In partnership with
Dictate code. Wispr tags the files.
Speak your PR description, bug reproduction, or Cursor prompt. Wispr Flow auto-tags file names, preserves variable names, and formats everything for immediate paste into GitHub, Jira, or your editor.
No re-typing. No context gaps. No mangled syntax. Works natively inside Cursor, Warp, and every IDE at the system level.
4x faster than typing. 89% of messages sent with zero edits. Used by engineering teams at OpenAI, Vercel, and Clay.
the big story: the limits of AI agents
A rare piece of primary-source research is pushing back against the agentic AI wave, and it's coming from inside the house. Danielle Perszyk, a cognitive scientist at Amazon AGI Lab (the group that absorbed Adept), is making an argument that most of the industry isn't structured to hear: current AI agents don't actually understand you, and the entire interaction paradigm we've built around them — turn-taking, batch processing, chatbot-style dialogue — is a accommodation of machine limitations, not a reflection of how intelligence actually works. Her framing is worth sitting with: we haven't built agents that meet humans where we are. We've quietly trained humans to meet agents where they are.
The core of Perszyk's argument draws from cognitive anthropology, not ML benchmarks. Human intelligence, she argues, is fundamentally collective and emergent — it arises from population-scale diversity, interconnectivity, and real-time negotiation of meaning. No individual human survives alone; our cognition is literally distributed across our social environment, our tools, and our shared context. Current agents lack several of the components that would make them genuinely aligned with this: they don't perceive digital environments the way humans do (which itself mirrors the physical world), they don't update understanding in real time mid-interaction, and they have nothing resembling episodic memory — the individuated, perspective-grounded retrieval system that makes human memory more than storage. On that last point, she's deliberately cagey: Amazon AGI Lab is actively working on something that changes how information is "contextualized at inference," but she won't call it memory, because the word carries the wrong connotations.
What makes this worth more than a typical "AI has limitations" take is the institutional context. Perszyk isn't an outside critic — she came in with the Adept team and helped seed Amazon AGI Lab's mission around "perception agents." The lab deliberately runs like a startup, insulated from the broader Amazon org, specifically to do frontier research without the pressures of the Nova product group. When someone with that vantage point says agents are "so unreliable that we ironically feel a little better" about automation displacing jobs, that's a meaningful signal. It's not pessimism — it's a precise diagnosis of where the capability gap actually sits, delivered by someone whose job is to close it.
The practical upshot for newsletter readers: the industry is in what Perszyk calls a "local attractor state" — chatbots, coding agents, batch interactions. The companies and builders who are going to matter in 18-36 months are the ones investing now in real-time interaction models, richer perception layers, and memory systems that aren't just RAG-over-a-database. She points to convergent independent work (Thinking Machines, full-duplex voice research, Moshi from Kyutai) as evidence that the field is slowly, quietly moving toward something more genuinely human-collaborative — but the mainstream conversation hasn't caught up. This is a really interesting watch if you’d like to be at the bleeding edge of AI.
check it out here:Why AI Agents Don't Actually Understand You — Danielle Perszyk, Amazon AGI Lab (Latent Space)
(https://www.youtube.com/watch?v=K796MYUgt0k)
The cognitive science framing of collective intelligence and the candid inside-view on where Amazon AGI Lab's perception agent research actually stands makes this one of the more substantive primary-source conversations on agent limitations you'll find anywhere right now.
*also worth watching:
The "clear-eyed" framing is becoming a practitioner rallying cry.
A sharp soundbite from Lenny's Podcast cuts through the AI-pilled vs. anti-AI binary that's dominated the discourse: the people who extract the most value from AI are the ones with a calibrated nose for what it's good at now and what it will be good at in two months — not the maximalists or the skeptics. It's a small point but an important editorial one: the useful frame isn't capability vs. hype, it's temporal specificity. Tools that are "remarkably bad" at something today may flip in a single model release. That's a genuinely different cognitive posture than either cheerleading or dismissal.
GPT-5.6 + Codex for video clipping is a real workflow unlock, not a demo trick.
Claire Vo’s latest How I AI episode is light on analysis but heavy on specificity: drag a long-form video in, prompt for social clips with format and pacing instructions ("hype video cuts, faster, tighter, horizontal"), and get usable output without touching a timeline. The signal here isn't the wow factor — it's that a non-technical creator is describing a loop that previously required either a video editor or significant personal time investment, and the friction is now genuinely low enough to disappear into a morning workflow. Watch for this pattern: GPT-5.6 + Codex as a composable media production layer is probably undercovered relative to the coding use cases.
Real-time interaction is the next capability frontier that mainstream discourse is sleeping on.
Perszyk's point about real-time AI interaction deserves its own callout beyond the big story. The full-duplex voice model thread — GPT-4o's launch, Kyutai's Moshi, Thinking Machines' recent work — represents a coherent research direction toward agents that update mid-conversation rather than waiting for a prompt to complete. This isn't just a UX nicety. Perszyk frames it as architecturally necessary for genuine understanding: meaning is negotiated in real time in human conversation, and batch turn-taking forecloses that entirely. The companies investing in this now are building toward a fundamentally different interaction model than the one most product teams are designing for today.
*contrarian corner
There's a real tension worth naming between the How I AI episode and the Latent Space + Lenny's framing — and it's not just vibes-level disagreement. The How I AI workflow demo implicitly treats GPT-5.6's video editing capabilities as a stable, reliable feature worth building around. The Latent Space episode, from someone actively building the next generation of agents, is essentially arguing that current reliability levels are still the core unsolved problem — agents are unreliable enough that automation-driven job displacement is less imminent than feared. Lenny's show adds the temporal caveat: capabilities shift fast enough that neither confident adoption nor confident dismissal is the right posture.
None of these three sources are wrong. But they're operating at different layers: one is workflow-level (does this save me time today?), one is systems-level (what's architecturally missing?), and one is strategic (how do I calibrate my bets over a rolling horizon?). For your newsletter, the most useful frame for readers is probably: enthusiastic adoption at the task level is fine and often correct — but don't let task-level wins update your beliefs about agent reliability or understanding at the systems level. Those are different questions, and conflating them is where a lot of the hype-vs-reality whiplash comes from.
The human mind is the original generative machine.