John Milinovich

Anthropic published research on what they call the “Assistant Axis” — basically a measurable direction in a model’s internal “persona space” that correlates with how “assistant-like” the model is. (https://lnkd.in/e_cNBWpy) For builders working on AI agents: Persona drift is a real problem. Models can drift and forget who they were meant to be (remember when you could ask a major automotive brand’s AI bot to do your math homework for you?) If you ship LLMs in production: would you rather (a) add more post-training rules, or (b) enforce a mechanistic “stay in character” constraint like this, or (c) never adopt a closed source solution? #ai #llm #alignment #safety #mlops

1

In a previous post, I discussed how #AgenticAI is removing the “steps” from our interfaces. When agents collapse multi-step workflows into hidden operations behind the scenes, flows don’t disappear—they go underground. That changes what designers actually need to design. In this post, I try to answer three questions: - If agents now “do the doing,” what is the designer’s job beyond drawing screens and flows? - How can we treat encapsulation not just as an engineering pattern, but as a governance construct that makes agent authority, guardrails, and oversight explicit? - What concrete artifacts and #metrics can teams use to turn vague ideas like “good agent behavior” and “trustworthy AI” into shared, testable #requirements? Curious how other teams are tackling this: are you still designing flows, or are you already designing human–agent contracts? As I wrote this post, I thank the thought leaders who came before me: Jon Yablonski, Saleema Amershi, Mihaela Vorvoreanu, Besmira Nushi, Penny Collisson, Jina Suh, Shamsi Iqbal, Jaime Teevan, Ruth kikin-Gil, Eric Horvitz, John Maeda, Adrian Bussone, Simone Stumpf, Dympna O'Sullivan, Richard Sutton, Doina Precup, Satinder Singh Baveja, Ben Shneiderman, Prof Brent Mittelstadt, Rafael A. Calvo, Dorian Peters, Myroslava Dzikovska, Lucian Galescu, Amanda Stent, Dr. Upol E., Mark Riedl, David McDonald, Dr. Joy Buolamwini, Paul Daugherty, H. James Wilson, Christopher Noessel and Robert Wood. If you want to learn more about Human-Agent Centered Design (H-ACD), check out the website: https://hacd.lovable.app/ The Human-Agent Centered Design (H-ACD) project continues to evolve, with ongoing collaboration from André Neves, Silvio Meira, Filipe Calegario, Marcello Bressan, Diego Credidio, Farley Fernandes, César França, and Rodrigo Medeiros. #HumanAgentCenteredDesign #HACD #AI #Artificialintelligence #UserExperience #UserExperienceDesign #InteractionDesign #CareerDevelopment #IndividualContributor #DesignStrategy #Facilitation #UX #aiux #uxai #mlux #Design #DesignProcesses #SoftwareEnginnering #WorkflowEncapsulation #ProductDiscovery

49

5 Comments

Perplexity just announced Perplexity Patents. Ask in plain language, get verified prior art with links you can open. A natural next move from Perplexity and Aravind Srinivas, shaped by Jerry Ma and team. It feels less like a product launch, more like a quiet redefinition of research literacy. India is an intriguing test case. Patent publications grew 60 percent in 2024, but examinations collapsed nearly 80 percent since 2020. R&D spending still sits at 0.64 percent of GDP. Many universities file patents because NIRF rankings reward volume, not validation. The result is a sea of forms, few defensible claims, and a generation that thinks innovation equals paperwork. The government’s 2024 amendments tried to speed things up, yet capacity gaps, procedural opacity, and the occasional rent-seeking habit remain. We made filing easier but not necessarily wiser. That’s why the airtel x Perplexity partnership feels like a smart play by Perplexity and a real opportunity for Indian users. If 390 million people get a year of Perplexity Pro, prior art literacy could expand fast, provided coverage and provenance hold up. In that case, we should ideally see better habits: fewer blind filings, fewer checkbox patents in classrooms, and more startups aiming for real differentiation instead of duplication. Potential ripple effects • Broader access for MSMEs, labs, and first-time filers • Cleaner claims and fewer speculative filings • Better go or no-go calls before lawyers enter the room Caveats • Verify coverage of Indian and non-English filings • Keep provenance logs and multiple reference points • Remember: AI assists judgment, it does not replace it Perplexity Patents could help India pivot from filing races to intellectual integrity. If curiosity becomes the default operating system, the next decade may finally reward original thought over bureaucratic stamina. Can we turn this cultural awakening into high-quality, evidence-backed research before the next wave of filings arrives?

7

We’re not just writing for people anymore — we’re writing for the models that choose what people see first 🤖 That’s why we need to design content that LLMs actually want to read 🎯 Loved this conversation with Dharmesh Shah about AEO, AI agents, and the coming “AI App Store” 💡 It’s a clear reminder that long‑form, unstructured posts are no longer enough 📚 If we want LLMs to select and quote our work, we need content that is precise, semantically consistent, and machine‑readable from day one — almost like an API for our ideas 🚀 That shift—from human‑only copy to dual‑audience content (humans + LLMs)—will decide who stays visible, trusted, and useful in the next wave of AI search ✨ Watch the full discussion here and ask yourself: is your content truly LLM‑ready, or just “nice to read” for humans? #AEO #AIO #LLM #AIContent #AIOptimization #DharmeshShah #AIAppStore #SVITAR #LLMNutritionist #MedAIMark https://lnkd.in/dVBzY75B

6

Exciting insights from the Nvidia partners event at JPM led by Jensen Huang (reposting now because original post was mysteriously deleted): - Patrick Collison, The visionary Founder and CEO of Stripe and co-founder of Arc Insitute (they have developed Evo (1, 2) genomic foundation models) - Christina Zorn, Chief Administrator Officer of Mayo Clinic, aims to enhance patient care by harnessing their vast repository pf 20 Billion images and clinical data sets - Jacob Thaysen, PhD, CEO of Illumina, is steering the company from their current status as a sequencing leader to a pioneering digital biology platform - Ari Bousbib, CEO of IQVIA, wants to develop AI agents to extract insights from a remarkable 64 petabytes of exclusive healthcare data Lots of great promise for AI in improving lives of patients! #NvidiaEvent #healthcareinnovation #AIfor healthcare

115

6 Comments

In Episode 1, we explored how NLP turns the overwhelming scale of biomedical literature into structured intelligence. Now, in Episode 2, we go one layer deeper - into the mechanics that let AI interpret dense, technical research with real semantic precision. Modern biomedical NLP doesn’t just “scan” text. It breaks down complex sentences into tokens, maps them into scientific embedding spaces, and identifies how genes, drugs, and pathways relate inside real-world studies. This is the foundation of machine comprehension - and the reason AI can now read research with accuracy that rivals domain experts. At Anshar AI (https://anshar.ai), we help teams build these high-fidelity understanding engines: • domain-tuned embeddings • scientific entity detection • interaction parsing across oncology and rare disease corpora The closer machines get to true understanding, the faster discovery accelerates. Let’s keep decoding the future of drug discovery together. #NLP #DrugDiscovery #BiomedicalAI #OncologyResearch #AIinPharma #LifeSciencesAI Anshar AI Anshar Labs, Inc.

21

2 Comments

“AI-Powered” Can’t Just Mean Smoke and Mirrors Everyone’s racing to slap “AI-powered” on their product. But recent headlines exposed just how hollow that label can be. Builder.ai—once valued at ~$1.5 billion—filed for bankruptcy. Repirts indicate that its “AI assistant” wasn’t powered by groundbreaking tech, but by 700+ human engineers in India, doing the work manually in real time. I’ve seen tools that barely function rush to market with polished branding and vague promises. Why? Because “AI” attracts attention—and funding. This isn’t anti-human. I’m building something too: an emotionally intelligent tool designed to support people through deep transitions. Human insight is the core. But stories like this? They put the “artificial” in artificial intelligence. If you’re marketing “AI-powered,” the bar has to be higher: - Is your AI actually doing the work? - Are you designing for trust—or just trying to ride the wave? - Does your product truly help people—or just perform the part? People deserve better than smoke and mirrors. Especially when the stakes are human. Curious: Have you seen any AI tools that over-promised and under-delivered?

3

Product development is becoming a living system where AI, causality, and experiments constantly reshape the roadmap. Check out TechArena.ai Voice of Innovation Banani Mohapatra's latest article about the new product DNA – AI, causality, experimentation. She breaks down how AI, causal inference, and always-on experimentation are reshaping the entire product lifecycle. Ideation shifts from blank-page sessions to generative exploration with LLMs and copilots. Design becomes data-infused, with tools that can anticipate engagement before anything ships. Engineering teams are building “experiment-ready” products with feature flags and metrics wired in from day one. And launch moves beyond metrics to causation. Banani uses the example of an AI-powered experimentation loop that can steadily improve adherence for people managing chronic conditions, turning a product into a living system that learns from every interaction and adapts in real time. Check it out: https://lnkd.in/gJU_Gdk7 #TechArena #VoicesofInnovation

5

“AI-friendly UX is about building interfaces with consistent syntax, machine-readable cues, and the kind of semantic rigidity that would make a programmer weep with joy.” Interesting take from our own Toni Wowtscherk on how product design is adapting to serve both, humans and AI agents.

12

2 Comments

This blog by Marty Cagan is for all the folks in Tech wondering if AI will enable everyone to build their tools and eliminate the SaaS industry. https://lnkd.in/ge5RQxbP

2

2 Comments

An interesting recent paper demonstrating how Semantic Similarity Ratings (SSR) let LLMs replicate human purchase intent with 90% reliability without fine-tuning. Working with agents has been transformative to our creative process, and this has the potential to unlock more precision in a connected workflow.

7

3 Comments