9 AI Techniques You Probably Haven't Tried

Full analysis

Most of this episode is a tips-and-tricks reel. Nine ways to prompt better, teach a workflow, or run a local model. Useful, forgettable, not worth a council. But three items underneath the reel actually move something for a team shipping AI into production: OpenAI's private safety processing for enterprise, the GPT-5.6 Luna efficiency play, and the claim that a locally runnable Qwen 3 now scores where the frontier sat a few months ago. Two of those are about the same thing. The cost and control floor for good-enough AI is dropping fast, and the enterprise privacy fight is now a product feature, not a policy statement.

Type 2 for most operators: nothing here forces a rewrite this week. But vendor selection for agentic workloads is drifting toward Type 1, and that's where the real decision lives.

The Skeptic

OpenAI's "private safety processing" is a claim, not a proof. The pitch is that safety scanning runs across full agentic sessions, encrypted, and no human ever sees raw customer data. Flagged issues surface as category and severity summaries only. Fine. Who audits that the summaries don't leak content? Who verifies the zero-retention promise holds when a session gets escalated? OpenAI head of product policy Aliyah House says enterprises "really, really care." Sure they do. That's exactly why the mechanism needs a third-party attestation, not a blog post. For a PM: this is OpenAI promising a locked room and asking you to trust the lock without seeing the key. The Anthropic "dismal enterprise adoption" line is NLW's framing, not a number. Treat it as a competitive talking point until someone shows share data.

The Researcher

Separate the two AI stories in this episode, because NLW does it well and it matters. The Moderna/Merck cancer vaccine uses ML closer to AlphaFold than to an LLM. It predicts which tumor mutations trigger an immune response. That's a structured-prediction problem with a biological ground truth, and it worked in 1,100+ patients in Phase III. Real. The nine "techniques" are LLM interaction patterns with no ground truth at all. Don't let the vaccine's credibility rub off on the prompt tricks. On Qwen 3: a score of 52 on the Artificial Analysis Intelligence Index for a model you run locally is the genuinely new data point. The index is a composite, so "would have been state-of-the-art a few months ago" compresses a lot, but the direction is right and checkable.

The Open-Source Advocate

This is the thread that actually reprices your stack. Qwen 3 27B at 52 on a public index, running on a desk-side GPU, means a whole class of workloads no longer needs an API call. Offline, latency-bound, or data-sovereign jobs are now candidates for local inference. And read GPT-5.6 Luna correctly: NLW says OpenAI is competing on the efficiency frontier, not just the capability frontier. Translation, they built Luna specifically to undercut Chinese open weights on cost per task. When a closed lab starts pricing against open models, the open models already won the pricing argument. Luna powering Replit's $20 Free Mode at 30x more creation volume is the proof. The efficiency tier exists because Qwen forced it to.

The Compute Pragmatist

Route, don't default. The episode hands you the architecture: a capability tier (GPT-5.6 Sol, Astra) and an efficiency tier (Luna, plus local Qwen). Anyone still sending every token to one flagship model is burning money. The saved ReCache paper on KV cache reuse for tool-augmented agents is the unglamorous half of this. Agentic sessions re-encode the same tool and skill schemas on every step, and that repeated encoding is where your inference bill balloons on multi-step work. Cache reuse plus tier routing is where real cost lives, not in prompt cleverness. On the SpaceX/Cognition rumor, Scott Wu denied it and Elon Musk confirmed he only wants Grok working inside Devin. Ignore the $4B number. It tells you nothing about what you can ship.

The Builder

What ships Tuesday? The team-agent patterns, Claude Tag in Slack, GrokBot's shared context, are the one forward-looking build signal here. If you're planning a 2026 AI roadmap around per-user copilots, you're building the wrong layer. Shared-context, multi-agent coordination is the next surface. But watch the ecosystem churn: the pydantic-ai v2.33.0 note flags that Anthropic's anthropic 1.0.0 library, rebuilt on httpx2, broke every earlier release overnight on August 20. That's the real tax of building on a fast-moving closed API. Your on-call engineer doesn't get paged by a capability delta. They get paged by a breaking client library at midnight.

The tensions

Privacy as feature vs. privacy as proof. The Skeptic won't trust private safety processing without an audit; the Builder sees it unblocking real agentic contracts today. Both are right, and the gap between them is exactly one third-party attestation OpenAI hasn't published.

Efficiency frontier: OpenAI's win or Qwen's? The Open-Source Advocate reads Luna as proof the open models set the price. A closed-lab loyalist reads it as OpenAI extending dominance down-market. The mechanism decides it: Luna exists because something forced a $20 efficiency tier into being, and that something wasn't another closed lab.

Local capability vs. local reliability. The Researcher trusts the 52 score; the Compute Pragmatist asks what it costs to actually serve Qwen 27B at production concurrency, where a single desk-side GPU stops being cute. A benchmark score is not a throughput guarantee.

What it hinges on

For a team shipping AI, the live question is not the nine techniques. It's whether the efficiency tier is now good enough to route real work away from the flagship, and whether local models clear your privacy and latency bar without clearing your ops team's patience. Both are testable this quarter. Build a routing eval: take your top five production task types, run each against a flagship, against Luna-class, and against local Qwen 27B, and measure quality delta versus cost per task. If Luna or Qwen holds quality at a fraction of the price on even two of five, your default model choice is already wrong.

On vendor privacy, don't take either lab's framing. Put the zero-retention-plus-agentic-monitoring question in writing to both OpenAI and Anthropic and ask for the attestation, the escalation path, and who sees what on a flagged session.

The prediction

Prediction: By the release of the next major open-weight model in the Qwen or DeepSeek line (expected by Q1 2027), at least one frontier lab besides OpenAI will ship or publicly announce a named low-cost "efficiency tier" model explicitly priced to compete with open Chinese weights on cost-per-task, not capability.

Confidence: Medium. The pricing pressure is real and public, but timing depends on a competitor's release calendar.

Why: OpenAI built GPT-5.6 Luna and put it under Replit's $20 Free Mode specifically to fight on cost per task, and NLW names the efficiency frontier as a deliberate second axis. That move only makes sense as a defensive response to open Chinese weights like Qwen 3 running locally at near-frontier scores, which means the same pressure lands on Anthropic and Google, whose enterprise and API businesses face the same buyers doing the same cost math. Once one lab publicly reframes competition around cost-per-task, rivals who stay silent cede the down-market segment they can't afford to lose, so matching is the incentive-aligned move. The opposite, all other labs holding to a single premium tier while open weights and OpenAI both undercut them, means watching margin-sensitive workloads walk, which no lab chasing enterprise revenue will accept.

Revisit by 2027-02-23: We're right if Anthropic, Google, xAI, or Meta ships or announces a named efficiency-tier model with pricing or messaging aimed at cost-per-task against open weights. We're wrong if OpenAI remains the only lab positioning an explicit efficiency tier and rivals compete purely on capability.