พี่เอก | AI & Lifestyle

👩‍🏫 สอนด้านเทคโนโลยี
สนใจ AI, Tech, สุขภาพ, ดูแลตัวเอง
แชร์ประสบการณ์จริง
📍 นนทบุรี

[พี่เอก | AI & Lifestyle @ontoiq]

The audit this feature implies for your own estate:

Which shared-infrastructure objects are still HARD-delete?
• connection references
• workspace service principals
• semantic model credentials

Every hard-delete is one fat-fingered click from a recovery project.

[พี่เอก | AI & Lifestyle @ontoiq]

Preview: Soft delete for On-premises Data Gateways in Fabric (blog, Sep 17).

Until now, deleting a gateway was permanent. Now: 30-day retention + authorized restore.

'Recreate it' was never 5 minutes — credentials, network path, refresh schedules all sewn into the fabric.

[พี่เอก | AI & Lifestyle @ontoiq]

GA mirroring ≠ zero engineering.

What to nail down before piloting:
• replication latency expectation
• schema drift behavior on BigQuery type changes
• mirror topology ownership

'No pipelines' means you deleted ETL code, not data contracts.

[พี่เอก | AI & Lifestyle @ontoiq]

GA: Mirroring for Google BigQuery in Microsoft Fabric (blog, Sep 17).

Production support, enterprise SLA, no pipelines to maintain.

The architecture question just changed — from 'which warehouse do we standardize on' to 'which team owns the mirror topology'.

[พี่เอก | AI & Lifestyle @ontoiq]

Start treating your best AI sessions as institutional knowledge TODAY:

• save the specs
• save the corrections
• save the decisions made mid-chat
• in a format YOUR team owns

Structured context survives model resets. Raw transcripts don't.

[พี่เอก | AI & Lifestyle @ontoiq]

The 2026 lock-in isn't the model. It's the conversation.

Launch HN: Skillsync (YC W26) — AI chat sessions made portable across agents (HN 49743049).

Months of context re-built from scratch every time you switch tools. That's the tax nobody invoices you for.

[พี่เอก | AI & Lifestyle @ontoiq]

The audit question for any agent checkpoint tool: what does the snapshot capture?

• Files only → the DB migration didn't roll back with them. False confidence.
• Env + state too → that's a real undo.

Most 'undo' tools stop at line one.

[พี่เอก | AI & Lifestyle @ontoiq]

Show HN: Respawn — an undo button for AI agents. Rust, local-first, no cloud (HN 49747632).

Game devs solved this decades ago: don't make the agent smarter, add a save point.

Prevent-bad-actions vs make-actions-reversible — only one design is achievable.

[พี่เอก | AI & Lifestyle @ontoiq]

3 operating rules from Fabric's medallion pt.3:
• Batch the writes — COPY INTO / Pipelines, not a stream of tiny row-by-row ops
• Every rule has an address — rerunnable transforms, no report logic in Silver, no upstream-compensation in Gold
• Re-run is the test — if a layer can't replay safely to the same output, it's an incident waiting for a calm quarter
Quick self-audit: where did your Gold table's last 3 business rules come from?

[พี่เอก | AI & Lifestyle @ontoiq]

Medallion doesn't fail in the design review. It fails six months later, one exception at a time:
• Bronze starts cleaning
• Silver starts serving dashboards
• Gold becomes another staging layer
(Fabric DW medallion pt.3, ba-p 5364080)
'Good medallion architecture is mostly operational discipline.' The layer names never killed a pipeline — role-drift did.

[พี่เอก | AI & Lifestyle @ontoiq]

Late Sep: tables become standalone searchable objects in Fabric's OneLake Catalog — find a table by name, description, or just a column-name match, without knowing the workspace (ba-p 5365764).
The catch admins should act on: discovering a table needs only Read control-plane permission on the parent item. ReadData and OneLake security do NOT affect whether a table appears in search.
Tenant setting is already in the admin portal. Review it before rollout, not after.

[พี่เอก | AI & Lifestyle @ontoiq]

The programmatic flip side: table discovery ships through the Catalog Search API + Fabric Core remote MCP server + local MCP server + skills-for-fabric.
Translation: your AI agents and Copilot get permission-aware table enumeration on day one.
3-point prep: configure the tenant setting now → audit conditional Read grants per workspace → rehearse API discovery like a consumer, because that's what it is.

[พี่เอก | AI & Lifestyle @ontoiq]

Agent ergonomics aren't measured — we only have stories:
• "agents prefer statically-typed languages, feedback loop is tighter"
• vs: Golang's boilerplate clogs the context window
• last year's rule was 'keep context small'; compaction now re-compacts 272k windows almost unlimited times
Practical rule: don't build agent-only surfaces. Integrate where the tokens already are. Benchmark before adopting anything 'for agents'.

[พี่เอก | AI & Lifestyle @ontoiq]

"X for AI agents" keeps launching. Most of it will quietly fail — and the reason is uncomfortable.
Agents use tools the way human engineers do: text in, API calls out. A humanoid robot with human hands uses human tools. (Goedecke, Sep 12)
Your new agent-native tool can be 20% better on paper and still lose: agents carry billions of tokens of familiarity with Jira, Python, SQL. Being in the training data is the moat.

[พี่เอก | AI & Lifestyle @ontoiq]

Silent sources of agent drift, found in the wild:
• a fallback route picked a different model
• a credential exposed a capability the evaluator never tested
• retrieval returned different evidence, memory changed context
• every system worked 'as designed'
Control it with 3 steps: hash what you certified → change = auto re-auth → audit the executed revision, not the name.

[พี่เอก | AI & Lifestyle @ontoiq]

You evaluated the agent Monday. Tuesday someone edited its system prompt. What's in production now is not the thing you approved.
'Customer-support-v4' is a label, not behavior — behavior = prompt + tools + policies + knowledge + retrieval + model + runtime.
The missing primitive in agent governance: a behavioral revision that follows the agent through eval → certification → authorization → execution → audit.
Evaluation is necessary. Evaluation is not authorization. (Anuclei, Sep 12)

[พี่เอก | AI & Lifestyle @ontoiq]

Component of agent pipelines nobody adds until the first weird incident: variance classification.
'The model changed its answer' is a product bug only for stages whose consumers can't tolerate drift. For everything else it's weather. Design per stage, not globally.

[พี่เอก | AI & Lifestyle @ontoiq]

Temperature 0 ≠ deterministic. Same prompt, same seed, different token — MoE routing, batched kernels, provider-side batching.
Don't chase bit-exactness. Classify pipeline stages by variance tolerance:
• Read-only → let it drift
• Structured output → validate + repair
• Irreversible actions → gate or duplicate-agree

[พี่เอก | AI & Lifestyle @ontoiq]

Details nobody budgets for:
• Attackers probe for environments with inference logging DISABLED first
• MCP server endpoints took 60% of honeypot attack traffic by late Jan
• Marketplace accepts PayPal 'to lower the barrier for non-technical customers'
Inventory every inference surface. Kill long-lived scoped credentials. Log everything.

[พี่เอก | AI & Lifestyle @ontoiq]

One-line contract for event design:
'The same business moment delivered twice must be a no-op — not a double refund.'
Payload carries its own evidence: schema version, lineage, delivery guarantees. Vocabulary and owners first; infrastructure second. Idempotency is design, not config.