Today we're sharing our work on interaction models. A new class of model trained from scratch to handle real-time interaction natively, instead of gluing it onto a turn-based one.
https://t.co/MoS5s4cm60
@Al_Grigor Agent are amazing but they should never have permissions to apply changes in production, be it infra of product code changes. Tests and checks are now more important than ever.
One agent makes a change another checks it, and then a human checks a second time.
@TheLaranox@SimonHoiberg Surely, there are countless ways they could have optimized Kubernetes, but they chose to move to a simpler platform instead.
My point was just to correct the notion that “Fargate should never be cheaper than EKS unless you’re doing something wrong.”
@TheLaranox@SimonHoiberg Also without AWS EKS add-ons, you'll still have to deploy replacements. Custom ingress, storage, cert management etc. every feature runs as a pod you pay for.
Kubernetes is great but it's not for every org.
@TheLaranox@SimonHoiberg There are no absolutes
ECS will infact be cheaper for most depending on scale and strategy.
Cloudwatch is not a must for monitoring, logs and metrics can use most observalibility platforms.
How are helm charts getting applied without a CD tool?
@TheLaranox@SimonHoiberg ECS Fargate is often cheaper for many workloads. On EKS, you pay for the EC2 running add-on pods, observability, ArgoCD, Kube-system pods and $72/month for managed control plane. There's also a pod limits per node. Cost ultimately depends on your workload, scale and architecture
You're never getting me back into an office full time, but I'd also hate to work remotely without these fantastic semiannual meetups. Spending a week together in person is invaluable for charging trust batteries and connecting as humans. Highly recommend it for any remote team.
Company spent $18k/month on Datadog. Leadership loved the dashboards.
I dug into what we actually used:
- 12 people logged in per month
- 89% of metrics never viewed
- Alerts had 97% false positive rate (team ignored them)
- We were ingesting 2.4TB of logs daily
- Retention: 180 days (compliance needed 30)
Moved to:
- Grafana + Prometheus (self-hosted)
- Loki for logs
- PagerDuty for critical alerts only
- 30-day retention
New cost: $340/month (hosting + PagerDuty).
The controversial part? Engineers initially hated it because I killed their "observability theater." But incidents got resolved faster because we only tracked what mattered.
Exactly! It’s easy to mess things up with general purpose languages. HCL is less flexible but maybe that’s actually a feature. The inconsistency with using other tools is too wide, some folks even write their own layer of abstraction ontop of Pulumi or CDK. Not worth the stress in my opinion.
@FrancoC137@mitchellh Most Ops folks don’t want to write infrastructure in imperative languages like Go/Python. This is the same reason I think Ansible won.
@oluvvafemi@jorgemanru@dhh@jasonfried This is massive, being top 6 out of 1600 is no joke. It’s truly a proof of your ability and skills. I’m sure any org will be lucky to have you.
Great job!
@GergelyOrosz@Pragmatic_Eng Thanks for clarifying!
I was indeed surprised at first. The cloud is so complex most devs don’t even want to deal with it. I was really curious to see how a startup that size functions without an infra/DevOps/SRE team or person.
freeCodeCamp just published a FREE Python AI course on Retrieval Augmented Generation.
Learn RAG from scratch, and combine your own custom data with the power of LLMs. Taught by LangChain dev Dr. Lance Martin.
[3 hour course]
https://t.co/1W1JUmfWeE