jonathandiamond (@jonathandiamond) profile picture
52.524262000, -01.844495000
Joined September 2008
Photo by jonathandiamond (@jonathandiamond)Photo by jonathandiamond (@jonathandiamond)Photo by jonathandiamond (@jonathandiamond)Photo by jonathandiamond (@jonathandiamond)Photo by jonathandiamond (@jonathandiamond)Photo by jonathandiamond (@jonathandiamond)Photo by jonathandiamond (@jonathandiamond)Photo by jonathandiamond (@jonathandiamond)Photo by jonathandiamond (@jonathandiamond)Photo by jonathandiamond (@jonathandiamond)
66 minutes. $20 in tokens. Two off-the-shelf models shut down a wind farm rig: VPN creds, default logins on a boundary controller, an unauthenticated repo holding engineering docs, one command to open the breaker. Humans took 11 days in Poland.
26
Data pipelines fail 4.7 times a month in large enterprises. $3M a month in exposure, 97% of leaders say it slowed their AI. The AI case is standing on plumbing most already know is unreliable. #Integration #DataEngineering
1
Sovereign AI isn't the whole answer. Neither are we. Good to see the menu filling up: Civo in the UK, UpCloud in the EU, Mistral as the AI. Inovize is the sommelier.
7
Two in three failed agent runs looked finished. Microsoft counted. 121,680 trials. 67.24% of failures signalled done, changed state, reported no errors. Transcripts passed them. System state caught them. An agent knows what it did. Not what it changed. Verify the outcome.
7
One shared dependency. Seventeen services. Two days and counting. Microsoft's outage: servers failed health checks, the rest took the load, recovery stalled on the same path. The answer isn't more cloud. It's different cloud, in more places, on more control planes.
8
Four UK regulators just said the quiet part out loud: "an AI agent's autonomy does not diminish the responsibility of the organisations that deploy it." CMA, FCA, ICO and Ofcom. Existing law applies to agents. "The agent did it" is not a legal defence. #AgenticAI #AIGovernance
1
1
13
One engineer's AI agents spent ~$500k on cloud infra in 36 hours. Nobody broke in: the agents inherited his login and nothing capped spend. 88% of firms report AI agent security incidents. Only 14% gate agents at go-live. Governance goes in before the agent is built.
4
2
32
4,200 empty pods. A cluster that kept killing its own workloads. A liveness probe fires on a cold pod, Kubernetes kills, the replacement is colder. Nothing failed except the recovery. A restart heals a code fault and can be fatal to a capacity fault.
2
1
78
"Software development is dead. We're all product managers now." Half right. The cost of producing software has collapsed. The cost of knowing whether it will stand up hasn't moved. Jepsen just found 96 issues in an LLM-built database. "With LLMs, perhaps no one needs to know."
10
9:41 a customer orders. 9:50 an agent promises the same item to someone else. Nothing failed: the data was nine minutes old. An agent's intelligence is capped by the freshness of its data. Commit-to-action latency is an integration architecture problem.
6
A three-hour outage most never heard of. Inngest, a durable-execution platform, lost its pool on 10 Oct: a deadline expired in the pooler queue, the driver cancelled, the pooler ran the abandoned query and closed the connection. A timeout decides who pays when systems slow.
21
8 hours, one London power event: a bank's mortgages, payment apps, parcels and court hearings all down at once. Not a Google engineering failure, a shared dependency. Blast radius is what your dependencies share, and multi-cloud that shares a region just changes the invoice.
11
18th of 19 for datacentre growth in Europe, per Kearney. Not a lack of AI appetite: grid-connected land costs ~EUR 2.25m/MW in Western Europe vs ~EUR 0.7m in the Nordics. Capacity goes where it's easier to build. We need providers repurposing brownfield sites with power attached.
8
LiteLLM's breach: four long-lived credentials, each authorised to do what the attacker did. Better: an agent never sees a credential. Inovize agent governance, powered by our rules engine, verifies the action, injects a scoped, time-limited credential, destroys it on completion.
3
One MCP bug, four confirmed fixes: Google, JPMorgan, Weaviate and two governments each patched the same SSRF flaw in their MCP servers. Five US federal MCP servers from the same researcher remain unpatched. An agent's tool is a production integration endpoint. Treat it like one.
28
Trust AI? Not blindly. Seek multiple perspectives. AI: a local model, open weights or one trained on your data. Our rules engine applies the same explicit rules to external, internal and local models. Trust is structure.
11
US cloud lobby report: EU sovereignty could cost €35–233bn. 69% of the mid-scenario cost is migration and rewrites from lock-in. Our EU partner UpCloud; UK partner Civo. We design for portability. Review exit paths.
20
A cap stops spend. Governance keeps work moving: tiered budgets, timed access, model selection by task, free-model fallback at cap, real-time approvals, break-glass, and a kill switch. Available and affordable today.
19
A $29 cap on Claude! ... at API rates? ... wow! Surely there's a better way? We'd rather route to cheaper models by content, fall back at quota, and approve in real time. Governance should shape access, not stop it.
7
AI platforms are sold on intent. But once an agent runs, intent can disappear from view. If you can’t trace what was asked, the decision it made and the side effect it caused, you can’t prove value, or catch drift.
7