A local model can still sit inside a cloud-dependent workflow.
Sovereign agentic AI needs control of tools, identity, state, traces and updates too. Test the whole workflow with external access removed.
voxelperfect.com/writing/sov…
I've worked through custom software, SaaS and now co-development.
I built Gaia for UBITECH. As that chapter closes today, I'll keep pursuing software built with the customer.
My reflection, prompted by Ed Byrne:
voxelperfect.com/writing/the…
One retry can help a service. Retries across a chain can overload the system.
As AI expands what we can build, I expect architects who understand those interactions to become more valuable.
voxelperfect.com/writing/ai-…
My bet for 2027: agentic AI will reward teams that understand the workflow.
An invoice explanation is useful. Resolving the case also needs evidence, approvals and reliable updates.
Measure cost per correctly completed case.
voxelperfect.com/writing/age…
I would evaluate an imported model through concrete controls: where data goes, who can inspect behavior, what can be switched, and how failures are contained.
The governance argument from my July Kimi essay:
voxelperfect.com/writing/the…
Hardware portability is a systems question.
I would examine libraries, kernels, tooling, operations and switching cost before declaring a compute ecosystem replaceable.
The argument from my July essay:
voxelperfect.com/writing/cud…
What gets harder to copy after a year of using your AI product?
My candidate: verified outcomes that improve evaluations, routing and workflow decisions.
An operating loop, not just an inventory of data.
voxelperfect.com/writing/the…
NVIDIA can benefit from cheap, interchangeable models—if its compute platform stays hard to replace.
Hugging Face helps with the first. Hardware portability complicates the second.
The strategic tension behind the acquisition agreement:
voxelperfect.com/writing/why…
A customer can stop depending on your product before cancelling it.
I call that functional churn: the work leaves before the contract does.
Which customer-health metric would notice?
voxelperfect.com/writing/wha…
I kept exploring a project despite better AI arriving later.
My test: what will I understand even if I discard the implementation?
A prototype can be temporary. The question it resolves need not be.
voxelperfect.com/writing/i-d…
A familiar entry point can be a distribution advantage for a new AI capability.
My July essay used Chat, Work and Codex to ask when product consolidation makes discovery easier—and when it makes the experience confusing.
voxelperfect.com/writing/ope…
A provenance signal is not an authorship verdict.
The product question: does the interface preserve that distinction, or invite a stronger conclusion than the evidence supports?
Revisiting my August essay:
voxelperfect.com/writing/ant…
A new capability can disappear inside an old evaluation framework.
Before asking whether an AI product is good enough, ask: good enough for which job, measured against which assumptions?
voxelperfect.com/writing/goo…
Cheaper production does not decide who gets scarce energy, reliable compute or desirable land.
Prices, quotas and access rules still allocate the bottlenecks. Abundance changes what money buys.
voxelperfect.com/writing/abu…
Jev was fastest on successful responses in our practical tests. It didn’t settle the choice.
77 intents, 17 questions per contract: batching helped every model. Cache hits changed costs; Jev had service failures.
voxelperfect.com/writing/jev…
Huang’s math remarks miss a distinction: a task can lose economic value while keeping educational value.
Delegating work you understand differs from skipping the practice that builds understanding.
voxelperfect.com/writing/jen…
A coding agent can implement an unclear idea very quickly.
The valuable skill is explaining what must stay true, what should change, and how you will check the result.
voxelperfect.com/writing/cod…
LLMs inherit culture through language.
Fluent German does not guarantee reliable German reasoning. Aleph Alpha’s experiment makes that gap visible.
Multilingual AI should be tested on local work, not translation quality alone.
voxelperfect.com/writing/llm…
Claude Marketplace borrows a familiar Azure playbook: let partner purchases use an existing commitment.
That can make switching models a budget question as well as an engineering question. Anthropic's purchasing path remains a limited preview.
voxelperfect.com/writing/cla…
LLMs may not have human emotions because they do not have human bodies or human lives.
But “not like us” is not the same as nothing. Anthropic found emotion-like internal states in Claude that causally change preferences and behavior.
voxelperfect.com/writing/llm…












