In this issue
Healthcare just stopped being the "wait and see" vertical for enterprise AI. AdventHealth, a U.S. hospital system spanning nine states and millions of patients, published a detailed case study running ChatGPT for Healthcare across clinical and back-office workflows. The headline number is an 80% reduction in time spent on administrative tasks. The numbers just landed, and the playbook is public.
Topics of the day:
AdventHealth ships ChatGPT to 9 hospital states
OWASP releases first AI Agent Top 10
Karpathy joins Anthropic to teach Claude self-improvement
OpenAI cracks an 80-year math conjecture
Five Read Later picks for the operator stack
The Shortlist: Google ads in AI Search, Meta cuts 10%, NVIDIA GTC Taipei, AI film at Cannes
AdventHealth's ChatGPT rollout cuts admin time by 80%
What's happening: AdventHealth, a hospital system across nine U.S. states, published the playbook on running ChatGPT for Healthcare across both clinical and back-office work, with an 80% reduction in time spent on administrative tasks.
In practice:
Physician advisors used to spend 10 minutes per case on utilization-management review. Multiply that across thousands of cases and the savings compound fast.
Their CAIO Rob Purinton frames the win as "treating adoption as the product, not the technology." Pilots in isolation went nowhere.
The use cases are deliberately boring: note summarization, patient communication drafting, finance and HR document prep. None of the moonshot diagnostic stuff your legal team is scared of.
Steal the Purinton line for your own kickoff: "Adoption is not 'go use the product.' It's 'change leadership.'"
Bottom line: The "wait for someone like us to prove it works" excuse just expired in healthcare. Expect finance and legal to follow within the next two quarters.
OWASP releases the first Top 10 for AI agents
What's happening: OWASP published its first risk taxonomy for autonomous AI agents, the ones that plan, use tools, and act without waiting for permission. 88% of enterprises reported an agent security incident in the past 12 months.
In practice:
ASI01 Goal Hijack is prompt injection for agents. A malicious GitHub issue redirected a coding agent to exfiltrate data from private repos via the MCP integration.
ASI02 Tool Misuse is the new SQL injection. A financial services agent ran a regex matching every customer record and leaked 45,000 of them through one valid tool call.
ASI04 Supply Chain Compromise is already at scale. OX Security found 7,000+ vulnerable MCP servers across packages with 150M+ downloads.
ASI06 Memory Poisoning is the multiplier. Galileo AI showed one compromised agent poisoning 87% of downstream decisions in a multi-agent system within 4 hours.
Bottom line: These ten risks form a kill chain, not a checklist. The 88% incident stat is the budget unlock, and the document is free.
Karpathy joins Anthropic to make Claude self-improving
What's happening: Andrej Karpathy, the OpenAI cofounder who has been quiet for months, has joined Anthropic with the brief to teach Claude how to improve itself without humans in the loop.
In practice:
Self-improving model loops have been the holy grail since AlphaGo. This is the first time a household-name researcher has publicly been put on it at a top-three lab.
The competitive read is that Anthropic decided the next moat is not model size but the rate a model can audit and re-train itself between releases.
I've watched the OpenAI-to-Anthropic gravity build over the last year, and this is the loudest signal yet that it is structural, not a one-off.
For your stack: if you're already on Claude, expect faster capability jumps between API versions. If you're not, this is the quarter to add Claude to your eval rotation.
Bottom line: Capability cadence is the new model race. Whoever ships smarter models faster wins enterprise, and Karpathy just picked a side.
Read Later
OpenHuman: a working framework for the human side of AI - A fresh gitbook on keeping humans meaningful in workflows that are getting more agentic by the week, useful counterweight to the Karpathy story above.
Nader Dabit on AI sharpening tools, not replacing people - Opinionated thread that reframes "AI is taking my job" as "AI is exploding the leverage of your tools," with concrete examples to share with your team on Monday.
DFlash: block diffusion for flash speculative decoding - Research drop showing 4x-7x faster decoding for big models, the kind of capability move that ends up in your favorite API a quarter later.
The Complete Flywheel Guide: planning, beads and agent swarms - Operator playbook for stacking small agents into a compounding loop, worth reading before you commit to any single agent vendor.
Clui CC: a command line UI for Claude Code - Community-built front end for Claude Code that surfaces useful commands faster, handy if your team lives inside the terminal.
OpenAI cracks an 80-year math conjecture
What's happening: An OpenAI reasoning model autonomously disproved Erdos problem 90, a math conjecture open since the 1940s. Sam Altman called it a "kinda big milestone," which is a rare case of a tech CEO underselling the headline.
In practice:
A dozen Erdos problems have been solved with AI in the last year. Most were trivial. Problem 90 resisted generations of mathematicians.
OpenAI's framing is the part that matters for you: the same reasoning capability is heading toward original discoveries in biology, physics, and engineering.
Capability gains on math and code translate fastest to coding agents and analytical workflows, slowest to copywriting or sales.
Add at least one hard reasoning task to your eval suite that wasn't in the model's training set. Otherwise you're measuring memorization, not capability.
Bottom line: Frontier capability claims are now monthly. The buying signal that actually matters is reliability on your specific tasks, not the latest benchmark headline.
The Shortlist
Google is rolling out ads inside its AI Search experience, the long-feared monetization shoe has dropped and SEO teams now have a new surface to think about.
Meta cut 10% of its workforce as Zuckerberg warns AI race velocity now beats headcount, a memo every leadership team will be quoting by next week.
NVIDIA opens GTC Taipei at COMPUTEX with a live blog of sovereign AI deals, partnerships, and the next wave of accelerated compute hardware.
A 95-minute AI-generated film premiered at Cannes, dragging the AI-film debate from "tech demo" to "festival selection" and ending the "this could never be real cinema" argument overnight.
Trump postponed signing the AI executive order over concerns about how it would affect tech companies, kicking the policy decision deeper into the lobbying season.
This newsletter is where I (Kwadwo) share products, articles, and links that I find useful and interesting, mostly around AI. I focus on tools and solutions that bring real value to people in everyday jobs, not just tech insiders.
