

NIH-funded AI model reads CT scans, predicts disease onset years early Stanford’s Merlin (Nature): 750+ tasks on 3D abdominal CTs from 15k scans/1M codes; single-GPU trained. 75% accuracy spotting high-risk diabetes/osteoporosis/heart dz 5yrs ahead (vs 68% baseline). Zero-shot chest CT rivals specialists.

5

🤖 How did the PRIDIT R package get built? Here's the behind-the-scenes story... The PRIDIT methodology has been available in my published research, and I've shared implementations with colleagues who requested them. But creating a robust, CRAN-ready R package? CRAN's rigorous quality standards require expertise I didn't have. Enter Claude.ai: 🔹 Helped structure the package correctly 🔹 Generated proper documentation 🔹 Created comprehensive examples 🔹 Fixed CRAN submission issues 🔹 Made the code more robust The collaboration was fascinating: 🔹 I provided the domain expertise and mathematical foundation 🔹 Claude handled R package best practices and CRAN requirements 🔹 Together, we created something neither could have done alone Key takeaway: AI isn't replacing domain experts - it's amplifying our ability to share our work with the world. The PRIDIT method (based on decades of statistical research) is now accessible to any R user thanks to modern AI assistance. Sometimes the best way forward is knowing when to ask for help! 🚀 Package: https://lnkd.in/eKFNdqkT #AIAssisted #RStats #PackageDevelopment #Collaboration #Innovation

10

2 Comments

Meet Nea Korvenlaita, Postdoctoral Research Fellow at Johns Hopkins Medicine. When she’s reading papers, drafting study plans, or doing peer review, her notes used to scatter across apps. With reconfigured, those highlights and scribbles become organized, searchable knowledge, automatically. Result: multiple hours saved every week - time she now spends on the science, not the scavenger hunt. Do your future self a favor (and your teammates, too). Get the app: reconfigured.io

22