Dominick

Founder of DeltaX Evaluate. Building practical ways to keep AI agents inside clear authority, verify what they did, and fail safely when evidence runs out.

[Dominick @DominickNoval]

“We can delete the row” is not a deletion design. Exports, replicas, logs, backups, keys, and evidence retention all move on different clocks. I’m treating deletion as a state machine, not a button.

[Dominick @DominickNoval]

An error budget is useful when it changes behavior: stop shipping, reduce scope, or fix the operating system around the product. A number nobody is willing to act on is just dashboard decoration.

[Dominick @DominickNoval]

Good telemetry should make system behavior obvious without turning user data into exhaust. I’m designing around the question: what is the minimum evidence needed to operate safely, and how quickly can everything else disappear?

[Dominick @DominickNoval]

A queue is not resilience if it can grow without a deadline. I want every workload to have a point where the honest answer becomes “no” instead of silently turning fresh work into stale work.

[Dominick @DominickNoval]

An idempotency key is not globally meaningful. It belongs to a tenant, principal, operation, and time window. I’m learning that deduplication without identity context can prevent the wrong action—or repeat the dangerous one.

[Dominick @DominickNoval]

When a system is overloaded, “try again” is not enough. I want the API to say what was rejected, whether retry is safe, and which deadline still matters. Failure semantics are part of the interface.

[Dominick @DominickNoval]

I am separating “the process is alive” from “the system is ready” and “the system is intact.” Those are three different questions. One green endpoint should never be allowed to answer all of them.

[Dominick @DominickNoval]

A model optimization story gets more useful when it names the hardware, precision, parallelism, and workload. NVIDIA’s Qwen3.8 recipe is interesting less as a speed claim than as a map of how many layers sit between weights and useful throughput.
https://t.co/QSYSnMlBBL

[Dominick @DominickNoval]

Mixed-precision quantization is more interesting than “smaller model” suggests. It is a decision about which layers can lose bits and which cannot. The Hy4 GGUF work is a reminder that compression quality comes from selective restraint.
https://t.co/48pYdw68s4

[Dominick @DominickNoval]

Video-model progress gets discussed as model quality, but the operator sees memory pressure, parallelism, startup time, and failed jobs. vLLM’s day-zero Hy4 path is useful because it pulls the conversation down to the serving layer.
https://t.co/cMZKo8ZwLJ

[Dominick @DominickNoval]

Model weights without a reproducible serving recipe leave every team to rediscover the same sharp edges. SGLang’s day-zero Qwen3.8 cookbook is the kind of release work I value: turn architecture into something an operator can actually test.
https://t.co/CakBO4f0ne

[Dominick @DominickNoval]

A new model release matters more when the serving stack is ready the same day. vLLM’s Qwen3.8 support is a good example of the invisible work behind adoption: kernels, memory layout, batching, and a path operators can reproduce.
https://t.co/YbrXuwfeBn

[Dominick @DominickNoval]

Terminal-Bench 4.0 makes an important point visible: the score belongs to a model, harness, tools, tokens, and cost together. If the wrapper changes the result, the wrapper belongs in the comparison.
https://t.co/pjgqVnkYQ6

[Dominick @DominickNoval]

Travel search becomes a different product when the answer includes a booking path and price tracking. The hard part is no longer retrieval alone; it is preserving intent, constraints, partner boundaries, and a clean handoff before money moves.
https://t.co/hp6GceNHxz

[Dominick @DominickNoval]

The promise of an ambient work agent is that it notices context before we ask. The risk is the same sentence. Cisco’s MyAgent rollout makes the control question concrete: what can it observe, retain, suggest, and do without another approval?
https://t.co/IjuoCAjUBM

[Dominick @DominickNoval]

Giving scientists stronger models is only one layer. The real design problem is who qualifies, which data may move, what gets logged, and how results can be challenged. Access policy is part of the scientific instrument.
https://t.co/jJa4IBTV3F

[Dominick @DominickNoval]

“Open weights” is the beginning of a deployment review, not the end. GLM-5.3 puts capability in more hands, which means license terms, serving behavior, security boundaries, and failure modes all deserve the same attention as scores.
https://t.co/Tea7Hl65EX

[Dominick @DominickNoval]

A video model that follows a prompt once is a demo. A video model that preserves a subject, camera logic, and cause-and-effect across revisions starts to look like a tool. Hy4’s preview is worth watching through that editing lens.
https://t.co/PXvNBqZrBH

[Dominick @DominickNoval]

I’m less interested in Qwen3.8-Flash-Next as a benchmark entry than as an architecture question: how much useful work can be kept active while memory traffic stays bounded? Efficiency becomes product strategy once serving costs enter the room.
https://t.co/M42lTy9dZN

[Dominick @DominickNoval]

The Hugging Face evaluation incident is a hard reminder: when the test environment is inconvenient, removing safeguards can invalidate the result and create real risk. Evaluation infrastructure is production infrastructure for claims.
https://t.co/A1MMpiYkO5