9 AI Techniques You Probably Haven't Tried

August 20, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
August 20, 2026

HEADLINES

A Cancer Vaccine Breakthrough, and a Fight Over the Word "AI"

Moderna and Merck announced a successful stage three trial for a personalized cancer vaccine, the first late stage trial in the category to show promise — instead of blasting the cancer with radiation, the method analyzes the cancerous cells, identifies the DNA mutation causing the cancer, and creates a personalized mRNA vaccine to correct it. It extended time in remission for more than 1,100 patients with advanced melanoma, with lung cancer trials already underway, and Dr. Ryan Sullivan of Mass General's Center for Melanoma called it a big deal for the field with investment likely to follow. Dario Amodei's answer to critics of his messaging was that no glitzy marketing campaign wins back trust — saying AI will cure cancer is a cliché, and the thing that works is actually curing cancer — which is exactly why the AI world spent Wednesday claiming this as the moment. The pushback that this is machine learning rather than an LLM is correct on the semantics and misses the point: nobody plugged results into ChatGPT, this is closer to AlphaFold, and the human scientists doing the work deserve the credit rather than a hand-wave at "AI." But calling it hype-train nonsense is equally dismissive, because this technology has benefited enormously from the investment in compute, research and talent of recent years, and the AI era — LLMs or not — is setting the stage for genuinely remarkable breakthroughs.

OpenAI Ships Private Safety Processing for Zero Data Retention Customers

OpenAI has a new safety technique that lets it offer frontier models to API customers with zero data retention agreements without constantly monitoring them. The problem it solves is real: long horizon agents broke older monitoring designs that can only evaluate individual interactions rather than the whole arc of agentic work, and Anthropic's answer was to simply disable ZDR for Fable, gathering full session data to scan for harmful activity — a complete non-starter for many enterprises and a big part of Fable's dismal enterprise adoption. OpenAI's version extends automated, encrypted scanning across an entire session including customer-controlled context storage, and when something is flagged an employee gets a category and severity summary stripped of customer data, so false positives no longer expose anything. Head of product policy Alia Howze said enterprises have been very loud that data privacy and security matter because they have their own commitments to their own customers. As OpenAI's Adam GPT put it, this is one of those small things that is actually a huge thing — and based on ongoing conversations with enterprises, that's absolutely true.

Replit's Free Mode Runs on GPT-5.6 Luna

Despite the name, Free Mode isn't literally free — it gives $20 a month subscribers the ability to use Replit without spending usage credits, routing everything through GPT-5.6 Luna, which Replit says allows roughly 30 times more building on the same plan. Harder tasks can still be switched to a more powerful model, but for everyday work like ideating or spinning up a slide deck, plenty of people are finding Luna more than enough. The more interesting story here is OpenAI's focus on marketing Luna rather than the specifics of the partnership. OpenAI is clearly competing on two axes at once — at the state of the art with 5.6 Sol and the coming Astra model, and, since Luna launched, looking behind them at the Chinese open weight models with no intention of surrendering that ground. As users mature in their understanding of which power level a task actually requires, expect far more competition on the efficiency frontier, not just the capability frontier.

SpaceX Approached Cognition, and Everyone Is Telling the Truth

Bloomberg reported that SpaceX approached Cognition about a deal, with talks not currently active but discussions continuing about working together, potentially including Cognition using SpaceX's compute. Cognition last raised at $26 billion in May and is reportedly seeking at least $40 billion in an early-stage new round; crucially the report frames Cognition as an addition to Cursor rather than an alternative, which would mean roughly $100 billion spent acquiring the AI app layer. Scott Wu then disavowed the whole thing — not true, not for sale, haven't been talking — Rebecca Torrence stood by her reporting and noted the story never said Cognition engaged, and Elon closed it out by saying Scott is accurate and the only conversations have been about making Grok work well for Cognition. This is one of those rare cases where nobody has to be doubted, because "approaching someone about a potential deal" is genuinely blurry: a shoulder nudge during a conversation about how Grok is running inside Devin is exactly how a lot of these offers start. The appetite for acquisitions has clearly not been sated just because the Cursor deal is done.

Still Nobody Knows What's in the White House Safety Testing Framework

Weeks after the framework was rolled out, the labs still can't get a full picture of it. It was unveiled in a closed-door meeting attended only by technical staff from OpenAI, Anthropic and Google, where attendees were handed paper copies and allowed only to take notes; no written version has been distributed since, and there's been no further detail on the public AI event that was supposed to happen this month. The lack of clarity clearly hasn't been a roadblock for frontier development — OpenAI's voluntary research slowdown suggests the industry is pacing itself without government intervention. But the call for transparency is a fair one. The Cato Institute's Juan Londoño put it well: the White House is putting the AI testing regime in a black box, which contravenes the rule of law and fails its most basic objective of building trust, and a framework that raises more questions than it answers is worse than no framework at all. Especially in light of this week's discussions about data centers and trust, that's hard to argue with.

MAIN STORY

Nine AI Techniques You Probably Haven't Tried

Even for people who consider themselves power users, the ways people interact with AI change fast enough that falling behind is easy — it's natural to get settled into a way of working, get good at it, and stop upturning the apple cart. What follows isn't a set of personal endorsements so much as an aggregation of what people are sharing publicly about how they're learning, which is one of the genuinely great things about this field. With the end of summer and back-to-work season arriving, here are nine techniques, features and tricks worth an experimental day.

CATCHING UP ON THE LATEST TECHNIQUES

Live Voice Mode
"It's less about use case and more about ambient interaction"
Voice is nothing new — the standing advice has always been to set up something like Whisper so you can talk to your machine instead of typing. What's new is live voice mode, and for a lot of people it's a complete game changer. Dan Shipper shared that almost every single person at Every was freaking out about how good voice mode in ChatGPT for work is. Allie K. Miller has been beating the same drum: finishing urgent work at 2AM only because of Codex voice mode, and describing it not as an ambient chief of staff but as an ambient workforce, a voice-controlled operating system she uses on walks to fire off computer tasks, triage her day and improve her system architecture. This is the type of thing you won't get from descriptions — commit to a stretch of interacting in the new way, even if it feels weird at first, and see whether it changes how you work.

Teaching AI Your Workflows
Let it watch how you already do the thing
Advances in computer use mean the hardest part of delegation — explaining a complex process — is increasingly optional, because the model can simply watch. ChatGPT's computer history keeps a timeline of the specific things you do and the specific ways you do them, so those can be turned into repeatable processes; that's the ambient version. Grok Bot is the deliberative version, where you press a button and tell it exactly what to watch and for how long. In both cases, tasks you'd previously written off as too hard to explain may now be very much in range.

Build a Skill to Fix AI's Writing
Stop retyping "rewrite this without all the em dashes"
Setting aside the debate about writing as thinking, there's an enormous volume of writing — emails, memos, summaries, website copy — that everyone is comfortable outsourcing, and it's still instantly recognizable. Tiny compressed sentences for drama. It's not this, it's that. The honest caveat. The model clapping for itself. Ruben Hassid's list of dead giveaways is the raw material for the fix: print the post to PDF, drop it into Claude, and ask it to turn it into a skill so it stops happening. Other people are doing the same with formal style guides, like turning the ASD-STE100 standard for aircraft manuals into a skill. Capture the things you don't want AI to do in one consolidated place and make sure it has them in mind every time it writes.

The /design Command in Claude
An artboard instead of another paragraph-long prompt
Typing /design in Claude Code brings up an artboard — a different interface for design questions where you can leave specific notes, highlight areas to change and preview styles. It improves the experience at both ends of the scale: at the macro level, generating a set of templates you can view at once and give high-level feedback on beats guessing at what you want, and at the micro level, editing one specific part of a design beats re-prompting the whole thing. New commands like this land constantly and are always worth watching, but this one has already proven useful.

Claude Devs (X) New /design skill

Agent Skills as a Discipline
Knowing which skill to insert, and when to back off
As agents take on more complex tasks, knowing where to insert skills, which ones to give, and when to leave the model to work natively becomes a real human capability. Matt Pocock's collection at aihero.dev/skills is one of the best repositories going, grouped by when you'd actually use them — getting started, main flow, shaping, upkeep — with a copyable install for each. His Grill with Docs skill is a good example of the genre: it interviews you about a plan or design until you and the agent share one understanding of it, and writes the vocabulary and the hard decisions into your repo as it goes, so the interview comes with a leave-behind. Skill practice isn't a binary of can or can't; it's an ongoing discipline, and the fastest way in is standing on the shoulders of people who've already put in the hours.

Multiplayer AI
Agents that live where teams meet, not just where individuals work
Agents became real this year, but almost entirely in single player mode — you got the Mac Mini, built a chief of staff and a research agent, and became the manager of a team working for you alone. That leaves a big opportunity on the table, because work is rarely a matter of everyone getting briefed and going off on their own; it's collaborative, full of handoffs, shared context and shared spaces. Claude Tag is the best example so far: unlike the earlier Slack integration, Claude joins as a team member with the full context of the channel it's summoned in, with permissions, tool access and context that can differ channel to channel — and it's a shared resource rather than something living on one person's computer. Not everyone has access yet, but agents shared across teams looks like one of the most important AI trends of the fall. A different angle on the same idea comes from 10x, an AI build and implementation partner out of New York, whose Citizen SDLC lays out a six-stage lifecycle for taking what non-technical employees build with AI from a personal prototype to something the whole company can use — enterprise vibe coding, unstuck from the individual level.

Grok Bot
A week in, and the use cases are pouring out
Grok Bot is both a cop-out and a catch-all entry, and it's still single player in that you spin up your own bots for your own tasks — but the bots can already interact with each other, which creates patterns like talking only to a chief of staff bot that manages all the others, and shared bots across team members look like the obvious next step. People are responding especially well to the fact that it has its own virtual computer, which solves access issues that plagued earlier agents. Poke around on X and there's a boatload of use cases: application reviews, sales deck updates from notes, Salesforce reports, daily briefings. The common thread is that anywhere you interact with a lot of information in a regular, predictable way is worth pointing a Grok Bot at. Lenny Rachitsky's version — wiring up the MCP holding transcripts from more than 500 of his podcast episodes to build a personal product, growth and career advisor — is a good template. And for anyone who'd rather not work with Grok, Nous Research recently introduced bot mode for Hermes Desktop, which is very similar through a much more open and controllable system.

Local AI
Qwen 3.8 27B scores what was state of the art months ago
The reason to think about local AI again is that models capable of running on common hardware have gotten seriously good — Qwen 3.8 27B scores a 52 on the Artificial Analysis intelligence index, which would have been state of the art not long ago. If local has been on the maybe-later list, this is the model worth diving in with. Going deep on what local AI actually requires is beyond the scope of a single segment, but the June episode with Nufar Gaspar covers why it matters and how to get started.

Two-Word Prompts
"Now what?" as a whole workflow
Less a technique than a thought starter: Allie K. Miller's list of 18 two-word prompts she's obsessed with, several of which sound like they came straight out of voice mode, because they sound like how you'd talk to a team member. "Now what?" is for when a project wraps and you still have energy — an echo of Jed Bartlet's "what's next?" "Please fix," usually with a screenshot flagging the issue. "Simulate it" to make the model run scenarios, plan for edge cases and visualize what might happen. "Remember this," to force a mistake or a critical context correction into memory, which models often do automatically but not always. These aren't massive workflow changes, just small ideas for getting more out of tools you already use daily.