It’s timely, given something that came up in the podcast, that AOL has announced they’re hanging up on dial-up internet.
If you miss the good old days, check out this site - then switch off your modem for the full effect of someone in the house making a phone call.
HOT TAKE
The Next Asset to Depreciate?
On the pod this week, today’s top models were called “the fastest depreciating assets in tech.” After reading this week’s blog, I’m wondering - how long before the cloud giants’ software layer gets the same treatment?
Curated finds to help you stay ahead
Multi-agent app design is examined in this blog highlighting why system architecture and orchestration often matter more than the code.
A Mixture-of-Experts coding model built for long-context, multi-turn agent workflows, with a focus on large-scale code generation, tool use, and repository-level tasks.
An unsupervised defense approach for LLM-based multi-agent systems, using interaction modeling and contrastive learning to detect attacks.
MCP vs A2A: this guide explains how protocol choice impacts interoperability, tool orchestration, and long-term maintainability in multi-agent architectures.
MLOPS COMMUNITY
The Truth About LLM Training
Ever wondered how to pick the right model when a new “best-ever” drops every week? Paul and Zulkuf share how they cut through hype with private eval sets, LLM-as-a-judge scoring, and leaderboards that expose real performance.
Model evaluation – task-specific benchmarks in multiple languages, kept fresh to avoid training set contamination.
GPU strategy – spiky training runs demand on-demand flexibility, while inference needs steady capacity with scale-down options.
Cost control – fine-tuned models replace costly APIs without sacrificing performance.
Together, these form a no-nonsense blueprint for picking winners, cutting costs, and staying ahead of the model churn.
The Agentic Cloud: Forging the Next Era of Infrastructure
The “Agentic Cloud” could shift power away from today’s hyperscalers by replacing their sprawling service catalogs with intent-driven agents that design, deploy, and run infrastructure across any provider. It’s a vision that trades centralization for intelligence-driven decentralization.
Here’s what that shift looks like:
Architecture - Agents handle specification, provisioning, deployment, and autonomous maintenance.
Business model - Cloud providers reduced to commodity hardware while agent intelligence captures the value.
Challenges - Reliability, security, state management, and integrating messy real-world systems.
A potential playbook for dismantling the cloud’s “fortress” and redistributing its power.
MEME OF THE WEEK
ML CONFESSIONS
One Bad Benchmark, Total Carnage
I was in a rush to get a new evaluation set in before a model release, so I copied the JSON from an old benchmark, changed the filename, and pushed it straight into the pipeline. I thought I’d updated all the labels. I had not.
Half the “correct” answers were wrong, half the wrong ones were “correct”. The leaderboard came back looking like every top model had been hit with a hammer. Took a couple of days before someone finally asked in Slack why their model had fallen off a cliff, and that’s when it clicked.
We had to re-run the whole lot and the infra team gently pointed out we’d basically paid twice for the same jobs. Painful lesson, but at least now we have a checklist so this particular flavor of chaos doesn’t happen again.
Share your confession here.