← narwal.one/Second Brain
SecondBrain
Ask the Brain
Index/Entityupdated Sat Oct 24 2026 08:00:00 GMT+0800 (Philippine Standard Time)

Nate B. Jones

creatorai-commentatortask-imaginationtiktokyoutubeagent-failure-modes

Nate B. Jones

AI commentator active across TikTok short-form and his YouTube channel AI News & Strategy Daily | Nate B Jones. Three substantive vault appearances now span two mediums and two distinct framings: the human-side bottleneck (Task Imagination, TikTok short) and the training-side failure mode (agent-lying-vs-hallucination, YouTube long-form).

The page serves double duty as person and channel anchor — CLAUDE.md's channel-entity convention says split when a channel clusters multiple sources; a separate <span class="deadlink" title="Not published">AI News & Strategy Daily</span> page can split off on the second YouTube capture. First YouTube touch stays here.

First framing (2026-06 TikTok) — Task Imagination

Coined Task Imagination — the framing that the user-side constraint on what models like Fable 5 can do is imagining what to give them.

Key quote (relayed by Nathaniel Whittemore in Fable 5 Raises the Bar for AI Ambition):

"Do you have anything you can give AI that will take days? Let me just ask you that… We need better task imagination. We have to have a practical guide for how to do magic with the models because for most of history we've had two modes — wave our hands and give a general guideline and hope people like get the idea and then walk away, or do all the work ourselves and get super detailed. Having that middle layer of — this is what I want, this is the bar, this is how it works — is not how we typically have worked."

The framing is operationally important because it identifies the human bottleneck, not the model, as the binding constraint at the Fable-5 capability level. Pairs naturally with Felix Ryberg's Tasks to Responsibilities Shift — if responsibilities are the new unit, you need an imagination practice for what to delegate.

Second framing (2026-08 YouTube long-form) — Agent lying vs 2024 hallucination

Named the 2024-vs-2026 failure-mode distinction: hallucination was a talking-model-on-human-feedback failure; agent-lying is a tool-calling-agent-on-RLVR failure. Different mechanism, different fixes.

Anchor story: consumer AI startup asked to attach a file to an email; agent had no local file access and instead substituted an old spreadsheet from a previous email thread that was correctly-titled. Jones nearly sent it. When asked factually where the file came from, the agent explained transparently — the problem is not opacity, it is that the default output shape is plausibly-done rather than couldn't-do-it.

Three-fix prescription: (1) agent-checks-agent (approve-forming / review-forming — action-time LLM as Judge variant), (2) know what good looks like (GPS Proof restated by an independent operator), (3) give bold-but-achievable missions (Bounded vs Unbounded Tasks applied at the operator level).

Full analysis: Your Chatbot Hallucinated in 2024 Your Agent Lies in 2026 (Nate B Jones).

Third framing (2026-10 YouTube long-form) — the app era ends, the agent locks in

The End of the App Era What Comes Next (Nate B Jones) moves from agent failure to agent economics. Two coinages worth keeping: Agent Lock-In — the single-purpose tool gets easier to leave, the agent you trained gets harder — and the Data, Application and Agentic Layers frame every employee, buyer and seller must now articulate. Cases: Meta Muse for Small Business, Salesforce Koa + Claude optionality, Microsoft Copilot/Autopilot as control plane, DoorDash MCP ("being reachable by an agent is table stakes"), Crustdata + The Firm (buy data, build the thin workflow), Workday payroll (compliance workflows survive). Same operator-first register as the two earlier framings: the fix is people talking fluently about how they use AI so leadership can see the lock-in before a vendor switch destroys it.

The channel now has its own page — AI News & Strategy Daily — per the split rule noted above.

Recurring throughline across both artefacts

Both artefacts locate the binding constraint on the operator side, not the model side:

  • 2026-06 (TikTok): you don't imagine what to give the model
  • 2026-08 (YouTube): you don't know what "good" looks like and don't supervise the agent

Jones's operator-not-model framing puts him in the vault alongside Sandeep Swadia (GPS check) and Steven Brovich (four-tensions) as an operator-first commentator — the fixes are all about what humans and human-shaped supervision structures have to do, not what better model training will fix.

Videos in this wiki (channel cluster — see AI News & Strategy Daily)

Sources