Jason Merkoski

AI results with quantified and compounding outcomes. Healthcare is one if the most inefficient and arcane systems perfectly positioned to harness the promise of AI. Congratulations Autonomize AI. Thrilled for ATX Venture Partners as debut and ongoing investor in this phenomenal team and market opportunity. Stay tuned for more raving customer results and case studies…

10

1 Comment

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

For those who haven’t heard, Anthropic leaked the entire Claude Code repository 2 days ago. Aside from being one of the most expensive leaks in history ($10B+ ?), there was also a lot to learn about how anthropic works and how they build agentic harnesses. From reading the code, it's clear that almost all of Anthropics code is produced by claude (the repo was 500k lines and pretty sloppy), and that harness engineering (ie, the tooling built around and integrated with the LLM in agentic workflows) is incredibly hard and deeply non-trivial. My hypothesis is twofold: (1) Proprietary agentic tooling is going to get exponentially better in the next 3 months. CC is arguably one of the best harnesses (across all products, not just coding) in the world right now, and its full implementation has just been exposed. (2) The quality gap between companies treating the harness as a thin wrapper over a model and companies doing real systems engineering will exacerbate to the point where no one uses the former's products. Without a deep understanding of AI/LLMs and the domain that you are building in, your product will fail the second it hits some type of ambiguity or edge case. Lots to learn from the leak... if you're a pub interested in applying agentic tooling to your monetization stack, hmu! We are running pilots right now

16

AI has entered its systems phase. Outcomes are no longer driven by model size alone, but by how well models are integrated into context, tools, routing, and feedback loops. A small set of primitives is defining this shift: - Tool use turns models into operators, not just assistants. - Dynamic context (e.g. MCPs) makes context portable, composable, and governed at runtime. - Intelligent routing allocates the right model to the right task under real-world constraints. - Telemetry closes the loop, turning production behavior into learning. On top of these primitives, new alignment strategies – synthetic data, verifiable rewards, and multi-turn objectives – are allowing systems to learn from real outcomes rather than static labels. The implication: As capabilities commoditize, differentiation moves up the stack to system design, domain grounding, and alignment with measurable value. Pricing follows the same path, shifting from labor proxies to outcomes, and in some cases even risk-sharing. For builders: think more about control planes that govern routing, context, and accountability. For buyers: evaluate AI as an operator, one that can be audited, improved, and ultimately trusted with responsibility 🔗 More here: https://lnkd.in/gYfF7cWu

23

1 Comment

Are benchmarks measuring reasoning, or just formatting? Our team at TriFetch replicated MedAgentBench, the first clinical AI workflow benchmark and found something surprising: most "model failures" weren't reasoning failures, they were a result of poor benchmark design. Here's what we discovered: 1. Models excelled at simple retrieval tasks, achieving >90% accuracy 2. For complex tasks, the models failed because of missing instructions and vague context However, this is where it gets interesting: - When we provided precise formatting rules and shared context, the error rates dropped dramatically - Many "wrong" answers were actually correct but didn't match expected JSON format, which is something tool calling can perfectly solve As PhD students turned founders building in healthcare, this matters. If we benchmark wrong, we teach the field the wrong lessons. We should penalize the model for failing to reason like a clinician, instead of failures due to missing context. At TriFetch, we build internal benchmarks and evals to surface true reasoning gaps representative of real world clinical use cases, not just formatting edge cases. If you're building for agentic hospital workflows, we'd love to connect and share insights from our research. Our takeaway: Benchmarks should be constructed carefully so that they showcase true model failure modes, instead of engineering friction that can be easily addressed. Read our full analysis here: https://lnkd.in/gQ3nDtRA

131

3 Comments

Unpopular opinion: Most health systems aren't ready for AI. And a lot of the vendors selling it to them know that. I'm not saying AI doesn't work. I'm saying that buying a platform doesn't make you ready for a platform. I've watched organizations invest millions in technology and get a fraction of the value- not because the product failed, but because nobody changed how they operated around it. The health systems I've seen actually transform with AI share three things: 1. A senior leader who owns the outcome- not just the budget 2. A willingness to change workflows, not just add tools 3. A clear definition of what success looks like before they buy The technology is ready. The question is whether the organization is. What's your experience and where do you see the biggest gap?

6

1 Comment

Today is a big day for SEQSTER and for clinical #AI. We just launched something we've been building towards for a decade. A data harmonization engine that takes raw, patient-consented EHR data and transforms it into clean, structured, AI-ready records, automated at scale, across virtually every U.S. provider. Why does this matter? The #1 reason clinical AI stalls isn't the model. It's the data. Raw EHR formats were built for record exchange, not machine reasoning. We fixed that. This is the foundation the industry has been missing. And we couldn't be more proud to bring it to life. Congratulations to the team on this milestone. Read the full announcement in the comments. 👇 #HealthcareAI #ClinicalTrials #LifeSciences

51

18 Comments

Slipstream is redefining new employee onboarding with AI agents that actually get people up to speed fast. If a software engineer costs $250k a year, every week saved in onboarding returns roughly $4,800 dollars in reclaimed productivity. Multiply that across a team and the savings become wild! The future of onboarding is automated, personal and finally efficient!

10

🐮 AI-powered cows worth $2 billion are using an algorithm called the "cowgorithm." I'm not joking. Halter makes AI-powered cow collars. They monitor health, location — and herd cattle. Farmers tap one button on an app and cows gather for milking. 600,000 collars already live. Peter Thiel just backed their latest round at a $2 billion valuation. Every time I think I've seen the most ridiculous application of AI — I'm proven wrong. But here's what the headline misses. This isn't a funny story about cows. It's the proof of concept for the entire physical world AI thesis. The collar is solar-powered. GPS. Health monitoring. Heat cycle detection. 6,000+ data points per minute. Per cow. That stream feeds the "cowgorithms" — models that translate farmer intent into signals the animal understands, and animal behavior into decisions the farmer can act on. No ranch hands. No horses. No physical fences. One app. One button. 600,000 cows moving. The number that reframes everything: Traditional fencing: ~$20,000 per mile. Halter customers have created 11,000 miles of virtual fencing in the US alone. $220 million in avoided costs. From collars. The valuation trajectory is the real signal: 2023 → $85M. Proof of concept. June 2025 → $1 billion. 300,000 collars. March 2026 → $2 billion. Founders Fund. 600,000 collars. Valuation doubled in 9 months. Collar count doubled in the same window. Thiel backed Palantir when defense software was unglamorous. SpaceX when rockets seemed too physical. Now Halter — when AI collars for cows seem too ridiculous. The $7 trillion global agriculture industry is the largest physical-world industry on the planet that has received the least AI investment relative to its economic weight. Halter isn't a cow collar company. It's the first proof that the physical AI stack — understand current state, predict future state, control future state — works on biological systems at industrial scale. The cow is the sensor. The collar is the compute. The cowgorithm is the model. If 600,000 collars can autonomously herd cattle — what does 6,000,000 look like? What does that dataset become when it's large enough to predict disease outbreaks, optimize global protein supply chains, and inform regenerative land management at continental scale? The cowgorithm today is a herding tool. The cowgorithm at scale is a biological world model. The physical world is 100% of the world. We've barely started digitizing it. What's the "cowgorithm" in your industry — the problem that seems too ridiculous to take seriously, but is actually the $2 billion opportunity nobody is looking at yet? https://lnkd.in/eiHTsrcC 🐮

5