A personalized cancer vaccine clears phase III and reignites the fight over what counts as 'AI' — then NLW switches into operator mode with nine techniques, from live voice mode to team agents to local Qwen, that even power users probably haven't tried yet.
AI is changing faster than your habits — nine techniques to catch up before fall.
The thing that makes AI exciting is the thing that makes it hard: it changes constantly, and even power users settle into comfortable workflows while the frontier moves on. With back-to-school season approaching, this episode aggregates what people are learning in public — live voice mode, teaching agents by letting them watch, writing skills, /design, multiplayer AI, GrokBot, local models, and two-word prompts — as a menu for your next experimental day.
— Dario Amodei, responding to critics of his messaging. Responding to critics of his negative messaging, the Anthropic CEO argued that no glitzy marketing campaign wins back public trust — 'saying that AI will cure cancer is more a cliché than it is inspiring.' A day later, a lot of people were saying that's exactly what had happened.
Moderna and Merck's treatment analyzes cancerous cells, identifies the DNA mutation, and creates a personalized mRNA vaccine — extending remission in more than 1,100 advanced melanoma patients in a successful phase III trial, with lung cancer trials underway. Mass General's Dr. Ryan Sullivan: 'these approaches may change the way we treat cancer more broadly.'
Critics are right that this isn't scientists plugging results into ChatGPT — it's advanced ML closer to AlphaFold, and the brilliant human scientists shouldn't be erased by attributing the breakthrough to 'AI.' But it would also be dismissive to call it hype: the AI era's investment in compute, research, and talent fed this, and the market noticed — Moderna jumped 70%, then 125%, within hours.
For eligible API customers with zero-data-retention agreements, fully automated, encrypted scanning now covers entire agentic sessions — no human ever sees sensitive data, and flagged issues reach staff only as anonymized category-and-severity summaries. OpenAI's Adam GPT: 'This is one of those small things that is actually a huge thing' — and per NLW's enterprise conversations, that's absolutely true.
Anthropic's answer to long-horizon agent safety was to simply disable zero data retention for Fable and scan full session data — a complete non-starter for many enterprise customers, showing up in fairly dismal enterprise adoption. OpenAI's new approach is aimed squarely at that gap.
Not literally free — $20/month subscribers can use Replit without burning usage credits, with all queries routed through Luna for a claimed 30× more output, and escalation to stronger models still available for complex tasks. Replit: 'AI models are now capable and affordable enough to make once unreachable outcomes practical.'
They're competing at the state of the art with 5.6 Sol and the Astra model to come — but since Luna launched, they're also clearly looking behind them at the Chinese open-weight models and refusing to surrender that ground. As users mature in matching power levels to tasks, expect much more competition on the efficiency frontier, not just the capability frontier.
Bloomberg reported SpaceX approached the coding-agent startup — which raised at $26B in May and is seeking $40B+ — before CEO Scott Wu shot it down ('Cognition is not for sale and we haven't been talking') and Elon backed him. NLW's read: 'approaching someone about a deal' is blurry enough that nobody needs to be lying — that shoulder-nudge conversation is genuinely how these offers start, and Elon's appetite for app-layer acquisitions wasn't sated by the Cursor deal.
OpenAI, Anthropic, and Google staff were handed paper copies at the closed-door unveiling and allowed only to take notes; written versions still haven't been distributed. Cato's Juan Ladano: 'a framework that raises more questions than the one it answers is worse than no framework at all' — and especially in light of the week's data-center trust discussions, NLW fully agrees.
This episode isn't NLW's own testing — it's an aggregation of what others are sharing, because one of the things that makes AI so cool is that people learn in public. For as cesspool-y as they can be, if you're not watching the AI conversation on X and LinkedIn, you're missing a lot of free R&D.
Dan Shipper says almost everyone at Every is 'freaking out' about voice mode in ChatGPT for work, and Allie K. Miller credits Codex voice mode for finishing all her urgent work at 2 AM — firing off thumbnail requests, doc retrieval, and email checks while she fills her water bottle. 'Holler if you need me' is now a prompt.
Miller's point: ambient assistant versus click-to-speak 'probably sounds like a negligible difference... but it is a world of distance in practice' — less a chief of staff than 'an ambient workforce, an ambient voice-controlled operating system.' NLW's advice: you won't get it from descriptions; commit to a period of working this way, even if it feels weird at first.
ChatGPT's computer history keeps an ambient timeline of how you actually work and turns it into repeatable processes, while GrokBot takes the deliberate approach: press a button and it watches a specific task as you do it. If you'd written something off as too hard to explain to AI, that may have changed significantly.
Instead of endlessly prompting 'rewrite without the tropes' — the tiny dramatic sentences, 'it's not this, it's that,' the AI clapping for itself — print Ruben Haseed's post of dead giveaways as a PDF, drop it into Claude, and make it a skill so it never happens again. Others are converting full style guides, like the ASD-STE-100 standard for plane manuals, the same way.
The new artboard interface lets you preview a batch of templates and give high-level feedback when you don't yet know what you want — and then highlight and edit specific regions rather than re-prompting the whole design. In NLW's experience, it improves the design loop at both ends.
Matt Pocock's repository groups skills by when you'd actually use them — getting started, main flow, shaping, upkeep — each with a copyable install. His 'Grill with Docs' skill interviews you about a plan until you and the agent 'share one understanding of it,' writing the vocabulary and hard decisions into your repo as a leave-behind.
Since Open Claw, we've run agents in single-player mode — personal chiefs of staff and research agents working for one person. But work is collaborative, full of handoffs and shared context, and the big opportunity is agentic tools that live where teams intersect rather than where individuals work alone.
Unlike the older tag-to-summon Slack integration, Claude Tag joins a channel as a team member with the entire channel's context — with permissions, tool access, and context that can differ channel by channel. Crucially, that Claude is a shared resource across the team, not something living on one person's computer.
10x, an AI build partner out of New York, published a six-stage lifecycle that takes what non-technical employees vibe-code with AI from a personal prototype to production the whole company can use — another attack on the same problem of individual AI wins never scaling to the team.
Its own virtual computer solves the access issues that plagued earlier agents, and X is full of use cases — application reviews, sales deck updates, Salesforce reports, daily briefings — plus patterns like a chief-of-staff bot managing all the others. Lenny Rachitsky suggests pointing it at an MCP of his 500+ podcast transcripts as a personal product-strategy advisor; for the Grok-averse, Nous Research's new Hermes Desktop bot mode is the open, more controllable alternative.
The model runs on common hardware and scores 52 on the Artificial Analysis Intelligence Index — a level that would have been state of the art just a few months ago. If you've been meaning to experiment with local AI, this is the moment to dive in.
Allie K. Miller's 18 two-word prompts read like talking to a teammate: 'Now what?' when you've finished a push and want more; 'Please fix' with a screenshot; 'Simulate it' to run scenarios and edge cases; 'Remember this' to force the AI to log a mistake in memory rather than hoping it does so automatically. Not a workflow overhaul — just simple ways to get more from tools you already use daily.