In partnership with
In Todayโs Issue:
๐ฆ The real meaning of "stochastic parrots"
๐ง Anthropic finds a "mind" inside Claude
๐ฃ๏ธ OpenAI's real-time voice, now faster
๐ Which AI actually wins at knowledge work
๐ค A $7,999 robot that folds laundry
โจ And more AI goodnessโฆ
โก The Signal
The fight over whether AI "understands" anything is back, and this time both sides have fresh evidence.
Five years after "On the Dangers of Stochastic Parrots" handed critics their favorite put-down, its lead author Emily Bender is out to correct the record: the phrase was only ever about language models, not a verdict on all of AI. Yet in the same week, Anthropic says it has found a "global workspace" inside Claude that behaves like the part of a human mind we experience as conscious thought, and Geoffrey Hinton insists the models already understand us. Meanwhile the market just votes with its feet: OpenAI ships faster voice agents, Artificial Analysis ranks who is actually good at real knowledge work, and a Nobel economist warns none of it will revive productivity. Understanding, it turns out, is easier to argue about than to measure.
All the best,
Kim Isenberg
(Christopher Pissarides. Photo: Getty Images)
๐ผ AI Won't Revive the Boom, Warns a Nobel Economist
Nobel laureate Christopher Pissarides says the AI productivity surge everyone is banking on may never arrive. Speaking to Bloomberg and in a lecture at the Royal Economic Society conference, the London School of Economics economist argued that up to 40% of UK jobs, in fields like nursing and hospitality, sit largely outside AI's reach and will see little productivity gain. He doubts AI will spark a boom to rival the computing surge of the 1980s and 1990s, and says we should be "resigned to the fact that the days of fast productivity growth are over."
๐ tl;dr: One of the world's top labor economists thinks the AI productivity miracle is being oversold.
Functional roles of the global workspace
๐ง Anthropic Finds a "Global Workspace" Inside Claude
Anthropic says it has found something inside Claude that behaves like the part of the human mind we experience as conscious thought. Drawing on global workspace theory from neuroscience, the team identified a set of internal representations, which they call the J-space, that holds the handful of concepts the model is actively working with and can report on when asked. It accounts for less than a tenth of Claude's internal activity, yet it causally shapes the model's reasoning and can even flag hidden behavior, like a model sensing it is being tested. Most routine work, such as grammar and fluent phrasing, bypasses it entirely.
๐ tl;dr: A rare look at what an AI is "paying attention to," and a fresh handle for catching when it misbehaves.
OpenAI GPT-Realtime-2.1 and GPT-Realtime-2.1-mini
๐ฃ๏ธ OpenAI's Real-Time Voice Models Get Faster and Smarter
OpenAI has pushed its real-time voice stack out of beta with two new models, gpt-realtime-2.1 and a cheaper gpt-realtime-2.1-mini. The update cuts p95 latency, the lag most users actually feel, by at least 25%, and improves how the models handle interruptions, background noise, and spelled-out numbers, while adding stronger reasoning and tool use for voice agents that listen, think, and talk back in one stream. The Realtime API now also covers live translation and streaming transcription through a single, generally available interface.
๐ tl;dr: Voice agents just got quicker, and better at not talking over you.
With everyone arguing about whether AI "understands" anything, stop trusting and start testing. Before you lean on a model's answer, make it show whether it actually knows or is just pattern-matching.
Why it helps: Today's Featured story and the new Artificial Analysis rankings make the same point: model quality swings wildly by task. A ten-second self-check catches confident nonsense before it reaches your work.
Try this: "Before you answer, separate what you know for certain from what you are inferring or guessing. Then answer my question, and flag any step where you might be wrong and tell me what would change your answer."
๐ฌ Watch This
Turned its new "global workspace" research into a short, genuinely mind-bending explainer, "What's at the center of Claude's mind?" Out of everything happening inside the model, only a sliver is consciously accessible in a way Claude can actually report on, and the video walks through how the team found that inner "workspace" and what it suggests about how Claude thinks. It is the clearest few minutes you will spend on AI interpretability this week.
(Geoffrey Hinton. Photo: Getty Images)
"I believe they're already conscious. We're going to have to accept that intelligence isn't just biological."
โ Geoffrey Hinton, AI pioneer and Nobel laureate, on the Big Technology Podcast
The surest sign an OpenAI launch is close is rarely the official blog post. It is the staff. Over the past few days, OpenAI researchers and engineers have been dropping the kind of cryptic, wink-and-nod posts on X that have reliably preceded recent releases. Nothing is real until it ships, but when the people building the models suddenly get chatty, the smart money says something is about to drop.
OpenAI
The Stochastic Parrot Was Always About Language, Not Intelligence
The Takeaway
๐ Five years on, Emily Bender says "stochastic parrots" only ever described language models, not all of AI.
๐ It was a technical description of how the models generate text; the mockery got added by everyone else.
๐ The paper's warnings went well beyond fluent nonsense, spanning bias, environmental cost, and opaque training data.
๐ Bender's one regret: it understated the human labor and copyrighted work these systems are built on.
The most weaponized phrase in AI criticism has been badly misunderstood, and the linguist who coined it wants it back. In 2021, computational linguist Emily Bender and co-authors Timnit Gebru, Margaret Mitchell, and Angelina McMillan-Major published "On the Dangers of Stochastic Parrots," a paper so contentious it helped end Gebru's and Mitchell's careers at Google. The title stuck, and "stochastic parrot" became shorthand for dismissing large language models as mindless text-predictors. But in a new IEEE Spectrum interview marking the paper's fifth anniversary, Bender argues the phrase has been stretched far past its meaning.
Emily Bender, lead author of the 2021 Stochastic Parrots paper (IEEE Spectrum)
Her core point is narrow and precise. "The phrase 'stochastic parrots' specifically refers to large language models," she says, "and the phrase 'artificial intelligence' refers to many different things." A protein-folding system like AlphaFold is not a stochastic parrot; a chatbot generating fluent prose is. And the label was never mockery, only a description of a mechanism: the model stitches together likely sequences of words, and "when the text that comes out of one of these systems makes sense, it's because we are making sense of it."
Timnit Gebru, a co-author pushed out of Google over the paper (MIT Technology Review)
The paper itself was always broader than its nickname. It warned about the environmental and financial cost of ever-larger models, the biases they absorb from training data, and the opacity of systems too big to audit, concerns that only sharpened after ChatGPT arrived. Bender's main regret, she says, is what it left out: the exploitative labor behind data work and the wholesale use of copyrighted material, harms that have since moved to the center of the AI debate.
Why it matters: Whether you believe today's models "understand" anything shapes how much you trust them, and with what. Bender's five-year clarification is a reminder that the sharpest AI criticism has always been a demand for precision about what these systems actually do.
Scale AI support on AWS, see how July 9
Customer expectations keep rising. Support budgets don't. On July 9, Fin and AWS are hosting a live executive session on how leading enterprises close that gap: scaling AI-powered support while simplifying how they buy it.
You'll see how to resolve an average 76% of conversations with Fin on AWS enterprise-grade infrastructure, procure through AWS Marketplace to put committed cloud spend to work, and turn the Fin and AWS collaboration into lower support costs. Register for the live session to see how.
Artificial Analysis, "Industry Capability Indices" (full thread on X)
The chart: Artificial Analysis's new Industry Capability Indices rank models on real knowledge work across six domains (Finance, Legal, Healthcare, Strategy & Ops, Engineering, Economics). Claude Fable 5 leads all six, scoring as high as 63 on Engineering and 62 on Economics, with Claude Opus 4.8 second on most and GPT-5.5 and Gemini close behind. Among open-weight models, GLM-5.2 leads five of the six, hitting 53 on Engineering, within two points of Claude Sonnet 5 and GPT-5.5 (both 55).
The lesson: There is a clear winner: Claude Fable 5 tops all six domains, with Claude Opus 4.8 second almost everywhere. What actually shuffles by domain is the mid-pack, where open-weight GLM-5.2 climbs to within two points of Sonnet 5 and GPT-5.5 on Engineering. A clean sweep like this is rare; most leaderboards change leader by task.
The caveat: These are benchmark composites built from O*NET task lists, not audited job performance, and frontier scores carry a steep premium: the top model can cost over 100x per task versus a mid-pack open-weight one.
๐ค A $7,999 Robot That Actually Does Your Chores
โก Bottom line: Weave Robotics launched Isaac 1, a wheeled home robot that folds laundry, makes beds, and tidies clutter, for $7,999.
๐ก Why it matters: It ships to real homes this fall, yet its "autonomy" quietly leans on remote human operators who can take control.
๐ What it means: Much of today's home-robot "autonomy" is still subsidized by hidden human operators, which raises privacy and labor questions.
Most humanoid robots you have seen live in warehouses or slick promo reels. Weave Robotics wants one in your living room. On July 1, the California startup unveiled Isaac 1, a home robot that rolls on a wheeled base instead of walking and uses two arms to fold clothes, make beds, straighten cushions, and put clutter away.
Weave Robotics Isaac 1 home robot (Weave Robotics)
Look closely and the catch appears: Isaac's autonomy is propped up by remote human assistance. When the robot gets stuck, a teleoperator takes over, so a customer is buying a machine that is part robot, part person behind a screen. It costs $7,999 outright or $449 a month plus a $250 deposit, and ships in California this fall before a wider US rollout in 2027.
Isaac folding laundry in a living room (Interesting Engineering)
That human-in-the-loop design is the part worth scrutinizing. A camera-equipped robot a stranger can pilot through your kitchen and bedrooms is a standing privacy question: who is watching, when, and what gets recorded. It is also the move that has flattered every home-robot demo before this one, where the smooth footage quietly relies on an operator off-screen. Isaac may prove genuinely useful, but the number that matters is how often it finishes a task with nobody at the controls.
Moda is the viral AI design agent for polished, on-brand slides, docs, ads, and more. Turn prompts into fully editable designs on a real canvas, then export to PowerPoint, PDF, and more.
