- The Gap Through Which We Praise the Machine: Thoughts about the difference between what ai agents are capable of and where they fit vs in practice. How we’re having to adapt ourselves to the misunderstanding / shortcomings of this new tool? (And how well that doesn’t usually go.) Listed are several new practices we can try to adopt to increase context available to llms but they’re often hit or miss. And discusses the pressure from above to get on this ride.
- AI Angst: More ai thoughts that are balanced. Good uses, bad uses. Hard to ignore smart people saying there’s something here. LLMs are going to be used for terrible things and there will be consequences. We should probably think about environmental impact.
- Series: How to think about throttling an application by planetscale (uses db as an example)
- What requests should be throttled?
- What metric do you use to kick in throttler? (eg replication lag)
- How often to check the metric that decides throttling is needed
- Queuing theory: Replication lag is “the amount of time a change event waits to be transfered to secondaries from primary and then applied”
- Source: Anatomy of a Throttler, part 1
- AI at Amazon: a case study of brittleness: An organization with the means to participate in the latest ai wave of exploration / innovation couldn’t because of org friction. They got a slow start.
- Thinking about facets of (cloud) identity providers: Authentication is hard and outsourcing it is a big decision. Should you?
- SREcon25 Americas - Network Flow Data in the Cloud: Slack srecon presentation about account for network flows between services in their system.
- SREcon25 Americas - An SRE Approach to Monitoring ML in Production: How to think about monitoring a ML service from data engineering to inference to end-to-end health of a system where ML is a part of the whole.
- Anthropic: How we built our multi-agent research system: Anthropic runs subagents in parallel and a leader agent merges results together. This apparently gets better results faster but is expensive.
- Things are different between system and application monitoring: An idea here that systems folks are more interested in metrics and logs vs app developers who use metrics less and traces and logs are a bit more emphasized. Maybe more interested isn’t right. Often all that’s available to us are metrics and logs. (And what we see is what we get!) We can’t go in and modify the program for better debugging in production.
- Expert Generalists: Soft skills like collaboration a solid grasp of the basics can often get one to a workable solution vs having expert level knowledge in a specific area.